| 1 | /***************************************************************************** |
|---|---|
| 2 | |
| 3 | Copyright (c) 1994, 2016, Oracle and/or its affiliates. All Rights Reserved. |
| 4 | Copyright (c) 2008, Google Inc. |
| 5 | Copyright (c) 2012, Facebook Inc. |
| 6 | Copyright (c) 2015, 2018, MariaDB Corporation. |
| 7 | |
| 8 | Portions of this file contain modifications contributed and copyrighted by |
| 9 | Google, Inc. Those modifications are gratefully acknowledged and are described |
| 10 | briefly in the InnoDB documentation. The contributions by Google are |
| 11 | incorporated with their permission, and subject to the conditions contained in |
| 12 | the file COPYING.Google. |
| 13 | |
| 14 | This program is free software; you can redistribute it and/or modify it under |
| 15 | the terms of the GNU General Public License as published by the Free Software |
| 16 | Foundation; version 2 of the License. |
| 17 | |
| 18 | This program is distributed in the hope that it will be useful, but WITHOUT |
| 19 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS |
| 20 | FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. |
| 21 | |
| 22 | You should have received a copy of the GNU General Public License along with |
| 23 | this program; if not, write to the Free Software Foundation, Inc., |
| 24 | 51 Franklin Street, Suite 500, Boston, MA 02110-1335 USA |
| 25 | |
| 26 | *****************************************************************************/ |
| 27 | |
| 28 | /**************************************************//** |
| 29 | @file btr/btr0cur.cc |
| 30 | The index tree cursor |
| 31 | |
| 32 | All changes that row operations make to a B-tree or the records |
| 33 | there must go through this module! Undo log records are written here |
| 34 | of every modify or insert of a clustered index record. |
| 35 | |
| 36 | NOTE!!! |
| 37 | To make sure we do not run out of disk space during a pessimistic |
| 38 | insert or update, we have to reserve 2 x the height of the index tree |
| 39 | many pages in the tablespace before we start the operation, because |
| 40 | if leaf splitting has been started, it is difficult to undo, except |
| 41 | by crashing the database and doing a roll-forward. |
| 42 | |
| 43 | Created 10/16/1994 Heikki Tuuri |
| 44 | *******************************************************/ |
| 45 | |
| 46 | #include "btr0cur.h" |
| 47 | #include "row0upd.h" |
| 48 | #include "mtr0log.h" |
| 49 | #include "page0page.h" |
| 50 | #include "page0zip.h" |
| 51 | #include "rem0rec.h" |
| 52 | #include "rem0cmp.h" |
| 53 | #include "buf0lru.h" |
| 54 | #include "btr0btr.h" |
| 55 | #include "btr0sea.h" |
| 56 | #include "row0log.h" |
| 57 | #include "row0purge.h" |
| 58 | #include "row0upd.h" |
| 59 | #include "trx0rec.h" |
| 60 | #include "trx0roll.h" |
| 61 | #include "que0que.h" |
| 62 | #include "row0row.h" |
| 63 | #include "srv0srv.h" |
| 64 | #include "ibuf0ibuf.h" |
| 65 | #include "lock0lock.h" |
| 66 | #include "zlib.h" |
| 67 | #include "srv0start.h" |
| 68 | |
| 69 | /** Buffered B-tree operation types, introduced as part of delete buffering. */ |
| 70 | enum btr_op_t { |
| 71 | BTR_NO_OP = 0, /*!< Not buffered */ |
| 72 | BTR_INSERT_OP, /*!< Insert, do not ignore UNIQUE */ |
| 73 | BTR_INSERT_IGNORE_UNIQUE_OP, /*!< Insert, ignoring UNIQUE */ |
| 74 | BTR_DELETE_OP, /*!< Purge a delete-marked record */ |
| 75 | BTR_DELMARK_OP /*!< Mark a record for deletion */ |
| 76 | }; |
| 77 | |
| 78 | /** Modification types for the B-tree operation. |
| 79 | Note that the order must be DELETE, BOTH, INSERT !! |
| 80 | */ |
| 81 | enum btr_intention_t { |
| 82 | BTR_INTENTION_DELETE, |
| 83 | BTR_INTENTION_BOTH, |
| 84 | BTR_INTENTION_INSERT |
| 85 | }; |
| 86 | |
| 87 | /** For the index->lock scalability improvement, only possibility of clear |
| 88 | performance regression observed was caused by grown huge history list length. |
| 89 | That is because the exclusive use of index->lock also worked as reserving |
| 90 | free blocks and read IO bandwidth with priority. To avoid huge glowing history |
| 91 | list as same level with previous implementation, prioritizes pessimistic tree |
| 92 | operations by purge as the previous, when it seems to be growing huge. |
| 93 | |
| 94 | Experimentally, the history list length starts to affect to performance |
| 95 | throughput clearly from about 100000. */ |
| 96 | #define BTR_CUR_FINE_HISTORY_LENGTH 100000 |
| 97 | |
| 98 | /** Number of searches down the B-tree in btr_cur_search_to_nth_level(). */ |
| 99 | ulint btr_cur_n_non_sea; |
| 100 | /** Old value of btr_cur_n_non_sea. Copied by |
| 101 | srv_refresh_innodb_monitor_stats(). Referenced by |
| 102 | srv_printf_innodb_monitor(). */ |
| 103 | ulint btr_cur_n_non_sea_old; |
| 104 | #ifdef BTR_CUR_HASH_ADAPT |
| 105 | /** Number of successful adaptive hash index lookups in |
| 106 | btr_cur_search_to_nth_level(). */ |
| 107 | ulint btr_cur_n_sea; |
| 108 | /** Old value of btr_cur_n_sea. Copied by |
| 109 | srv_refresh_innodb_monitor_stats(). Referenced by |
| 110 | srv_printf_innodb_monitor(). */ |
| 111 | ulint btr_cur_n_sea_old; |
| 112 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 113 | |
| 114 | #ifdef UNIV_DEBUG |
| 115 | /* Flag to limit optimistic insert records */ |
| 116 | uint btr_cur_limit_optimistic_insert_debug; |
| 117 | #endif /* UNIV_DEBUG */ |
| 118 | |
| 119 | /** In the optimistic insert, if the insert does not fit, but this much space |
| 120 | can be released by page reorganize, then it is reorganized */ |
| 121 | #define BTR_CUR_PAGE_REORGANIZE_LIMIT (srv_page_size / 32) |
| 122 | |
| 123 | /** The structure of a BLOB part header */ |
| 124 | /* @{ */ |
| 125 | /*--------------------------------------*/ |
| 126 | #define BTR_BLOB_HDR_PART_LEN 0 /*!< BLOB part len on this |
| 127 | page */ |
| 128 | #define BTR_BLOB_HDR_NEXT_PAGE_NO 4 /*!< next BLOB part page no, |
| 129 | FIL_NULL if none */ |
| 130 | /*--------------------------------------*/ |
| 131 | #define BTR_BLOB_HDR_SIZE 8 /*!< Size of a BLOB |
| 132 | part header, in bytes */ |
| 133 | |
| 134 | /** Estimated table level stats from sampled value. |
| 135 | @param value sampled stats |
| 136 | @param index index being sampled |
| 137 | @param sample number of sampled rows |
| 138 | @param ext_size external stored data size |
| 139 | @param not_empty table not empty |
| 140 | @return estimated table wide stats from sampled value */ |
| 141 | #define BTR_TABLE_STATS_FROM_SAMPLE(value, index, sample, ext_size, not_empty) \ |
| 142 | (((value) * static_cast<ib_uint64_t>(index->stat_n_leaf_pages) \ |
| 143 | + (sample) - 1 + (ext_size) + (not_empty)) / ((sample) + (ext_size))) |
| 144 | |
| 145 | /* @} */ |
| 146 | |
| 147 | /*******************************************************************//** |
| 148 | Marks all extern fields in a record as owned by the record. This function |
| 149 | should be called if the delete mark of a record is removed: a not delete |
| 150 | marked record always owns all its extern fields. */ |
| 151 | static |
| 152 | void |
| 153 | btr_cur_unmark_extern_fields( |
| 154 | /*=========================*/ |
| 155 | page_zip_des_t* page_zip,/*!< in/out: compressed page whose uncompressed |
| 156 | part will be updated, or NULL */ |
| 157 | rec_t* rec, /*!< in/out: record in a clustered index */ |
| 158 | dict_index_t* index, /*!< in: index of the page */ |
| 159 | const ulint* offsets,/*!< in: array returned by rec_get_offsets() */ |
| 160 | mtr_t* mtr); /*!< in: mtr, or NULL if not logged */ |
| 161 | /*******************************************************************//** |
| 162 | Adds path information to the cursor for the current page, for which |
| 163 | the binary search has been performed. */ |
| 164 | static |
| 165 | void |
| 166 | btr_cur_add_path_info( |
| 167 | /*==================*/ |
| 168 | btr_cur_t* cursor, /*!< in: cursor positioned on a page */ |
| 169 | ulint height, /*!< in: height of the page in tree; |
| 170 | 0 means leaf node */ |
| 171 | ulint root_height); /*!< in: root node height in tree */ |
| 172 | /***********************************************************//** |
| 173 | Frees the externally stored fields for a record, if the field is mentioned |
| 174 | in the update vector. */ |
| 175 | static |
| 176 | void |
| 177 | btr_rec_free_updated_extern_fields( |
| 178 | /*===============================*/ |
| 179 | dict_index_t* index, /*!< in: index of rec; the index tree MUST be |
| 180 | X-latched */ |
| 181 | rec_t* rec, /*!< in: record */ |
| 182 | page_zip_des_t* page_zip,/*!< in: compressed page whose uncompressed |
| 183 | part will be updated, or NULL */ |
| 184 | const ulint* offsets,/*!< in: rec_get_offsets(rec, index) */ |
| 185 | const upd_t* update, /*!< in: update vector */ |
| 186 | bool rollback,/*!< in: performing rollback? */ |
| 187 | mtr_t* mtr); /*!< in: mini-transaction handle which contains |
| 188 | an X-latch to record page and to the tree */ |
| 189 | /***********************************************************//** |
| 190 | Frees the externally stored fields for a record. */ |
| 191 | static |
| 192 | void |
| 193 | btr_rec_free_externally_stored_fields( |
| 194 | /*==================================*/ |
| 195 | dict_index_t* index, /*!< in: index of the data, the index |
| 196 | tree MUST be X-latched */ |
| 197 | rec_t* rec, /*!< in: record */ |
| 198 | const ulint* offsets,/*!< in: rec_get_offsets(rec, index) */ |
| 199 | page_zip_des_t* page_zip,/*!< in: compressed page whose uncompressed |
| 200 | part will be updated, or NULL */ |
| 201 | bool rollback,/*!< in: performing rollback? */ |
| 202 | mtr_t* mtr); /*!< in: mini-transaction handle which contains |
| 203 | an X-latch to record page and to the index |
| 204 | tree */ |
| 205 | |
| 206 | /*==================== B-TREE SEARCH =========================*/ |
| 207 | |
| 208 | /** Latches the leaf page or pages requested. |
| 209 | @param[in] block leaf page where the search converged |
| 210 | @param[in] page_id page id of the leaf |
| 211 | @param[in] latch_mode BTR_SEARCH_LEAF, ... |
| 212 | @param[in] cursor cursor |
| 213 | @param[in] mtr mini-transaction |
| 214 | @return blocks and savepoints which actually latched. */ |
| 215 | btr_latch_leaves_t |
| 216 | btr_cur_latch_leaves( |
| 217 | buf_block_t* block, |
| 218 | const page_id_t& page_id, |
| 219 | const page_size_t& page_size, |
| 220 | ulint latch_mode, |
| 221 | btr_cur_t* cursor, |
| 222 | mtr_t* mtr) |
| 223 | { |
| 224 | ulint mode; |
| 225 | ulint left_page_no; |
| 226 | ulint right_page_no; |
| 227 | buf_block_t* get_block; |
| 228 | page_t* page = buf_block_get_frame(block); |
| 229 | bool spatial; |
| 230 | btr_latch_leaves_t latch_leaves = {{NULL, NULL, NULL}, {0, 0, 0}}; |
| 231 | |
| 232 | compile_time_assert(int(MTR_MEMO_PAGE_S_FIX) == int(RW_S_LATCH)); |
| 233 | compile_time_assert(int(MTR_MEMO_PAGE_X_FIX) == int(RW_X_LATCH)); |
| 234 | compile_time_assert(int(MTR_MEMO_PAGE_SX_FIX) == int(RW_SX_LATCH)); |
| 235 | |
| 236 | spatial = dict_index_is_spatial(cursor->index) && cursor->rtr_info; |
| 237 | ut_ad(buf_page_in_file(&block->page)); |
| 238 | |
| 239 | switch (latch_mode) { |
| 240 | case BTR_SEARCH_LEAF: |
| 241 | case BTR_MODIFY_LEAF: |
| 242 | case BTR_SEARCH_TREE: |
| 243 | if (spatial) { |
| 244 | cursor->rtr_info->tree_savepoints[RTR_MAX_LEVELS] |
| 245 | = mtr_set_savepoint(mtr); |
| 246 | } |
| 247 | |
| 248 | mode = latch_mode == BTR_MODIFY_LEAF ? RW_X_LATCH : RW_S_LATCH; |
| 249 | latch_leaves.savepoints[1] = mtr_set_savepoint(mtr); |
| 250 | get_block = btr_block_get(page_id, page_size, mode, |
| 251 | cursor->index, mtr); |
| 252 | latch_leaves.blocks[1] = get_block; |
| 253 | #ifdef UNIV_BTR_DEBUG |
| 254 | ut_a(page_is_comp(get_block->frame) == page_is_comp(page)); |
| 255 | #endif /* UNIV_BTR_DEBUG */ |
| 256 | if (spatial) { |
| 257 | cursor->rtr_info->tree_blocks[RTR_MAX_LEVELS] |
| 258 | = get_block; |
| 259 | } |
| 260 | |
| 261 | return(latch_leaves); |
| 262 | case BTR_MODIFY_TREE: |
| 263 | /* It is exclusive for other operations which calls |
| 264 | btr_page_set_prev() */ |
| 265 | ut_ad(mtr_memo_contains_flagged( |
| 266 | mtr, |
| 267 | dict_index_get_lock(cursor->index), |
| 268 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK)); |
| 269 | /* x-latch also siblings from left to right */ |
| 270 | left_page_no = btr_page_get_prev(page, mtr); |
| 271 | mode = latch_mode; |
| 272 | |
| 273 | if (left_page_no != FIL_NULL) { |
| 274 | |
| 275 | if (spatial) { |
| 276 | cursor->rtr_info->tree_savepoints[ |
| 277 | RTR_MAX_LEVELS] = mtr_set_savepoint(mtr); |
| 278 | } |
| 279 | |
| 280 | latch_leaves.savepoints[0] = mtr_set_savepoint(mtr); |
| 281 | get_block = btr_block_get( |
| 282 | page_id_t(page_id.space(), left_page_no), |
| 283 | page_size, RW_X_LATCH, cursor->index, mtr); |
| 284 | latch_leaves.blocks[0] = get_block; |
| 285 | |
| 286 | if (spatial) { |
| 287 | cursor->rtr_info->tree_blocks[RTR_MAX_LEVELS] |
| 288 | = get_block; |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | if (spatial) { |
| 293 | cursor->rtr_info->tree_savepoints[RTR_MAX_LEVELS + 1] |
| 294 | = mtr_set_savepoint(mtr); |
| 295 | } |
| 296 | |
| 297 | latch_leaves.savepoints[1] = mtr_set_savepoint(mtr); |
| 298 | get_block = btr_block_get( |
| 299 | page_id, page_size, RW_X_LATCH, cursor->index, mtr); |
| 300 | latch_leaves.blocks[1] = get_block; |
| 301 | |
| 302 | #ifdef UNIV_BTR_DEBUG |
| 303 | /* Sanity check only after both the blocks are latched. */ |
| 304 | if (latch_leaves.blocks[0] != NULL) { |
| 305 | ut_a(page_is_comp(latch_leaves.blocks[0]->frame) |
| 306 | == page_is_comp(page)); |
| 307 | ut_a(btr_page_get_next( |
| 308 | latch_leaves.blocks[0]->frame, mtr) |
| 309 | == page_get_page_no(page)); |
| 310 | } |
| 311 | ut_a(page_is_comp(get_block->frame) == page_is_comp(page)); |
| 312 | #endif /* UNIV_BTR_DEBUG */ |
| 313 | |
| 314 | if (spatial) { |
| 315 | cursor->rtr_info->tree_blocks[RTR_MAX_LEVELS + 1] |
| 316 | = get_block; |
| 317 | } |
| 318 | |
| 319 | right_page_no = btr_page_get_next(page, mtr); |
| 320 | |
| 321 | if (right_page_no != FIL_NULL) { |
| 322 | if (spatial) { |
| 323 | cursor->rtr_info->tree_savepoints[ |
| 324 | RTR_MAX_LEVELS + 2] = mtr_set_savepoint( |
| 325 | mtr); |
| 326 | } |
| 327 | latch_leaves.savepoints[2] = mtr_set_savepoint(mtr); |
| 328 | get_block = btr_block_get( |
| 329 | page_id_t(page_id.space(), right_page_no), |
| 330 | page_size, RW_X_LATCH, cursor->index, mtr); |
| 331 | latch_leaves.blocks[2] = get_block; |
| 332 | #ifdef UNIV_BTR_DEBUG |
| 333 | ut_a(page_is_comp(get_block->frame) |
| 334 | == page_is_comp(page)); |
| 335 | ut_a(btr_page_get_prev(get_block->frame, mtr) |
| 336 | == page_get_page_no(page)); |
| 337 | #endif /* UNIV_BTR_DEBUG */ |
| 338 | if (spatial) { |
| 339 | cursor->rtr_info->tree_blocks[ |
| 340 | RTR_MAX_LEVELS + 2] = get_block; |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | return(latch_leaves); |
| 345 | |
| 346 | case BTR_SEARCH_PREV: |
| 347 | case BTR_MODIFY_PREV: |
| 348 | mode = latch_mode == BTR_SEARCH_PREV ? RW_S_LATCH : RW_X_LATCH; |
| 349 | /* latch also left sibling */ |
| 350 | rw_lock_s_lock(&block->lock); |
| 351 | left_page_no = btr_page_get_prev(page, mtr); |
| 352 | rw_lock_s_unlock(&block->lock); |
| 353 | |
| 354 | if (left_page_no != FIL_NULL) { |
| 355 | latch_leaves.savepoints[0] = mtr_set_savepoint(mtr); |
| 356 | get_block = btr_block_get( |
| 357 | page_id_t(page_id.space(), left_page_no), |
| 358 | page_size, mode, cursor->index, mtr); |
| 359 | latch_leaves.blocks[0] = get_block; |
| 360 | cursor->left_block = get_block; |
| 361 | #ifdef UNIV_BTR_DEBUG |
| 362 | ut_a(page_is_comp(get_block->frame) |
| 363 | == page_is_comp(page)); |
| 364 | ut_a(btr_page_get_next(get_block->frame, mtr) |
| 365 | == page_get_page_no(page)); |
| 366 | #endif /* UNIV_BTR_DEBUG */ |
| 367 | } |
| 368 | |
| 369 | latch_leaves.savepoints[1] = mtr_set_savepoint(mtr); |
| 370 | get_block = btr_block_get(page_id, page_size, mode, |
| 371 | cursor->index, mtr); |
| 372 | latch_leaves.blocks[1] = get_block; |
| 373 | #ifdef UNIV_BTR_DEBUG |
| 374 | ut_a(page_is_comp(get_block->frame) == page_is_comp(page)); |
| 375 | #endif /* UNIV_BTR_DEBUG */ |
| 376 | return(latch_leaves); |
| 377 | case BTR_CONT_MODIFY_TREE: |
| 378 | ut_ad(dict_index_is_spatial(cursor->index)); |
| 379 | return(latch_leaves); |
| 380 | } |
| 381 | |
| 382 | ut_error; |
| 383 | return(latch_leaves); |
| 384 | } |
| 385 | |
| 386 | /** Load the instant ALTER TABLE metadata from the clustered index |
| 387 | when loading a table definition. |
| 388 | @param[in,out] index clustered index definition |
| 389 | @param[in,out] mtr mini-transaction |
| 390 | @return error code |
| 391 | @retval DB_SUCCESS if no error occurred |
| 392 | @retval DB_CORRUPTION if any corruption was noticed */ |
| 393 | static |
| 394 | dberr_t |
| 395 | btr_cur_instant_init_low(dict_index_t* index, mtr_t* mtr) |
| 396 | { |
| 397 | ut_ad(index->is_primary()); |
| 398 | ut_ad(index->n_core_null_bytes == dict_index_t::NO_CORE_NULL_BYTES); |
| 399 | ut_ad(index->table->supports_instant()); |
| 400 | ut_ad(index->table->is_readable()); |
| 401 | |
| 402 | page_t* root = btr_root_get(index, mtr); |
| 403 | |
| 404 | if (!root || btr_cur_instant_root_init(index, root)) { |
| 405 | ib::error() << "Table "<< index->table->name |
| 406 | << " has an unreadable root page"; |
| 407 | index->table->corrupted = true; |
| 408 | return DB_CORRUPTION; |
| 409 | } |
| 410 | |
| 411 | ut_ad(index->n_core_null_bytes != dict_index_t::NO_CORE_NULL_BYTES); |
| 412 | |
| 413 | if (!index->is_instant()) { |
| 414 | return DB_SUCCESS; |
| 415 | } |
| 416 | |
| 417 | btr_cur_t cur; |
| 418 | dberr_t err = btr_cur_open_at_index_side(true, index, BTR_SEARCH_LEAF, |
| 419 | &cur, 0, mtr); |
| 420 | if (err != DB_SUCCESS) { |
| 421 | index->table->corrupted = true; |
| 422 | return err; |
| 423 | } |
| 424 | |
| 425 | ut_ad(page_cur_is_before_first(&cur.page_cur)); |
| 426 | ut_ad(page_is_leaf(cur.page_cur.block->frame)); |
| 427 | |
| 428 | page_cur_move_to_next(&cur.page_cur); |
| 429 | |
| 430 | const rec_t* rec = cur.page_cur.rec; |
| 431 | |
| 432 | if (page_rec_is_supremum(rec) || !rec_is_default_row(rec, index)) { |
| 433 | ib::error() << "Table "<< index->table->name |
| 434 | << " is missing instant ALTER metadata"; |
| 435 | index->table->corrupted = true; |
| 436 | return DB_CORRUPTION; |
| 437 | } |
| 438 | |
| 439 | if (dict_table_is_comp(index->table)) { |
| 440 | if (rec_get_info_bits(rec, true) != REC_INFO_MIN_REC_FLAG |
| 441 | && rec_get_status(rec) != REC_STATUS_COLUMNS_ADDED) { |
| 442 | incompatible: |
| 443 | ib::error() << "Table "<< index->table->name |
| 444 | << " contains unrecognizable " |
| 445 | "instant ALTER metadata"; |
| 446 | index->table->corrupted = true; |
| 447 | return DB_CORRUPTION; |
| 448 | } |
| 449 | } else if (rec_get_info_bits(rec, false) != REC_INFO_MIN_REC_FLAG) { |
| 450 | goto incompatible; |
| 451 | } |
| 452 | |
| 453 | /* Read the 'default row'. We can get here on server restart |
| 454 | or when the table was evicted from the data dictionary cache |
| 455 | and is now being accessed again. |
| 456 | |
| 457 | Here, READ COMMITTED and REPEATABLE READ should be equivalent. |
| 458 | Committing the ADD COLUMN operation would acquire |
| 459 | MDL_EXCLUSIVE and LOCK_X|LOCK_TABLE, which would prevent any |
| 460 | concurrent operations on the table, including table eviction |
| 461 | from the cache. */ |
| 462 | |
| 463 | mem_heap_t* heap = NULL; |
| 464 | ulint* offsets = rec_get_offsets(rec, index, NULL, true, |
| 465 | ULINT_UNDEFINED, &heap); |
| 466 | if (rec_offs_any_default(offsets)) { |
| 467 | inconsistent: |
| 468 | mem_heap_free(heap); |
| 469 | goto incompatible; |
| 470 | } |
| 471 | |
| 472 | /* In fact, because we only ever append fields to the 'default |
| 473 | value' record, it is also OK to perform READ UNCOMMITTED and |
| 474 | then ignore any extra fields, provided that |
| 475 | trx_sys.is_registered(DB_TRX_ID). */ |
| 476 | if (rec_offs_n_fields(offsets) > index->n_fields |
| 477 | && !trx_sys.is_registered(current_trx(), |
| 478 | row_get_rec_trx_id(rec, index, |
| 479 | offsets))) { |
| 480 | goto inconsistent; |
| 481 | } |
| 482 | |
| 483 | for (unsigned i = index->n_core_fields; i < index->n_fields; i++) { |
| 484 | ulint len; |
| 485 | const byte* data = rec_get_nth_field(rec, offsets, i, &len); |
| 486 | dict_col_t* col = index->fields[i].col; |
| 487 | ut_ad(!col->is_instant()); |
| 488 | ut_ad(!col->def_val.data); |
| 489 | col->def_val.len = len; |
| 490 | switch (len) { |
| 491 | case UNIV_SQL_NULL: |
| 492 | continue; |
| 493 | case 0: |
| 494 | col->def_val.data = field_ref_zero; |
| 495 | continue; |
| 496 | } |
| 497 | ut_ad(len != UNIV_SQL_DEFAULT); |
| 498 | if (!rec_offs_nth_extern(offsets, i)) { |
| 499 | col->def_val.data = mem_heap_dup( |
| 500 | index->table->heap, data, len); |
| 501 | } else if (len < BTR_EXTERN_FIELD_REF_SIZE |
| 502 | || !memcmp(data + len - BTR_EXTERN_FIELD_REF_SIZE, |
| 503 | field_ref_zero, |
| 504 | BTR_EXTERN_FIELD_REF_SIZE)) { |
| 505 | col->def_val.len = UNIV_SQL_DEFAULT; |
| 506 | goto inconsistent; |
| 507 | } else { |
| 508 | col->def_val.data = btr_copy_externally_stored_field( |
| 509 | &col->def_val.len, data, |
| 510 | dict_table_page_size(index->table), |
| 511 | len, index->table->heap); |
| 512 | } |
| 513 | } |
| 514 | |
| 515 | mem_heap_free(heap); |
| 516 | return DB_SUCCESS; |
| 517 | } |
| 518 | |
| 519 | /** Load the instant ALTER TABLE metadata from the clustered index |
| 520 | when loading a table definition. |
| 521 | @param[in,out] table table definition from the data dictionary |
| 522 | @return error code |
| 523 | @retval DB_SUCCESS if no error occurred */ |
| 524 | dberr_t |
| 525 | btr_cur_instant_init(dict_table_t* table) |
| 526 | { |
| 527 | mtr_t mtr; |
| 528 | dict_index_t* index = dict_table_get_first_index(table); |
| 529 | mtr.start(); |
| 530 | dberr_t err = index |
| 531 | ? btr_cur_instant_init_low(index, &mtr) |
| 532 | : DB_CORRUPTION; |
| 533 | mtr.commit(); |
| 534 | return(err); |
| 535 | } |
| 536 | |
| 537 | /** Initialize the n_core_null_bytes on first access to a clustered |
| 538 | index root page. |
| 539 | @param[in] index clustered index that is on its first access |
| 540 | @param[in] page clustered index root page |
| 541 | @return whether the page is corrupted */ |
| 542 | bool |
| 543 | btr_cur_instant_root_init(dict_index_t* index, const page_t* page) |
| 544 | { |
| 545 | ut_ad(page_is_root(page)); |
| 546 | ut_ad(!page_is_comp(page) == !dict_table_is_comp(index->table)); |
| 547 | ut_ad(index->is_primary()); |
| 548 | ut_ad(!index->is_instant()); |
| 549 | ut_ad(index->table->supports_instant()); |
| 550 | /* This is normally executed as part of btr_cur_instant_init() |
| 551 | when dict_load_table_one() is loading a table definition. |
| 552 | Other threads should not access or modify the n_core_null_bytes, |
| 553 | n_core_fields before dict_load_table_one() returns. |
| 554 | |
| 555 | This can also be executed during IMPORT TABLESPACE, where the |
| 556 | table definition is exclusively locked. */ |
| 557 | |
| 558 | switch (fil_page_get_type(page)) { |
| 559 | default: |
| 560 | ut_ad(!"wrong page type"); |
| 561 | return true; |
| 562 | case FIL_PAGE_INDEX: |
| 563 | /* The field PAGE_INSTANT is guaranteed 0 on clustered |
| 564 | index root pages of ROW_FORMAT=COMPACT or |
| 565 | ROW_FORMAT=DYNAMIC when instant ADD COLUMN is not used. */ |
| 566 | ut_ad(!page_is_comp(page) || !page_get_instant(page)); |
| 567 | index->n_core_null_bytes = UT_BITS_IN_BYTES( |
| 568 | unsigned(index->n_nullable)); |
| 569 | return false; |
| 570 | case FIL_PAGE_TYPE_INSTANT: |
| 571 | break; |
| 572 | } |
| 573 | |
| 574 | uint16_t n = page_get_instant(page); |
| 575 | if (n < index->n_uniq + DATA_ROLL_PTR || n > index->n_fields) { |
| 576 | /* The PRIMARY KEY (or hidden DB_ROW_ID) and |
| 577 | DB_TRX_ID,DB_ROLL_PTR columns must always be present |
| 578 | as 'core' fields. All fields, including those for |
| 579 | instantly added columns, must be present in the data |
| 580 | dictionary. */ |
| 581 | return true; |
| 582 | } |
| 583 | index->n_core_fields = n; |
| 584 | ut_ad(!index->is_dummy); |
| 585 | ut_d(index->is_dummy = true); |
| 586 | index->n_core_null_bytes = n == index->n_fields |
| 587 | ? UT_BITS_IN_BYTES(unsigned(index->n_nullable)) |
| 588 | : UT_BITS_IN_BYTES(index->get_n_nullable(n)); |
| 589 | ut_d(index->is_dummy = false); |
| 590 | return false; |
| 591 | } |
| 592 | |
| 593 | /** Optimistically latches the leaf page or pages requested. |
| 594 | @param[in] block guessed buffer block |
| 595 | @param[in] modify_clock modify clock value |
| 596 | @param[in,out] latch_mode BTR_SEARCH_LEAF, ... |
| 597 | @param[in,out] cursor cursor |
| 598 | @param[in] file file name |
| 599 | @param[in] line line where called |
| 600 | @param[in] mtr mini-transaction |
| 601 | @return true if success */ |
| 602 | bool |
| 603 | btr_cur_optimistic_latch_leaves( |
| 604 | buf_block_t* block, |
| 605 | ib_uint64_t modify_clock, |
| 606 | ulint* latch_mode, |
| 607 | btr_cur_t* cursor, |
| 608 | const char* file, |
| 609 | unsigned line, |
| 610 | mtr_t* mtr) |
| 611 | { |
| 612 | ulint mode; |
| 613 | ulint left_page_no; |
| 614 | |
| 615 | switch (*latch_mode) { |
| 616 | case BTR_SEARCH_LEAF: |
| 617 | case BTR_MODIFY_LEAF: |
| 618 | return(buf_page_optimistic_get(*latch_mode, block, |
| 619 | modify_clock, file, line, mtr)); |
| 620 | case BTR_SEARCH_PREV: |
| 621 | case BTR_MODIFY_PREV: |
| 622 | mode = *latch_mode == BTR_SEARCH_PREV |
| 623 | ? RW_S_LATCH : RW_X_LATCH; |
| 624 | |
| 625 | buf_page_mutex_enter(block); |
| 626 | if (buf_block_get_state(block) != BUF_BLOCK_FILE_PAGE) { |
| 627 | buf_page_mutex_exit(block); |
| 628 | return(false); |
| 629 | } |
| 630 | /* pin the block not to be relocated */ |
| 631 | buf_block_buf_fix_inc(block, file, line); |
| 632 | buf_page_mutex_exit(block); |
| 633 | |
| 634 | rw_lock_s_lock(&block->lock); |
| 635 | if (block->modify_clock != modify_clock) { |
| 636 | rw_lock_s_unlock(&block->lock); |
| 637 | |
| 638 | goto unpin_failed; |
| 639 | } |
| 640 | left_page_no = btr_page_get_prev( |
| 641 | buf_block_get_frame(block), mtr); |
| 642 | rw_lock_s_unlock(&block->lock); |
| 643 | |
| 644 | if (left_page_no != FIL_NULL) { |
| 645 | cursor->left_block = btr_block_get( |
| 646 | page_id_t(cursor->index->table->space->id, |
| 647 | left_page_no), |
| 648 | page_size_t(cursor->index->table->space |
| 649 | ->flags), |
| 650 | mode, cursor->index, mtr); |
| 651 | } else { |
| 652 | cursor->left_block = NULL; |
| 653 | } |
| 654 | |
| 655 | if (buf_page_optimistic_get(mode, block, modify_clock, |
| 656 | file, line, mtr)) { |
| 657 | if (btr_page_get_prev(buf_block_get_frame(block), mtr) |
| 658 | == left_page_no) { |
| 659 | /* adjust buf_fix_count */ |
| 660 | buf_page_mutex_enter(block); |
| 661 | buf_block_buf_fix_dec(block); |
| 662 | buf_page_mutex_exit(block); |
| 663 | |
| 664 | *latch_mode = mode; |
| 665 | return(true); |
| 666 | } else { |
| 667 | /* release the block */ |
| 668 | btr_leaf_page_release(block, mode, mtr); |
| 669 | } |
| 670 | } |
| 671 | |
| 672 | /* release the left block */ |
| 673 | if (cursor->left_block != NULL) { |
| 674 | btr_leaf_page_release(cursor->left_block, |
| 675 | mode, mtr); |
| 676 | } |
| 677 | unpin_failed: |
| 678 | /* unpin the block */ |
| 679 | buf_page_mutex_enter(block); |
| 680 | buf_block_buf_fix_dec(block); |
| 681 | buf_page_mutex_exit(block); |
| 682 | |
| 683 | return(false); |
| 684 | |
| 685 | default: |
| 686 | ut_error; |
| 687 | return(false); |
| 688 | } |
| 689 | } |
| 690 | |
| 691 | /** |
| 692 | Gets intention in btr_intention_t from latch_mode, and cleares the intention |
| 693 | at the latch_mode. |
| 694 | @param latch_mode in/out: pointer to latch_mode |
| 695 | @return intention for latching tree */ |
| 696 | static |
| 697 | btr_intention_t |
| 698 | btr_cur_get_and_clear_intention( |
| 699 | ulint *latch_mode) |
| 700 | { |
| 701 | btr_intention_t intention; |
| 702 | |
| 703 | switch (*latch_mode & (BTR_LATCH_FOR_INSERT | BTR_LATCH_FOR_DELETE)) { |
| 704 | case BTR_LATCH_FOR_INSERT: |
| 705 | intention = BTR_INTENTION_INSERT; |
| 706 | break; |
| 707 | case BTR_LATCH_FOR_DELETE: |
| 708 | intention = BTR_INTENTION_DELETE; |
| 709 | break; |
| 710 | default: |
| 711 | /* both or unknown */ |
| 712 | intention = BTR_INTENTION_BOTH; |
| 713 | } |
| 714 | *latch_mode &= ulint(~(BTR_LATCH_FOR_INSERT | BTR_LATCH_FOR_DELETE)); |
| 715 | |
| 716 | return(intention); |
| 717 | } |
| 718 | |
| 719 | /** |
| 720 | Gets the desired latch type for the root leaf (root page is root leaf) |
| 721 | at the latch mode. |
| 722 | @param latch_mode in: BTR_SEARCH_LEAF, ... |
| 723 | @return latch type */ |
| 724 | static |
| 725 | rw_lock_type_t |
| 726 | btr_cur_latch_for_root_leaf( |
| 727 | ulint latch_mode) |
| 728 | { |
| 729 | switch (latch_mode) { |
| 730 | case BTR_SEARCH_LEAF: |
| 731 | case BTR_SEARCH_TREE: |
| 732 | case BTR_SEARCH_PREV: |
| 733 | return(RW_S_LATCH); |
| 734 | case BTR_MODIFY_LEAF: |
| 735 | case BTR_MODIFY_TREE: |
| 736 | case BTR_MODIFY_PREV: |
| 737 | return(RW_X_LATCH); |
| 738 | case BTR_CONT_MODIFY_TREE: |
| 739 | case BTR_CONT_SEARCH_TREE: |
| 740 | /* A root page should be latched already, |
| 741 | and don't need to be latched here. |
| 742 | fall through (RW_NO_LATCH) */ |
| 743 | case BTR_NO_LATCHES: |
| 744 | return(RW_NO_LATCH); |
| 745 | } |
| 746 | |
| 747 | ut_error; |
| 748 | return(RW_NO_LATCH); /* avoid compiler warnings */ |
| 749 | } |
| 750 | |
| 751 | /** Detects whether the modifying record might need a modifying tree structure. |
| 752 | @param[in] index index |
| 753 | @param[in] page page |
| 754 | @param[in] lock_intention lock intention for the tree operation |
| 755 | @param[in] rec record (current node_ptr) |
| 756 | @param[in] rec_size size of the record or max size of node_ptr |
| 757 | @param[in] page_size page size |
| 758 | @param[in] mtr mtr |
| 759 | @return true if tree modification is needed */ |
| 760 | static |
| 761 | bool |
| 762 | btr_cur_will_modify_tree( |
| 763 | dict_index_t* index, |
| 764 | const page_t* page, |
| 765 | btr_intention_t lock_intention, |
| 766 | const rec_t* rec, |
| 767 | ulint rec_size, |
| 768 | const page_size_t& page_size, |
| 769 | mtr_t* mtr) |
| 770 | { |
| 771 | ut_ad(!page_is_leaf(page)); |
| 772 | ut_ad(mtr_memo_contains_flagged(mtr, dict_index_get_lock(index), |
| 773 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK)); |
| 774 | |
| 775 | /* Pessimistic delete of the first record causes delete & insert |
| 776 | of node_ptr at upper level. And a subsequent page shrink is |
| 777 | possible. It causes delete of node_ptr at the upper level. |
| 778 | So we should pay attention also to 2nd record not only |
| 779 | first record and last record. Because if the "delete & insert" are |
| 780 | done for the different page, the 2nd record become |
| 781 | first record and following compress might delete the record and causes |
| 782 | the uppper level node_ptr modification. */ |
| 783 | |
| 784 | if (lock_intention <= BTR_INTENTION_BOTH) { |
| 785 | ulint margin; |
| 786 | |
| 787 | /* check delete will cause. (BTR_INTENTION_BOTH |
| 788 | or BTR_INTENTION_DELETE) */ |
| 789 | /* first, 2nd, 2nd-last and last records are 4 records */ |
| 790 | if (page_get_n_recs(page) < 5) { |
| 791 | return(true); |
| 792 | } |
| 793 | |
| 794 | /* is first, 2nd or last record */ |
| 795 | if (page_rec_is_first(rec, page) |
| 796 | || (page_has_next(page) |
| 797 | && (page_rec_is_last(rec, page) |
| 798 | || page_rec_is_second_last(rec, page))) |
| 799 | || (page_has_prev(page) |
| 800 | && page_rec_is_second(rec, page))) { |
| 801 | return(true); |
| 802 | } |
| 803 | |
| 804 | if (lock_intention == BTR_INTENTION_BOTH) { |
| 805 | /* Delete at leftmost record in a page causes delete |
| 806 | & insert at its parent page. After that, the delete |
| 807 | might cause btr_compress() and delete record at its |
| 808 | parent page. Thus we should consider max 2 deletes. */ |
| 809 | |
| 810 | margin = rec_size * 2; |
| 811 | } else { |
| 812 | ut_ad(lock_intention == BTR_INTENTION_DELETE); |
| 813 | |
| 814 | margin = rec_size; |
| 815 | } |
| 816 | /* NOTE: call mach_read_from_4() directly to avoid assertion |
| 817 | failure. It is safe because we already have SX latch of the |
| 818 | index tree */ |
| 819 | if (page_get_data_size(page) |
| 820 | < margin + BTR_CUR_PAGE_COMPRESS_LIMIT(index) |
| 821 | || (mach_read_from_4(page + FIL_PAGE_NEXT) |
| 822 | == FIL_NULL |
| 823 | && mach_read_from_4(page + FIL_PAGE_PREV) |
| 824 | == FIL_NULL)) { |
| 825 | return(true); |
| 826 | } |
| 827 | } |
| 828 | |
| 829 | if (lock_intention >= BTR_INTENTION_BOTH) { |
| 830 | /* check insert will cause. BTR_INTENTION_BOTH |
| 831 | or BTR_INTENTION_INSERT*/ |
| 832 | |
| 833 | /* Once we invoke the btr_cur_limit_optimistic_insert_debug, |
| 834 | we should check it here in advance, since the max allowable |
| 835 | records in a page is limited. */ |
| 836 | LIMIT_OPTIMISTIC_INSERT_DEBUG(page_get_n_recs(page), |
| 837 | return(true)); |
| 838 | |
| 839 | /* needs 2 records' space for the case the single split and |
| 840 | insert cannot fit. |
| 841 | page_get_max_insert_size_after_reorganize() includes space |
| 842 | for page directory already */ |
| 843 | ulint max_size |
| 844 | = page_get_max_insert_size_after_reorganize(page, 2); |
| 845 | |
| 846 | if (max_size < BTR_CUR_PAGE_REORGANIZE_LIMIT + rec_size |
| 847 | || max_size < rec_size * 2) { |
| 848 | return(true); |
| 849 | } |
| 850 | /* TODO: optimize this condition for compressed page. |
| 851 | this is based on the worst compress rate. |
| 852 | currently looking only uncompressed page, but we can look |
| 853 | also compressed page page_zip_available() if already in the |
| 854 | buffer pool */ |
| 855 | /* needs 2 records' space also for worst compress rate. */ |
| 856 | if (page_size.is_compressed() |
| 857 | && page_zip_empty_size(index->n_fields, |
| 858 | page_size.physical()) |
| 859 | < rec_size * 2 + page_get_data_size(page) |
| 860 | + page_dir_calc_reserved_space( |
| 861 | ulint(page_get_n_recs(page)) + 2) + 1) { |
| 862 | return(true); |
| 863 | } |
| 864 | } |
| 865 | |
| 866 | return(false); |
| 867 | } |
| 868 | |
| 869 | /** Detects whether the modifying record might need a opposite modification |
| 870 | to the intention. |
| 871 | @param[in] page page |
| 872 | @param[in] lock_intention lock intention for the tree operation |
| 873 | @param[in] rec record (current node_ptr) |
| 874 | @return true if tree modification is needed */ |
| 875 | static |
| 876 | bool |
| 877 | btr_cur_need_opposite_intention( |
| 878 | const page_t* page, |
| 879 | btr_intention_t lock_intention, |
| 880 | const rec_t* rec) |
| 881 | { |
| 882 | switch (lock_intention) { |
| 883 | case BTR_INTENTION_DELETE: |
| 884 | return (page_has_prev(page) && page_rec_is_first(rec, page)) || |
| 885 | (page_has_next(page) && page_rec_is_last(rec, page)); |
| 886 | case BTR_INTENTION_INSERT: |
| 887 | return page_has_next(page) && page_rec_is_last(rec, page); |
| 888 | case BTR_INTENTION_BOTH: |
| 889 | return(false); |
| 890 | } |
| 891 | |
| 892 | ut_error; |
| 893 | return(false); |
| 894 | } |
| 895 | |
| 896 | /********************************************************************//** |
| 897 | Searches an index tree and positions a tree cursor on a given level. |
| 898 | NOTE: n_fields_cmp in tuple must be set so that it cannot be compared |
| 899 | to node pointer page number fields on the upper levels of the tree! |
| 900 | Note that if mode is PAGE_CUR_LE, which is used in inserts, then |
| 901 | cursor->up_match and cursor->low_match both will have sensible values. |
| 902 | If mode is PAGE_CUR_GE, then up_match will a have a sensible value. |
| 903 | |
| 904 | If mode is PAGE_CUR_LE , cursor is left at the place where an insert of the |
| 905 | search tuple should be performed in the B-tree. InnoDB does an insert |
| 906 | immediately after the cursor. Thus, the cursor may end up on a user record, |
| 907 | or on a page infimum record. */ |
| 908 | dberr_t |
| 909 | btr_cur_search_to_nth_level_func( |
| 910 | dict_index_t* index, /*!< in: index */ |
| 911 | ulint level, /*!< in: the tree level of search */ |
| 912 | const dtuple_t* tuple, /*!< in: data tuple; NOTE: n_fields_cmp in |
| 913 | tuple must be set so that it cannot get |
| 914 | compared to the node ptr page number field! */ |
| 915 | page_cur_mode_t mode, /*!< in: PAGE_CUR_L, ...; |
| 916 | Inserts should always be made using |
| 917 | PAGE_CUR_LE to search the position! */ |
| 918 | ulint latch_mode, /*!< in: BTR_SEARCH_LEAF, ..., ORed with |
| 919 | at most one of BTR_INSERT, BTR_DELETE_MARK, |
| 920 | BTR_DELETE, or BTR_ESTIMATE; |
| 921 | cursor->left_block is used to store a pointer |
| 922 | to the left neighbor page, in the cases |
| 923 | BTR_SEARCH_PREV and BTR_MODIFY_PREV; |
| 924 | NOTE that if ahi_latch, we might not have a |
| 925 | cursor page latch, we assume that ahi_latch |
| 926 | protects the record! */ |
| 927 | btr_cur_t* cursor, /*!< in/out: tree cursor; the cursor page is |
| 928 | s- or x-latched, but see also above! */ |
| 929 | #ifdef BTR_CUR_HASH_ADAPT |
| 930 | rw_lock_t* ahi_latch, |
| 931 | /*!< in: currently held btr_search_latch |
| 932 | (in RW_S_LATCH mode), or NULL */ |
| 933 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 934 | const char* file, /*!< in: file name */ |
| 935 | unsigned line, /*!< in: line where called */ |
| 936 | mtr_t* mtr, /*!< in: mtr */ |
| 937 | ib_uint64_t autoinc)/*!< in: PAGE_ROOT_AUTO_INC to be written |
| 938 | (0 if none) */ |
| 939 | { |
| 940 | page_t* page = NULL; /* remove warning */ |
| 941 | buf_block_t* block; |
| 942 | buf_block_t* guess; |
| 943 | ulint height; |
| 944 | ulint up_match; |
| 945 | ulint up_bytes; |
| 946 | ulint low_match; |
| 947 | ulint low_bytes; |
| 948 | ulint savepoint; |
| 949 | ulint rw_latch; |
| 950 | page_cur_mode_t page_mode; |
| 951 | page_cur_mode_t search_mode = PAGE_CUR_UNSUPP; |
| 952 | ulint buf_mode; |
| 953 | ulint estimate; |
| 954 | ulint node_ptr_max_size = srv_page_size / 2; |
| 955 | page_cur_t* page_cursor; |
| 956 | btr_op_t btr_op; |
| 957 | ulint root_height = 0; /* remove warning */ |
| 958 | dberr_t err = DB_SUCCESS; |
| 959 | |
| 960 | ulint upper_rw_latch, root_leaf_rw_latch; |
| 961 | btr_intention_t lock_intention; |
| 962 | bool modify_external; |
| 963 | buf_block_t* tree_blocks[BTR_MAX_LEVELS]; |
| 964 | ulint tree_savepoints[BTR_MAX_LEVELS]; |
| 965 | ulint n_blocks = 0; |
| 966 | ulint n_releases = 0; |
| 967 | bool detected_same_key_root = false; |
| 968 | |
| 969 | bool retrying_for_search_prev = false; |
| 970 | ulint leftmost_from_level = 0; |
| 971 | buf_block_t** prev_tree_blocks = NULL; |
| 972 | ulint* prev_tree_savepoints = NULL; |
| 973 | ulint prev_n_blocks = 0; |
| 974 | ulint prev_n_releases = 0; |
| 975 | bool need_path = true; |
| 976 | bool rtree_parent_modified = false; |
| 977 | bool mbr_adj = false; |
| 978 | bool found = false; |
| 979 | |
| 980 | DBUG_ENTER("btr_cur_search_to_nth_level"); |
| 981 | |
| 982 | #ifdef BTR_CUR_ADAPT |
| 983 | btr_search_t* info; |
| 984 | #endif /* BTR_CUR_ADAPT */ |
| 985 | mem_heap_t* heap = NULL; |
| 986 | ulint offsets_[REC_OFFS_NORMAL_SIZE]; |
| 987 | ulint* offsets = offsets_; |
| 988 | ulint offsets2_[REC_OFFS_NORMAL_SIZE]; |
| 989 | ulint* offsets2 = offsets2_; |
| 990 | rec_offs_init(offsets_); |
| 991 | rec_offs_init(offsets2_); |
| 992 | /* Currently, PAGE_CUR_LE is the only search mode used for searches |
| 993 | ending to upper levels */ |
| 994 | |
| 995 | ut_ad(level == 0 || mode == PAGE_CUR_LE |
| 996 | || RTREE_SEARCH_MODE(mode)); |
| 997 | ut_ad(dict_index_check_search_tuple(index, tuple)); |
| 998 | ut_ad(!dict_index_is_ibuf(index) || ibuf_inside(mtr)); |
| 999 | ut_ad(dtuple_check_typed(tuple)); |
| 1000 | ut_ad(!(index->type & DICT_FTS)); |
| 1001 | ut_ad(index->page != FIL_NULL); |
| 1002 | |
| 1003 | UNIV_MEM_INVALID(&cursor->up_match, sizeof cursor->up_match); |
| 1004 | UNIV_MEM_INVALID(&cursor->up_bytes, sizeof cursor->up_bytes); |
| 1005 | UNIV_MEM_INVALID(&cursor->low_match, sizeof cursor->low_match); |
| 1006 | UNIV_MEM_INVALID(&cursor->low_bytes, sizeof cursor->low_bytes); |
| 1007 | #ifdef UNIV_DEBUG |
| 1008 | cursor->up_match = ULINT_UNDEFINED; |
| 1009 | cursor->low_match = ULINT_UNDEFINED; |
| 1010 | #endif /* UNIV_DEBUG */ |
| 1011 | |
| 1012 | ibool s_latch_by_caller; |
| 1013 | |
| 1014 | s_latch_by_caller = latch_mode & BTR_ALREADY_S_LATCHED; |
| 1015 | |
| 1016 | ut_ad(!s_latch_by_caller |
| 1017 | || srv_read_only_mode |
| 1018 | || mtr_memo_contains_flagged(mtr, |
| 1019 | dict_index_get_lock(index), |
| 1020 | MTR_MEMO_S_LOCK |
| 1021 | | MTR_MEMO_SX_LOCK)); |
| 1022 | |
| 1023 | /* These flags are mutually exclusive, they are lumped together |
| 1024 | with the latch mode for historical reasons. It's possible for |
| 1025 | none of the flags to be set. */ |
| 1026 | switch (UNIV_EXPECT(latch_mode |
| 1027 | & (BTR_INSERT | BTR_DELETE | BTR_DELETE_MARK), |
| 1028 | 0)) { |
| 1029 | case 0: |
| 1030 | btr_op = BTR_NO_OP; |
| 1031 | break; |
| 1032 | case BTR_INSERT: |
| 1033 | btr_op = (latch_mode & BTR_IGNORE_SEC_UNIQUE) |
| 1034 | ? BTR_INSERT_IGNORE_UNIQUE_OP |
| 1035 | : BTR_INSERT_OP; |
| 1036 | break; |
| 1037 | case BTR_DELETE: |
| 1038 | btr_op = BTR_DELETE_OP; |
| 1039 | ut_a(cursor->purge_node); |
| 1040 | break; |
| 1041 | case BTR_DELETE_MARK: |
| 1042 | btr_op = BTR_DELMARK_OP; |
| 1043 | break; |
| 1044 | default: |
| 1045 | /* only one of BTR_INSERT, BTR_DELETE, BTR_DELETE_MARK |
| 1046 | should be specified at a time */ |
| 1047 | ut_error; |
| 1048 | } |
| 1049 | |
| 1050 | /* Operations on the insert buffer tree cannot be buffered. */ |
| 1051 | ut_ad(btr_op == BTR_NO_OP || !dict_index_is_ibuf(index)); |
| 1052 | /* Operations on the clustered index cannot be buffered. */ |
| 1053 | ut_ad(btr_op == BTR_NO_OP || !dict_index_is_clust(index)); |
| 1054 | /* Operations on the temporary table(indexes) cannot be buffered. */ |
| 1055 | ut_ad(btr_op == BTR_NO_OP || !index->table->is_temporary()); |
| 1056 | /* Operation on the spatial index cannot be buffered. */ |
| 1057 | ut_ad(btr_op == BTR_NO_OP || !dict_index_is_spatial(index)); |
| 1058 | |
| 1059 | estimate = latch_mode & BTR_ESTIMATE; |
| 1060 | |
| 1061 | lock_intention = btr_cur_get_and_clear_intention(&latch_mode); |
| 1062 | |
| 1063 | modify_external = latch_mode & BTR_MODIFY_EXTERNAL; |
| 1064 | |
| 1065 | /* Turn the flags unrelated to the latch mode off. */ |
| 1066 | latch_mode = BTR_LATCH_MODE_WITHOUT_FLAGS(latch_mode); |
| 1067 | |
| 1068 | ut_ad(!modify_external || latch_mode == BTR_MODIFY_LEAF); |
| 1069 | |
| 1070 | ut_ad(!s_latch_by_caller |
| 1071 | || latch_mode == BTR_SEARCH_LEAF |
| 1072 | || latch_mode == BTR_SEARCH_TREE |
| 1073 | || latch_mode == BTR_MODIFY_LEAF); |
| 1074 | |
| 1075 | ut_ad(autoinc == 0 || dict_index_is_clust(index)); |
| 1076 | ut_ad(autoinc == 0 |
| 1077 | || latch_mode == BTR_MODIFY_TREE |
| 1078 | || latch_mode == BTR_MODIFY_LEAF); |
| 1079 | ut_ad(autoinc == 0 || level == 0); |
| 1080 | |
| 1081 | cursor->flag = BTR_CUR_BINARY; |
| 1082 | cursor->index = index; |
| 1083 | |
| 1084 | #ifndef BTR_CUR_ADAPT |
| 1085 | guess = NULL; |
| 1086 | #else |
| 1087 | info = btr_search_get_info(index); |
| 1088 | |
| 1089 | if (!buf_pool_is_obsolete(info->withdraw_clock)) { |
| 1090 | guess = info->root_guess; |
| 1091 | } else { |
| 1092 | guess = NULL; |
| 1093 | } |
| 1094 | |
| 1095 | #ifdef BTR_CUR_HASH_ADAPT |
| 1096 | |
| 1097 | # ifdef UNIV_SEARCH_PERF_STAT |
| 1098 | info->n_searches++; |
| 1099 | # endif |
| 1100 | if (autoinc == 0 |
| 1101 | && latch_mode <= BTR_MODIFY_LEAF |
| 1102 | && info->last_hash_succ |
| 1103 | # ifdef MYSQL_INDEX_DISABLE_AHI |
| 1104 | && !index->disable_ahi |
| 1105 | # endif |
| 1106 | && !estimate |
| 1107 | # ifdef PAGE_CUR_LE_OR_EXTENDS |
| 1108 | && mode != PAGE_CUR_LE_OR_EXTENDS |
| 1109 | # endif /* PAGE_CUR_LE_OR_EXTENDS */ |
| 1110 | && !dict_index_is_spatial(index) |
| 1111 | /* If !ahi_latch, we do a dirty read of |
| 1112 | btr_search_enabled below, and btr_search_guess_on_hash() |
| 1113 | will have to check it again. */ |
| 1114 | && btr_search_enabled |
| 1115 | && !modify_external |
| 1116 | && !(tuple->info_bits & REC_INFO_MIN_REC_FLAG) |
| 1117 | && btr_search_guess_on_hash(index, info, tuple, mode, |
| 1118 | latch_mode, cursor, |
| 1119 | ahi_latch, mtr)) { |
| 1120 | |
| 1121 | /* Search using the hash index succeeded */ |
| 1122 | |
| 1123 | ut_ad(cursor->up_match != ULINT_UNDEFINED |
| 1124 | || mode != PAGE_CUR_GE); |
| 1125 | ut_ad(cursor->up_match != ULINT_UNDEFINED |
| 1126 | || mode != PAGE_CUR_LE); |
| 1127 | ut_ad(cursor->low_match != ULINT_UNDEFINED |
| 1128 | || mode != PAGE_CUR_LE); |
| 1129 | btr_cur_n_sea++; |
| 1130 | |
| 1131 | DBUG_RETURN(err); |
| 1132 | } |
| 1133 | # endif /* BTR_CUR_HASH_ADAPT */ |
| 1134 | #endif /* BTR_CUR_ADAPT */ |
| 1135 | btr_cur_n_non_sea++; |
| 1136 | |
| 1137 | /* If the hash search did not succeed, do binary search down the |
| 1138 | tree */ |
| 1139 | |
| 1140 | #ifdef BTR_CUR_HASH_ADAPT |
| 1141 | if (ahi_latch) { |
| 1142 | /* Release possible search latch to obey latching order */ |
| 1143 | rw_lock_s_unlock(ahi_latch); |
| 1144 | } |
| 1145 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 1146 | |
| 1147 | /* Store the position of the tree latch we push to mtr so that we |
| 1148 | know how to release it when we have latched leaf node(s) */ |
| 1149 | |
| 1150 | savepoint = mtr_set_savepoint(mtr); |
| 1151 | |
| 1152 | switch (latch_mode) { |
| 1153 | case BTR_MODIFY_TREE: |
| 1154 | /* Most of delete-intended operations are purging. |
| 1155 | Free blocks and read IO bandwidth should be prior |
| 1156 | for them, when the history list is glowing huge. */ |
| 1157 | if (lock_intention == BTR_INTENTION_DELETE |
| 1158 | && trx_sys.history_size() > BTR_CUR_FINE_HISTORY_LENGTH |
| 1159 | && buf_get_n_pending_read_ios()) { |
| 1160 | mtr_x_lock(dict_index_get_lock(index), mtr); |
| 1161 | } else if (dict_index_is_spatial(index) |
| 1162 | && lock_intention <= BTR_INTENTION_BOTH) { |
| 1163 | /* X lock the if there is possibility of |
| 1164 | pessimistic delete on spatial index. As we could |
| 1165 | lock upward for the tree */ |
| 1166 | |
| 1167 | mtr_x_lock(dict_index_get_lock(index), mtr); |
| 1168 | } else { |
| 1169 | mtr_sx_lock(dict_index_get_lock(index), mtr); |
| 1170 | } |
| 1171 | upper_rw_latch = RW_X_LATCH; |
| 1172 | break; |
| 1173 | case BTR_CONT_MODIFY_TREE: |
| 1174 | case BTR_CONT_SEARCH_TREE: |
| 1175 | /* Do nothing */ |
| 1176 | ut_ad(srv_read_only_mode |
| 1177 | || mtr_memo_contains_flagged(mtr, |
| 1178 | dict_index_get_lock(index), |
| 1179 | MTR_MEMO_X_LOCK |
| 1180 | | MTR_MEMO_SX_LOCK)); |
| 1181 | if (dict_index_is_spatial(index) |
| 1182 | && latch_mode == BTR_CONT_MODIFY_TREE) { |
| 1183 | /* If we are about to locating parent page for split |
| 1184 | and/or merge operation for R-Tree index, X latch |
| 1185 | the parent */ |
| 1186 | upper_rw_latch = RW_X_LATCH; |
| 1187 | } else { |
| 1188 | upper_rw_latch = RW_NO_LATCH; |
| 1189 | } |
| 1190 | break; |
| 1191 | default: |
| 1192 | if (!srv_read_only_mode) { |
| 1193 | if (s_latch_by_caller) { |
| 1194 | ut_ad(rw_lock_own(dict_index_get_lock(index), |
| 1195 | RW_LOCK_S)); |
| 1196 | } else if (!modify_external) { |
| 1197 | /* BTR_SEARCH_TREE is intended to be used with |
| 1198 | BTR_ALREADY_S_LATCHED */ |
| 1199 | ut_ad(latch_mode != BTR_SEARCH_TREE); |
| 1200 | |
| 1201 | mtr_s_lock(dict_index_get_lock(index), mtr); |
| 1202 | } else { |
| 1203 | /* BTR_MODIFY_EXTERNAL needs to be excluded */ |
| 1204 | mtr_sx_lock(dict_index_get_lock(index), mtr); |
| 1205 | } |
| 1206 | upper_rw_latch = RW_S_LATCH; |
| 1207 | } else { |
| 1208 | upper_rw_latch = RW_NO_LATCH; |
| 1209 | } |
| 1210 | } |
| 1211 | root_leaf_rw_latch = btr_cur_latch_for_root_leaf(latch_mode); |
| 1212 | |
| 1213 | page_cursor = btr_cur_get_page_cur(cursor); |
| 1214 | |
| 1215 | const page_size_t page_size(index->table->space->flags); |
| 1216 | |
| 1217 | /* Start with the root page. */ |
| 1218 | page_id_t page_id(index->table->space->id, index->page); |
| 1219 | |
| 1220 | if (root_leaf_rw_latch == RW_X_LATCH) { |
| 1221 | node_ptr_max_size = dict_index_node_ptr_max_size(index); |
| 1222 | } |
| 1223 | |
| 1224 | up_match = 0; |
| 1225 | up_bytes = 0; |
| 1226 | low_match = 0; |
| 1227 | low_bytes = 0; |
| 1228 | |
| 1229 | height = ULINT_UNDEFINED; |
| 1230 | |
| 1231 | /* We use these modified search modes on non-leaf levels of the |
| 1232 | B-tree. These let us end up in the right B-tree leaf. In that leaf |
| 1233 | we use the original search mode. */ |
| 1234 | |
| 1235 | switch (mode) { |
| 1236 | case PAGE_CUR_GE: |
| 1237 | page_mode = PAGE_CUR_L; |
| 1238 | break; |
| 1239 | case PAGE_CUR_G: |
| 1240 | page_mode = PAGE_CUR_LE; |
| 1241 | break; |
| 1242 | default: |
| 1243 | #ifdef PAGE_CUR_LE_OR_EXTENDS |
| 1244 | ut_ad(mode == PAGE_CUR_L || mode == PAGE_CUR_LE |
| 1245 | || RTREE_SEARCH_MODE(mode) |
| 1246 | || mode == PAGE_CUR_LE_OR_EXTENDS); |
| 1247 | #else /* PAGE_CUR_LE_OR_EXTENDS */ |
| 1248 | ut_ad(mode == PAGE_CUR_L || mode == PAGE_CUR_LE |
| 1249 | || RTREE_SEARCH_MODE(mode)); |
| 1250 | #endif /* PAGE_CUR_LE_OR_EXTENDS */ |
| 1251 | page_mode = mode; |
| 1252 | break; |
| 1253 | } |
| 1254 | |
| 1255 | /* Loop and search until we arrive at the desired level */ |
| 1256 | btr_latch_leaves_t latch_leaves = {{NULL, NULL, NULL}, {0, 0, 0}}; |
| 1257 | |
| 1258 | search_loop: |
| 1259 | buf_mode = BUF_GET; |
| 1260 | rw_latch = RW_NO_LATCH; |
| 1261 | rtree_parent_modified = false; |
| 1262 | |
| 1263 | if (height != 0) { |
| 1264 | /* We are about to fetch the root or a non-leaf page. */ |
| 1265 | if ((latch_mode != BTR_MODIFY_TREE || height == level) |
| 1266 | && !retrying_for_search_prev) { |
| 1267 | /* If doesn't have SX or X latch of index, |
| 1268 | each pages should be latched before reading. */ |
| 1269 | if (height == ULINT_UNDEFINED |
| 1270 | && upper_rw_latch == RW_S_LATCH |
| 1271 | && (modify_external || autoinc)) { |
| 1272 | /* needs sx-latch of root page |
| 1273 | for fseg operation or for writing |
| 1274 | PAGE_ROOT_AUTO_INC */ |
| 1275 | rw_latch = RW_SX_LATCH; |
| 1276 | } else { |
| 1277 | rw_latch = upper_rw_latch; |
| 1278 | } |
| 1279 | } |
| 1280 | } else if (latch_mode <= BTR_MODIFY_LEAF) { |
| 1281 | rw_latch = latch_mode; |
| 1282 | |
| 1283 | if (btr_op != BTR_NO_OP |
| 1284 | && ibuf_should_try(index, btr_op != BTR_INSERT_OP)) { |
| 1285 | |
| 1286 | /* Try to buffer the operation if the leaf |
| 1287 | page is not in the buffer pool. */ |
| 1288 | |
| 1289 | buf_mode = btr_op == BTR_DELETE_OP |
| 1290 | ? BUF_GET_IF_IN_POOL_OR_WATCH |
| 1291 | : BUF_GET_IF_IN_POOL; |
| 1292 | } |
| 1293 | } |
| 1294 | |
| 1295 | retry_page_get: |
| 1296 | ut_ad(n_blocks < BTR_MAX_LEVELS); |
| 1297 | tree_savepoints[n_blocks] = mtr_set_savepoint(mtr); |
| 1298 | block = buf_page_get_gen(page_id, page_size, rw_latch, guess, |
| 1299 | buf_mode, file, line, mtr, &err); |
| 1300 | tree_blocks[n_blocks] = block; |
| 1301 | |
| 1302 | /* Note that block==NULL signifies either an error or change |
| 1303 | buffering. */ |
| 1304 | |
| 1305 | if (err != DB_SUCCESS) { |
| 1306 | ut_ad(block == NULL); |
| 1307 | if (err == DB_DECRYPTION_FAILED) { |
| 1308 | ib_push_warning((void *)NULL, |
| 1309 | DB_DECRYPTION_FAILED, |
| 1310 | "Table %s is encrypted but encryption service or" |
| 1311 | " used key_id is not available. " |
| 1312 | " Can't continue reading table.", |
| 1313 | index->table->name); |
| 1314 | index->table->file_unreadable = true; |
| 1315 | } |
| 1316 | |
| 1317 | goto func_exit; |
| 1318 | } |
| 1319 | |
| 1320 | if (block == NULL) { |
| 1321 | /* This must be a search to perform an insert/delete |
| 1322 | mark/ delete; try using the insert/delete buffer */ |
| 1323 | |
| 1324 | ut_ad(height == 0); |
| 1325 | ut_ad(cursor->thr); |
| 1326 | |
| 1327 | switch (btr_op) { |
| 1328 | case BTR_INSERT_OP: |
| 1329 | case BTR_INSERT_IGNORE_UNIQUE_OP: |
| 1330 | ut_ad(buf_mode == BUF_GET_IF_IN_POOL); |
| 1331 | ut_ad(!dict_index_is_spatial(index)); |
| 1332 | |
| 1333 | if (ibuf_insert(IBUF_OP_INSERT, tuple, index, |
| 1334 | page_id, page_size, cursor->thr)) { |
| 1335 | |
| 1336 | cursor->flag = BTR_CUR_INSERT_TO_IBUF; |
| 1337 | |
| 1338 | goto func_exit; |
| 1339 | } |
| 1340 | break; |
| 1341 | |
| 1342 | case BTR_DELMARK_OP: |
| 1343 | ut_ad(buf_mode == BUF_GET_IF_IN_POOL); |
| 1344 | ut_ad(!dict_index_is_spatial(index)); |
| 1345 | |
| 1346 | if (ibuf_insert(IBUF_OP_DELETE_MARK, tuple, |
| 1347 | index, page_id, page_size, |
| 1348 | cursor->thr)) { |
| 1349 | |
| 1350 | cursor->flag = BTR_CUR_DEL_MARK_IBUF; |
| 1351 | |
| 1352 | goto func_exit; |
| 1353 | } |
| 1354 | |
| 1355 | break; |
| 1356 | |
| 1357 | case BTR_DELETE_OP: |
| 1358 | ut_ad(buf_mode == BUF_GET_IF_IN_POOL_OR_WATCH); |
| 1359 | ut_ad(!dict_index_is_spatial(index)); |
| 1360 | |
| 1361 | if (!row_purge_poss_sec(cursor->purge_node, |
| 1362 | index, tuple)) { |
| 1363 | |
| 1364 | /* The record cannot be purged yet. */ |
| 1365 | cursor->flag = BTR_CUR_DELETE_REF; |
| 1366 | } else if (ibuf_insert(IBUF_OP_DELETE, tuple, |
| 1367 | index, page_id, page_size, |
| 1368 | cursor->thr)) { |
| 1369 | |
| 1370 | /* The purge was buffered. */ |
| 1371 | cursor->flag = BTR_CUR_DELETE_IBUF; |
| 1372 | } else { |
| 1373 | /* The purge could not be buffered. */ |
| 1374 | buf_pool_watch_unset(page_id); |
| 1375 | break; |
| 1376 | } |
| 1377 | |
| 1378 | buf_pool_watch_unset(page_id); |
| 1379 | goto func_exit; |
| 1380 | |
| 1381 | default: |
| 1382 | ut_error; |
| 1383 | } |
| 1384 | |
| 1385 | /* Insert to the insert/delete buffer did not succeed, we |
| 1386 | must read the page from disk. */ |
| 1387 | |
| 1388 | buf_mode = BUF_GET; |
| 1389 | |
| 1390 | goto retry_page_get; |
| 1391 | } |
| 1392 | |
| 1393 | if (retrying_for_search_prev && height != 0) { |
| 1394 | /* also latch left sibling */ |
| 1395 | ulint left_page_no; |
| 1396 | buf_block_t* get_block; |
| 1397 | |
| 1398 | ut_ad(rw_latch == RW_NO_LATCH); |
| 1399 | |
| 1400 | rw_latch = upper_rw_latch; |
| 1401 | |
| 1402 | rw_lock_s_lock(&block->lock); |
| 1403 | left_page_no = btr_page_get_prev( |
| 1404 | buf_block_get_frame(block), mtr); |
| 1405 | rw_lock_s_unlock(&block->lock); |
| 1406 | |
| 1407 | if (left_page_no != FIL_NULL) { |
| 1408 | ut_ad(prev_n_blocks < leftmost_from_level); |
| 1409 | |
| 1410 | prev_tree_savepoints[prev_n_blocks] |
| 1411 | = mtr_set_savepoint(mtr); |
| 1412 | get_block = buf_page_get_gen( |
| 1413 | page_id_t(page_id.space(), left_page_no), |
| 1414 | page_size, rw_latch, NULL, buf_mode, |
| 1415 | file, line, mtr, &err); |
| 1416 | prev_tree_blocks[prev_n_blocks] = get_block; |
| 1417 | prev_n_blocks++; |
| 1418 | |
| 1419 | if (err != DB_SUCCESS) { |
| 1420 | if (err == DB_DECRYPTION_FAILED) { |
| 1421 | ib_push_warning((void *)NULL, |
| 1422 | DB_DECRYPTION_FAILED, |
| 1423 | "Table %s is encrypted but encryption service or" |
| 1424 | " used key_id is not available. " |
| 1425 | " Can't continue reading table.", |
| 1426 | index->table->name); |
| 1427 | index->table->file_unreadable = true; |
| 1428 | } |
| 1429 | |
| 1430 | goto func_exit; |
| 1431 | } |
| 1432 | |
| 1433 | /* BTR_MODIFY_TREE doesn't update prev/next_page_no, |
| 1434 | without their parent page's lock. So, not needed to |
| 1435 | retry here, because we have the parent page's lock. */ |
| 1436 | } |
| 1437 | |
| 1438 | /* release RW_NO_LATCH page and lock with RW_S_LATCH */ |
| 1439 | mtr_release_block_at_savepoint( |
| 1440 | mtr, tree_savepoints[n_blocks], |
| 1441 | tree_blocks[n_blocks]); |
| 1442 | |
| 1443 | tree_savepoints[n_blocks] = mtr_set_savepoint(mtr); |
| 1444 | block = buf_page_get_gen(page_id, page_size, rw_latch, NULL, |
| 1445 | buf_mode, file, line, mtr, &err); |
| 1446 | tree_blocks[n_blocks] = block; |
| 1447 | |
| 1448 | if (err != DB_SUCCESS) { |
| 1449 | if (err == DB_DECRYPTION_FAILED) { |
| 1450 | ib_push_warning((void *)NULL, |
| 1451 | DB_DECRYPTION_FAILED, |
| 1452 | "Table %s is encrypted but encryption service or" |
| 1453 | " used key_id is not available. " |
| 1454 | " Can't continue reading table.", |
| 1455 | index->table->name); |
| 1456 | index->table->file_unreadable = true; |
| 1457 | } |
| 1458 | |
| 1459 | goto func_exit; |
| 1460 | } |
| 1461 | } |
| 1462 | |
| 1463 | page = buf_block_get_frame(block); |
| 1464 | |
| 1465 | if (height == ULINT_UNDEFINED |
| 1466 | && page_is_leaf(page) |
| 1467 | && rw_latch != RW_NO_LATCH |
| 1468 | && rw_latch != root_leaf_rw_latch) { |
| 1469 | /* The root page is also a leaf page (root_leaf). |
| 1470 | We should reacquire the page, because the root page |
| 1471 | is latched differently from leaf pages. */ |
| 1472 | ut_ad(root_leaf_rw_latch != RW_NO_LATCH); |
| 1473 | ut_ad(rw_latch == RW_S_LATCH || rw_latch == RW_SX_LATCH); |
| 1474 | ut_ad(rw_latch == RW_S_LATCH || modify_external || autoinc); |
| 1475 | ut_ad(!autoinc || root_leaf_rw_latch == RW_X_LATCH); |
| 1476 | |
| 1477 | ut_ad(n_blocks == 0); |
| 1478 | mtr_release_block_at_savepoint( |
| 1479 | mtr, tree_savepoints[n_blocks], |
| 1480 | tree_blocks[n_blocks]); |
| 1481 | |
| 1482 | upper_rw_latch = root_leaf_rw_latch; |
| 1483 | goto search_loop; |
| 1484 | } |
| 1485 | |
| 1486 | if (rw_latch != RW_NO_LATCH) { |
| 1487 | #ifdef UNIV_ZIP_DEBUG |
| 1488 | const page_zip_des_t* page_zip |
| 1489 | = buf_block_get_page_zip(block); |
| 1490 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 1491 | #endif /* UNIV_ZIP_DEBUG */ |
| 1492 | |
| 1493 | buf_block_dbg_add_level( |
| 1494 | block, dict_index_is_ibuf(index) |
| 1495 | ? SYNC_IBUF_TREE_NODE : SYNC_TREE_NODE); |
| 1496 | } |
| 1497 | |
| 1498 | ut_ad(fil_page_index_page_check(page)); |
| 1499 | ut_ad(index->id == btr_page_get_index_id(page)); |
| 1500 | |
| 1501 | if (height == ULINT_UNDEFINED) { |
| 1502 | /* We are in the root node */ |
| 1503 | |
| 1504 | height = btr_page_get_level(page); |
| 1505 | root_height = height; |
| 1506 | cursor->tree_height = root_height + 1; |
| 1507 | |
| 1508 | if (dict_index_is_spatial(index)) { |
| 1509 | ut_ad(cursor->rtr_info); |
| 1510 | |
| 1511 | node_seq_t seq_no = rtr_get_current_ssn_id(index); |
| 1512 | |
| 1513 | /* If SSN in memory is not initialized, fetch |
| 1514 | it from root page */ |
| 1515 | if (seq_no < 1) { |
| 1516 | node_seq_t root_seq_no; |
| 1517 | |
| 1518 | root_seq_no = page_get_ssn_id(page); |
| 1519 | |
| 1520 | mutex_enter(&(index->rtr_ssn.mutex)); |
| 1521 | index->rtr_ssn.seq_no = root_seq_no + 1; |
| 1522 | mutex_exit(&(index->rtr_ssn.mutex)); |
| 1523 | } |
| 1524 | |
| 1525 | /* Save the MBR */ |
| 1526 | cursor->rtr_info->thr = cursor->thr; |
| 1527 | rtr_get_mbr_from_tuple(tuple, &cursor->rtr_info->mbr); |
| 1528 | } |
| 1529 | |
| 1530 | #ifdef BTR_CUR_ADAPT |
| 1531 | if (block != guess) { |
| 1532 | info->root_guess = block; |
| 1533 | info->withdraw_clock = buf_withdraw_clock; |
| 1534 | } |
| 1535 | #endif |
| 1536 | } |
| 1537 | |
| 1538 | if (height == 0) { |
| 1539 | if (rw_latch == RW_NO_LATCH) { |
| 1540 | |
| 1541 | latch_leaves = btr_cur_latch_leaves( |
| 1542 | block, page_id, page_size, latch_mode, |
| 1543 | cursor, mtr); |
| 1544 | } |
| 1545 | |
| 1546 | switch (latch_mode) { |
| 1547 | case BTR_MODIFY_TREE: |
| 1548 | case BTR_CONT_MODIFY_TREE: |
| 1549 | case BTR_CONT_SEARCH_TREE: |
| 1550 | break; |
| 1551 | default: |
| 1552 | if (!s_latch_by_caller |
| 1553 | && !srv_read_only_mode |
| 1554 | && !modify_external) { |
| 1555 | /* Release the tree s-latch */ |
| 1556 | /* NOTE: BTR_MODIFY_EXTERNAL |
| 1557 | needs to keep tree sx-latch */ |
| 1558 | mtr_release_s_latch_at_savepoint( |
| 1559 | mtr, savepoint, |
| 1560 | dict_index_get_lock(index)); |
| 1561 | } |
| 1562 | |
| 1563 | /* release upper blocks */ |
| 1564 | if (retrying_for_search_prev) { |
| 1565 | ut_ad(!autoinc); |
| 1566 | for (; |
| 1567 | prev_n_releases < prev_n_blocks; |
| 1568 | prev_n_releases++) { |
| 1569 | mtr_release_block_at_savepoint( |
| 1570 | mtr, |
| 1571 | prev_tree_savepoints[ |
| 1572 | prev_n_releases], |
| 1573 | prev_tree_blocks[ |
| 1574 | prev_n_releases]); |
| 1575 | } |
| 1576 | } |
| 1577 | |
| 1578 | for (; n_releases < n_blocks; n_releases++) { |
| 1579 | if (n_releases == 0 |
| 1580 | && (modify_external || autoinc)) { |
| 1581 | /* keep the root page latch */ |
| 1582 | ut_ad(mtr_memo_contains_flagged( |
| 1583 | mtr, tree_blocks[n_releases], |
| 1584 | MTR_MEMO_PAGE_SX_FIX |
| 1585 | | MTR_MEMO_PAGE_X_FIX)); |
| 1586 | continue; |
| 1587 | } |
| 1588 | |
| 1589 | mtr_release_block_at_savepoint( |
| 1590 | mtr, tree_savepoints[n_releases], |
| 1591 | tree_blocks[n_releases]); |
| 1592 | } |
| 1593 | } |
| 1594 | |
| 1595 | page_mode = mode; |
| 1596 | } |
| 1597 | |
| 1598 | if (dict_index_is_spatial(index)) { |
| 1599 | /* Remember the page search mode */ |
| 1600 | search_mode = page_mode; |
| 1601 | |
| 1602 | /* Some adjustment on search mode, when the |
| 1603 | page search mode is PAGE_CUR_RTREE_LOCATE |
| 1604 | or PAGE_CUR_RTREE_INSERT, as we are searching |
| 1605 | with MBRs. When it is not the target level, we |
| 1606 | should search all sub-trees that "CONTAIN" the |
| 1607 | search range/MBR. When it is at the target |
| 1608 | level, the search becomes PAGE_CUR_LE */ |
| 1609 | if (page_mode == PAGE_CUR_RTREE_LOCATE |
| 1610 | && level == height) { |
| 1611 | if (level == 0) { |
| 1612 | page_mode = PAGE_CUR_LE; |
| 1613 | } else { |
| 1614 | page_mode = PAGE_CUR_RTREE_GET_FATHER; |
| 1615 | } |
| 1616 | } |
| 1617 | |
| 1618 | if (page_mode == PAGE_CUR_RTREE_INSERT) { |
| 1619 | page_mode = (level == height) |
| 1620 | ? PAGE_CUR_LE |
| 1621 | : PAGE_CUR_RTREE_INSERT; |
| 1622 | |
| 1623 | ut_ad(!page_is_leaf(page) || page_mode == PAGE_CUR_LE); |
| 1624 | } |
| 1625 | |
| 1626 | /* "need_path" indicates if we need to tracking the parent |
| 1627 | pages, if it is not spatial comparison, then no need to |
| 1628 | track it */ |
| 1629 | if (page_mode < PAGE_CUR_CONTAIN) { |
| 1630 | need_path = false; |
| 1631 | } |
| 1632 | |
| 1633 | up_match = 0; |
| 1634 | low_match = 0; |
| 1635 | |
| 1636 | if (latch_mode == BTR_MODIFY_TREE |
| 1637 | || latch_mode == BTR_CONT_MODIFY_TREE |
| 1638 | || latch_mode == BTR_CONT_SEARCH_TREE) { |
| 1639 | /* Tree are locked, no need for Page Lock to protect |
| 1640 | the "path" */ |
| 1641 | cursor->rtr_info->need_page_lock = false; |
| 1642 | } |
| 1643 | } |
| 1644 | |
| 1645 | if (dict_index_is_spatial(index) && page_mode >= PAGE_CUR_CONTAIN) { |
| 1646 | ut_ad(need_path); |
| 1647 | found = rtr_cur_search_with_match( |
| 1648 | block, index, tuple, page_mode, page_cursor, |
| 1649 | cursor->rtr_info); |
| 1650 | |
| 1651 | /* Need to use BTR_MODIFY_TREE to do the MBR adjustment */ |
| 1652 | if (search_mode == PAGE_CUR_RTREE_INSERT |
| 1653 | && cursor->rtr_info->mbr_adj) { |
| 1654 | if (latch_mode & BTR_MODIFY_LEAF) { |
| 1655 | /* Parent MBR needs updated, should retry |
| 1656 | with BTR_MODIFY_TREE */ |
| 1657 | goto func_exit; |
| 1658 | } else if (latch_mode & BTR_MODIFY_TREE) { |
| 1659 | rtree_parent_modified = true; |
| 1660 | cursor->rtr_info->mbr_adj = false; |
| 1661 | mbr_adj = true; |
| 1662 | } else { |
| 1663 | ut_ad(0); |
| 1664 | } |
| 1665 | } |
| 1666 | |
| 1667 | if (found && page_mode == PAGE_CUR_RTREE_GET_FATHER) { |
| 1668 | cursor->low_match = |
| 1669 | DICT_INDEX_SPATIAL_NODEPTR_SIZE + 1; |
| 1670 | } |
| 1671 | #ifdef BTR_CUR_HASH_ADAPT |
| 1672 | } else if (height == 0 && btr_search_enabled |
| 1673 | && !(tuple->info_bits & REC_INFO_MIN_REC_FLAG) |
| 1674 | && !dict_index_is_spatial(index)) { |
| 1675 | /* The adaptive hash index is only used when searching |
| 1676 | for leaf pages (height==0), but not in r-trees. |
| 1677 | We only need the byte prefix comparison for the purpose |
| 1678 | of updating the adaptive hash index. */ |
| 1679 | page_cur_search_with_match_bytes( |
| 1680 | block, index, tuple, page_mode, &up_match, &up_bytes, |
| 1681 | &low_match, &low_bytes, page_cursor); |
| 1682 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 1683 | } else { |
| 1684 | /* Search for complete index fields. */ |
| 1685 | up_bytes = low_bytes = 0; |
| 1686 | page_cur_search_with_match( |
| 1687 | block, index, tuple, page_mode, &up_match, |
| 1688 | &low_match, page_cursor, |
| 1689 | need_path ? cursor->rtr_info : NULL); |
| 1690 | } |
| 1691 | |
| 1692 | if (estimate) { |
| 1693 | btr_cur_add_path_info(cursor, height, root_height); |
| 1694 | } |
| 1695 | |
| 1696 | /* If this is the desired level, leave the loop */ |
| 1697 | |
| 1698 | ut_ad(height == btr_page_get_level(page_cur_get_page(page_cursor))); |
| 1699 | |
| 1700 | /* Add Predicate lock if it is serializable isolation |
| 1701 | and only if it is in the search case */ |
| 1702 | if (dict_index_is_spatial(index) |
| 1703 | && cursor->rtr_info->need_prdt_lock |
| 1704 | && mode != PAGE_CUR_RTREE_INSERT |
| 1705 | && mode != PAGE_CUR_RTREE_LOCATE |
| 1706 | && mode >= PAGE_CUR_CONTAIN) { |
| 1707 | trx_t* trx = thr_get_trx(cursor->thr); |
| 1708 | lock_prdt_t prdt; |
| 1709 | |
| 1710 | lock_mutex_enter(); |
| 1711 | lock_init_prdt_from_mbr( |
| 1712 | &prdt, &cursor->rtr_info->mbr, mode, |
| 1713 | trx->lock.lock_heap); |
| 1714 | lock_mutex_exit(); |
| 1715 | |
| 1716 | if (rw_latch == RW_NO_LATCH && height != 0) { |
| 1717 | rw_lock_s_lock(&(block->lock)); |
| 1718 | } |
| 1719 | |
| 1720 | lock_prdt_lock(block, &prdt, index, LOCK_S, |
| 1721 | LOCK_PREDICATE, cursor->thr); |
| 1722 | |
| 1723 | if (rw_latch == RW_NO_LATCH && height != 0) { |
| 1724 | rw_lock_s_unlock(&(block->lock)); |
| 1725 | } |
| 1726 | } |
| 1727 | |
| 1728 | if (level != height) { |
| 1729 | |
| 1730 | const rec_t* node_ptr; |
| 1731 | ut_ad(height > 0); |
| 1732 | |
| 1733 | height--; |
| 1734 | guess = NULL; |
| 1735 | |
| 1736 | node_ptr = page_cur_get_rec(page_cursor); |
| 1737 | |
| 1738 | offsets = rec_get_offsets(node_ptr, index, offsets, false, |
| 1739 | ULINT_UNDEFINED, &heap); |
| 1740 | |
| 1741 | /* If the rec is the first or last in the page for |
| 1742 | pessimistic delete intention, it might cause node_ptr insert |
| 1743 | for the upper level. We should change the intention and retry. |
| 1744 | */ |
| 1745 | if (latch_mode == BTR_MODIFY_TREE |
| 1746 | && btr_cur_need_opposite_intention( |
| 1747 | page, lock_intention, node_ptr)) { |
| 1748 | |
| 1749 | need_opposite_intention: |
| 1750 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 1751 | |
| 1752 | if (n_releases > 0) { |
| 1753 | /* release root block */ |
| 1754 | mtr_release_block_at_savepoint( |
| 1755 | mtr, tree_savepoints[0], |
| 1756 | tree_blocks[0]); |
| 1757 | } |
| 1758 | |
| 1759 | /* release all blocks */ |
| 1760 | for (; n_releases <= n_blocks; n_releases++) { |
| 1761 | mtr_release_block_at_savepoint( |
| 1762 | mtr, tree_savepoints[n_releases], |
| 1763 | tree_blocks[n_releases]); |
| 1764 | } |
| 1765 | |
| 1766 | lock_intention = BTR_INTENTION_BOTH; |
| 1767 | |
| 1768 | page_id.set_page_no(index->page); |
| 1769 | up_match = 0; |
| 1770 | low_match = 0; |
| 1771 | height = ULINT_UNDEFINED; |
| 1772 | |
| 1773 | n_blocks = 0; |
| 1774 | n_releases = 0; |
| 1775 | |
| 1776 | goto search_loop; |
| 1777 | } |
| 1778 | |
| 1779 | if (dict_index_is_spatial(index)) { |
| 1780 | if (page_rec_is_supremum(node_ptr)) { |
| 1781 | cursor->low_match = 0; |
| 1782 | cursor->up_match = 0; |
| 1783 | goto func_exit; |
| 1784 | } |
| 1785 | |
| 1786 | /* If we are doing insertion or record locating, |
| 1787 | remember the tree nodes we visited */ |
| 1788 | if (page_mode == PAGE_CUR_RTREE_INSERT |
| 1789 | || (search_mode == PAGE_CUR_RTREE_LOCATE |
| 1790 | && (latch_mode != BTR_MODIFY_LEAF))) { |
| 1791 | bool add_latch = false; |
| 1792 | |
| 1793 | if (latch_mode == BTR_MODIFY_TREE |
| 1794 | && rw_latch == RW_NO_LATCH) { |
| 1795 | ut_ad(mtr_memo_contains_flagged( |
| 1796 | mtr, dict_index_get_lock(index), |
| 1797 | MTR_MEMO_X_LOCK |
| 1798 | | MTR_MEMO_SX_LOCK)); |
| 1799 | rw_lock_s_lock(&block->lock); |
| 1800 | add_latch = true; |
| 1801 | } |
| 1802 | |
| 1803 | /* Store the parent cursor location */ |
| 1804 | #ifdef UNIV_DEBUG |
| 1805 | ulint num_stored = rtr_store_parent_path( |
| 1806 | block, cursor, latch_mode, |
| 1807 | height + 1, mtr); |
| 1808 | #else |
| 1809 | rtr_store_parent_path( |
| 1810 | block, cursor, latch_mode, |
| 1811 | height + 1, mtr); |
| 1812 | #endif |
| 1813 | |
| 1814 | if (page_mode == PAGE_CUR_RTREE_INSERT) { |
| 1815 | btr_pcur_t* r_cursor = |
| 1816 | rtr_get_parent_cursor( |
| 1817 | cursor, height + 1, |
| 1818 | true); |
| 1819 | /* If it is insertion, there should |
| 1820 | be only one parent for each level |
| 1821 | traverse */ |
| 1822 | #ifdef UNIV_DEBUG |
| 1823 | ut_ad(num_stored == 1); |
| 1824 | #endif |
| 1825 | |
| 1826 | node_ptr = btr_pcur_get_rec(r_cursor); |
| 1827 | |
| 1828 | } |
| 1829 | |
| 1830 | if (add_latch) { |
| 1831 | rw_lock_s_unlock(&block->lock); |
| 1832 | } |
| 1833 | |
| 1834 | ut_ad(!page_rec_is_supremum(node_ptr)); |
| 1835 | } |
| 1836 | |
| 1837 | ut_ad(page_mode == search_mode |
| 1838 | || (page_mode == PAGE_CUR_WITHIN |
| 1839 | && search_mode == PAGE_CUR_RTREE_LOCATE)); |
| 1840 | |
| 1841 | page_mode = search_mode; |
| 1842 | } |
| 1843 | |
| 1844 | /* If the first or the last record of the page |
| 1845 | or the same key value to the first record or last record, |
| 1846 | the another page might be choosen when BTR_CONT_MODIFY_TREE. |
| 1847 | So, the parent page should not released to avoiding deadlock |
| 1848 | with blocking the another search with the same key value. */ |
| 1849 | if (!detected_same_key_root |
| 1850 | && lock_intention == BTR_INTENTION_BOTH |
| 1851 | && !dict_index_is_unique(index) |
| 1852 | && latch_mode == BTR_MODIFY_TREE |
| 1853 | && (up_match >= rec_offs_n_fields(offsets) - 1 |
| 1854 | || low_match >= rec_offs_n_fields(offsets) - 1)) { |
| 1855 | const rec_t* first_rec = page_rec_get_next_const( |
| 1856 | page_get_infimum_rec(page)); |
| 1857 | ulint matched_fields; |
| 1858 | |
| 1859 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 1860 | |
| 1861 | if (node_ptr == first_rec |
| 1862 | || page_rec_is_last(node_ptr, page)) { |
| 1863 | detected_same_key_root = true; |
| 1864 | } else { |
| 1865 | matched_fields = 0; |
| 1866 | |
| 1867 | offsets2 = rec_get_offsets( |
| 1868 | first_rec, index, offsets2, |
| 1869 | false, ULINT_UNDEFINED, &heap); |
| 1870 | cmp_rec_rec_with_match(node_ptr, first_rec, |
| 1871 | offsets, offsets2, index, FALSE, |
| 1872 | &matched_fields); |
| 1873 | |
| 1874 | if (matched_fields |
| 1875 | >= rec_offs_n_fields(offsets) - 1) { |
| 1876 | detected_same_key_root = true; |
| 1877 | } else { |
| 1878 | const rec_t* last_rec; |
| 1879 | |
| 1880 | last_rec = page_rec_get_prev_const( |
| 1881 | page_get_supremum_rec(page)); |
| 1882 | |
| 1883 | matched_fields = 0; |
| 1884 | |
| 1885 | offsets2 = rec_get_offsets( |
| 1886 | last_rec, index, offsets2, |
| 1887 | false, ULINT_UNDEFINED, &heap); |
| 1888 | cmp_rec_rec_with_match( |
| 1889 | node_ptr, last_rec, |
| 1890 | offsets, offsets2, index, |
| 1891 | FALSE, &matched_fields); |
| 1892 | if (matched_fields |
| 1893 | >= rec_offs_n_fields(offsets) - 1) { |
| 1894 | detected_same_key_root = true; |
| 1895 | } |
| 1896 | } |
| 1897 | } |
| 1898 | } |
| 1899 | |
| 1900 | /* If the page might cause modify_tree, |
| 1901 | we should not release the parent page's lock. */ |
| 1902 | if (!detected_same_key_root |
| 1903 | && latch_mode == BTR_MODIFY_TREE |
| 1904 | && !btr_cur_will_modify_tree( |
| 1905 | index, page, lock_intention, node_ptr, |
| 1906 | node_ptr_max_size, page_size, mtr) |
| 1907 | && !rtree_parent_modified) { |
| 1908 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 1909 | ut_ad(n_releases <= n_blocks); |
| 1910 | |
| 1911 | /* we can release upper blocks */ |
| 1912 | for (; n_releases < n_blocks; n_releases++) { |
| 1913 | if (n_releases == 0) { |
| 1914 | /* we should not release root page |
| 1915 | to pin to same block. */ |
| 1916 | continue; |
| 1917 | } |
| 1918 | |
| 1919 | /* release unused blocks to unpin */ |
| 1920 | mtr_release_block_at_savepoint( |
| 1921 | mtr, tree_savepoints[n_releases], |
| 1922 | tree_blocks[n_releases]); |
| 1923 | } |
| 1924 | } |
| 1925 | |
| 1926 | if (height == level |
| 1927 | && latch_mode == BTR_MODIFY_TREE) { |
| 1928 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 1929 | /* we should sx-latch root page, if released already. |
| 1930 | It contains seg_header. */ |
| 1931 | if (n_releases > 0) { |
| 1932 | mtr_block_sx_latch_at_savepoint( |
| 1933 | mtr, tree_savepoints[0], |
| 1934 | tree_blocks[0]); |
| 1935 | } |
| 1936 | |
| 1937 | /* x-latch the branch blocks not released yet. */ |
| 1938 | for (ulint i = n_releases; i <= n_blocks; i++) { |
| 1939 | mtr_block_x_latch_at_savepoint( |
| 1940 | mtr, tree_savepoints[i], |
| 1941 | tree_blocks[i]); |
| 1942 | } |
| 1943 | } |
| 1944 | |
| 1945 | /* We should consider prev_page of parent page, if the node_ptr |
| 1946 | is the leftmost of the page. because BTR_SEARCH_PREV and |
| 1947 | BTR_MODIFY_PREV latches prev_page of the leaf page. */ |
| 1948 | if ((latch_mode == BTR_SEARCH_PREV |
| 1949 | || latch_mode == BTR_MODIFY_PREV) |
| 1950 | && !retrying_for_search_prev) { |
| 1951 | /* block should be latched for consistent |
| 1952 | btr_page_get_prev() */ |
| 1953 | ut_ad(mtr_memo_contains_flagged(mtr, block, |
| 1954 | MTR_MEMO_PAGE_S_FIX |
| 1955 | | MTR_MEMO_PAGE_X_FIX)); |
| 1956 | |
| 1957 | if (page_has_prev(page) |
| 1958 | && page_rec_is_first(node_ptr, page)) { |
| 1959 | |
| 1960 | if (leftmost_from_level == 0) { |
| 1961 | leftmost_from_level = height + 1; |
| 1962 | } |
| 1963 | } else { |
| 1964 | leftmost_from_level = 0; |
| 1965 | } |
| 1966 | |
| 1967 | if (height == 0 && leftmost_from_level > 0) { |
| 1968 | /* should retry to get also prev_page |
| 1969 | from level==leftmost_from_level. */ |
| 1970 | retrying_for_search_prev = true; |
| 1971 | |
| 1972 | prev_tree_blocks = static_cast<buf_block_t**>( |
| 1973 | ut_malloc_nokey(sizeof(buf_block_t*) |
| 1974 | * leftmost_from_level)); |
| 1975 | |
| 1976 | prev_tree_savepoints = static_cast<ulint*>( |
| 1977 | ut_malloc_nokey(sizeof(ulint) |
| 1978 | * leftmost_from_level)); |
| 1979 | |
| 1980 | /* back to the level (leftmost_from_level+1) */ |
| 1981 | ulint idx = n_blocks |
| 1982 | - (leftmost_from_level - 1); |
| 1983 | |
| 1984 | page_id.set_page_no( |
| 1985 | tree_blocks[idx]->page.id.page_no()); |
| 1986 | |
| 1987 | for (ulint i = n_blocks |
| 1988 | - (leftmost_from_level - 1); |
| 1989 | i <= n_blocks; i++) { |
| 1990 | mtr_release_block_at_savepoint( |
| 1991 | mtr, tree_savepoints[i], |
| 1992 | tree_blocks[i]); |
| 1993 | } |
| 1994 | |
| 1995 | n_blocks -= (leftmost_from_level - 1); |
| 1996 | height = leftmost_from_level; |
| 1997 | ut_ad(n_releases == 0); |
| 1998 | |
| 1999 | /* replay up_match, low_match */ |
| 2000 | up_match = 0; |
| 2001 | low_match = 0; |
| 2002 | rtr_info_t* rtr_info = need_path |
| 2003 | ? cursor->rtr_info : NULL; |
| 2004 | |
| 2005 | for (ulint i = 0; i < n_blocks; i++) { |
| 2006 | page_cur_search_with_match( |
| 2007 | tree_blocks[i], index, tuple, |
| 2008 | page_mode, &up_match, |
| 2009 | &low_match, page_cursor, |
| 2010 | rtr_info); |
| 2011 | } |
| 2012 | |
| 2013 | goto search_loop; |
| 2014 | } |
| 2015 | } |
| 2016 | |
| 2017 | /* Go to the child node */ |
| 2018 | page_id.set_page_no( |
| 2019 | btr_node_ptr_get_child_page_no(node_ptr, offsets)); |
| 2020 | |
| 2021 | n_blocks++; |
| 2022 | |
| 2023 | if (UNIV_UNLIKELY(height == 0 && dict_index_is_ibuf(index))) { |
| 2024 | /* We're doing a search on an ibuf tree and we're one |
| 2025 | level above the leaf page. */ |
| 2026 | |
| 2027 | ut_ad(level == 0); |
| 2028 | |
| 2029 | buf_mode = BUF_GET; |
| 2030 | rw_latch = RW_NO_LATCH; |
| 2031 | goto retry_page_get; |
| 2032 | } |
| 2033 | |
| 2034 | if (dict_index_is_spatial(index) |
| 2035 | && page_mode >= PAGE_CUR_CONTAIN |
| 2036 | && page_mode != PAGE_CUR_RTREE_INSERT) { |
| 2037 | ut_ad(need_path); |
| 2038 | rtr_node_path_t* path = |
| 2039 | cursor->rtr_info->path; |
| 2040 | |
| 2041 | if (!path->empty() && found) { |
| 2042 | ut_ad(path->back().page_no |
| 2043 | == page_id.page_no()); |
| 2044 | path->pop_back(); |
| 2045 | #ifdef UNIV_DEBUG |
| 2046 | if (page_mode == PAGE_CUR_RTREE_LOCATE |
| 2047 | && (latch_mode != BTR_MODIFY_LEAF)) { |
| 2048 | btr_pcur_t* cur |
| 2049 | = cursor->rtr_info->parent_path->back( |
| 2050 | ).cursor; |
| 2051 | rec_t* my_node_ptr |
| 2052 | = btr_pcur_get_rec(cur); |
| 2053 | |
| 2054 | offsets = rec_get_offsets( |
| 2055 | my_node_ptr, index, offsets, |
| 2056 | false, ULINT_UNDEFINED, &heap); |
| 2057 | |
| 2058 | ulint my_page_no |
| 2059 | = btr_node_ptr_get_child_page_no( |
| 2060 | my_node_ptr, offsets); |
| 2061 | |
| 2062 | ut_ad(page_id.page_no() == my_page_no); |
| 2063 | } |
| 2064 | #endif |
| 2065 | } |
| 2066 | } |
| 2067 | |
| 2068 | goto search_loop; |
| 2069 | } else if (!dict_index_is_spatial(index) |
| 2070 | && latch_mode == BTR_MODIFY_TREE |
| 2071 | && lock_intention == BTR_INTENTION_INSERT |
| 2072 | && page_has_next(page) |
| 2073 | && page_rec_is_last(page_cur_get_rec(page_cursor), page)) { |
| 2074 | |
| 2075 | /* btr_insert_into_right_sibling() might cause |
| 2076 | deleting node_ptr at upper level */ |
| 2077 | |
| 2078 | guess = NULL; |
| 2079 | |
| 2080 | if (height == 0) { |
| 2081 | /* release the leaf pages if latched */ |
| 2082 | for (uint i = 0; i < 3; i++) { |
| 2083 | if (latch_leaves.blocks[i] != NULL) { |
| 2084 | mtr_release_block_at_savepoint( |
| 2085 | mtr, latch_leaves.savepoints[i], |
| 2086 | latch_leaves.blocks[i]); |
| 2087 | latch_leaves.blocks[i] = NULL; |
| 2088 | } |
| 2089 | } |
| 2090 | } |
| 2091 | |
| 2092 | goto need_opposite_intention; |
| 2093 | } |
| 2094 | |
| 2095 | if (level != 0) { |
| 2096 | ut_ad(!autoinc); |
| 2097 | |
| 2098 | if (upper_rw_latch == RW_NO_LATCH) { |
| 2099 | /* latch the page */ |
| 2100 | buf_block_t* child_block; |
| 2101 | |
| 2102 | if (latch_mode == BTR_CONT_MODIFY_TREE) { |
| 2103 | child_block = btr_block_get( |
| 2104 | page_id, page_size, RW_X_LATCH, |
| 2105 | index, mtr); |
| 2106 | } else { |
| 2107 | ut_ad(latch_mode == BTR_CONT_SEARCH_TREE); |
| 2108 | child_block = btr_block_get( |
| 2109 | page_id, page_size, RW_SX_LATCH, |
| 2110 | index, mtr); |
| 2111 | } |
| 2112 | |
| 2113 | btr_assert_not_corrupted(child_block, index); |
| 2114 | } else { |
| 2115 | ut_ad(mtr_memo_contains(mtr, block, upper_rw_latch)); |
| 2116 | btr_assert_not_corrupted(block, index); |
| 2117 | |
| 2118 | if (s_latch_by_caller) { |
| 2119 | ut_ad(latch_mode == BTR_SEARCH_TREE); |
| 2120 | /* to exclude modifying tree operations |
| 2121 | should sx-latch the index. */ |
| 2122 | ut_ad(mtr_memo_contains( |
| 2123 | mtr, dict_index_get_lock(index), |
| 2124 | MTR_MEMO_SX_LOCK)); |
| 2125 | /* because has sx-latch of index, |
| 2126 | can release upper blocks. */ |
| 2127 | for (; n_releases < n_blocks; n_releases++) { |
| 2128 | mtr_release_block_at_savepoint( |
| 2129 | mtr, |
| 2130 | tree_savepoints[n_releases], |
| 2131 | tree_blocks[n_releases]); |
| 2132 | } |
| 2133 | } |
| 2134 | } |
| 2135 | |
| 2136 | if (page_mode <= PAGE_CUR_LE) { |
| 2137 | cursor->low_match = low_match; |
| 2138 | cursor->up_match = up_match; |
| 2139 | } |
| 2140 | } else { |
| 2141 | cursor->low_match = low_match; |
| 2142 | cursor->low_bytes = low_bytes; |
| 2143 | cursor->up_match = up_match; |
| 2144 | cursor->up_bytes = up_bytes; |
| 2145 | |
| 2146 | if (autoinc) { |
| 2147 | page_set_autoinc(tree_blocks[0], |
| 2148 | index, autoinc, mtr, false); |
| 2149 | } |
| 2150 | |
| 2151 | #ifdef BTR_CUR_HASH_ADAPT |
| 2152 | /* We do a dirty read of btr_search_enabled here. We |
| 2153 | will properly check btr_search_enabled again in |
| 2154 | btr_search_build_page_hash_index() before building a |
| 2155 | page hash index, while holding search latch. */ |
| 2156 | if (!btr_search_enabled) { |
| 2157 | # ifdef MYSQL_INDEX_DISABLE_AHI |
| 2158 | } else if (index->disable_ahi) { |
| 2159 | # endif |
| 2160 | } else if (tuple->info_bits & REC_INFO_MIN_REC_FLAG) { |
| 2161 | ut_ad(index->is_instant()); |
| 2162 | /* This may be a search tuple for |
| 2163 | btr_pcur_restore_position(). */ |
| 2164 | ut_ad(tuple->info_bits == REC_INFO_DEFAULT_ROW |
| 2165 | || tuple->info_bits == REC_INFO_MIN_REC_FLAG); |
| 2166 | } else if (rec_is_default_row(btr_cur_get_rec(cursor), |
| 2167 | index)) { |
| 2168 | /* Only user records belong in the adaptive |
| 2169 | hash index. */ |
| 2170 | } else { |
| 2171 | btr_search_info_update(index, cursor); |
| 2172 | } |
| 2173 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 2174 | ut_ad(cursor->up_match != ULINT_UNDEFINED |
| 2175 | || mode != PAGE_CUR_GE); |
| 2176 | ut_ad(cursor->up_match != ULINT_UNDEFINED |
| 2177 | || mode != PAGE_CUR_LE); |
| 2178 | ut_ad(cursor->low_match != ULINT_UNDEFINED |
| 2179 | || mode != PAGE_CUR_LE); |
| 2180 | } |
| 2181 | |
| 2182 | /* For spatial index, remember what blocks are still latched */ |
| 2183 | if (dict_index_is_spatial(index) |
| 2184 | && (latch_mode == BTR_MODIFY_TREE |
| 2185 | || latch_mode == BTR_MODIFY_LEAF)) { |
| 2186 | for (ulint i = 0; i < n_releases; i++) { |
| 2187 | cursor->rtr_info->tree_blocks[i] = NULL; |
| 2188 | cursor->rtr_info->tree_savepoints[i] = 0; |
| 2189 | } |
| 2190 | |
| 2191 | for (ulint i = n_releases; i <= n_blocks; i++) { |
| 2192 | cursor->rtr_info->tree_blocks[i] = tree_blocks[i]; |
| 2193 | cursor->rtr_info->tree_savepoints[i] = tree_savepoints[i]; |
| 2194 | } |
| 2195 | } |
| 2196 | |
| 2197 | func_exit: |
| 2198 | |
| 2199 | if (UNIV_LIKELY_NULL(heap)) { |
| 2200 | mem_heap_free(heap); |
| 2201 | } |
| 2202 | |
| 2203 | if (retrying_for_search_prev) { |
| 2204 | ut_free(prev_tree_blocks); |
| 2205 | ut_free(prev_tree_savepoints); |
| 2206 | } |
| 2207 | |
| 2208 | if (mbr_adj) { |
| 2209 | /* remember that we will need to adjust parent MBR */ |
| 2210 | cursor->rtr_info->mbr_adj = true; |
| 2211 | } |
| 2212 | |
| 2213 | #ifdef BTR_CUR_HASH_ADAPT |
| 2214 | if (ahi_latch) { |
| 2215 | rw_lock_s_lock(ahi_latch); |
| 2216 | } |
| 2217 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 2218 | |
| 2219 | DBUG_RETURN(err); |
| 2220 | } |
| 2221 | |
| 2222 | /*****************************************************************//** |
| 2223 | Opens a cursor at either end of an index. */ |
| 2224 | dberr_t |
| 2225 | btr_cur_open_at_index_side_func( |
| 2226 | /*============================*/ |
| 2227 | bool from_left, /*!< in: true if open to the low end, |
| 2228 | false if to the high end */ |
| 2229 | dict_index_t* index, /*!< in: index */ |
| 2230 | ulint latch_mode, /*!< in: latch mode */ |
| 2231 | btr_cur_t* cursor, /*!< in/out: cursor */ |
| 2232 | ulint level, /*!< in: level to search for |
| 2233 | (0=leaf). */ |
| 2234 | const char* file, /*!< in: file name */ |
| 2235 | unsigned line, /*!< in: line where called */ |
| 2236 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 2237 | { |
| 2238 | page_cur_t* page_cursor; |
| 2239 | ulint node_ptr_max_size = srv_page_size / 2; |
| 2240 | ulint height; |
| 2241 | ulint root_height = 0; /* remove warning */ |
| 2242 | rec_t* node_ptr; |
| 2243 | ulint estimate; |
| 2244 | ulint savepoint; |
| 2245 | ulint upper_rw_latch, root_leaf_rw_latch; |
| 2246 | btr_intention_t lock_intention; |
| 2247 | buf_block_t* tree_blocks[BTR_MAX_LEVELS]; |
| 2248 | ulint tree_savepoints[BTR_MAX_LEVELS]; |
| 2249 | ulint n_blocks = 0; |
| 2250 | ulint n_releases = 0; |
| 2251 | mem_heap_t* heap = NULL; |
| 2252 | ulint offsets_[REC_OFFS_NORMAL_SIZE]; |
| 2253 | ulint* offsets = offsets_; |
| 2254 | dberr_t err = DB_SUCCESS; |
| 2255 | |
| 2256 | rec_offs_init(offsets_); |
| 2257 | |
| 2258 | estimate = latch_mode & BTR_ESTIMATE; |
| 2259 | latch_mode &= ulint(~BTR_ESTIMATE); |
| 2260 | |
| 2261 | ut_ad(level != ULINT_UNDEFINED); |
| 2262 | |
| 2263 | bool s_latch_by_caller; |
| 2264 | |
| 2265 | s_latch_by_caller = latch_mode & BTR_ALREADY_S_LATCHED; |
| 2266 | latch_mode &= ulint(~BTR_ALREADY_S_LATCHED); |
| 2267 | |
| 2268 | lock_intention = btr_cur_get_and_clear_intention(&latch_mode); |
| 2269 | |
| 2270 | ut_ad(!(latch_mode & BTR_MODIFY_EXTERNAL)); |
| 2271 | |
| 2272 | /* This function doesn't need to lock left page of the leaf page */ |
| 2273 | if (latch_mode == BTR_SEARCH_PREV) { |
| 2274 | latch_mode = BTR_SEARCH_LEAF; |
| 2275 | } else if (latch_mode == BTR_MODIFY_PREV) { |
| 2276 | latch_mode = BTR_MODIFY_LEAF; |
| 2277 | } |
| 2278 | |
| 2279 | /* Store the position of the tree latch we push to mtr so that we |
| 2280 | know how to release it when we have latched the leaf node */ |
| 2281 | |
| 2282 | savepoint = mtr_set_savepoint(mtr); |
| 2283 | |
| 2284 | switch (latch_mode) { |
| 2285 | case BTR_CONT_MODIFY_TREE: |
| 2286 | case BTR_CONT_SEARCH_TREE: |
| 2287 | upper_rw_latch = RW_NO_LATCH; |
| 2288 | break; |
| 2289 | case BTR_MODIFY_TREE: |
| 2290 | /* Most of delete-intended operations are purging. |
| 2291 | Free blocks and read IO bandwidth should be prior |
| 2292 | for them, when the history list is glowing huge. */ |
| 2293 | if (lock_intention == BTR_INTENTION_DELETE |
| 2294 | && trx_sys.history_size() > BTR_CUR_FINE_HISTORY_LENGTH |
| 2295 | && buf_get_n_pending_read_ios()) { |
| 2296 | mtr_x_lock(dict_index_get_lock(index), mtr); |
| 2297 | } else { |
| 2298 | mtr_sx_lock(dict_index_get_lock(index), mtr); |
| 2299 | } |
| 2300 | upper_rw_latch = RW_X_LATCH; |
| 2301 | break; |
| 2302 | default: |
| 2303 | ut_ad(!s_latch_by_caller |
| 2304 | || mtr_memo_contains_flagged(mtr, |
| 2305 | dict_index_get_lock(index), |
| 2306 | MTR_MEMO_SX_LOCK |
| 2307 | | MTR_MEMO_S_LOCK)); |
| 2308 | if (!srv_read_only_mode) { |
| 2309 | if (!s_latch_by_caller) { |
| 2310 | /* BTR_SEARCH_TREE is intended to be used with |
| 2311 | BTR_ALREADY_S_LATCHED */ |
| 2312 | ut_ad(latch_mode != BTR_SEARCH_TREE); |
| 2313 | |
| 2314 | mtr_s_lock(dict_index_get_lock(index), mtr); |
| 2315 | } |
| 2316 | upper_rw_latch = RW_S_LATCH; |
| 2317 | } else { |
| 2318 | upper_rw_latch = RW_NO_LATCH; |
| 2319 | } |
| 2320 | } |
| 2321 | root_leaf_rw_latch = btr_cur_latch_for_root_leaf(latch_mode); |
| 2322 | |
| 2323 | page_cursor = btr_cur_get_page_cur(cursor); |
| 2324 | cursor->index = index; |
| 2325 | |
| 2326 | page_id_t page_id(index->table->space->id, index->page); |
| 2327 | const page_size_t page_size(index->table->space->flags); |
| 2328 | |
| 2329 | if (root_leaf_rw_latch == RW_X_LATCH) { |
| 2330 | node_ptr_max_size = dict_index_node_ptr_max_size(index); |
| 2331 | } |
| 2332 | |
| 2333 | height = ULINT_UNDEFINED; |
| 2334 | |
| 2335 | for (;;) { |
| 2336 | buf_block_t* block; |
| 2337 | ulint rw_latch; |
| 2338 | |
| 2339 | ut_ad(n_blocks < BTR_MAX_LEVELS); |
| 2340 | |
| 2341 | if (height != 0 |
| 2342 | && (latch_mode != BTR_MODIFY_TREE |
| 2343 | || height == level)) { |
| 2344 | rw_latch = upper_rw_latch; |
| 2345 | } else { |
| 2346 | rw_latch = RW_NO_LATCH; |
| 2347 | } |
| 2348 | |
| 2349 | tree_savepoints[n_blocks] = mtr_set_savepoint(mtr); |
| 2350 | block = buf_page_get_gen(page_id, page_size, rw_latch, NULL, |
| 2351 | BUF_GET, file, line, mtr, &err); |
| 2352 | ut_ad((block != NULL) == (err == DB_SUCCESS)); |
| 2353 | tree_blocks[n_blocks] = block; |
| 2354 | |
| 2355 | if (err != DB_SUCCESS) { |
| 2356 | if (err == DB_DECRYPTION_FAILED) { |
| 2357 | ib_push_warning((void *)NULL, |
| 2358 | DB_DECRYPTION_FAILED, |
| 2359 | "Table %s is encrypted but encryption service or" |
| 2360 | " used key_id is not available. " |
| 2361 | " Can't continue reading table.", |
| 2362 | index->table->name); |
| 2363 | index->table->file_unreadable = true; |
| 2364 | } |
| 2365 | |
| 2366 | goto exit_loop; |
| 2367 | } |
| 2368 | |
| 2369 | const page_t* page = buf_block_get_frame(block); |
| 2370 | |
| 2371 | if (height == ULINT_UNDEFINED |
| 2372 | && page_is_leaf(page) |
| 2373 | && rw_latch != RW_NO_LATCH |
| 2374 | && rw_latch != root_leaf_rw_latch) { |
| 2375 | /* We should retry to get the page, because the root page |
| 2376 | is latched with different level as a leaf page. */ |
| 2377 | ut_ad(root_leaf_rw_latch != RW_NO_LATCH); |
| 2378 | ut_ad(rw_latch == RW_S_LATCH); |
| 2379 | |
| 2380 | ut_ad(n_blocks == 0); |
| 2381 | mtr_release_block_at_savepoint( |
| 2382 | mtr, tree_savepoints[n_blocks], |
| 2383 | tree_blocks[n_blocks]); |
| 2384 | |
| 2385 | upper_rw_latch = root_leaf_rw_latch; |
| 2386 | continue; |
| 2387 | } |
| 2388 | |
| 2389 | ut_ad(fil_page_index_page_check(page)); |
| 2390 | ut_ad(index->id == btr_page_get_index_id(page)); |
| 2391 | |
| 2392 | if (height == ULINT_UNDEFINED) { |
| 2393 | /* We are in the root node */ |
| 2394 | |
| 2395 | height = btr_page_get_level(page); |
| 2396 | root_height = height; |
| 2397 | ut_a(height >= level); |
| 2398 | } else { |
| 2399 | /* TODO: flag the index corrupted if this fails */ |
| 2400 | ut_ad(height == btr_page_get_level(page)); |
| 2401 | } |
| 2402 | |
| 2403 | if (height == level) { |
| 2404 | if (srv_read_only_mode) { |
| 2405 | btr_cur_latch_leaves( |
| 2406 | block, page_id, page_size, |
| 2407 | latch_mode, cursor, mtr); |
| 2408 | } else if (height == 0) { |
| 2409 | if (rw_latch == RW_NO_LATCH) { |
| 2410 | btr_cur_latch_leaves( |
| 2411 | block, page_id, page_size, |
| 2412 | latch_mode, cursor, mtr); |
| 2413 | } |
| 2414 | /* In versions <= 3.23.52 we had |
| 2415 | forgotten to release the tree latch |
| 2416 | here. If in an index scan we had to |
| 2417 | scan far to find a record visible to |
| 2418 | the current transaction, that could |
| 2419 | starve others waiting for the tree |
| 2420 | latch. */ |
| 2421 | |
| 2422 | switch (latch_mode) { |
| 2423 | case BTR_MODIFY_TREE: |
| 2424 | case BTR_CONT_MODIFY_TREE: |
| 2425 | case BTR_CONT_SEARCH_TREE: |
| 2426 | break; |
| 2427 | default: |
| 2428 | if (!s_latch_by_caller) { |
| 2429 | /* Release the tree s-latch */ |
| 2430 | mtr_release_s_latch_at_savepoint( |
| 2431 | mtr, savepoint, |
| 2432 | dict_index_get_lock( |
| 2433 | index)); |
| 2434 | } |
| 2435 | |
| 2436 | /* release upper blocks */ |
| 2437 | for (; n_releases < n_blocks; |
| 2438 | n_releases++) { |
| 2439 | mtr_release_block_at_savepoint( |
| 2440 | mtr, |
| 2441 | tree_savepoints[ |
| 2442 | n_releases], |
| 2443 | tree_blocks[ |
| 2444 | n_releases]); |
| 2445 | } |
| 2446 | } |
| 2447 | } else { /* height != 0 */ |
| 2448 | /* We already have the block latched. */ |
| 2449 | ut_ad(latch_mode == BTR_SEARCH_TREE); |
| 2450 | ut_ad(s_latch_by_caller); |
| 2451 | ut_ad(upper_rw_latch == RW_S_LATCH); |
| 2452 | |
| 2453 | ut_ad(mtr_memo_contains(mtr, block, |
| 2454 | upper_rw_latch)); |
| 2455 | |
| 2456 | if (s_latch_by_caller) { |
| 2457 | /* to exclude modifying tree operations |
| 2458 | should sx-latch the index. */ |
| 2459 | ut_ad(mtr_memo_contains( |
| 2460 | mtr, |
| 2461 | dict_index_get_lock(index), |
| 2462 | MTR_MEMO_SX_LOCK)); |
| 2463 | /* because has sx-latch of index, |
| 2464 | can release upper blocks. */ |
| 2465 | for (; n_releases < n_blocks; |
| 2466 | n_releases++) { |
| 2467 | mtr_release_block_at_savepoint( |
| 2468 | mtr, |
| 2469 | tree_savepoints[ |
| 2470 | n_releases], |
| 2471 | tree_blocks[ |
| 2472 | n_releases]); |
| 2473 | } |
| 2474 | } |
| 2475 | } |
| 2476 | } |
| 2477 | |
| 2478 | if (from_left) { |
| 2479 | page_cur_set_before_first(block, page_cursor); |
| 2480 | } else { |
| 2481 | page_cur_set_after_last(block, page_cursor); |
| 2482 | } |
| 2483 | |
| 2484 | if (height == level) { |
| 2485 | if (estimate) { |
| 2486 | btr_cur_add_path_info(cursor, height, |
| 2487 | root_height); |
| 2488 | } |
| 2489 | |
| 2490 | break; |
| 2491 | } |
| 2492 | |
| 2493 | ut_ad(height > 0); |
| 2494 | |
| 2495 | if (from_left) { |
| 2496 | page_cur_move_to_next(page_cursor); |
| 2497 | } else { |
| 2498 | page_cur_move_to_prev(page_cursor); |
| 2499 | } |
| 2500 | |
| 2501 | if (estimate) { |
| 2502 | btr_cur_add_path_info(cursor, height, root_height); |
| 2503 | } |
| 2504 | |
| 2505 | height--; |
| 2506 | |
| 2507 | node_ptr = page_cur_get_rec(page_cursor); |
| 2508 | offsets = rec_get_offsets(node_ptr, cursor->index, offsets, |
| 2509 | false, ULINT_UNDEFINED, &heap); |
| 2510 | |
| 2511 | /* If the rec is the first or last in the page for |
| 2512 | pessimistic delete intention, it might cause node_ptr insert |
| 2513 | for the upper level. We should change the intention and retry. |
| 2514 | */ |
| 2515 | if (latch_mode == BTR_MODIFY_TREE |
| 2516 | && btr_cur_need_opposite_intention( |
| 2517 | page, lock_intention, node_ptr)) { |
| 2518 | |
| 2519 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 2520 | /* release all blocks */ |
| 2521 | for (; n_releases <= n_blocks; n_releases++) { |
| 2522 | mtr_release_block_at_savepoint( |
| 2523 | mtr, tree_savepoints[n_releases], |
| 2524 | tree_blocks[n_releases]); |
| 2525 | } |
| 2526 | |
| 2527 | lock_intention = BTR_INTENTION_BOTH; |
| 2528 | |
| 2529 | page_id.set_page_no(dict_index_get_page(index)); |
| 2530 | |
| 2531 | height = ULINT_UNDEFINED; |
| 2532 | |
| 2533 | n_blocks = 0; |
| 2534 | n_releases = 0; |
| 2535 | |
| 2536 | continue; |
| 2537 | } |
| 2538 | |
| 2539 | if (latch_mode == BTR_MODIFY_TREE |
| 2540 | && !btr_cur_will_modify_tree( |
| 2541 | cursor->index, page, lock_intention, node_ptr, |
| 2542 | node_ptr_max_size, page_size, mtr)) { |
| 2543 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 2544 | ut_ad(n_releases <= n_blocks); |
| 2545 | |
| 2546 | /* we can release upper blocks */ |
| 2547 | for (; n_releases < n_blocks; n_releases++) { |
| 2548 | if (n_releases == 0) { |
| 2549 | /* we should not release root page |
| 2550 | to pin to same block. */ |
| 2551 | continue; |
| 2552 | } |
| 2553 | |
| 2554 | /* release unused blocks to unpin */ |
| 2555 | mtr_release_block_at_savepoint( |
| 2556 | mtr, tree_savepoints[n_releases], |
| 2557 | tree_blocks[n_releases]); |
| 2558 | } |
| 2559 | } |
| 2560 | |
| 2561 | if (height == level |
| 2562 | && latch_mode == BTR_MODIFY_TREE) { |
| 2563 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 2564 | /* we should sx-latch root page, if released already. |
| 2565 | It contains seg_header. */ |
| 2566 | if (n_releases > 0) { |
| 2567 | mtr_block_sx_latch_at_savepoint( |
| 2568 | mtr, tree_savepoints[0], |
| 2569 | tree_blocks[0]); |
| 2570 | } |
| 2571 | |
| 2572 | /* x-latch the branch blocks not released yet. */ |
| 2573 | for (ulint i = n_releases; i <= n_blocks; i++) { |
| 2574 | mtr_block_x_latch_at_savepoint( |
| 2575 | mtr, tree_savepoints[i], |
| 2576 | tree_blocks[i]); |
| 2577 | } |
| 2578 | } |
| 2579 | |
| 2580 | /* Go to the child node */ |
| 2581 | page_id.set_page_no( |
| 2582 | btr_node_ptr_get_child_page_no(node_ptr, offsets)); |
| 2583 | |
| 2584 | n_blocks++; |
| 2585 | } |
| 2586 | |
| 2587 | exit_loop: |
| 2588 | if (heap) { |
| 2589 | mem_heap_free(heap); |
| 2590 | } |
| 2591 | |
| 2592 | return err; |
| 2593 | } |
| 2594 | |
| 2595 | /**********************************************************************//** |
| 2596 | Positions a cursor at a randomly chosen position within a B-tree. |
| 2597 | @return true if the index is available and we have put the cursor, false |
| 2598 | if the index is unavailable */ |
| 2599 | bool |
| 2600 | btr_cur_open_at_rnd_pos_func( |
| 2601 | /*=========================*/ |
| 2602 | dict_index_t* index, /*!< in: index */ |
| 2603 | ulint latch_mode, /*!< in: BTR_SEARCH_LEAF, ... */ |
| 2604 | btr_cur_t* cursor, /*!< in/out: B-tree cursor */ |
| 2605 | const char* file, /*!< in: file name */ |
| 2606 | unsigned line, /*!< in: line where called */ |
| 2607 | mtr_t* mtr) /*!< in: mtr */ |
| 2608 | { |
| 2609 | page_cur_t* page_cursor; |
| 2610 | ulint node_ptr_max_size = srv_page_size / 2; |
| 2611 | ulint height; |
| 2612 | rec_t* node_ptr; |
| 2613 | ulint savepoint; |
| 2614 | ulint upper_rw_latch, root_leaf_rw_latch; |
| 2615 | btr_intention_t lock_intention; |
| 2616 | buf_block_t* tree_blocks[BTR_MAX_LEVELS]; |
| 2617 | ulint tree_savepoints[BTR_MAX_LEVELS]; |
| 2618 | ulint n_blocks = 0; |
| 2619 | ulint n_releases = 0; |
| 2620 | mem_heap_t* heap = NULL; |
| 2621 | ulint offsets_[REC_OFFS_NORMAL_SIZE]; |
| 2622 | ulint* offsets = offsets_; |
| 2623 | rec_offs_init(offsets_); |
| 2624 | |
| 2625 | ut_ad(!dict_index_is_spatial(index)); |
| 2626 | |
| 2627 | lock_intention = btr_cur_get_and_clear_intention(&latch_mode); |
| 2628 | |
| 2629 | ut_ad(!(latch_mode & BTR_MODIFY_EXTERNAL)); |
| 2630 | |
| 2631 | savepoint = mtr_set_savepoint(mtr); |
| 2632 | |
| 2633 | switch (latch_mode) { |
| 2634 | case BTR_MODIFY_TREE: |
| 2635 | /* Most of delete-intended operations are purging. |
| 2636 | Free blocks and read IO bandwidth should be prior |
| 2637 | for them, when the history list is glowing huge. */ |
| 2638 | if (lock_intention == BTR_INTENTION_DELETE |
| 2639 | && trx_sys.history_size() > BTR_CUR_FINE_HISTORY_LENGTH |
| 2640 | && buf_get_n_pending_read_ios()) { |
| 2641 | mtr_x_lock(dict_index_get_lock(index), mtr); |
| 2642 | } else { |
| 2643 | mtr_sx_lock(dict_index_get_lock(index), mtr); |
| 2644 | } |
| 2645 | upper_rw_latch = RW_X_LATCH; |
| 2646 | break; |
| 2647 | case BTR_SEARCH_PREV: |
| 2648 | case BTR_MODIFY_PREV: |
| 2649 | /* This function doesn't support left uncle |
| 2650 | page lock for left leaf page lock, when |
| 2651 | needed. */ |
| 2652 | case BTR_SEARCH_TREE: |
| 2653 | case BTR_CONT_MODIFY_TREE: |
| 2654 | case BTR_CONT_SEARCH_TREE: |
| 2655 | ut_ad(0); |
| 2656 | /* fall through */ |
| 2657 | default: |
| 2658 | if (!srv_read_only_mode) { |
| 2659 | mtr_s_lock(dict_index_get_lock(index), mtr); |
| 2660 | upper_rw_latch = RW_S_LATCH; |
| 2661 | } else { |
| 2662 | upper_rw_latch = RW_NO_LATCH; |
| 2663 | } |
| 2664 | } |
| 2665 | |
| 2666 | DBUG_EXECUTE_IF("test_index_is_unavailable", |
| 2667 | return(false);); |
| 2668 | |
| 2669 | if (index->page == FIL_NULL) { |
| 2670 | /* Since we don't hold index lock until just now, the index |
| 2671 | could be modified by others, for example, if this is a |
| 2672 | statistics updater for referenced table, it could be marked |
| 2673 | as unavailable by 'DROP TABLE' in the mean time, since |
| 2674 | we don't hold lock for statistics updater */ |
| 2675 | return(false); |
| 2676 | } |
| 2677 | |
| 2678 | root_leaf_rw_latch = btr_cur_latch_for_root_leaf(latch_mode); |
| 2679 | |
| 2680 | page_cursor = btr_cur_get_page_cur(cursor); |
| 2681 | cursor->index = index; |
| 2682 | |
| 2683 | page_id_t page_id(index->table->space->id, index->page); |
| 2684 | const page_size_t page_size(index->table->space->flags); |
| 2685 | dberr_t err = DB_SUCCESS; |
| 2686 | |
| 2687 | if (root_leaf_rw_latch == RW_X_LATCH) { |
| 2688 | node_ptr_max_size = dict_index_node_ptr_max_size(index); |
| 2689 | } |
| 2690 | |
| 2691 | height = ULINT_UNDEFINED; |
| 2692 | |
| 2693 | for (;;) { |
| 2694 | buf_block_t* block; |
| 2695 | page_t* page; |
| 2696 | ulint rw_latch; |
| 2697 | |
| 2698 | ut_ad(n_blocks < BTR_MAX_LEVELS); |
| 2699 | |
| 2700 | if (height != 0 |
| 2701 | && latch_mode != BTR_MODIFY_TREE) { |
| 2702 | rw_latch = upper_rw_latch; |
| 2703 | } else { |
| 2704 | rw_latch = RW_NO_LATCH; |
| 2705 | } |
| 2706 | |
| 2707 | tree_savepoints[n_blocks] = mtr_set_savepoint(mtr); |
| 2708 | block = buf_page_get_gen(page_id, page_size, rw_latch, NULL, |
| 2709 | BUF_GET, file, line, mtr, &err); |
| 2710 | tree_blocks[n_blocks] = block; |
| 2711 | |
| 2712 | ut_ad((block != NULL) == (err == DB_SUCCESS)); |
| 2713 | |
| 2714 | if (err != DB_SUCCESS) { |
| 2715 | if (err == DB_DECRYPTION_FAILED) { |
| 2716 | ib_push_warning((void *)NULL, |
| 2717 | DB_DECRYPTION_FAILED, |
| 2718 | "Table %s is encrypted but encryption service or" |
| 2719 | " used key_id is not available. " |
| 2720 | " Can't continue reading table.", |
| 2721 | index->table->name); |
| 2722 | index->table->file_unreadable = true; |
| 2723 | } |
| 2724 | |
| 2725 | goto exit_loop; |
| 2726 | } |
| 2727 | |
| 2728 | page = buf_block_get_frame(block); |
| 2729 | |
| 2730 | if (height == ULINT_UNDEFINED |
| 2731 | && page_is_leaf(page) |
| 2732 | && rw_latch != RW_NO_LATCH |
| 2733 | && rw_latch != root_leaf_rw_latch) { |
| 2734 | /* We should retry to get the page, because the root page |
| 2735 | is latched with different level as a leaf page. */ |
| 2736 | ut_ad(root_leaf_rw_latch != RW_NO_LATCH); |
| 2737 | ut_ad(rw_latch == RW_S_LATCH); |
| 2738 | |
| 2739 | ut_ad(n_blocks == 0); |
| 2740 | mtr_release_block_at_savepoint( |
| 2741 | mtr, tree_savepoints[n_blocks], |
| 2742 | tree_blocks[n_blocks]); |
| 2743 | |
| 2744 | upper_rw_latch = root_leaf_rw_latch; |
| 2745 | continue; |
| 2746 | } |
| 2747 | |
| 2748 | ut_ad(fil_page_index_page_check(page)); |
| 2749 | ut_ad(index->id == btr_page_get_index_id(page)); |
| 2750 | |
| 2751 | if (height == ULINT_UNDEFINED) { |
| 2752 | /* We are in the root node */ |
| 2753 | |
| 2754 | height = btr_page_get_level(page); |
| 2755 | } |
| 2756 | |
| 2757 | if (height == 0) { |
| 2758 | if (rw_latch == RW_NO_LATCH |
| 2759 | || srv_read_only_mode) { |
| 2760 | btr_cur_latch_leaves( |
| 2761 | block, page_id, page_size, |
| 2762 | latch_mode, cursor, mtr); |
| 2763 | } |
| 2764 | |
| 2765 | /* btr_cur_open_at_index_side_func() and |
| 2766 | btr_cur_search_to_nth_level() release |
| 2767 | tree s-latch here.*/ |
| 2768 | switch (latch_mode) { |
| 2769 | case BTR_MODIFY_TREE: |
| 2770 | case BTR_CONT_MODIFY_TREE: |
| 2771 | case BTR_CONT_SEARCH_TREE: |
| 2772 | break; |
| 2773 | default: |
| 2774 | /* Release the tree s-latch */ |
| 2775 | if (!srv_read_only_mode) { |
| 2776 | mtr_release_s_latch_at_savepoint( |
| 2777 | mtr, savepoint, |
| 2778 | dict_index_get_lock(index)); |
| 2779 | } |
| 2780 | |
| 2781 | /* release upper blocks */ |
| 2782 | for (; n_releases < n_blocks; n_releases++) { |
| 2783 | mtr_release_block_at_savepoint( |
| 2784 | mtr, |
| 2785 | tree_savepoints[n_releases], |
| 2786 | tree_blocks[n_releases]); |
| 2787 | } |
| 2788 | } |
| 2789 | } |
| 2790 | |
| 2791 | page_cur_open_on_rnd_user_rec(block, page_cursor); |
| 2792 | |
| 2793 | if (height == 0) { |
| 2794 | |
| 2795 | break; |
| 2796 | } |
| 2797 | |
| 2798 | ut_ad(height > 0); |
| 2799 | |
| 2800 | height--; |
| 2801 | |
| 2802 | node_ptr = page_cur_get_rec(page_cursor); |
| 2803 | offsets = rec_get_offsets(node_ptr, cursor->index, offsets, |
| 2804 | false, ULINT_UNDEFINED, &heap); |
| 2805 | |
| 2806 | /* If the rec is the first or last in the page for |
| 2807 | pessimistic delete intention, it might cause node_ptr insert |
| 2808 | for the upper level. We should change the intention and retry. |
| 2809 | */ |
| 2810 | if (latch_mode == BTR_MODIFY_TREE |
| 2811 | && btr_cur_need_opposite_intention( |
| 2812 | page, lock_intention, node_ptr)) { |
| 2813 | |
| 2814 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 2815 | /* release all blocks */ |
| 2816 | for (; n_releases <= n_blocks; n_releases++) { |
| 2817 | mtr_release_block_at_savepoint( |
| 2818 | mtr, tree_savepoints[n_releases], |
| 2819 | tree_blocks[n_releases]); |
| 2820 | } |
| 2821 | |
| 2822 | lock_intention = BTR_INTENTION_BOTH; |
| 2823 | |
| 2824 | page_id.set_page_no(dict_index_get_page(index)); |
| 2825 | |
| 2826 | height = ULINT_UNDEFINED; |
| 2827 | |
| 2828 | n_blocks = 0; |
| 2829 | n_releases = 0; |
| 2830 | |
| 2831 | continue; |
| 2832 | } |
| 2833 | |
| 2834 | if (latch_mode == BTR_MODIFY_TREE |
| 2835 | && !btr_cur_will_modify_tree( |
| 2836 | cursor->index, page, lock_intention, node_ptr, |
| 2837 | node_ptr_max_size, page_size, mtr)) { |
| 2838 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 2839 | ut_ad(n_releases <= n_blocks); |
| 2840 | |
| 2841 | /* we can release upper blocks */ |
| 2842 | for (; n_releases < n_blocks; n_releases++) { |
| 2843 | if (n_releases == 0) { |
| 2844 | /* we should not release root page |
| 2845 | to pin to same block. */ |
| 2846 | continue; |
| 2847 | } |
| 2848 | |
| 2849 | /* release unused blocks to unpin */ |
| 2850 | mtr_release_block_at_savepoint( |
| 2851 | mtr, tree_savepoints[n_releases], |
| 2852 | tree_blocks[n_releases]); |
| 2853 | } |
| 2854 | } |
| 2855 | |
| 2856 | if (height == 0 |
| 2857 | && latch_mode == BTR_MODIFY_TREE) { |
| 2858 | ut_ad(upper_rw_latch == RW_X_LATCH); |
| 2859 | /* we should sx-latch root page, if released already. |
| 2860 | It contains seg_header. */ |
| 2861 | if (n_releases > 0) { |
| 2862 | mtr_block_sx_latch_at_savepoint( |
| 2863 | mtr, tree_savepoints[0], |
| 2864 | tree_blocks[0]); |
| 2865 | } |
| 2866 | |
| 2867 | /* x-latch the branch blocks not released yet. */ |
| 2868 | for (ulint i = n_releases; i <= n_blocks; i++) { |
| 2869 | mtr_block_x_latch_at_savepoint( |
| 2870 | mtr, tree_savepoints[i], |
| 2871 | tree_blocks[i]); |
| 2872 | } |
| 2873 | } |
| 2874 | |
| 2875 | /* Go to the child node */ |
| 2876 | page_id.set_page_no( |
| 2877 | btr_node_ptr_get_child_page_no(node_ptr, offsets)); |
| 2878 | |
| 2879 | n_blocks++; |
| 2880 | } |
| 2881 | |
| 2882 | exit_loop: |
| 2883 | if (UNIV_LIKELY_NULL(heap)) { |
| 2884 | mem_heap_free(heap); |
| 2885 | } |
| 2886 | |
| 2887 | return(true); |
| 2888 | } |
| 2889 | |
| 2890 | /*==================== B-TREE INSERT =========================*/ |
| 2891 | |
| 2892 | /*************************************************************//** |
| 2893 | Inserts a record if there is enough space, or if enough space can |
| 2894 | be freed by reorganizing. Differs from btr_cur_optimistic_insert because |
| 2895 | no heuristics is applied to whether it pays to use CPU time for |
| 2896 | reorganizing the page or not. |
| 2897 | |
| 2898 | IMPORTANT: The caller will have to update IBUF_BITMAP_FREE |
| 2899 | if this is a compressed leaf page in a secondary index. |
| 2900 | This has to be done either within the same mini-transaction, |
| 2901 | or by invoking ibuf_reset_free_bits() before mtr_commit(). |
| 2902 | |
| 2903 | @return pointer to inserted record if succeed, else NULL */ |
| 2904 | static MY_ATTRIBUTE((nonnull, warn_unused_result)) |
| 2905 | rec_t* |
| 2906 | btr_cur_insert_if_possible( |
| 2907 | /*=======================*/ |
| 2908 | btr_cur_t* cursor, /*!< in: cursor on page after which to insert; |
| 2909 | cursor stays valid */ |
| 2910 | const dtuple_t* tuple, /*!< in: tuple to insert; the size info need not |
| 2911 | have been stored to tuple */ |
| 2912 | ulint** offsets,/*!< out: offsets on *rec */ |
| 2913 | mem_heap_t** heap, /*!< in/out: pointer to memory heap, or NULL */ |
| 2914 | ulint n_ext, /*!< in: number of externally stored columns */ |
| 2915 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 2916 | { |
| 2917 | page_cur_t* page_cursor; |
| 2918 | rec_t* rec; |
| 2919 | |
| 2920 | ut_ad(dtuple_check_typed(tuple)); |
| 2921 | |
| 2922 | ut_ad(mtr_memo_contains(mtr, btr_cur_get_block(cursor), |
| 2923 | MTR_MEMO_PAGE_X_FIX)); |
| 2924 | page_cursor = btr_cur_get_page_cur(cursor); |
| 2925 | |
| 2926 | /* Now, try the insert */ |
| 2927 | rec = page_cur_tuple_insert(page_cursor, tuple, cursor->index, |
| 2928 | offsets, heap, n_ext, mtr); |
| 2929 | |
| 2930 | /* If the record did not fit, reorganize. |
| 2931 | For compressed pages, page_cur_tuple_insert() |
| 2932 | attempted this already. */ |
| 2933 | if (!rec && !page_cur_get_page_zip(page_cursor) |
| 2934 | && btr_page_reorganize(page_cursor, cursor->index, mtr)) { |
| 2935 | rec = page_cur_tuple_insert( |
| 2936 | page_cursor, tuple, cursor->index, |
| 2937 | offsets, heap, n_ext, mtr); |
| 2938 | } |
| 2939 | |
| 2940 | ut_ad(!rec || rec_offs_validate(rec, cursor->index, *offsets)); |
| 2941 | return(rec); |
| 2942 | } |
| 2943 | |
| 2944 | /*************************************************************//** |
| 2945 | For an insert, checks the locks and does the undo logging if desired. |
| 2946 | @return DB_SUCCESS, DB_WAIT_LOCK, DB_FAIL, or error number */ |
| 2947 | UNIV_INLINE MY_ATTRIBUTE((warn_unused_result, nonnull(2,3,5,6))) |
| 2948 | dberr_t |
| 2949 | btr_cur_ins_lock_and_undo( |
| 2950 | /*======================*/ |
| 2951 | ulint flags, /*!< in: undo logging and locking flags: if |
| 2952 | not zero, the parameters index and thr |
| 2953 | should be specified */ |
| 2954 | btr_cur_t* cursor, /*!< in: cursor on page after which to insert */ |
| 2955 | dtuple_t* entry, /*!< in/out: entry to insert */ |
| 2956 | que_thr_t* thr, /*!< in: query thread or NULL */ |
| 2957 | mtr_t* mtr, /*!< in/out: mini-transaction */ |
| 2958 | bool* inherit)/*!< out: true if the inserted new record maybe |
| 2959 | should inherit LOCK_GAP type locks from the |
| 2960 | successor record */ |
| 2961 | { |
| 2962 | dict_index_t* index; |
| 2963 | dberr_t err = DB_SUCCESS; |
| 2964 | rec_t* rec; |
| 2965 | roll_ptr_t roll_ptr; |
| 2966 | |
| 2967 | /* Check if we have to wait for a lock: enqueue an explicit lock |
| 2968 | request if yes */ |
| 2969 | |
| 2970 | rec = btr_cur_get_rec(cursor); |
| 2971 | index = cursor->index; |
| 2972 | |
| 2973 | ut_ad(!dict_index_is_online_ddl(index) |
| 2974 | || dict_index_is_clust(index) |
| 2975 | || (flags & BTR_CREATE_FLAG)); |
| 2976 | ut_ad(mtr->is_named_space(index->table->space)); |
| 2977 | |
| 2978 | /* Check if there is predicate or GAP lock preventing the insertion */ |
| 2979 | if (!(flags & BTR_NO_LOCKING_FLAG)) { |
| 2980 | if (dict_index_is_spatial(index)) { |
| 2981 | lock_prdt_t prdt; |
| 2982 | rtr_mbr_t mbr; |
| 2983 | |
| 2984 | rtr_get_mbr_from_tuple(entry, &mbr); |
| 2985 | |
| 2986 | /* Use on stack MBR variable to test if a lock is |
| 2987 | needed. If so, the predicate (MBR) will be allocated |
| 2988 | from lock heap in lock_prdt_insert_check_and_lock() */ |
| 2989 | lock_init_prdt_from_mbr( |
| 2990 | &prdt, &mbr, 0, NULL); |
| 2991 | |
| 2992 | err = lock_prdt_insert_check_and_lock( |
| 2993 | flags, rec, btr_cur_get_block(cursor), |
| 2994 | index, thr, mtr, &prdt); |
| 2995 | *inherit = false; |
| 2996 | } else { |
| 2997 | err = lock_rec_insert_check_and_lock( |
| 2998 | flags, rec, btr_cur_get_block(cursor), |
| 2999 | index, thr, mtr, inherit); |
| 3000 | } |
| 3001 | } |
| 3002 | |
| 3003 | if (err != DB_SUCCESS |
| 3004 | || !(~flags | (BTR_NO_UNDO_LOG_FLAG | BTR_KEEP_SYS_FLAG)) |
| 3005 | || !dict_index_is_clust(index) || dict_index_is_ibuf(index)) { |
| 3006 | |
| 3007 | return(err); |
| 3008 | } |
| 3009 | |
| 3010 | if (flags & BTR_NO_UNDO_LOG_FLAG) { |
| 3011 | roll_ptr = roll_ptr_t(1) << ROLL_PTR_INSERT_FLAG_POS; |
| 3012 | if (!(flags & BTR_KEEP_SYS_FLAG)) { |
| 3013 | upd_sys: |
| 3014 | row_upd_index_entry_sys_field(entry, index, |
| 3015 | DATA_ROLL_PTR, roll_ptr); |
| 3016 | } |
| 3017 | } else { |
| 3018 | err = trx_undo_report_row_operation(thr, index, entry, |
| 3019 | NULL, 0, NULL, NULL, |
| 3020 | &roll_ptr); |
| 3021 | if (err == DB_SUCCESS) { |
| 3022 | goto upd_sys; |
| 3023 | } |
| 3024 | } |
| 3025 | |
| 3026 | return(err); |
| 3027 | } |
| 3028 | |
| 3029 | /** |
| 3030 | Prefetch siblings of the leaf for the pessimistic operation. |
| 3031 | @param block leaf page */ |
| 3032 | static |
| 3033 | void |
| 3034 | btr_cur_prefetch_siblings( |
| 3035 | buf_block_t* block) |
| 3036 | { |
| 3037 | page_t* page = buf_block_get_frame(block); |
| 3038 | |
| 3039 | ut_ad(page_is_leaf(page)); |
| 3040 | |
| 3041 | ulint left_page_no = fil_page_get_prev(page); |
| 3042 | ulint right_page_no = fil_page_get_next(page); |
| 3043 | |
| 3044 | if (left_page_no != FIL_NULL) { |
| 3045 | buf_read_page_background( |
| 3046 | page_id_t(block->page.id.space(), left_page_no), |
| 3047 | block->page.size, false); |
| 3048 | } |
| 3049 | if (right_page_no != FIL_NULL) { |
| 3050 | buf_read_page_background( |
| 3051 | page_id_t(block->page.id.space(), right_page_no), |
| 3052 | block->page.size, false); |
| 3053 | } |
| 3054 | if (left_page_no != FIL_NULL |
| 3055 | || right_page_no != FIL_NULL) { |
| 3056 | os_aio_simulated_wake_handler_threads(); |
| 3057 | } |
| 3058 | } |
| 3059 | |
| 3060 | /*************************************************************//** |
| 3061 | Tries to perform an insert to a page in an index tree, next to cursor. |
| 3062 | It is assumed that mtr holds an x-latch on the page. The operation does |
| 3063 | not succeed if there is too little space on the page. If there is just |
| 3064 | one record on the page, the insert will always succeed; this is to |
| 3065 | prevent trying to split a page with just one record. |
| 3066 | @return DB_SUCCESS, DB_WAIT_LOCK, DB_FAIL, or error number */ |
| 3067 | dberr_t |
| 3068 | btr_cur_optimistic_insert( |
| 3069 | /*======================*/ |
| 3070 | ulint flags, /*!< in: undo logging and locking flags: if not |
| 3071 | zero, the parameters index and thr should be |
| 3072 | specified */ |
| 3073 | btr_cur_t* cursor, /*!< in: cursor on page after which to insert; |
| 3074 | cursor stays valid */ |
| 3075 | ulint** offsets,/*!< out: offsets on *rec */ |
| 3076 | mem_heap_t** heap, /*!< in/out: pointer to memory heap */ |
| 3077 | dtuple_t* entry, /*!< in/out: entry to insert */ |
| 3078 | rec_t** rec, /*!< out: pointer to inserted record if |
| 3079 | succeed */ |
| 3080 | big_rec_t** big_rec,/*!< out: big rec vector whose fields have to |
| 3081 | be stored externally by the caller */ |
| 3082 | ulint n_ext, /*!< in: number of externally stored columns */ |
| 3083 | que_thr_t* thr, /*!< in/out: query thread; can be NULL if |
| 3084 | !(~flags |
| 3085 | & (BTR_NO_LOCKING_FLAG |
| 3086 | | BTR_NO_UNDO_LOG_FLAG)) */ |
| 3087 | mtr_t* mtr) /*!< in/out: mini-transaction; |
| 3088 | if this function returns DB_SUCCESS on |
| 3089 | a leaf page of a secondary index in a |
| 3090 | compressed tablespace, the caller must |
| 3091 | mtr_commit(mtr) before latching |
| 3092 | any further pages */ |
| 3093 | { |
| 3094 | big_rec_t* big_rec_vec = NULL; |
| 3095 | dict_index_t* index; |
| 3096 | page_cur_t* page_cursor; |
| 3097 | buf_block_t* block; |
| 3098 | page_t* page; |
| 3099 | rec_t* dummy; |
| 3100 | bool leaf; |
| 3101 | bool reorg; |
| 3102 | bool inherit = true; |
| 3103 | ulint rec_size; |
| 3104 | dberr_t err; |
| 3105 | |
| 3106 | ut_ad(thr || !(~flags & (BTR_NO_LOCKING_FLAG | BTR_NO_UNDO_LOG_FLAG))); |
| 3107 | *big_rec = NULL; |
| 3108 | |
| 3109 | block = btr_cur_get_block(cursor); |
| 3110 | page = buf_block_get_frame(block); |
| 3111 | index = cursor->index; |
| 3112 | |
| 3113 | ut_ad(mtr_memo_contains(mtr, block, MTR_MEMO_PAGE_X_FIX)); |
| 3114 | ut_ad(!dict_index_is_online_ddl(index) |
| 3115 | || dict_index_is_clust(index) |
| 3116 | || (flags & BTR_CREATE_FLAG)); |
| 3117 | ut_ad(dtuple_check_typed(entry)); |
| 3118 | |
| 3119 | const page_size_t& page_size = block->page.size; |
| 3120 | |
| 3121 | #ifdef UNIV_DEBUG_VALGRIND |
| 3122 | if (page_size.is_compressed()) { |
| 3123 | UNIV_MEM_ASSERT_RW(page, page_size.logical()); |
| 3124 | UNIV_MEM_ASSERT_RW(block->page.zip.data, page_size.physical()); |
| 3125 | } |
| 3126 | #endif /* UNIV_DEBUG_VALGRIND */ |
| 3127 | |
| 3128 | leaf = page_is_leaf(page); |
| 3129 | |
| 3130 | /* Calculate the record size when entry is converted to a record */ |
| 3131 | rec_size = rec_get_converted_size(index, entry, n_ext); |
| 3132 | |
| 3133 | if (page_zip_rec_needs_ext(rec_size, page_is_comp(page), |
| 3134 | dtuple_get_n_fields(entry), page_size)) { |
| 3135 | |
| 3136 | /* The record is so big that we have to store some fields |
| 3137 | externally on separate database pages */ |
| 3138 | big_rec_vec = dtuple_convert_big_rec(index, 0, entry, &n_ext); |
| 3139 | |
| 3140 | if (UNIV_UNLIKELY(big_rec_vec == NULL)) { |
| 3141 | |
| 3142 | return(DB_TOO_BIG_RECORD); |
| 3143 | } |
| 3144 | |
| 3145 | rec_size = rec_get_converted_size(index, entry, n_ext); |
| 3146 | } |
| 3147 | |
| 3148 | if (page_size.is_compressed() && page_zip_is_too_big(index, entry)) { |
| 3149 | if (big_rec_vec != NULL) { |
| 3150 | dtuple_convert_back_big_rec(index, entry, big_rec_vec); |
| 3151 | } |
| 3152 | |
| 3153 | return(DB_TOO_BIG_RECORD); |
| 3154 | } |
| 3155 | |
| 3156 | LIMIT_OPTIMISTIC_INSERT_DEBUG(page_get_n_recs(page), |
| 3157 | goto fail); |
| 3158 | |
| 3159 | if (leaf && page_size.is_compressed() |
| 3160 | && (page_get_data_size(page) + rec_size |
| 3161 | >= dict_index_zip_pad_optimal_page_size(index))) { |
| 3162 | /* If compression padding tells us that insertion will |
| 3163 | result in too packed up page i.e.: which is likely to |
| 3164 | cause compression failure then don't do an optimistic |
| 3165 | insertion. */ |
| 3166 | fail: |
| 3167 | err = DB_FAIL; |
| 3168 | |
| 3169 | /* prefetch siblings of the leaf for the pessimistic |
| 3170 | operation, if the page is leaf. */ |
| 3171 | if (page_is_leaf(page)) { |
| 3172 | btr_cur_prefetch_siblings(block); |
| 3173 | } |
| 3174 | fail_err: |
| 3175 | |
| 3176 | if (big_rec_vec) { |
| 3177 | dtuple_convert_back_big_rec(index, entry, big_rec_vec); |
| 3178 | } |
| 3179 | |
| 3180 | return(err); |
| 3181 | } |
| 3182 | |
| 3183 | ulint max_size = page_get_max_insert_size_after_reorganize(page, 1); |
| 3184 | |
| 3185 | if (page_has_garbage(page)) { |
| 3186 | if ((max_size < rec_size |
| 3187 | || max_size < BTR_CUR_PAGE_REORGANIZE_LIMIT) |
| 3188 | && page_get_n_recs(page) > 1 |
| 3189 | && page_get_max_insert_size(page, 1) < rec_size) { |
| 3190 | |
| 3191 | goto fail; |
| 3192 | } |
| 3193 | } else if (max_size < rec_size) { |
| 3194 | goto fail; |
| 3195 | } |
| 3196 | |
| 3197 | /* If there have been many consecutive inserts to the |
| 3198 | clustered index leaf page of an uncompressed table, check if |
| 3199 | we have to split the page to reserve enough free space for |
| 3200 | future updates of records. */ |
| 3201 | |
| 3202 | if (leaf && !page_size.is_compressed() && dict_index_is_clust(index) |
| 3203 | && page_get_n_recs(page) >= 2 |
| 3204 | && dict_index_get_space_reserve() + rec_size > max_size |
| 3205 | && (btr_page_get_split_rec_to_right(cursor, &dummy) |
| 3206 | || btr_page_get_split_rec_to_left(cursor, &dummy))) { |
| 3207 | goto fail; |
| 3208 | } |
| 3209 | |
| 3210 | page_cursor = btr_cur_get_page_cur(cursor); |
| 3211 | |
| 3212 | DBUG_LOG("ib_cur", |
| 3213 | "insert "<< index->name << " ("<< index->id << ") by " |
| 3214 | << ib::hex(thr ? thr->graph->trx->id : 0) |
| 3215 | << ' ' << rec_printer(entry).str()); |
| 3216 | DBUG_EXECUTE_IF("do_page_reorganize", |
| 3217 | btr_page_reorganize(page_cursor, index, mtr);); |
| 3218 | |
| 3219 | /* Now, try the insert */ |
| 3220 | { |
| 3221 | const rec_t* page_cursor_rec = page_cur_get_rec(page_cursor); |
| 3222 | |
| 3223 | /* Check locks and write to the undo log, |
| 3224 | if specified */ |
| 3225 | err = btr_cur_ins_lock_and_undo(flags, cursor, entry, |
| 3226 | thr, mtr, &inherit); |
| 3227 | if (err != DB_SUCCESS) { |
| 3228 | goto fail_err; |
| 3229 | } |
| 3230 | |
| 3231 | #ifdef UNIV_DEBUG |
| 3232 | if (!(flags & BTR_CREATE_FLAG) |
| 3233 | && index->is_primary() && page_is_leaf(page)) { |
| 3234 | const dfield_t* trx_id = dtuple_get_nth_field( |
| 3235 | entry, dict_col_get_clust_pos( |
| 3236 | dict_table_get_sys_col(index->table, |
| 3237 | DATA_TRX_ID), |
| 3238 | index)); |
| 3239 | |
| 3240 | ut_ad(trx_id->len == DATA_TRX_ID_LEN); |
| 3241 | ut_ad(trx_id[1].len == DATA_ROLL_PTR_LEN); |
| 3242 | ut_ad(*static_cast<const byte*> |
| 3243 | (trx_id[1].data) & 0x80); |
| 3244 | if (flags & BTR_NO_UNDO_LOG_FLAG) { |
| 3245 | ut_ad(!memcmp(trx_id->data, reset_trx_id, |
| 3246 | DATA_TRX_ID_LEN)); |
| 3247 | } else { |
| 3248 | ut_ad(thr->graph->trx->id); |
| 3249 | ut_ad(thr->graph->trx->id |
| 3250 | == trx_read_trx_id( |
| 3251 | static_cast<const byte*>( |
| 3252 | trx_id->data))); |
| 3253 | } |
| 3254 | } |
| 3255 | #endif |
| 3256 | |
| 3257 | *rec = page_cur_tuple_insert( |
| 3258 | page_cursor, entry, index, offsets, heap, |
| 3259 | n_ext, mtr); |
| 3260 | |
| 3261 | reorg = page_cursor_rec != page_cur_get_rec(page_cursor); |
| 3262 | } |
| 3263 | |
| 3264 | if (*rec) { |
| 3265 | } else if (page_size.is_compressed()) { |
| 3266 | ut_ad(!index->table->is_temporary()); |
| 3267 | /* Reset the IBUF_BITMAP_FREE bits, because |
| 3268 | page_cur_tuple_insert() will have attempted page |
| 3269 | reorganize before failing. */ |
| 3270 | if (leaf |
| 3271 | && !dict_index_is_clust(index)) { |
| 3272 | ibuf_reset_free_bits(block); |
| 3273 | } |
| 3274 | |
| 3275 | goto fail; |
| 3276 | } else { |
| 3277 | ut_ad(!reorg); |
| 3278 | |
| 3279 | /* If the record did not fit, reorganize */ |
| 3280 | if (!btr_page_reorganize(page_cursor, index, mtr)) { |
| 3281 | ut_ad(0); |
| 3282 | goto fail; |
| 3283 | } |
| 3284 | |
| 3285 | ut_ad(page_get_max_insert_size(page, 1) == max_size); |
| 3286 | |
| 3287 | reorg = TRUE; |
| 3288 | |
| 3289 | *rec = page_cur_tuple_insert(page_cursor, entry, index, |
| 3290 | offsets, heap, n_ext, mtr); |
| 3291 | |
| 3292 | if (UNIV_UNLIKELY(!*rec)) { |
| 3293 | ib::fatal() << "Cannot insert tuple "<< *entry |
| 3294 | << "into index "<< index->name |
| 3295 | << " of table "<< index->table->name |
| 3296 | << ". Max size: "<< max_size; |
| 3297 | } |
| 3298 | } |
| 3299 | |
| 3300 | #ifdef BTR_CUR_HASH_ADAPT |
| 3301 | if (!leaf) { |
| 3302 | # ifdef MYSQL_INDEX_DISABLE_AHI |
| 3303 | } else if (index->disable_ahi) { |
| 3304 | # endif |
| 3305 | } else if (entry->info_bits & REC_INFO_MIN_REC_FLAG) { |
| 3306 | ut_ad(entry->info_bits == REC_INFO_DEFAULT_ROW); |
| 3307 | ut_ad(index->is_instant()); |
| 3308 | ut_ad(flags == BTR_NO_LOCKING_FLAG); |
| 3309 | } else { |
| 3310 | rw_lock_t* ahi_latch = btr_get_search_latch(index); |
| 3311 | if (!reorg && cursor->flag == BTR_CUR_HASH) { |
| 3312 | btr_search_update_hash_node_on_insert( |
| 3313 | cursor, ahi_latch); |
| 3314 | } else { |
| 3315 | btr_search_update_hash_on_insert(cursor, ahi_latch); |
| 3316 | } |
| 3317 | } |
| 3318 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 3319 | |
| 3320 | if (!(flags & BTR_NO_LOCKING_FLAG) && inherit) { |
| 3321 | |
| 3322 | lock_update_insert(block, *rec); |
| 3323 | } |
| 3324 | |
| 3325 | if (leaf |
| 3326 | && !dict_index_is_clust(index) |
| 3327 | && !index->table->is_temporary()) { |
| 3328 | /* Update the free bits of the B-tree page in the |
| 3329 | insert buffer bitmap. */ |
| 3330 | |
| 3331 | /* The free bits in the insert buffer bitmap must |
| 3332 | never exceed the free space on a page. It is safe to |
| 3333 | decrement or reset the bits in the bitmap in a |
| 3334 | mini-transaction that is committed before the |
| 3335 | mini-transaction that affects the free space. */ |
| 3336 | |
| 3337 | /* It is unsafe to increment the bits in a separately |
| 3338 | committed mini-transaction, because in crash recovery, |
| 3339 | the free bits could momentarily be set too high. */ |
| 3340 | |
| 3341 | if (page_size.is_compressed()) { |
| 3342 | /* Update the bits in the same mini-transaction. */ |
| 3343 | ibuf_update_free_bits_zip(block, mtr); |
| 3344 | } else { |
| 3345 | /* Decrement the bits in a separate |
| 3346 | mini-transaction. */ |
| 3347 | ibuf_update_free_bits_if_full( |
| 3348 | block, max_size, |
| 3349 | rec_size + PAGE_DIR_SLOT_SIZE); |
| 3350 | } |
| 3351 | } |
| 3352 | |
| 3353 | *big_rec = big_rec_vec; |
| 3354 | |
| 3355 | return(DB_SUCCESS); |
| 3356 | } |
| 3357 | |
| 3358 | /*************************************************************//** |
| 3359 | Performs an insert on a page of an index tree. It is assumed that mtr |
| 3360 | holds an x-latch on the tree and on the cursor page. If the insert is |
| 3361 | made on the leaf level, to avoid deadlocks, mtr must also own x-latches |
| 3362 | to brothers of page, if those brothers exist. |
| 3363 | @return DB_SUCCESS or error number */ |
| 3364 | dberr_t |
| 3365 | btr_cur_pessimistic_insert( |
| 3366 | /*=======================*/ |
| 3367 | ulint flags, /*!< in: undo logging and locking flags: if not |
| 3368 | zero, the parameter thr should be |
| 3369 | specified; if no undo logging is specified, |
| 3370 | then the caller must have reserved enough |
| 3371 | free extents in the file space so that the |
| 3372 | insertion will certainly succeed */ |
| 3373 | btr_cur_t* cursor, /*!< in: cursor after which to insert; |
| 3374 | cursor stays valid */ |
| 3375 | ulint** offsets,/*!< out: offsets on *rec */ |
| 3376 | mem_heap_t** heap, /*!< in/out: pointer to memory heap |
| 3377 | that can be emptied */ |
| 3378 | dtuple_t* entry, /*!< in/out: entry to insert */ |
| 3379 | rec_t** rec, /*!< out: pointer to inserted record if |
| 3380 | succeed */ |
| 3381 | big_rec_t** big_rec,/*!< out: big rec vector whose fields have to |
| 3382 | be stored externally by the caller */ |
| 3383 | ulint n_ext, /*!< in: number of externally stored columns */ |
| 3384 | que_thr_t* thr, /*!< in/out: query thread; can be NULL if |
| 3385 | !(~flags |
| 3386 | & (BTR_NO_LOCKING_FLAG |
| 3387 | | BTR_NO_UNDO_LOG_FLAG)) */ |
| 3388 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 3389 | { |
| 3390 | dict_index_t* index = cursor->index; |
| 3391 | big_rec_t* big_rec_vec = NULL; |
| 3392 | dberr_t err; |
| 3393 | bool inherit = false; |
| 3394 | bool success; |
| 3395 | ulint n_reserved = 0; |
| 3396 | |
| 3397 | ut_ad(dtuple_check_typed(entry)); |
| 3398 | ut_ad(thr || !(~flags & (BTR_NO_LOCKING_FLAG | BTR_NO_UNDO_LOG_FLAG))); |
| 3399 | |
| 3400 | *big_rec = NULL; |
| 3401 | |
| 3402 | ut_ad(mtr_memo_contains_flagged( |
| 3403 | mtr, dict_index_get_lock(btr_cur_get_index(cursor)), |
| 3404 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK)); |
| 3405 | ut_ad(mtr_memo_contains(mtr, btr_cur_get_block(cursor), |
| 3406 | MTR_MEMO_PAGE_X_FIX)); |
| 3407 | ut_ad(!dict_index_is_online_ddl(index) |
| 3408 | || dict_index_is_clust(index) |
| 3409 | || (flags & BTR_CREATE_FLAG)); |
| 3410 | |
| 3411 | cursor->flag = BTR_CUR_BINARY; |
| 3412 | |
| 3413 | /* Check locks and write to undo log, if specified */ |
| 3414 | |
| 3415 | err = btr_cur_ins_lock_and_undo(flags, cursor, entry, |
| 3416 | thr, mtr, &inherit); |
| 3417 | |
| 3418 | if (err != DB_SUCCESS) { |
| 3419 | |
| 3420 | return(err); |
| 3421 | } |
| 3422 | |
| 3423 | if (!(flags & BTR_NO_UNDO_LOG_FLAG)) { |
| 3424 | /* First reserve enough free space for the file segments |
| 3425 | of the index tree, so that the insert will not fail because |
| 3426 | of lack of space */ |
| 3427 | |
| 3428 | ulint n_extents = cursor->tree_height / 16 + 3; |
| 3429 | |
| 3430 | success = fsp_reserve_free_extents(&n_reserved, |
| 3431 | index->table->space, |
| 3432 | n_extents, FSP_NORMAL, mtr); |
| 3433 | if (!success) { |
| 3434 | return(DB_OUT_OF_FILE_SPACE); |
| 3435 | } |
| 3436 | } |
| 3437 | |
| 3438 | if (page_zip_rec_needs_ext(rec_get_converted_size(index, entry, n_ext), |
| 3439 | dict_table_is_comp(index->table), |
| 3440 | dtuple_get_n_fields(entry), |
| 3441 | dict_table_page_size(index->table))) { |
| 3442 | /* The record is so big that we have to store some fields |
| 3443 | externally on separate database pages */ |
| 3444 | |
| 3445 | if (UNIV_LIKELY_NULL(big_rec_vec)) { |
| 3446 | /* This should never happen, but we handle |
| 3447 | the situation in a robust manner. */ |
| 3448 | ut_ad(0); |
| 3449 | dtuple_convert_back_big_rec(index, entry, big_rec_vec); |
| 3450 | } |
| 3451 | |
| 3452 | big_rec_vec = dtuple_convert_big_rec(index, 0, entry, &n_ext); |
| 3453 | |
| 3454 | if (big_rec_vec == NULL) { |
| 3455 | |
| 3456 | index->table->space->release_free_extents(n_reserved); |
| 3457 | return(DB_TOO_BIG_RECORD); |
| 3458 | } |
| 3459 | } |
| 3460 | |
| 3461 | if (dict_index_get_page(index) |
| 3462 | == btr_cur_get_block(cursor)->page.id.page_no()) { |
| 3463 | |
| 3464 | /* The page is the root page */ |
| 3465 | *rec = btr_root_raise_and_insert( |
| 3466 | flags, cursor, offsets, heap, entry, n_ext, mtr); |
| 3467 | } else { |
| 3468 | *rec = btr_page_split_and_insert( |
| 3469 | flags, cursor, offsets, heap, entry, n_ext, mtr); |
| 3470 | } |
| 3471 | |
| 3472 | if (*rec == NULL && os_has_said_disk_full) { |
| 3473 | return(DB_OUT_OF_FILE_SPACE); |
| 3474 | } |
| 3475 | |
| 3476 | ut_ad(page_rec_get_next(btr_cur_get_rec(cursor)) == *rec |
| 3477 | || dict_index_is_spatial(index)); |
| 3478 | |
| 3479 | if (!(flags & BTR_NO_LOCKING_FLAG)) { |
| 3480 | ut_ad(!index->table->is_temporary()); |
| 3481 | if (dict_index_is_spatial(index)) { |
| 3482 | /* Do nothing */ |
| 3483 | } else { |
| 3484 | /* The cursor might be moved to the other page |
| 3485 | and the max trx id field should be updated after |
| 3486 | the cursor was fixed. */ |
| 3487 | if (!dict_index_is_clust(index)) { |
| 3488 | page_update_max_trx_id( |
| 3489 | btr_cur_get_block(cursor), |
| 3490 | btr_cur_get_page_zip(cursor), |
| 3491 | thr_get_trx(thr)->id, mtr); |
| 3492 | } |
| 3493 | |
| 3494 | if (!page_rec_is_infimum(btr_cur_get_rec(cursor)) |
| 3495 | || btr_page_get_prev( |
| 3496 | buf_block_get_frame( |
| 3497 | btr_cur_get_block(cursor)), mtr) |
| 3498 | == FIL_NULL) { |
| 3499 | /* split and inserted need to call |
| 3500 | lock_update_insert() always. */ |
| 3501 | inherit = true; |
| 3502 | } |
| 3503 | } |
| 3504 | } |
| 3505 | |
| 3506 | if (!page_is_leaf(btr_cur_get_page(cursor))) { |
| 3507 | ut_ad(!big_rec_vec); |
| 3508 | } else { |
| 3509 | #ifdef BTR_CUR_HASH_ADAPT |
| 3510 | # ifdef MYSQL_INDEX_DISABLE_AHI |
| 3511 | if (index->disable_ahi); else |
| 3512 | # endif |
| 3513 | if (entry->info_bits & REC_INFO_MIN_REC_FLAG) { |
| 3514 | ut_ad(entry->info_bits == REC_INFO_DEFAULT_ROW); |
| 3515 | ut_ad(index->is_instant()); |
| 3516 | ut_ad((flags & ulint(~BTR_KEEP_IBUF_BITMAP)) |
| 3517 | == BTR_NO_LOCKING_FLAG); |
| 3518 | } else { |
| 3519 | btr_search_update_hash_on_insert( |
| 3520 | cursor, btr_get_search_latch(index)); |
| 3521 | } |
| 3522 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 3523 | if (inherit && !(flags & BTR_NO_LOCKING_FLAG)) { |
| 3524 | |
| 3525 | lock_update_insert(btr_cur_get_block(cursor), *rec); |
| 3526 | } |
| 3527 | } |
| 3528 | |
| 3529 | index->table->space->release_free_extents(n_reserved); |
| 3530 | *big_rec = big_rec_vec; |
| 3531 | |
| 3532 | return(DB_SUCCESS); |
| 3533 | } |
| 3534 | |
| 3535 | /*==================== B-TREE UPDATE =========================*/ |
| 3536 | |
| 3537 | /*************************************************************//** |
| 3538 | For an update, checks the locks and does the undo logging. |
| 3539 | @return DB_SUCCESS, DB_WAIT_LOCK, or error number */ |
| 3540 | UNIV_INLINE MY_ATTRIBUTE((warn_unused_result)) |
| 3541 | dberr_t |
| 3542 | btr_cur_upd_lock_and_undo( |
| 3543 | /*======================*/ |
| 3544 | ulint flags, /*!< in: undo logging and locking flags */ |
| 3545 | btr_cur_t* cursor, /*!< in: cursor on record to update */ |
| 3546 | const ulint* offsets,/*!< in: rec_get_offsets() on cursor */ |
| 3547 | const upd_t* update, /*!< in: update vector */ |
| 3548 | ulint cmpl_info,/*!< in: compiler info on secondary index |
| 3549 | updates */ |
| 3550 | que_thr_t* thr, /*!< in: query thread |
| 3551 | (can be NULL if BTR_NO_LOCKING_FLAG) */ |
| 3552 | mtr_t* mtr, /*!< in/out: mini-transaction */ |
| 3553 | roll_ptr_t* roll_ptr)/*!< out: roll pointer */ |
| 3554 | { |
| 3555 | dict_index_t* index; |
| 3556 | const rec_t* rec; |
| 3557 | dberr_t err; |
| 3558 | |
| 3559 | ut_ad((thr != NULL) || (flags & BTR_NO_LOCKING_FLAG)); |
| 3560 | |
| 3561 | rec = btr_cur_get_rec(cursor); |
| 3562 | index = cursor->index; |
| 3563 | |
| 3564 | ut_ad(rec_offs_validate(rec, index, offsets)); |
| 3565 | ut_ad(mtr->is_named_space(index->table->space)); |
| 3566 | |
| 3567 | if (!dict_index_is_clust(index)) { |
| 3568 | ut_ad(dict_index_is_online_ddl(index) |
| 3569 | == !!(flags & BTR_CREATE_FLAG)); |
| 3570 | |
| 3571 | /* We do undo logging only when we update a clustered index |
| 3572 | record */ |
| 3573 | return(lock_sec_rec_modify_check_and_lock( |
| 3574 | flags, btr_cur_get_block(cursor), rec, |
| 3575 | index, thr, mtr)); |
| 3576 | } |
| 3577 | |
| 3578 | /* Check if we have to wait for a lock: enqueue an explicit lock |
| 3579 | request if yes */ |
| 3580 | |
| 3581 | if (!(flags & BTR_NO_LOCKING_FLAG)) { |
| 3582 | err = lock_clust_rec_modify_check_and_lock( |
| 3583 | flags, btr_cur_get_block(cursor), rec, index, |
| 3584 | offsets, thr); |
| 3585 | if (err != DB_SUCCESS) { |
| 3586 | return(err); |
| 3587 | } |
| 3588 | } |
| 3589 | |
| 3590 | /* Append the info about the update in the undo log */ |
| 3591 | |
| 3592 | return((flags & BTR_NO_UNDO_LOG_FLAG) |
| 3593 | ? DB_SUCCESS |
| 3594 | : trx_undo_report_row_operation( |
| 3595 | thr, index, NULL, update, |
| 3596 | cmpl_info, rec, offsets, roll_ptr)); |
| 3597 | } |
| 3598 | |
| 3599 | /***********************************************************//** |
| 3600 | Writes a redo log record of updating a record in-place. */ |
| 3601 | void |
| 3602 | btr_cur_update_in_place_log( |
| 3603 | /*========================*/ |
| 3604 | ulint flags, /*!< in: flags */ |
| 3605 | const rec_t* rec, /*!< in: record */ |
| 3606 | dict_index_t* index, /*!< in: index of the record */ |
| 3607 | const upd_t* update, /*!< in: update vector */ |
| 3608 | trx_id_t trx_id, /*!< in: transaction id */ |
| 3609 | roll_ptr_t roll_ptr, /*!< in: roll ptr */ |
| 3610 | mtr_t* mtr) /*!< in: mtr */ |
| 3611 | { |
| 3612 | byte* log_ptr; |
| 3613 | const page_t* page = page_align(rec); |
| 3614 | ut_ad(flags < 256); |
| 3615 | ut_ad(!!page_is_comp(page) == dict_table_is_comp(index->table)); |
| 3616 | |
| 3617 | log_ptr = mlog_open_and_write_index(mtr, rec, index, page_is_comp(page) |
| 3618 | ? MLOG_COMP_REC_UPDATE_IN_PLACE |
| 3619 | : MLOG_REC_UPDATE_IN_PLACE, |
| 3620 | 1 + DATA_ROLL_PTR_LEN + 14 + 2 |
| 3621 | + MLOG_BUF_MARGIN); |
| 3622 | |
| 3623 | if (!log_ptr) { |
| 3624 | /* Logging in mtr is switched off during crash recovery */ |
| 3625 | return; |
| 3626 | } |
| 3627 | |
| 3628 | /* For secondary indexes, we could skip writing the dummy system fields |
| 3629 | to the redo log but we have to change redo log parsing of |
| 3630 | MLOG_REC_UPDATE_IN_PLACE/MLOG_COMP_REC_UPDATE_IN_PLACE or we have to add |
| 3631 | new redo log record. For now, just write dummy sys fields to the redo |
| 3632 | log if we are updating a secondary index record. |
| 3633 | */ |
| 3634 | mach_write_to_1(log_ptr, flags); |
| 3635 | log_ptr++; |
| 3636 | |
| 3637 | if (dict_index_is_clust(index)) { |
| 3638 | log_ptr = row_upd_write_sys_vals_to_log( |
| 3639 | index, trx_id, roll_ptr, log_ptr, mtr); |
| 3640 | } else { |
| 3641 | /* Dummy system fields for a secondary index */ |
| 3642 | /* TRX_ID Position */ |
| 3643 | log_ptr += mach_write_compressed(log_ptr, 0); |
| 3644 | /* ROLL_PTR */ |
| 3645 | trx_write_roll_ptr(log_ptr, 0); |
| 3646 | log_ptr += DATA_ROLL_PTR_LEN; |
| 3647 | /* TRX_ID */ |
| 3648 | log_ptr += mach_u64_write_compressed(log_ptr, 0); |
| 3649 | } |
| 3650 | |
| 3651 | mach_write_to_2(log_ptr, page_offset(rec)); |
| 3652 | log_ptr += 2; |
| 3653 | |
| 3654 | row_upd_index_write_log(update, log_ptr, mtr); |
| 3655 | } |
| 3656 | |
| 3657 | /***********************************************************//** |
| 3658 | Parses a redo log record of updating a record in-place. |
| 3659 | @return end of log record or NULL */ |
| 3660 | byte* |
| 3661 | btr_cur_parse_update_in_place( |
| 3662 | /*==========================*/ |
| 3663 | byte* ptr, /*!< in: buffer */ |
| 3664 | byte* end_ptr,/*!< in: buffer end */ |
| 3665 | page_t* page, /*!< in/out: page or NULL */ |
| 3666 | page_zip_des_t* page_zip,/*!< in/out: compressed page, or NULL */ |
| 3667 | dict_index_t* index) /*!< in: index corresponding to page */ |
| 3668 | { |
| 3669 | ulint flags; |
| 3670 | rec_t* rec; |
| 3671 | upd_t* update; |
| 3672 | ulint pos; |
| 3673 | trx_id_t trx_id; |
| 3674 | roll_ptr_t roll_ptr; |
| 3675 | ulint rec_offset; |
| 3676 | mem_heap_t* heap; |
| 3677 | ulint* offsets; |
| 3678 | |
| 3679 | if (end_ptr < ptr + 1) { |
| 3680 | |
| 3681 | return(NULL); |
| 3682 | } |
| 3683 | |
| 3684 | flags = mach_read_from_1(ptr); |
| 3685 | ptr++; |
| 3686 | |
| 3687 | ptr = row_upd_parse_sys_vals(ptr, end_ptr, &pos, &trx_id, &roll_ptr); |
| 3688 | |
| 3689 | if (ptr == NULL) { |
| 3690 | |
| 3691 | return(NULL); |
| 3692 | } |
| 3693 | |
| 3694 | if (end_ptr < ptr + 2) { |
| 3695 | |
| 3696 | return(NULL); |
| 3697 | } |
| 3698 | |
| 3699 | rec_offset = mach_read_from_2(ptr); |
| 3700 | ptr += 2; |
| 3701 | |
| 3702 | ut_a(rec_offset <= srv_page_size); |
| 3703 | |
| 3704 | heap = mem_heap_create(256); |
| 3705 | |
| 3706 | ptr = row_upd_index_parse(ptr, end_ptr, heap, &update); |
| 3707 | |
| 3708 | if (!ptr || !page) { |
| 3709 | |
| 3710 | goto func_exit; |
| 3711 | } |
| 3712 | |
| 3713 | ut_a((ibool)!!page_is_comp(page) == dict_table_is_comp(index->table)); |
| 3714 | rec = page + rec_offset; |
| 3715 | |
| 3716 | /* We do not need to reserve search latch, as the page is only |
| 3717 | being recovered, and there cannot be a hash index to it. */ |
| 3718 | |
| 3719 | /* The function rtr_update_mbr_field_in_place() is generating |
| 3720 | these records on node pointer pages; therefore we have to |
| 3721 | check if this is a leaf page. */ |
| 3722 | |
| 3723 | offsets = rec_get_offsets(rec, index, NULL, |
| 3724 | flags != (BTR_NO_UNDO_LOG_FLAG |
| 3725 | | BTR_NO_LOCKING_FLAG |
| 3726 | | BTR_KEEP_SYS_FLAG) |
| 3727 | || page_is_leaf(page), |
| 3728 | ULINT_UNDEFINED, &heap); |
| 3729 | |
| 3730 | if (!(flags & BTR_KEEP_SYS_FLAG)) { |
| 3731 | row_upd_rec_sys_fields_in_recovery(rec, page_zip, offsets, |
| 3732 | pos, trx_id, roll_ptr); |
| 3733 | } |
| 3734 | |
| 3735 | row_upd_rec_in_place(rec, index, offsets, update, page_zip); |
| 3736 | |
| 3737 | func_exit: |
| 3738 | mem_heap_free(heap); |
| 3739 | |
| 3740 | return(ptr); |
| 3741 | } |
| 3742 | |
| 3743 | /*************************************************************//** |
| 3744 | See if there is enough place in the page modification log to log |
| 3745 | an update-in-place. |
| 3746 | |
| 3747 | @retval false if out of space; IBUF_BITMAP_FREE will be reset |
| 3748 | outside mtr if the page was recompressed |
| 3749 | @retval true if enough place; |
| 3750 | |
| 3751 | IMPORTANT: The caller will have to update IBUF_BITMAP_FREE if this is |
| 3752 | a secondary index leaf page. This has to be done either within the |
| 3753 | same mini-transaction, or by invoking ibuf_reset_free_bits() before |
| 3754 | mtr_commit(mtr). */ |
| 3755 | bool |
| 3756 | btr_cur_update_alloc_zip_func( |
| 3757 | /*==========================*/ |
| 3758 | page_zip_des_t* page_zip,/*!< in/out: compressed page */ |
| 3759 | page_cur_t* cursor, /*!< in/out: B-tree page cursor */ |
| 3760 | dict_index_t* index, /*!< in: the index corresponding to cursor */ |
| 3761 | #ifdef UNIV_DEBUG |
| 3762 | ulint* offsets,/*!< in/out: offsets of the cursor record */ |
| 3763 | #endif /* UNIV_DEBUG */ |
| 3764 | ulint length, /*!< in: size needed */ |
| 3765 | bool create, /*!< in: true=delete-and-insert, |
| 3766 | false=update-in-place */ |
| 3767 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 3768 | { |
| 3769 | |
| 3770 | /* Have a local copy of the variables as these can change |
| 3771 | dynamically. */ |
| 3772 | const page_t* page = page_cur_get_page(cursor); |
| 3773 | |
| 3774 | ut_ad(page_zip == page_cur_get_page_zip(cursor)); |
| 3775 | ut_ad(!dict_index_is_ibuf(index)); |
| 3776 | ut_ad(rec_offs_validate(page_cur_get_rec(cursor), index, offsets)); |
| 3777 | |
| 3778 | if (page_zip_available(page_zip, dict_index_is_clust(index), |
| 3779 | length, create)) { |
| 3780 | return(true); |
| 3781 | } |
| 3782 | |
| 3783 | if (!page_zip->m_nonempty && !page_has_garbage(page)) { |
| 3784 | /* The page has been freshly compressed, so |
| 3785 | reorganizing it will not help. */ |
| 3786 | return(false); |
| 3787 | } |
| 3788 | |
| 3789 | if (create && page_is_leaf(page) |
| 3790 | && (length + page_get_data_size(page) |
| 3791 | >= dict_index_zip_pad_optimal_page_size(index))) { |
| 3792 | return(false); |
| 3793 | } |
| 3794 | |
| 3795 | if (!btr_page_reorganize(cursor, index, mtr)) { |
| 3796 | goto out_of_space; |
| 3797 | } |
| 3798 | |
| 3799 | rec_offs_make_valid(page_cur_get_rec(cursor), index, |
| 3800 | page_is_leaf(page), offsets); |
| 3801 | |
| 3802 | /* After recompressing a page, we must make sure that the free |
| 3803 | bits in the insert buffer bitmap will not exceed the free |
| 3804 | space on the page. Because this function will not attempt |
| 3805 | recompression unless page_zip_available() fails above, it is |
| 3806 | safe to reset the free bits if page_zip_available() fails |
| 3807 | again, below. The free bits can safely be reset in a separate |
| 3808 | mini-transaction. If page_zip_available() succeeds below, we |
| 3809 | can be sure that the btr_page_reorganize() above did not reduce |
| 3810 | the free space available on the page. */ |
| 3811 | |
| 3812 | if (page_zip_available(page_zip, dict_index_is_clust(index), |
| 3813 | length, create)) { |
| 3814 | return(true); |
| 3815 | } |
| 3816 | |
| 3817 | out_of_space: |
| 3818 | ut_ad(rec_offs_validate(page_cur_get_rec(cursor), index, offsets)); |
| 3819 | |
| 3820 | /* Out of space: reset the free bits. */ |
| 3821 | if (!dict_index_is_clust(index) |
| 3822 | && !index->table->is_temporary() |
| 3823 | && page_is_leaf(page)) { |
| 3824 | ibuf_reset_free_bits(page_cur_get_block(cursor)); |
| 3825 | } |
| 3826 | |
| 3827 | return(false); |
| 3828 | } |
| 3829 | |
| 3830 | /*************************************************************//** |
| 3831 | Updates a record when the update causes no size changes in its fields. |
| 3832 | We assume here that the ordering fields of the record do not change. |
| 3833 | @return locking or undo log related error code, or |
| 3834 | @retval DB_SUCCESS on success |
| 3835 | @retval DB_ZIP_OVERFLOW if there is not enough space left |
| 3836 | on the compressed page (IBUF_BITMAP_FREE was reset outside mtr) */ |
| 3837 | dberr_t |
| 3838 | btr_cur_update_in_place( |
| 3839 | /*====================*/ |
| 3840 | ulint flags, /*!< in: undo logging and locking flags */ |
| 3841 | btr_cur_t* cursor, /*!< in: cursor on the record to update; |
| 3842 | cursor stays valid and positioned on the |
| 3843 | same record */ |
| 3844 | ulint* offsets,/*!< in/out: offsets on cursor->page_cur.rec */ |
| 3845 | const upd_t* update, /*!< in: update vector */ |
| 3846 | ulint cmpl_info,/*!< in: compiler info on secondary index |
| 3847 | updates */ |
| 3848 | que_thr_t* thr, /*!< in: query thread */ |
| 3849 | trx_id_t trx_id, /*!< in: transaction id */ |
| 3850 | mtr_t* mtr) /*!< in/out: mini-transaction; if this |
| 3851 | is a secondary index, the caller must |
| 3852 | mtr_commit(mtr) before latching any |
| 3853 | further pages */ |
| 3854 | { |
| 3855 | dict_index_t* index; |
| 3856 | buf_block_t* block; |
| 3857 | page_zip_des_t* page_zip; |
| 3858 | dberr_t err; |
| 3859 | rec_t* rec; |
| 3860 | roll_ptr_t roll_ptr = 0; |
| 3861 | ulint was_delete_marked; |
| 3862 | |
| 3863 | ut_ad(page_is_leaf(cursor->page_cur.block->frame)); |
| 3864 | rec = btr_cur_get_rec(cursor); |
| 3865 | index = cursor->index; |
| 3866 | ut_ad(rec_offs_validate(rec, index, offsets)); |
| 3867 | ut_ad(!!page_rec_is_comp(rec) == dict_table_is_comp(index->table)); |
| 3868 | ut_ad(trx_id > 0 || (flags & BTR_KEEP_SYS_FLAG)); |
| 3869 | /* The insert buffer tree should never be updated in place. */ |
| 3870 | ut_ad(!dict_index_is_ibuf(index)); |
| 3871 | ut_ad(dict_index_is_online_ddl(index) == !!(flags & BTR_CREATE_FLAG) |
| 3872 | || dict_index_is_clust(index)); |
| 3873 | ut_ad(thr_get_trx(thr)->id == trx_id |
| 3874 | || (flags & ulint(~(BTR_KEEP_POS_FLAG | BTR_KEEP_IBUF_BITMAP))) |
| 3875 | == (BTR_NO_UNDO_LOG_FLAG | BTR_NO_LOCKING_FLAG |
| 3876 | | BTR_CREATE_FLAG | BTR_KEEP_SYS_FLAG)); |
| 3877 | ut_ad(fil_page_index_page_check(btr_cur_get_page(cursor))); |
| 3878 | ut_ad(btr_page_get_index_id(btr_cur_get_page(cursor)) == index->id); |
| 3879 | ut_ad(!(update->info_bits & REC_INFO_MIN_REC_FLAG)); |
| 3880 | |
| 3881 | DBUG_LOG("ib_cur", |
| 3882 | "update-in-place "<< index->name << " ("<< index->id |
| 3883 | << ") by "<< ib::hex(trx_id) << ": " |
| 3884 | << rec_printer(rec, offsets).str()); |
| 3885 | |
| 3886 | block = btr_cur_get_block(cursor); |
| 3887 | page_zip = buf_block_get_page_zip(block); |
| 3888 | |
| 3889 | /* Check that enough space is available on the compressed page. */ |
| 3890 | if (page_zip) { |
| 3891 | ut_ad(!index->table->is_temporary()); |
| 3892 | |
| 3893 | if (!btr_cur_update_alloc_zip( |
| 3894 | page_zip, btr_cur_get_page_cur(cursor), |
| 3895 | index, offsets, rec_offs_size(offsets), |
| 3896 | false, mtr)) { |
| 3897 | return(DB_ZIP_OVERFLOW); |
| 3898 | } |
| 3899 | |
| 3900 | rec = btr_cur_get_rec(cursor); |
| 3901 | } |
| 3902 | |
| 3903 | /* Do lock checking and undo logging */ |
| 3904 | err = btr_cur_upd_lock_and_undo(flags, cursor, offsets, |
| 3905 | update, cmpl_info, |
| 3906 | thr, mtr, &roll_ptr); |
| 3907 | if (UNIV_UNLIKELY(err != DB_SUCCESS)) { |
| 3908 | /* We may need to update the IBUF_BITMAP_FREE |
| 3909 | bits after a reorganize that was done in |
| 3910 | btr_cur_update_alloc_zip(). */ |
| 3911 | goto func_exit; |
| 3912 | } |
| 3913 | |
| 3914 | if (!(flags & BTR_KEEP_SYS_FLAG)) { |
| 3915 | row_upd_rec_sys_fields(rec, NULL, index, offsets, |
| 3916 | thr_get_trx(thr), roll_ptr); |
| 3917 | } |
| 3918 | |
| 3919 | was_delete_marked = rec_get_deleted_flag( |
| 3920 | rec, page_is_comp(buf_block_get_frame(block))); |
| 3921 | /* In delete-marked records, DB_TRX_ID must always refer to an |
| 3922 | existing undo log record. */ |
| 3923 | ut_ad(!was_delete_marked |
| 3924 | || !dict_index_is_clust(index) |
| 3925 | || row_get_rec_trx_id(rec, index, offsets)); |
| 3926 | |
| 3927 | #ifdef BTR_CUR_HASH_ADAPT |
| 3928 | { |
| 3929 | rw_lock_t* ahi_latch = block->index |
| 3930 | ? btr_get_search_latch(index) : NULL; |
| 3931 | if (ahi_latch) { |
| 3932 | /* TO DO: Can we skip this if none of the fields |
| 3933 | index->search_info->curr_n_fields |
| 3934 | are being updated? */ |
| 3935 | |
| 3936 | /* The function row_upd_changes_ord_field_binary |
| 3937 | does not work on a secondary index. */ |
| 3938 | |
| 3939 | if (!dict_index_is_clust(index) |
| 3940 | || row_upd_changes_ord_field_binary( |
| 3941 | index, update, thr, NULL, NULL)) { |
| 3942 | ut_ad(!(update->info_bits |
| 3943 | & REC_INFO_MIN_REC_FLAG)); |
| 3944 | /* Remove possible hash index pointer |
| 3945 | to this record */ |
| 3946 | btr_search_update_hash_on_delete(cursor); |
| 3947 | } |
| 3948 | |
| 3949 | rw_lock_x_lock(ahi_latch); |
| 3950 | } |
| 3951 | |
| 3952 | assert_block_ahi_valid(block); |
| 3953 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 3954 | |
| 3955 | row_upd_rec_in_place(rec, index, offsets, update, page_zip); |
| 3956 | |
| 3957 | #ifdef BTR_CUR_HASH_ADAPT |
| 3958 | if (ahi_latch) { |
| 3959 | rw_lock_x_unlock(ahi_latch); |
| 3960 | } |
| 3961 | } |
| 3962 | #endif /* BTR_CUR_HASH_ADAPT */ |
| 3963 | |
| 3964 | btr_cur_update_in_place_log(flags, rec, index, update, |
| 3965 | trx_id, roll_ptr, mtr); |
| 3966 | |
| 3967 | if (was_delete_marked |
| 3968 | && !rec_get_deleted_flag( |
| 3969 | rec, page_is_comp(buf_block_get_frame(block)))) { |
| 3970 | /* The new updated record owns its possible externally |
| 3971 | stored fields */ |
| 3972 | |
| 3973 | btr_cur_unmark_extern_fields(page_zip, |
| 3974 | rec, index, offsets, mtr); |
| 3975 | } |
| 3976 | |
| 3977 | ut_ad(err == DB_SUCCESS); |
| 3978 | |
| 3979 | func_exit: |
| 3980 | if (page_zip |
| 3981 | && !(flags & BTR_KEEP_IBUF_BITMAP) |
| 3982 | && !dict_index_is_clust(index) |
| 3983 | && page_is_leaf(buf_block_get_frame(block))) { |
| 3984 | /* Update the free bits in the insert buffer. */ |
| 3985 | ut_ad(!index->table->is_temporary()); |
| 3986 | ibuf_update_free_bits_zip(block, mtr); |
| 3987 | } |
| 3988 | |
| 3989 | return(err); |
| 3990 | } |
| 3991 | |
| 3992 | /** Trim an update tuple due to instant ADD COLUMN, if needed. |
| 3993 | For normal records, the trailing instantly added fields that match |
| 3994 | the 'default row' are omitted. |
| 3995 | |
| 3996 | For the special 'default row' record on a table on which instant |
| 3997 | ADD COLUMN has already been executed, both ADD COLUMN and the |
| 3998 | rollback of ADD COLUMN need to be handled specially. |
| 3999 | |
| 4000 | @param[in,out] entry index entry |
| 4001 | @param[in] index index |
| 4002 | @param[in] update update vector |
| 4003 | @param[in] thr execution thread */ |
| 4004 | static inline |
| 4005 | void |
| 4006 | btr_cur_trim( |
| 4007 | dtuple_t* entry, |
| 4008 | const dict_index_t* index, |
| 4009 | const upd_t* update, |
| 4010 | const que_thr_t* thr) |
| 4011 | { |
| 4012 | if (!index->is_instant()) { |
| 4013 | } else if (UNIV_UNLIKELY(update->info_bits == REC_INFO_DEFAULT_ROW)) { |
| 4014 | /* We are either updating a 'default row' |
| 4015 | (instantly adding columns to a table where instant ADD was |
| 4016 | already executed) or rolling back such an operation. */ |
| 4017 | ut_ad(!upd_get_nth_field(update, 0)->orig_len); |
| 4018 | ut_ad(upd_get_nth_field(update, 0)->field_no |
| 4019 | > index->n_core_fields); |
| 4020 | |
| 4021 | if (thr->graph->trx->in_rollback) { |
| 4022 | /* This rollback can occur either as part of |
| 4023 | ha_innobase::commit_inplace_alter_table() rolling |
| 4024 | back after a failed innobase_add_instant_try(), |
| 4025 | or as part of crash recovery. Either way, the |
| 4026 | table will be in the data dictionary cache, with |
| 4027 | the instantly added columns going to be removed |
| 4028 | later in the rollback. */ |
| 4029 | ut_ad(index->table->cached); |
| 4030 | /* The DB_TRX_ID,DB_ROLL_PTR are always last, |
| 4031 | and there should be some change to roll back. |
| 4032 | The first field in the update vector is the |
| 4033 | first instantly added column logged by |
| 4034 | innobase_add_instant_try(). */ |
| 4035 | ut_ad(update->n_fields > 2); |
| 4036 | ulint n_fields = upd_get_nth_field(update, 0) |
| 4037 | ->field_no; |
| 4038 | ut_ad(n_fields + 1 >= entry->n_fields); |
| 4039 | entry->n_fields = n_fields; |
| 4040 | } |
| 4041 | } else { |
| 4042 | entry->trim(*index); |
| 4043 | } |
| 4044 | } |
| 4045 | |
| 4046 | /*************************************************************//** |
| 4047 | Tries to update a record on a page in an index tree. It is assumed that mtr |
| 4048 | holds an x-latch on the page. The operation does not succeed if there is too |
| 4049 | little space on the page or if the update would result in too empty a page, |
| 4050 | so that tree compression is recommended. We assume here that the ordering |
| 4051 | fields of the record do not change. |
| 4052 | @return error code, including |
| 4053 | @retval DB_SUCCESS on success |
| 4054 | @retval DB_OVERFLOW if the updated record does not fit |
| 4055 | @retval DB_UNDERFLOW if the page would become too empty |
| 4056 | @retval DB_ZIP_OVERFLOW if there is not enough space left |
| 4057 | on the compressed page (IBUF_BITMAP_FREE was reset outside mtr) */ |
| 4058 | dberr_t |
| 4059 | btr_cur_optimistic_update( |
| 4060 | /*======================*/ |
| 4061 | ulint flags, /*!< in: undo logging and locking flags */ |
| 4062 | btr_cur_t* cursor, /*!< in: cursor on the record to update; |
| 4063 | cursor stays valid and positioned on the |
| 4064 | same record */ |
| 4065 | ulint** offsets,/*!< out: offsets on cursor->page_cur.rec */ |
| 4066 | mem_heap_t** heap, /*!< in/out: pointer to NULL or memory heap */ |
| 4067 | const upd_t* update, /*!< in: update vector; this must also |
| 4068 | contain trx id and roll ptr fields */ |
| 4069 | ulint cmpl_info,/*!< in: compiler info on secondary index |
| 4070 | updates */ |
| 4071 | que_thr_t* thr, /*!< in: query thread */ |
| 4072 | trx_id_t trx_id, /*!< in: transaction id */ |
| 4073 | mtr_t* mtr) /*!< in/out: mini-transaction; if this |
| 4074 | is a secondary index, the caller must |
| 4075 | mtr_commit(mtr) before latching any |
| 4076 | further pages */ |
| 4077 | { |
| 4078 | dict_index_t* index; |
| 4079 | page_cur_t* page_cursor; |
| 4080 | dberr_t err; |
| 4081 | buf_block_t* block; |
| 4082 | page_t* page; |
| 4083 | page_zip_des_t* page_zip; |
| 4084 | rec_t* rec; |
| 4085 | ulint max_size; |
| 4086 | ulint new_rec_size; |
| 4087 | ulint old_rec_size; |
| 4088 | ulint max_ins_size = 0; |
| 4089 | dtuple_t* new_entry; |
| 4090 | roll_ptr_t roll_ptr; |
| 4091 | ulint i; |
| 4092 | ulint n_ext; |
| 4093 | |
| 4094 | block = btr_cur_get_block(cursor); |
| 4095 | page = buf_block_get_frame(block); |
| 4096 | rec = btr_cur_get_rec(cursor); |
| 4097 | index = cursor->index; |
| 4098 | ut_ad(trx_id > 0 || (flags & BTR_KEEP_SYS_FLAG)); |
| 4099 | ut_ad(!!page_rec_is_comp(rec) == dict_table_is_comp(index->table)); |
| 4100 | ut_ad(mtr_memo_contains(mtr, block, MTR_MEMO_PAGE_X_FIX)); |
| 4101 | /* This is intended only for leaf page updates */ |
| 4102 | ut_ad(page_is_leaf(page)); |
| 4103 | /* The insert buffer tree should never be updated in place. */ |
| 4104 | ut_ad(!dict_index_is_ibuf(index)); |
| 4105 | ut_ad(dict_index_is_online_ddl(index) == !!(flags & BTR_CREATE_FLAG) |
| 4106 | || dict_index_is_clust(index)); |
| 4107 | ut_ad(thr_get_trx(thr)->id == trx_id |
| 4108 | || (flags & ulint(~(BTR_KEEP_POS_FLAG | BTR_KEEP_IBUF_BITMAP))) |
| 4109 | == (BTR_NO_UNDO_LOG_FLAG | BTR_NO_LOCKING_FLAG |
| 4110 | | BTR_CREATE_FLAG | BTR_KEEP_SYS_FLAG)); |
| 4111 | ut_ad(fil_page_index_page_check(page)); |
| 4112 | ut_ad(btr_page_get_index_id(page) == index->id); |
| 4113 | |
| 4114 | *offsets = rec_get_offsets(rec, index, *offsets, true, |
| 4115 | ULINT_UNDEFINED, heap); |
| 4116 | #if defined UNIV_DEBUG || defined UNIV_BLOB_LIGHT_DEBUG |
| 4117 | ut_a(!rec_offs_any_null_extern(rec, *offsets) |
| 4118 | || trx_is_recv(thr_get_trx(thr))); |
| 4119 | #endif /* UNIV_DEBUG || UNIV_BLOB_LIGHT_DEBUG */ |
| 4120 | |
| 4121 | const bool is_default_row = update->info_bits == REC_INFO_DEFAULT_ROW; |
| 4122 | |
| 4123 | if (UNIV_LIKELY(!is_default_row) |
| 4124 | && !row_upd_changes_field_size_or_external(index, *offsets, |
| 4125 | update)) { |
| 4126 | |
| 4127 | /* The simplest and the most common case: the update does not |
| 4128 | change the size of any field and none of the updated fields is |
| 4129 | externally stored in rec or update, and there is enough space |
| 4130 | on the compressed page to log the update. */ |
| 4131 | |
| 4132 | return(btr_cur_update_in_place( |
| 4133 | flags, cursor, *offsets, update, |
| 4134 | cmpl_info, thr, trx_id, mtr)); |
| 4135 | } |
| 4136 | |
| 4137 | if (rec_offs_any_extern(*offsets)) { |
| 4138 | any_extern: |
| 4139 | /* Externally stored fields are treated in pessimistic |
| 4140 | update */ |
| 4141 | |
| 4142 | /* prefetch siblings of the leaf for the pessimistic |
| 4143 | operation. */ |
| 4144 | btr_cur_prefetch_siblings(block); |
| 4145 | |
| 4146 | return(DB_OVERFLOW); |
| 4147 | } |
| 4148 | |
| 4149 | for (i = 0; i < upd_get_n_fields(update); i++) { |
| 4150 | if (dfield_is_ext(&upd_get_nth_field(update, i)->new_val)) { |
| 4151 | |
| 4152 | goto any_extern; |
| 4153 | } |
| 4154 | } |
| 4155 | |
| 4156 | DBUG_LOG("ib_cur", |
| 4157 | "update "<< index->name << " ("<< index->id << ") by " |
| 4158 | << ib::hex(trx_id) << ": " |
| 4159 | << rec_printer(rec, *offsets).str()); |
| 4160 | |
| 4161 | page_cursor = btr_cur_get_page_cur(cursor); |
| 4162 | |
| 4163 | if (!*heap) { |
| 4164 | *heap = mem_heap_create( |
| 4165 | rec_offs_size(*offsets) |
| 4166 | + DTUPLE_EST_ALLOC(rec_offs_n_fields(*offsets))); |
| 4167 | } |
| 4168 | |
| 4169 | new_entry = row_rec_to_index_entry(rec, index, *offsets, |
| 4170 | &n_ext, *heap); |
| 4171 | /* We checked above that there are no externally stored fields. */ |
| 4172 | ut_a(!n_ext); |
| 4173 | |
| 4174 | /* The page containing the clustered index record |
| 4175 | corresponding to new_entry is latched in mtr. |
| 4176 | Thus the following call is safe. */ |
| 4177 | row_upd_index_replace_new_col_vals_index_pos(new_entry, index, update, |
| 4178 | *heap); |
| 4179 | btr_cur_trim(new_entry, index, update, thr); |
| 4180 | old_rec_size = rec_offs_size(*offsets); |
| 4181 | new_rec_size = rec_get_converted_size(index, new_entry, 0); |
| 4182 | |
| 4183 | page_zip = buf_block_get_page_zip(block); |
| 4184 | #ifdef UNIV_ZIP_DEBUG |
| 4185 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 4186 | #endif /* UNIV_ZIP_DEBUG */ |
| 4187 | |
| 4188 | if (page_zip) { |
| 4189 | ut_ad(!index->table->is_temporary()); |
| 4190 | |
| 4191 | if (page_zip_rec_needs_ext(new_rec_size, page_is_comp(page), |
| 4192 | dict_index_get_n_fields(index), |
| 4193 | dict_table_page_size(index->table))) { |
| 4194 | goto any_extern; |
| 4195 | } |
| 4196 | |
| 4197 | if (!btr_cur_update_alloc_zip( |
| 4198 | page_zip, page_cursor, index, *offsets, |
| 4199 | new_rec_size, true, mtr)) { |
| 4200 | return(DB_ZIP_OVERFLOW); |
| 4201 | } |
| 4202 | |
| 4203 | rec = page_cur_get_rec(page_cursor); |
| 4204 | } |
| 4205 | |
| 4206 | /* We limit max record size to 16k even for 64k page size. */ |
| 4207 | if (new_rec_size >= COMPRESSED_REC_MAX_DATA_SIZE || |
| 4208 | (!dict_table_is_comp(index->table) |
| 4209 | && new_rec_size >= REDUNDANT_REC_MAX_DATA_SIZE)) { |
| 4210 | err = DB_OVERFLOW; |
| 4211 | |
| 4212 | goto func_exit; |
| 4213 | } |
| 4214 | |
| 4215 | if (UNIV_UNLIKELY(new_rec_size |
| 4216 | >= (page_get_free_space_of_empty(page_is_comp(page)) |
| 4217 | / 2))) { |
| 4218 | /* We may need to update the IBUF_BITMAP_FREE |
| 4219 | bits after a reorganize that was done in |
| 4220 | btr_cur_update_alloc_zip(). */ |
| 4221 | err = DB_OVERFLOW; |
| 4222 | goto func_exit; |
| 4223 | } |
| 4224 | |
| 4225 | if (UNIV_UNLIKELY(page_get_data_size(page) |
| 4226 | - old_rec_size + new_rec_size |
| 4227 | < BTR_CUR_PAGE_COMPRESS_LIMIT(index))) { |
| 4228 | /* We may need to update the IBUF_BITMAP_FREE |
| 4229 | bits after a reorganize that was done in |
| 4230 | btr_cur_update_alloc_zip(). */ |
| 4231 | |
| 4232 | /* The page would become too empty */ |
| 4233 | err = DB_UNDERFLOW; |
| 4234 | goto func_exit; |
| 4235 | } |
| 4236 | |
| 4237 | /* We do not attempt to reorganize if the page is compressed. |
| 4238 | This is because the page may fail to compress after reorganization. */ |
| 4239 | max_size = page_zip |
| 4240 | ? page_get_max_insert_size(page, 1) |
| 4241 | : (old_rec_size |
| 4242 | + page_get_max_insert_size_after_reorganize(page, 1)); |
| 4243 | |
| 4244 | if (!page_zip) { |
| 4245 | max_ins_size = page_get_max_insert_size_after_reorganize( |
| 4246 | page, 1); |
| 4247 | } |
| 4248 | |
| 4249 | if (!(((max_size >= BTR_CUR_PAGE_REORGANIZE_LIMIT) |
| 4250 | && (max_size >= new_rec_size)) |
| 4251 | || (page_get_n_recs(page) <= 1))) { |
| 4252 | |
| 4253 | /* We may need to update the IBUF_BITMAP_FREE |
| 4254 | bits after a reorganize that was done in |
| 4255 | btr_cur_update_alloc_zip(). */ |
| 4256 | |
| 4257 | /* There was not enough space, or it did not pay to |
| 4258 | reorganize: for simplicity, we decide what to do assuming a |
| 4259 | reorganization is needed, though it might not be necessary */ |
| 4260 | |
| 4261 | err = DB_OVERFLOW; |
| 4262 | goto func_exit; |
| 4263 | } |
| 4264 | |
| 4265 | /* Do lock checking and undo logging */ |
| 4266 | err = btr_cur_upd_lock_and_undo(flags, cursor, *offsets, |
| 4267 | update, cmpl_info, |
| 4268 | thr, mtr, &roll_ptr); |
| 4269 | if (err != DB_SUCCESS) { |
| 4270 | /* We may need to update the IBUF_BITMAP_FREE |
| 4271 | bits after a reorganize that was done in |
| 4272 | btr_cur_update_alloc_zip(). */ |
| 4273 | goto func_exit; |
| 4274 | } |
| 4275 | |
| 4276 | /* Ok, we may do the replacement. Store on the page infimum the |
| 4277 | explicit locks on rec, before deleting rec (see the comment in |
| 4278 | btr_cur_pessimistic_update). */ |
| 4279 | if (!dict_table_is_locking_disabled(index->table)) { |
| 4280 | lock_rec_store_on_page_infimum(block, rec); |
| 4281 | } |
| 4282 | |
| 4283 | if (UNIV_UNLIKELY(is_default_row)) { |
| 4284 | ut_ad(new_entry->info_bits == REC_INFO_DEFAULT_ROW); |
| 4285 | ut_ad(index->is_instant()); |
| 4286 | /* This can be innobase_add_instant_try() performing a |
| 4287 | subsequent instant ADD COLUMN, or its rollback by |
| 4288 | row_undo_mod_clust_low(). */ |
| 4289 | ut_ad(flags & BTR_NO_LOCKING_FLAG); |
| 4290 | } else { |
| 4291 | btr_search_update_hash_on_delete(cursor); |
| 4292 | } |
| 4293 | |
| 4294 | page_cur_delete_rec(page_cursor, index, *offsets, mtr); |
| 4295 | |
| 4296 | page_cur_move_to_prev(page_cursor); |
| 4297 | |
| 4298 | if (!(flags & BTR_KEEP_SYS_FLAG)) { |
| 4299 | row_upd_index_entry_sys_field(new_entry, index, DATA_ROLL_PTR, |
| 4300 | roll_ptr); |
| 4301 | row_upd_index_entry_sys_field(new_entry, index, DATA_TRX_ID, |
| 4302 | trx_id); |
| 4303 | } |
| 4304 | |
| 4305 | /* There are no externally stored columns in new_entry */ |
| 4306 | rec = btr_cur_insert_if_possible( |
| 4307 | cursor, new_entry, offsets, heap, 0/*n_ext*/, mtr); |
| 4308 | ut_a(rec); /* <- We calculated above the insert would fit */ |
| 4309 | |
| 4310 | if (UNIV_UNLIKELY(is_default_row)) { |
| 4311 | /* We must empty the PAGE_FREE list, because if this |
| 4312 | was a rollback, the shortened 'default row' record |
| 4313 | would have too many fields, and we would be unable to |
| 4314 | know the size of the freed record. */ |
| 4315 | btr_page_reorganize(page_cursor, index, mtr); |
| 4316 | } else if (!dict_table_is_locking_disabled(index->table)) { |
| 4317 | /* Restore the old explicit lock state on the record */ |
| 4318 | lock_rec_restore_from_page_infimum(block, rec, block); |
| 4319 | } |
| 4320 | |
| 4321 | page_cur_move_to_next(page_cursor); |
| 4322 | ut_ad(err == DB_SUCCESS); |
| 4323 | |
| 4324 | func_exit: |
| 4325 | if (!(flags & BTR_KEEP_IBUF_BITMAP) |
| 4326 | && !dict_index_is_clust(index)) { |
| 4327 | /* Update the free bits in the insert buffer. */ |
| 4328 | if (page_zip) { |
| 4329 | ut_ad(!index->table->is_temporary()); |
| 4330 | ibuf_update_free_bits_zip(block, mtr); |
| 4331 | } else if (!index->table->is_temporary()) { |
| 4332 | ibuf_update_free_bits_low(block, max_ins_size, mtr); |
| 4333 | } |
| 4334 | } |
| 4335 | |
| 4336 | if (err != DB_SUCCESS) { |
| 4337 | /* prefetch siblings of the leaf for the pessimistic |
| 4338 | operation. */ |
| 4339 | btr_cur_prefetch_siblings(block); |
| 4340 | } |
| 4341 | |
| 4342 | return(err); |
| 4343 | } |
| 4344 | |
| 4345 | /*************************************************************//** |
| 4346 | If, in a split, a new supremum record was created as the predecessor of the |
| 4347 | updated record, the supremum record must inherit exactly the locks on the |
| 4348 | updated record. In the split it may have inherited locks from the successor |
| 4349 | of the updated record, which is not correct. This function restores the |
| 4350 | right locks for the new supremum. */ |
| 4351 | static |
| 4352 | void |
| 4353 | btr_cur_pess_upd_restore_supremum( |
| 4354 | /*==============================*/ |
| 4355 | buf_block_t* block, /*!< in: buffer block of rec */ |
| 4356 | const rec_t* rec, /*!< in: updated record */ |
| 4357 | mtr_t* mtr) /*!< in: mtr */ |
| 4358 | { |
| 4359 | page_t* page; |
| 4360 | buf_block_t* prev_block; |
| 4361 | |
| 4362 | page = buf_block_get_frame(block); |
| 4363 | |
| 4364 | if (page_rec_get_next(page_get_infimum_rec(page)) != rec) { |
| 4365 | /* Updated record is not the first user record on its page */ |
| 4366 | |
| 4367 | return; |
| 4368 | } |
| 4369 | |
| 4370 | const ulint prev_page_no = btr_page_get_prev(page, mtr); |
| 4371 | |
| 4372 | const page_id_t page_id(block->page.id.space(), prev_page_no); |
| 4373 | |
| 4374 | ut_ad(prev_page_no != FIL_NULL); |
| 4375 | prev_block = buf_page_get_with_no_latch(page_id, block->page.size, mtr); |
| 4376 | #ifdef UNIV_BTR_DEBUG |
| 4377 | ut_a(btr_page_get_next(prev_block->frame, mtr) |
| 4378 | == page_get_page_no(page)); |
| 4379 | #endif /* UNIV_BTR_DEBUG */ |
| 4380 | |
| 4381 | /* We must already have an x-latch on prev_block! */ |
| 4382 | ut_ad(mtr_memo_contains(mtr, prev_block, MTR_MEMO_PAGE_X_FIX)); |
| 4383 | |
| 4384 | lock_rec_reset_and_inherit_gap_locks(prev_block, block, |
| 4385 | PAGE_HEAP_NO_SUPREMUM, |
| 4386 | page_rec_get_heap_no(rec)); |
| 4387 | } |
| 4388 | |
| 4389 | /*************************************************************//** |
| 4390 | Performs an update of a record on a page of a tree. It is assumed |
| 4391 | that mtr holds an x-latch on the tree and on the cursor page. If the |
| 4392 | update is made on the leaf level, to avoid deadlocks, mtr must also |
| 4393 | own x-latches to brothers of page, if those brothers exist. We assume |
| 4394 | here that the ordering fields of the record do not change. |
| 4395 | @return DB_SUCCESS or error code */ |
| 4396 | dberr_t |
| 4397 | btr_cur_pessimistic_update( |
| 4398 | /*=======================*/ |
| 4399 | ulint flags, /*!< in: undo logging, locking, and rollback |
| 4400 | flags */ |
| 4401 | btr_cur_t* cursor, /*!< in/out: cursor on the record to update; |
| 4402 | cursor may become invalid if *big_rec == NULL |
| 4403 | || !(flags & BTR_KEEP_POS_FLAG) */ |
| 4404 | ulint** offsets,/*!< out: offsets on cursor->page_cur.rec */ |
| 4405 | mem_heap_t** offsets_heap, |
| 4406 | /*!< in/out: pointer to memory heap |
| 4407 | that can be emptied */ |
| 4408 | mem_heap_t* entry_heap, |
| 4409 | /*!< in/out: memory heap for allocating |
| 4410 | big_rec and the index tuple */ |
| 4411 | big_rec_t** big_rec,/*!< out: big rec vector whose fields have to |
| 4412 | be stored externally by the caller */ |
| 4413 | upd_t* update, /*!< in/out: update vector; this is allowed to |
| 4414 | also contain trx id and roll ptr fields. |
| 4415 | Non-updated columns that are moved offpage will |
| 4416 | be appended to this. */ |
| 4417 | ulint cmpl_info,/*!< in: compiler info on secondary index |
| 4418 | updates */ |
| 4419 | que_thr_t* thr, /*!< in: query thread */ |
| 4420 | trx_id_t trx_id, /*!< in: transaction id */ |
| 4421 | mtr_t* mtr) /*!< in/out: mini-transaction; must be |
| 4422 | committed before latching any further pages */ |
| 4423 | { |
| 4424 | big_rec_t* big_rec_vec = NULL; |
| 4425 | big_rec_t* dummy_big_rec; |
| 4426 | dict_index_t* index; |
| 4427 | buf_block_t* block; |
| 4428 | page_t* page; |
| 4429 | page_zip_des_t* page_zip; |
| 4430 | rec_t* rec; |
| 4431 | page_cur_t* page_cursor; |
| 4432 | dberr_t err; |
| 4433 | dberr_t optim_err; |
| 4434 | roll_ptr_t roll_ptr; |
| 4435 | ibool was_first; |
| 4436 | ulint n_reserved = 0; |
| 4437 | ulint n_ext; |
| 4438 | ulint max_ins_size = 0; |
| 4439 | |
| 4440 | *offsets = NULL; |
| 4441 | *big_rec = NULL; |
| 4442 | |
| 4443 | block = btr_cur_get_block(cursor); |
| 4444 | page = buf_block_get_frame(block); |
| 4445 | page_zip = buf_block_get_page_zip(block); |
| 4446 | index = cursor->index; |
| 4447 | |
| 4448 | ut_ad(mtr_memo_contains_flagged(mtr, dict_index_get_lock(index), |
| 4449 | MTR_MEMO_X_LOCK | |
| 4450 | MTR_MEMO_SX_LOCK)); |
| 4451 | ut_ad(mtr_memo_contains(mtr, block, MTR_MEMO_PAGE_X_FIX)); |
| 4452 | #ifdef UNIV_ZIP_DEBUG |
| 4453 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 4454 | #endif /* UNIV_ZIP_DEBUG */ |
| 4455 | ut_ad(!page_zip || !index->table->is_temporary()); |
| 4456 | /* The insert buffer tree should never be updated in place. */ |
| 4457 | ut_ad(!dict_index_is_ibuf(index)); |
| 4458 | ut_ad(trx_id > 0 |
| 4459 | || (flags & BTR_KEEP_SYS_FLAG)); |
| 4460 | ut_ad(dict_index_is_online_ddl(index) == !!(flags & BTR_CREATE_FLAG) |
| 4461 | || dict_index_is_clust(index)); |
| 4462 | ut_ad(thr_get_trx(thr)->id == trx_id |
| 4463 | || (flags & ulint(~BTR_KEEP_POS_FLAG)) |
| 4464 | == (BTR_NO_UNDO_LOG_FLAG | BTR_NO_LOCKING_FLAG |
| 4465 | | BTR_CREATE_FLAG | BTR_KEEP_SYS_FLAG)); |
| 4466 | |
| 4467 | err = optim_err = btr_cur_optimistic_update( |
| 4468 | flags | BTR_KEEP_IBUF_BITMAP, |
| 4469 | cursor, offsets, offsets_heap, update, |
| 4470 | cmpl_info, thr, trx_id, mtr); |
| 4471 | |
| 4472 | switch (err) { |
| 4473 | case DB_ZIP_OVERFLOW: |
| 4474 | case DB_UNDERFLOW: |
| 4475 | case DB_OVERFLOW: |
| 4476 | break; |
| 4477 | default: |
| 4478 | err_exit: |
| 4479 | /* We suppressed this with BTR_KEEP_IBUF_BITMAP. |
| 4480 | For DB_ZIP_OVERFLOW, the IBUF_BITMAP_FREE bits were |
| 4481 | already reset by btr_cur_update_alloc_zip() if the |
| 4482 | page was recompressed. */ |
| 4483 | if (page_zip |
| 4484 | && optim_err != DB_ZIP_OVERFLOW |
| 4485 | && !dict_index_is_clust(index) |
| 4486 | && page_is_leaf(page)) { |
| 4487 | ut_ad(!index->table->is_temporary()); |
| 4488 | ibuf_update_free_bits_zip(block, mtr); |
| 4489 | } |
| 4490 | |
| 4491 | if (big_rec_vec != NULL) { |
| 4492 | dtuple_big_rec_free(big_rec_vec); |
| 4493 | } |
| 4494 | |
| 4495 | return(err); |
| 4496 | } |
| 4497 | |
| 4498 | rec = btr_cur_get_rec(cursor); |
| 4499 | |
| 4500 | *offsets = rec_get_offsets( |
| 4501 | rec, index, *offsets, page_is_leaf(page), |
| 4502 | ULINT_UNDEFINED, offsets_heap); |
| 4503 | |
| 4504 | dtuple_t* new_entry = row_rec_to_index_entry( |
| 4505 | rec, index, *offsets, &n_ext, entry_heap); |
| 4506 | |
| 4507 | /* The page containing the clustered index record |
| 4508 | corresponding to new_entry is latched in mtr. If the |
| 4509 | clustered index record is delete-marked, then its externally |
| 4510 | stored fields cannot have been purged yet, because then the |
| 4511 | purge would also have removed the clustered index record |
| 4512 | itself. Thus the following call is safe. */ |
| 4513 | row_upd_index_replace_new_col_vals_index_pos(new_entry, index, update, |
| 4514 | entry_heap); |
| 4515 | btr_cur_trim(new_entry, index, update, thr); |
| 4516 | |
| 4517 | const bool is_default_row = new_entry->info_bits |
| 4518 | & REC_INFO_MIN_REC_FLAG; |
| 4519 | |
| 4520 | /* We have to set appropriate extern storage bits in the new |
| 4521 | record to be inserted: we have to remember which fields were such */ |
| 4522 | |
| 4523 | ut_ad(!page_is_comp(page) || !rec_get_node_ptr_flag(rec)); |
| 4524 | ut_ad(rec_offs_validate(rec, index, *offsets)); |
| 4525 | n_ext += btr_push_update_extern_fields(new_entry, update, entry_heap); |
| 4526 | |
| 4527 | if ((flags & BTR_NO_UNDO_LOG_FLAG) |
| 4528 | && rec_offs_any_extern(*offsets)) { |
| 4529 | /* We are in a transaction rollback undoing a row |
| 4530 | update: we must free possible externally stored fields |
| 4531 | which got new values in the update, if they are not |
| 4532 | inherited values. They can be inherited if we have |
| 4533 | updated the primary key to another value, and then |
| 4534 | update it back again. */ |
| 4535 | |
| 4536 | ut_ad(big_rec_vec == NULL); |
| 4537 | ut_ad(dict_index_is_clust(index)); |
| 4538 | ut_ad(thr_get_trx(thr)->in_rollback); |
| 4539 | |
| 4540 | DBUG_EXECUTE_IF("ib_blob_update_rollback", DBUG_SUICIDE();); |
| 4541 | |
| 4542 | btr_rec_free_updated_extern_fields( |
| 4543 | index, rec, page_zip, *offsets, update, true, mtr); |
| 4544 | } |
| 4545 | |
| 4546 | if (page_zip_rec_needs_ext( |
| 4547 | rec_get_converted_size(index, new_entry, n_ext), |
| 4548 | page_is_comp(page), |
| 4549 | dict_index_get_n_fields(index), |
| 4550 | block->page.size)) { |
| 4551 | |
| 4552 | big_rec_vec = dtuple_convert_big_rec(index, update, new_entry, &n_ext); |
| 4553 | if (UNIV_UNLIKELY(big_rec_vec == NULL)) { |
| 4554 | |
| 4555 | /* We cannot goto return_after_reservations, |
| 4556 | because we may need to update the |
| 4557 | IBUF_BITMAP_FREE bits, which was suppressed by |
| 4558 | BTR_KEEP_IBUF_BITMAP. */ |
| 4559 | #ifdef UNIV_ZIP_DEBUG |
| 4560 | ut_a(!page_zip |
| 4561 | || page_zip_validate(page_zip, page, index)); |
| 4562 | #endif /* UNIV_ZIP_DEBUG */ |
| 4563 | index->table->space->release_free_extents(n_reserved); |
| 4564 | err = DB_TOO_BIG_RECORD; |
| 4565 | goto err_exit; |
| 4566 | } |
| 4567 | |
| 4568 | ut_ad(page_is_leaf(page)); |
| 4569 | ut_ad(dict_index_is_clust(index)); |
| 4570 | ut_ad(flags & BTR_KEEP_POS_FLAG); |
| 4571 | } |
| 4572 | |
| 4573 | /* Do lock checking and undo logging */ |
| 4574 | err = btr_cur_upd_lock_and_undo(flags, cursor, *offsets, |
| 4575 | update, cmpl_info, |
| 4576 | thr, mtr, &roll_ptr); |
| 4577 | if (err != DB_SUCCESS) { |
| 4578 | goto err_exit; |
| 4579 | } |
| 4580 | |
| 4581 | if (optim_err == DB_OVERFLOW) { |
| 4582 | |
| 4583 | /* First reserve enough free space for the file segments |
| 4584 | of the index tree, so that the update will not fail because |
| 4585 | of lack of space */ |
| 4586 | |
| 4587 | ulint n_extents = cursor->tree_height / 16 + 3; |
| 4588 | |
| 4589 | if (!fsp_reserve_free_extents( |
| 4590 | &n_reserved, index->table->space, n_extents, |
| 4591 | flags & BTR_NO_UNDO_LOG_FLAG |
| 4592 | ? FSP_CLEANING : FSP_NORMAL, |
| 4593 | mtr)) { |
| 4594 | err = DB_OUT_OF_FILE_SPACE; |
| 4595 | goto err_exit; |
| 4596 | } |
| 4597 | } |
| 4598 | |
| 4599 | if (!(flags & BTR_KEEP_SYS_FLAG)) { |
| 4600 | row_upd_index_entry_sys_field(new_entry, index, DATA_ROLL_PTR, |
| 4601 | roll_ptr); |
| 4602 | row_upd_index_entry_sys_field(new_entry, index, DATA_TRX_ID, |
| 4603 | trx_id); |
| 4604 | } |
| 4605 | |
| 4606 | if (!page_zip) { |
| 4607 | max_ins_size = page_get_max_insert_size_after_reorganize( |
| 4608 | page, 1); |
| 4609 | } |
| 4610 | |
| 4611 | if (UNIV_UNLIKELY(is_default_row)) { |
| 4612 | ut_ad(new_entry->info_bits == REC_INFO_DEFAULT_ROW); |
| 4613 | ut_ad(index->is_instant()); |
| 4614 | /* This can be innobase_add_instant_try() performing a |
| 4615 | subsequent instant ADD COLUMN, or its rollback by |
| 4616 | row_undo_mod_clust_low(). */ |
| 4617 | ut_ad(flags & BTR_NO_LOCKING_FLAG); |
| 4618 | } else { |
| 4619 | btr_search_update_hash_on_delete(cursor); |
| 4620 | |
| 4621 | /* Store state of explicit locks on rec on the page |
| 4622 | infimum record, before deleting rec. The page infimum |
| 4623 | acts as a dummy carrier of the locks, taking care also |
| 4624 | of lock releases, before we can move the locks back on |
| 4625 | the actual record. There is a special case: if we are |
| 4626 | inserting on the root page and the insert causes a |
| 4627 | call of btr_root_raise_and_insert. Therefore we cannot |
| 4628 | in the lock system delete the lock structs set on the |
| 4629 | root page even if the root page carries just node |
| 4630 | pointers. */ |
| 4631 | if (!dict_table_is_locking_disabled(index->table)) { |
| 4632 | lock_rec_store_on_page_infimum(block, rec); |
| 4633 | } |
| 4634 | } |
| 4635 | |
| 4636 | #ifdef UNIV_ZIP_DEBUG |
| 4637 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 4638 | #endif /* UNIV_ZIP_DEBUG */ |
| 4639 | page_cursor = btr_cur_get_page_cur(cursor); |
| 4640 | |
| 4641 | page_cur_delete_rec(page_cursor, index, *offsets, mtr); |
| 4642 | |
| 4643 | page_cur_move_to_prev(page_cursor); |
| 4644 | |
| 4645 | rec = btr_cur_insert_if_possible(cursor, new_entry, |
| 4646 | offsets, offsets_heap, n_ext, mtr); |
| 4647 | |
| 4648 | if (rec) { |
| 4649 | page_cursor->rec = rec; |
| 4650 | |
| 4651 | if (UNIV_UNLIKELY(is_default_row)) { |
| 4652 | /* We must empty the PAGE_FREE list, because if this |
| 4653 | was a rollback, the shortened 'default row' record |
| 4654 | would have too many fields, and we would be unable to |
| 4655 | know the size of the freed record. */ |
| 4656 | btr_page_reorganize(page_cursor, index, mtr); |
| 4657 | rec = page_cursor->rec; |
| 4658 | } else if (!dict_table_is_locking_disabled(index->table)) { |
| 4659 | lock_rec_restore_from_page_infimum( |
| 4660 | btr_cur_get_block(cursor), rec, block); |
| 4661 | } |
| 4662 | |
| 4663 | if (!rec_get_deleted_flag(rec, rec_offs_comp(*offsets))) { |
| 4664 | /* The new inserted record owns its possible externally |
| 4665 | stored fields */ |
| 4666 | btr_cur_unmark_extern_fields( |
| 4667 | page_zip, rec, index, *offsets, mtr); |
| 4668 | } else { |
| 4669 | /* In delete-marked records, DB_TRX_ID must |
| 4670 | always refer to an existing undo log record. */ |
| 4671 | ut_ad(row_get_rec_trx_id(rec, index, *offsets)); |
| 4672 | } |
| 4673 | |
| 4674 | bool adjust = big_rec_vec && (flags & BTR_KEEP_POS_FLAG); |
| 4675 | ut_ad(!adjust || page_is_leaf(page)); |
| 4676 | |
| 4677 | if (btr_cur_compress_if_useful(cursor, adjust, mtr)) { |
| 4678 | if (adjust) { |
| 4679 | rec_offs_make_valid(page_cursor->rec, index, |
| 4680 | true, *offsets); |
| 4681 | } |
| 4682 | } else if (!dict_index_is_clust(index) |
| 4683 | && page_is_leaf(page)) { |
| 4684 | /* Update the free bits in the insert buffer. |
| 4685 | This is the same block which was skipped by |
| 4686 | BTR_KEEP_IBUF_BITMAP. */ |
| 4687 | if (page_zip) { |
| 4688 | ut_ad(!index->table->is_temporary()); |
| 4689 | ibuf_update_free_bits_zip(block, mtr); |
| 4690 | } else if (!index->table->is_temporary()) { |
| 4691 | ibuf_update_free_bits_low(block, max_ins_size, |
| 4692 | mtr); |
| 4693 | } |
| 4694 | } |
| 4695 | |
| 4696 | if (!srv_read_only_mode |
| 4697 | && !big_rec_vec |
| 4698 | && page_is_leaf(page) |
| 4699 | && !dict_index_is_online_ddl(index)) { |
| 4700 | |
| 4701 | mtr_memo_release(mtr, dict_index_get_lock(index), |
| 4702 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK); |
| 4703 | |
| 4704 | /* NOTE: We cannot release root block latch here, because it |
| 4705 | has segment header and already modified in most of cases.*/ |
| 4706 | } |
| 4707 | |
| 4708 | err = DB_SUCCESS; |
| 4709 | goto return_after_reservations; |
| 4710 | } else { |
| 4711 | /* If the page is compressed and it initially |
| 4712 | compresses very well, and there is a subsequent insert |
| 4713 | of a badly-compressing record, it is possible for |
| 4714 | btr_cur_optimistic_update() to return DB_UNDERFLOW and |
| 4715 | btr_cur_insert_if_possible() to return FALSE. */ |
| 4716 | ut_a(page_zip || optim_err != DB_UNDERFLOW); |
| 4717 | |
| 4718 | /* Out of space: reset the free bits. |
| 4719 | This is the same block which was skipped by |
| 4720 | BTR_KEEP_IBUF_BITMAP. */ |
| 4721 | if (!dict_index_is_clust(index) |
| 4722 | && !index->table->is_temporary() |
| 4723 | && page_is_leaf(page)) { |
| 4724 | ibuf_reset_free_bits(block); |
| 4725 | } |
| 4726 | } |
| 4727 | |
| 4728 | if (big_rec_vec != NULL) { |
| 4729 | ut_ad(page_is_leaf(page)); |
| 4730 | ut_ad(dict_index_is_clust(index)); |
| 4731 | ut_ad(flags & BTR_KEEP_POS_FLAG); |
| 4732 | |
| 4733 | /* btr_page_split_and_insert() in |
| 4734 | btr_cur_pessimistic_insert() invokes |
| 4735 | mtr_memo_release(mtr, index->lock, MTR_MEMO_SX_LOCK). |
| 4736 | We must keep the index->lock when we created a |
| 4737 | big_rec, so that row_upd_clust_rec() can store the |
| 4738 | big_rec in the same mini-transaction. */ |
| 4739 | |
| 4740 | ut_ad(mtr_memo_contains_flagged(mtr, |
| 4741 | dict_index_get_lock(index), |
| 4742 | MTR_MEMO_X_LOCK | |
| 4743 | MTR_MEMO_SX_LOCK)); |
| 4744 | |
| 4745 | mtr_sx_lock(dict_index_get_lock(index), mtr); |
| 4746 | } |
| 4747 | |
| 4748 | /* Was the record to be updated positioned as the first user |
| 4749 | record on its page? */ |
| 4750 | was_first = page_cur_is_before_first(page_cursor); |
| 4751 | |
| 4752 | /* Lock checks and undo logging were already performed by |
| 4753 | btr_cur_upd_lock_and_undo(). We do not try |
| 4754 | btr_cur_optimistic_insert() because |
| 4755 | btr_cur_insert_if_possible() already failed above. */ |
| 4756 | |
| 4757 | err = btr_cur_pessimistic_insert(BTR_NO_UNDO_LOG_FLAG |
| 4758 | | BTR_NO_LOCKING_FLAG |
| 4759 | | BTR_KEEP_SYS_FLAG, |
| 4760 | cursor, offsets, offsets_heap, |
| 4761 | new_entry, &rec, |
| 4762 | &dummy_big_rec, n_ext, NULL, mtr); |
| 4763 | ut_a(rec); |
| 4764 | ut_a(err == DB_SUCCESS); |
| 4765 | ut_a(dummy_big_rec == NULL); |
| 4766 | ut_ad(rec_offs_validate(rec, cursor->index, *offsets)); |
| 4767 | page_cursor->rec = rec; |
| 4768 | |
| 4769 | /* Multiple transactions cannot simultaneously operate on the |
| 4770 | same temp-table in parallel. |
| 4771 | max_trx_id is ignored for temp tables because it not required |
| 4772 | for MVCC. */ |
| 4773 | if (dict_index_is_sec_or_ibuf(index) |
| 4774 | && !index->table->is_temporary()) { |
| 4775 | /* Update PAGE_MAX_TRX_ID in the index page header. |
| 4776 | It was not updated by btr_cur_pessimistic_insert() |
| 4777 | because of BTR_NO_LOCKING_FLAG. */ |
| 4778 | buf_block_t* rec_block; |
| 4779 | |
| 4780 | rec_block = btr_cur_get_block(cursor); |
| 4781 | |
| 4782 | page_update_max_trx_id(rec_block, |
| 4783 | buf_block_get_page_zip(rec_block), |
| 4784 | trx_id, mtr); |
| 4785 | } |
| 4786 | |
| 4787 | if (!rec_get_deleted_flag(rec, rec_offs_comp(*offsets))) { |
| 4788 | /* The new inserted record owns its possible externally |
| 4789 | stored fields */ |
| 4790 | buf_block_t* rec_block = btr_cur_get_block(cursor); |
| 4791 | |
| 4792 | #ifdef UNIV_ZIP_DEBUG |
| 4793 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 4794 | page = buf_block_get_frame(rec_block); |
| 4795 | #endif /* UNIV_ZIP_DEBUG */ |
| 4796 | page_zip = buf_block_get_page_zip(rec_block); |
| 4797 | |
| 4798 | btr_cur_unmark_extern_fields(page_zip, |
| 4799 | rec, index, *offsets, mtr); |
| 4800 | } else { |
| 4801 | /* In delete-marked records, DB_TRX_ID must |
| 4802 | always refer to an existing undo log record. */ |
| 4803 | ut_ad(row_get_rec_trx_id(rec, index, *offsets)); |
| 4804 | } |
| 4805 | |
| 4806 | if (UNIV_UNLIKELY(is_default_row)) { |
| 4807 | /* We must empty the PAGE_FREE list, because if this |
| 4808 | was a rollback, the shortened 'default row' record |
| 4809 | would have too many fields, and we would be unable to |
| 4810 | know the size of the freed record. */ |
| 4811 | btr_page_reorganize(page_cursor, index, mtr); |
| 4812 | rec = page_cursor->rec; |
| 4813 | } else if (!dict_table_is_locking_disabled(index->table)) { |
| 4814 | lock_rec_restore_from_page_infimum( |
| 4815 | btr_cur_get_block(cursor), rec, block); |
| 4816 | } |
| 4817 | |
| 4818 | /* If necessary, restore also the correct lock state for a new, |
| 4819 | preceding supremum record created in a page split. While the old |
| 4820 | record was nonexistent, the supremum might have inherited its locks |
| 4821 | from a wrong record. */ |
| 4822 | |
| 4823 | if (!was_first && !dict_table_is_locking_disabled(index->table)) { |
| 4824 | btr_cur_pess_upd_restore_supremum(btr_cur_get_block(cursor), |
| 4825 | rec, mtr); |
| 4826 | } |
| 4827 | |
| 4828 | return_after_reservations: |
| 4829 | #ifdef UNIV_ZIP_DEBUG |
| 4830 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 4831 | #endif /* UNIV_ZIP_DEBUG */ |
| 4832 | |
| 4833 | index->table->space->release_free_extents(n_reserved); |
| 4834 | *big_rec = big_rec_vec; |
| 4835 | return(err); |
| 4836 | } |
| 4837 | |
| 4838 | /*==================== B-TREE DELETE MARK AND UNMARK ===============*/ |
| 4839 | |
| 4840 | /****************************************************************//** |
| 4841 | Writes the redo log record for delete marking or unmarking of an index |
| 4842 | record. */ |
| 4843 | UNIV_INLINE |
| 4844 | void |
| 4845 | btr_cur_del_mark_set_clust_rec_log( |
| 4846 | /*===============================*/ |
| 4847 | rec_t* rec, /*!< in: record */ |
| 4848 | dict_index_t* index, /*!< in: index of the record */ |
| 4849 | trx_id_t trx_id, /*!< in: transaction id */ |
| 4850 | roll_ptr_t roll_ptr,/*!< in: roll ptr to the undo log record */ |
| 4851 | mtr_t* mtr) /*!< in: mtr */ |
| 4852 | { |
| 4853 | byte* log_ptr; |
| 4854 | |
| 4855 | ut_ad(!!page_rec_is_comp(rec) == dict_table_is_comp(index->table)); |
| 4856 | ut_ad(mtr->is_named_space(index->table->space)); |
| 4857 | |
| 4858 | log_ptr = mlog_open_and_write_index(mtr, rec, index, |
| 4859 | page_rec_is_comp(rec) |
| 4860 | ? MLOG_COMP_REC_CLUST_DELETE_MARK |
| 4861 | : MLOG_REC_CLUST_DELETE_MARK, |
| 4862 | 1 + 1 + DATA_ROLL_PTR_LEN |
| 4863 | + 14 + 2); |
| 4864 | |
| 4865 | if (!log_ptr) { |
| 4866 | /* Logging in mtr is switched off during crash recovery */ |
| 4867 | return; |
| 4868 | } |
| 4869 | |
| 4870 | *log_ptr++ = 0; |
| 4871 | *log_ptr++ = 1; |
| 4872 | |
| 4873 | log_ptr = row_upd_write_sys_vals_to_log( |
| 4874 | index, trx_id, roll_ptr, log_ptr, mtr); |
| 4875 | mach_write_to_2(log_ptr, page_offset(rec)); |
| 4876 | log_ptr += 2; |
| 4877 | |
| 4878 | mlog_close(mtr, log_ptr); |
| 4879 | } |
| 4880 | |
| 4881 | /****************************************************************//** |
| 4882 | Parses the redo log record for delete marking or unmarking of a clustered |
| 4883 | index record. |
| 4884 | @return end of log record or NULL */ |
| 4885 | byte* |
| 4886 | btr_cur_parse_del_mark_set_clust_rec( |
| 4887 | /*=================================*/ |
| 4888 | byte* ptr, /*!< in: buffer */ |
| 4889 | byte* end_ptr,/*!< in: buffer end */ |
| 4890 | page_t* page, /*!< in/out: page or NULL */ |
| 4891 | page_zip_des_t* page_zip,/*!< in/out: compressed page, or NULL */ |
| 4892 | dict_index_t* index) /*!< in: index corresponding to page */ |
| 4893 | { |
| 4894 | ulint flags; |
| 4895 | ulint val; |
| 4896 | ulint pos; |
| 4897 | trx_id_t trx_id; |
| 4898 | roll_ptr_t roll_ptr; |
| 4899 | ulint offset; |
| 4900 | rec_t* rec; |
| 4901 | |
| 4902 | ut_ad(!page |
| 4903 | || !!page_is_comp(page) == dict_table_is_comp(index->table)); |
| 4904 | |
| 4905 | if (end_ptr < ptr + 2) { |
| 4906 | |
| 4907 | return(NULL); |
| 4908 | } |
| 4909 | |
| 4910 | flags = mach_read_from_1(ptr); |
| 4911 | ptr++; |
| 4912 | val = mach_read_from_1(ptr); |
| 4913 | ptr++; |
| 4914 | |
| 4915 | ptr = row_upd_parse_sys_vals(ptr, end_ptr, &pos, &trx_id, &roll_ptr); |
| 4916 | |
| 4917 | if (ptr == NULL) { |
| 4918 | |
| 4919 | return(NULL); |
| 4920 | } |
| 4921 | |
| 4922 | if (end_ptr < ptr + 2) { |
| 4923 | |
| 4924 | return(NULL); |
| 4925 | } |
| 4926 | |
| 4927 | offset = mach_read_from_2(ptr); |
| 4928 | ptr += 2; |
| 4929 | |
| 4930 | ut_a(offset <= srv_page_size); |
| 4931 | |
| 4932 | /* In delete-marked records, DB_TRX_ID must |
| 4933 | always refer to an existing undo log record. */ |
| 4934 | ut_ad(trx_id || (flags & BTR_KEEP_SYS_FLAG)); |
| 4935 | |
| 4936 | if (page) { |
| 4937 | rec = page + offset; |
| 4938 | |
| 4939 | /* We do not need to reserve search latch, as the page |
| 4940 | is only being recovered, and there cannot be a hash index to |
| 4941 | it. Besides, these fields are being updated in place |
| 4942 | and the adaptive hash index does not depend on them. */ |
| 4943 | |
| 4944 | btr_rec_set_deleted_flag(rec, page_zip, val); |
| 4945 | /* pos is the offset of DB_TRX_ID in the clustered index. |
| 4946 | Debug assertions may also access DB_ROLL_PTR at pos+1. |
| 4947 | Therefore, we must compute offsets for the first pos+2 |
| 4948 | clustered index fields. */ |
| 4949 | ut_ad(pos <= MAX_REF_PARTS); |
| 4950 | |
| 4951 | ulint offsets[REC_OFFS_HEADER_SIZE + MAX_REF_PARTS + 2]; |
| 4952 | rec_offs_init(offsets); |
| 4953 | mem_heap_t* heap = NULL; |
| 4954 | |
| 4955 | if (!(flags & BTR_KEEP_SYS_FLAG)) { |
| 4956 | row_upd_rec_sys_fields_in_recovery( |
| 4957 | rec, page_zip, |
| 4958 | rec_get_offsets(rec, index, offsets, true, |
| 4959 | pos + 2, &heap), |
| 4960 | pos, trx_id, roll_ptr); |
| 4961 | } else { |
| 4962 | /* In delete-marked records, DB_TRX_ID must |
| 4963 | always refer to an existing undo log record. */ |
| 4964 | ut_ad(memcmp(rec_get_nth_field( |
| 4965 | rec, |
| 4966 | rec_get_offsets(rec, index, |
| 4967 | offsets, true, |
| 4968 | pos, &heap), |
| 4969 | pos, &offset), |
| 4970 | field_ref_zero, DATA_TRX_ID_LEN)); |
| 4971 | ut_ad(offset == DATA_TRX_ID_LEN); |
| 4972 | } |
| 4973 | |
| 4974 | if (UNIV_LIKELY_NULL(heap)) { |
| 4975 | mem_heap_free(heap); |
| 4976 | } |
| 4977 | } |
| 4978 | |
| 4979 | return(ptr); |
| 4980 | } |
| 4981 | |
| 4982 | /***********************************************************//** |
| 4983 | Marks a clustered index record deleted. Writes an undo log record to |
| 4984 | undo log on this delete marking. Writes in the trx id field the id |
| 4985 | of the deleting transaction, and in the roll ptr field pointer to the |
| 4986 | undo log record created. |
| 4987 | @return DB_SUCCESS, DB_LOCK_WAIT, or error number */ |
| 4988 | dberr_t |
| 4989 | btr_cur_del_mark_set_clust_rec( |
| 4990 | /*===========================*/ |
| 4991 | buf_block_t* block, /*!< in/out: buffer block of the record */ |
| 4992 | rec_t* rec, /*!< in/out: record */ |
| 4993 | dict_index_t* index, /*!< in: clustered index of the record */ |
| 4994 | const ulint* offsets,/*!< in: rec_get_offsets(rec) */ |
| 4995 | que_thr_t* thr, /*!< in: query thread */ |
| 4996 | const dtuple_t* entry, /*!< in: dtuple for the deleting record, also |
| 4997 | contains the virtual cols if there are any */ |
| 4998 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 4999 | { |
| 5000 | roll_ptr_t roll_ptr; |
| 5001 | dberr_t err; |
| 5002 | page_zip_des_t* page_zip; |
| 5003 | trx_t* trx; |
| 5004 | |
| 5005 | ut_ad(dict_index_is_clust(index)); |
| 5006 | ut_ad(rec_offs_validate(rec, index, offsets)); |
| 5007 | ut_ad(!!page_rec_is_comp(rec) == dict_table_is_comp(index->table)); |
| 5008 | ut_ad(buf_block_get_frame(block) == page_align(rec)); |
| 5009 | ut_ad(page_rec_is_leaf(rec)); |
| 5010 | ut_ad(mtr->is_named_space(index->table->space)); |
| 5011 | |
| 5012 | if (rec_get_deleted_flag(rec, rec_offs_comp(offsets))) { |
| 5013 | /* We may already have delete-marked this record |
| 5014 | when executing an ON DELETE CASCADE operation. */ |
| 5015 | ut_ad(row_get_rec_trx_id(rec, index, offsets) |
| 5016 | == thr_get_trx(thr)->id); |
| 5017 | return(DB_SUCCESS); |
| 5018 | } |
| 5019 | |
| 5020 | err = lock_clust_rec_modify_check_and_lock(BTR_NO_LOCKING_FLAG, block, |
| 5021 | rec, index, offsets, thr); |
| 5022 | |
| 5023 | if (err != DB_SUCCESS) { |
| 5024 | |
| 5025 | return(err); |
| 5026 | } |
| 5027 | |
| 5028 | err = trx_undo_report_row_operation(thr, index, |
| 5029 | entry, NULL, 0, rec, offsets, |
| 5030 | &roll_ptr); |
| 5031 | if (err != DB_SUCCESS) { |
| 5032 | |
| 5033 | return(err); |
| 5034 | } |
| 5035 | |
| 5036 | /* The search latch is not needed here, because |
| 5037 | the adaptive hash index does not depend on the delete-mark |
| 5038 | and the delete-mark is being updated in place. */ |
| 5039 | |
| 5040 | page_zip = buf_block_get_page_zip(block); |
| 5041 | |
| 5042 | btr_rec_set_deleted_flag(rec, page_zip, TRUE); |
| 5043 | |
| 5044 | trx = thr_get_trx(thr); |
| 5045 | |
| 5046 | DBUG_LOG("ib_cur", |
| 5047 | "delete-mark clust "<< index->table->name |
| 5048 | << " ("<< index->id << ") by " |
| 5049 | << ib::hex(trx_get_id_for_print(trx)) << ": " |
| 5050 | << rec_printer(rec, offsets).str()); |
| 5051 | |
| 5052 | if (dict_index_is_online_ddl(index)) { |
| 5053 | row_log_table_delete(rec, index, offsets, NULL); |
| 5054 | } |
| 5055 | |
| 5056 | row_upd_rec_sys_fields(rec, page_zip, index, offsets, trx, roll_ptr); |
| 5057 | |
| 5058 | btr_cur_del_mark_set_clust_rec_log(rec, index, trx->id, |
| 5059 | roll_ptr, mtr); |
| 5060 | |
| 5061 | return(err); |
| 5062 | } |
| 5063 | |
| 5064 | /****************************************************************//** |
| 5065 | Writes the redo log record for a delete mark setting of a secondary |
| 5066 | index record. */ |
| 5067 | UNIV_INLINE |
| 5068 | void |
| 5069 | btr_cur_del_mark_set_sec_rec_log( |
| 5070 | /*=============================*/ |
| 5071 | rec_t* rec, /*!< in: record */ |
| 5072 | ibool val, /*!< in: value to set */ |
| 5073 | mtr_t* mtr) /*!< in: mtr */ |
| 5074 | { |
| 5075 | byte* log_ptr; |
| 5076 | ut_ad(val <= 1); |
| 5077 | |
| 5078 | log_ptr = mlog_open(mtr, 11 + 1 + 2); |
| 5079 | |
| 5080 | if (!log_ptr) { |
| 5081 | /* Logging in mtr is switched off during crash recovery: |
| 5082 | in that case mlog_open returns NULL */ |
| 5083 | return; |
| 5084 | } |
| 5085 | |
| 5086 | log_ptr = mlog_write_initial_log_record_fast( |
| 5087 | rec, MLOG_REC_SEC_DELETE_MARK, log_ptr, mtr); |
| 5088 | mach_write_to_1(log_ptr, val); |
| 5089 | log_ptr++; |
| 5090 | |
| 5091 | mach_write_to_2(log_ptr, page_offset(rec)); |
| 5092 | log_ptr += 2; |
| 5093 | |
| 5094 | mlog_close(mtr, log_ptr); |
| 5095 | } |
| 5096 | |
| 5097 | /****************************************************************//** |
| 5098 | Parses the redo log record for delete marking or unmarking of a secondary |
| 5099 | index record. |
| 5100 | @return end of log record or NULL */ |
| 5101 | byte* |
| 5102 | btr_cur_parse_del_mark_set_sec_rec( |
| 5103 | /*===============================*/ |
| 5104 | byte* ptr, /*!< in: buffer */ |
| 5105 | byte* end_ptr,/*!< in: buffer end */ |
| 5106 | page_t* page, /*!< in/out: page or NULL */ |
| 5107 | page_zip_des_t* page_zip)/*!< in/out: compressed page, or NULL */ |
| 5108 | { |
| 5109 | ulint val; |
| 5110 | ulint offset; |
| 5111 | rec_t* rec; |
| 5112 | |
| 5113 | if (end_ptr < ptr + 3) { |
| 5114 | |
| 5115 | return(NULL); |
| 5116 | } |
| 5117 | |
| 5118 | val = mach_read_from_1(ptr); |
| 5119 | ptr++; |
| 5120 | |
| 5121 | offset = mach_read_from_2(ptr); |
| 5122 | ptr += 2; |
| 5123 | |
| 5124 | ut_a(offset <= srv_page_size); |
| 5125 | |
| 5126 | if (page) { |
| 5127 | rec = page + offset; |
| 5128 | |
| 5129 | /* We do not need to reserve search latch, as the page |
| 5130 | is only being recovered, and there cannot be a hash index to |
| 5131 | it. Besides, the delete-mark flag is being updated in place |
| 5132 | and the adaptive hash index does not depend on it. */ |
| 5133 | |
| 5134 | btr_rec_set_deleted_flag(rec, page_zip, val); |
| 5135 | } |
| 5136 | |
| 5137 | return(ptr); |
| 5138 | } |
| 5139 | |
| 5140 | /***********************************************************//** |
| 5141 | Sets a secondary index record delete mark to TRUE or FALSE. |
| 5142 | @return DB_SUCCESS, DB_LOCK_WAIT, or error number */ |
| 5143 | dberr_t |
| 5144 | btr_cur_del_mark_set_sec_rec( |
| 5145 | /*=========================*/ |
| 5146 | ulint flags, /*!< in: locking flag */ |
| 5147 | btr_cur_t* cursor, /*!< in: cursor */ |
| 5148 | ibool val, /*!< in: value to set */ |
| 5149 | que_thr_t* thr, /*!< in: query thread */ |
| 5150 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 5151 | { |
| 5152 | buf_block_t* block; |
| 5153 | rec_t* rec; |
| 5154 | dberr_t err; |
| 5155 | |
| 5156 | block = btr_cur_get_block(cursor); |
| 5157 | rec = btr_cur_get_rec(cursor); |
| 5158 | |
| 5159 | err = lock_sec_rec_modify_check_and_lock(flags, |
| 5160 | btr_cur_get_block(cursor), |
| 5161 | rec, cursor->index, thr, mtr); |
| 5162 | if (err != DB_SUCCESS) { |
| 5163 | |
| 5164 | return(err); |
| 5165 | } |
| 5166 | |
| 5167 | ut_ad(!!page_rec_is_comp(rec) |
| 5168 | == dict_table_is_comp(cursor->index->table)); |
| 5169 | |
| 5170 | DBUG_PRINT("ib_cur", ( "delete-mark=%u sec %u:%u:%u in %s(" |
| 5171 | IB_ID_FMT ") by "TRX_ID_FMT, |
| 5172 | unsigned(val), |
| 5173 | block->page.id.space(), block->page.id.page_no(), |
| 5174 | unsigned(page_rec_get_heap_no(rec)), |
| 5175 | cursor->index->name(), cursor->index->id, |
| 5176 | trx_get_id_for_print(thr_get_trx(thr)))); |
| 5177 | |
| 5178 | /* We do not need to reserve search latch, as the |
| 5179 | delete-mark flag is being updated in place and the adaptive |
| 5180 | hash index does not depend on it. */ |
| 5181 | btr_rec_set_deleted_flag(rec, buf_block_get_page_zip(block), val); |
| 5182 | |
| 5183 | btr_cur_del_mark_set_sec_rec_log(rec, val, mtr); |
| 5184 | |
| 5185 | return(DB_SUCCESS); |
| 5186 | } |
| 5187 | |
| 5188 | /***********************************************************//** |
| 5189 | Sets a secondary index record's delete mark to the given value. This |
| 5190 | function is only used by the insert buffer merge mechanism. */ |
| 5191 | void |
| 5192 | btr_cur_set_deleted_flag_for_ibuf( |
| 5193 | /*==============================*/ |
| 5194 | rec_t* rec, /*!< in/out: record */ |
| 5195 | page_zip_des_t* page_zip, /*!< in/out: compressed page |
| 5196 | corresponding to rec, or NULL |
| 5197 | when the tablespace is |
| 5198 | uncompressed */ |
| 5199 | ibool val, /*!< in: value to set */ |
| 5200 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 5201 | { |
| 5202 | /* We do not need to reserve search latch, as the page |
| 5203 | has just been read to the buffer pool and there cannot be |
| 5204 | a hash index to it. Besides, the delete-mark flag is being |
| 5205 | updated in place and the adaptive hash index does not depend |
| 5206 | on it. */ |
| 5207 | |
| 5208 | btr_rec_set_deleted_flag(rec, page_zip, val); |
| 5209 | |
| 5210 | btr_cur_del_mark_set_sec_rec_log(rec, val, mtr); |
| 5211 | } |
| 5212 | |
| 5213 | /*==================== B-TREE RECORD REMOVE =========================*/ |
| 5214 | |
| 5215 | /*************************************************************//** |
| 5216 | Tries to compress a page of the tree if it seems useful. It is assumed |
| 5217 | that mtr holds an x-latch on the tree and on the cursor page. To avoid |
| 5218 | deadlocks, mtr must also own x-latches to brothers of page, if those |
| 5219 | brothers exist. NOTE: it is assumed that the caller has reserved enough |
| 5220 | free extents so that the compression will always succeed if done! |
| 5221 | @return TRUE if compression occurred */ |
| 5222 | ibool |
| 5223 | btr_cur_compress_if_useful( |
| 5224 | /*=======================*/ |
| 5225 | btr_cur_t* cursor, /*!< in/out: cursor on the page to compress; |
| 5226 | cursor does not stay valid if !adjust and |
| 5227 | compression occurs */ |
| 5228 | ibool adjust, /*!< in: TRUE if should adjust the |
| 5229 | cursor position even if compression occurs */ |
| 5230 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 5231 | { |
| 5232 | ut_ad(mtr_memo_contains_flagged( |
| 5233 | mtr, dict_index_get_lock(btr_cur_get_index(cursor)), |
| 5234 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK)); |
| 5235 | ut_ad(mtr_memo_contains(mtr, btr_cur_get_block(cursor), |
| 5236 | MTR_MEMO_PAGE_X_FIX)); |
| 5237 | |
| 5238 | if (dict_index_is_spatial(cursor->index)) { |
| 5239 | const page_t* page = btr_cur_get_page(cursor); |
| 5240 | const trx_t* trx = NULL; |
| 5241 | |
| 5242 | if (cursor->rtr_info->thr != NULL) { |
| 5243 | trx = thr_get_trx(cursor->rtr_info->thr); |
| 5244 | } |
| 5245 | |
| 5246 | /* Check whether page lock prevents the compression */ |
| 5247 | if (!lock_test_prdt_page_lock(trx, page_get_space_id(page), |
| 5248 | page_get_page_no(page))) { |
| 5249 | return(false); |
| 5250 | } |
| 5251 | } |
| 5252 | |
| 5253 | return(btr_cur_compress_recommendation(cursor, mtr) |
| 5254 | && btr_compress(cursor, adjust, mtr)); |
| 5255 | } |
| 5256 | |
| 5257 | /*******************************************************//** |
| 5258 | Removes the record on which the tree cursor is positioned on a leaf page. |
| 5259 | It is assumed that the mtr has an x-latch on the page where the cursor is |
| 5260 | positioned, but no latch on the whole tree. |
| 5261 | @return TRUE if success, i.e., the page did not become too empty */ |
| 5262 | ibool |
| 5263 | btr_cur_optimistic_delete_func( |
| 5264 | /*===========================*/ |
| 5265 | btr_cur_t* cursor, /*!< in: cursor on leaf page, on the record to |
| 5266 | delete; cursor stays valid: if deletion |
| 5267 | succeeds, on function exit it points to the |
| 5268 | successor of the deleted record */ |
| 5269 | #ifdef UNIV_DEBUG |
| 5270 | ulint flags, /*!< in: BTR_CREATE_FLAG or 0 */ |
| 5271 | #endif /* UNIV_DEBUG */ |
| 5272 | mtr_t* mtr) /*!< in: mtr; if this function returns |
| 5273 | TRUE on a leaf page of a secondary |
| 5274 | index, the mtr must be committed |
| 5275 | before latching any further pages */ |
| 5276 | { |
| 5277 | buf_block_t* block; |
| 5278 | rec_t* rec; |
| 5279 | mem_heap_t* heap = NULL; |
| 5280 | ulint offsets_[REC_OFFS_NORMAL_SIZE]; |
| 5281 | ulint* offsets = offsets_; |
| 5282 | ibool no_compress_needed; |
| 5283 | rec_offs_init(offsets_); |
| 5284 | |
| 5285 | ut_ad(flags == 0 || flags == BTR_CREATE_FLAG); |
| 5286 | ut_ad(mtr_memo_contains(mtr, btr_cur_get_block(cursor), |
| 5287 | MTR_MEMO_PAGE_X_FIX)); |
| 5288 | ut_ad(mtr_memo_contains(mtr, btr_cur_get_block(cursor), |
| 5289 | MTR_MEMO_PAGE_X_FIX)); |
| 5290 | ut_ad(mtr->is_named_space(cursor->index->table->space)); |
| 5291 | |
| 5292 | /* This is intended only for leaf page deletions */ |
| 5293 | |
| 5294 | block = btr_cur_get_block(cursor); |
| 5295 | |
| 5296 | ut_ad(page_is_leaf(buf_block_get_frame(block))); |
| 5297 | ut_ad(!dict_index_is_online_ddl(cursor->index) |
| 5298 | || dict_index_is_clust(cursor->index) |
| 5299 | || (flags & BTR_CREATE_FLAG)); |
| 5300 | |
| 5301 | rec = btr_cur_get_rec(cursor); |
| 5302 | |
| 5303 | if (UNIV_UNLIKELY(page_is_root(block->frame) |
| 5304 | && page_get_n_recs(block->frame) == 1 |
| 5305 | + (cursor->index->is_instant() |
| 5306 | && !rec_is_default_row(rec, cursor->index)))) { |
| 5307 | /* The whole index (and table) becomes logically empty. |
| 5308 | Empty the whole page. That is, if we are deleting the |
| 5309 | only user record, also delete the 'default row' record |
| 5310 | if one exists (it exists if and only if is_instant()). |
| 5311 | If we are deleting the 'default row' record and the |
| 5312 | table becomes empty, clean up the whole page. */ |
| 5313 | dict_index_t* index = cursor->index; |
| 5314 | ut_ad(!index->is_instant() |
| 5315 | || rec_is_default_row( |
| 5316 | page_rec_get_next_const( |
| 5317 | page_get_infimum_rec(block->frame)), |
| 5318 | index)); |
| 5319 | if (UNIV_UNLIKELY(rec_get_info_bits(rec, page_rec_is_comp(rec)) |
| 5320 | & REC_INFO_MIN_REC_FLAG)) { |
| 5321 | /* This should be rolling back instant ADD COLUMN. |
| 5322 | If this is a recovered transaction, then |
| 5323 | index->is_instant() will hold until the |
| 5324 | insert into SYS_COLUMNS is rolled back. */ |
| 5325 | ut_ad(index->table->supports_instant()); |
| 5326 | ut_ad(index->is_primary()); |
| 5327 | } else { |
| 5328 | lock_update_delete(block, rec); |
| 5329 | } |
| 5330 | btr_page_empty(block, buf_block_get_page_zip(block), |
| 5331 | index, 0, mtr); |
| 5332 | page_cur_set_after_last(block, btr_cur_get_page_cur(cursor)); |
| 5333 | |
| 5334 | if (index->is_primary()) { |
| 5335 | /* Concurrent access is prevented by |
| 5336 | root_block->lock X-latch, so this should be |
| 5337 | safe. */ |
| 5338 | index->remove_instant(); |
| 5339 | } |
| 5340 | |
| 5341 | return true; |
| 5342 | } |
| 5343 | |
| 5344 | offsets = rec_get_offsets(rec, cursor->index, offsets, true, |
| 5345 | ULINT_UNDEFINED, &heap); |
| 5346 | |
| 5347 | no_compress_needed = !rec_offs_any_extern(offsets) |
| 5348 | && btr_cur_can_delete_without_compress( |
| 5349 | cursor, rec_offs_size(offsets), mtr); |
| 5350 | |
| 5351 | if (no_compress_needed) { |
| 5352 | |
| 5353 | page_t* page = buf_block_get_frame(block); |
| 5354 | page_zip_des_t* page_zip= buf_block_get_page_zip(block); |
| 5355 | |
| 5356 | if (UNIV_UNLIKELY(rec_get_info_bits(rec, page_rec_is_comp(rec)) |
| 5357 | & REC_INFO_MIN_REC_FLAG)) { |
| 5358 | /* This should be rolling back instant ADD COLUMN. |
| 5359 | If this is a recovered transaction, then |
| 5360 | index->is_instant() will hold until the |
| 5361 | insert into SYS_COLUMNS is rolled back. */ |
| 5362 | ut_ad(cursor->index->table->supports_instant()); |
| 5363 | ut_ad(cursor->index->is_primary()); |
| 5364 | ut_ad(!page_zip); |
| 5365 | page_cur_delete_rec(btr_cur_get_page_cur(cursor), |
| 5366 | cursor->index, offsets, mtr); |
| 5367 | /* We must empty the PAGE_FREE list, because |
| 5368 | after rollback, this deleted 'default row' record |
| 5369 | would have too many fields, and we would be |
| 5370 | unable to know the size of the freed record. */ |
| 5371 | btr_page_reorganize(btr_cur_get_page_cur(cursor), |
| 5372 | cursor->index, mtr); |
| 5373 | goto func_exit; |
| 5374 | } else { |
| 5375 | lock_update_delete(block, rec); |
| 5376 | |
| 5377 | btr_search_update_hash_on_delete(cursor); |
| 5378 | } |
| 5379 | |
| 5380 | if (page_zip) { |
| 5381 | #ifdef UNIV_ZIP_DEBUG |
| 5382 | ut_a(page_zip_validate(page_zip, page, cursor->index)); |
| 5383 | #endif /* UNIV_ZIP_DEBUG */ |
| 5384 | page_cur_delete_rec(btr_cur_get_page_cur(cursor), |
| 5385 | cursor->index, offsets, mtr); |
| 5386 | #ifdef UNIV_ZIP_DEBUG |
| 5387 | ut_a(page_zip_validate(page_zip, page, cursor->index)); |
| 5388 | #endif /* UNIV_ZIP_DEBUG */ |
| 5389 | |
| 5390 | /* On compressed pages, the IBUF_BITMAP_FREE |
| 5391 | space is not affected by deleting (purging) |
| 5392 | records, because it is defined as the minimum |
| 5393 | of space available *without* reorganize, and |
| 5394 | space available in the modification log. */ |
| 5395 | } else { |
| 5396 | const ulint max_ins |
| 5397 | = page_get_max_insert_size_after_reorganize( |
| 5398 | page, 1); |
| 5399 | |
| 5400 | page_cur_delete_rec(btr_cur_get_page_cur(cursor), |
| 5401 | cursor->index, offsets, mtr); |
| 5402 | |
| 5403 | /* The change buffer does not handle inserts |
| 5404 | into non-leaf pages, into clustered indexes, |
| 5405 | or into the change buffer. */ |
| 5406 | if (!dict_index_is_clust(cursor->index) |
| 5407 | && !cursor->index->table->is_temporary() |
| 5408 | && !dict_index_is_ibuf(cursor->index)) { |
| 5409 | ibuf_update_free_bits_low(block, max_ins, mtr); |
| 5410 | } |
| 5411 | } |
| 5412 | } else { |
| 5413 | /* prefetch siblings of the leaf for the pessimistic |
| 5414 | operation. */ |
| 5415 | btr_cur_prefetch_siblings(block); |
| 5416 | } |
| 5417 | |
| 5418 | func_exit: |
| 5419 | if (UNIV_LIKELY_NULL(heap)) { |
| 5420 | mem_heap_free(heap); |
| 5421 | } |
| 5422 | |
| 5423 | return(no_compress_needed); |
| 5424 | } |
| 5425 | |
| 5426 | /*************************************************************//** |
| 5427 | Removes the record on which the tree cursor is positioned. Tries |
| 5428 | to compress the page if its fillfactor drops below a threshold |
| 5429 | or if it is the only page on the level. It is assumed that mtr holds |
| 5430 | an x-latch on the tree and on the cursor page. To avoid deadlocks, |
| 5431 | mtr must also own x-latches to brothers of page, if those brothers |
| 5432 | exist. |
| 5433 | @return TRUE if compression occurred and FALSE if not or something |
| 5434 | wrong. */ |
| 5435 | ibool |
| 5436 | btr_cur_pessimistic_delete( |
| 5437 | /*=======================*/ |
| 5438 | dberr_t* err, /*!< out: DB_SUCCESS or DB_OUT_OF_FILE_SPACE; |
| 5439 | the latter may occur because we may have |
| 5440 | to update node pointers on upper levels, |
| 5441 | and in the case of variable length keys |
| 5442 | these may actually grow in size */ |
| 5443 | ibool has_reserved_extents, /*!< in: TRUE if the |
| 5444 | caller has already reserved enough free |
| 5445 | extents so that he knows that the operation |
| 5446 | will succeed */ |
| 5447 | btr_cur_t* cursor, /*!< in: cursor on the record to delete; |
| 5448 | if compression does not occur, the cursor |
| 5449 | stays valid: it points to successor of |
| 5450 | deleted record on function exit */ |
| 5451 | ulint flags, /*!< in: BTR_CREATE_FLAG or 0 */ |
| 5452 | bool rollback,/*!< in: performing rollback? */ |
| 5453 | mtr_t* mtr) /*!< in: mtr */ |
| 5454 | { |
| 5455 | buf_block_t* block; |
| 5456 | page_t* page; |
| 5457 | page_zip_des_t* page_zip; |
| 5458 | dict_index_t* index; |
| 5459 | rec_t* rec; |
| 5460 | ulint n_reserved = 0; |
| 5461 | bool success; |
| 5462 | ibool ret = FALSE; |
| 5463 | mem_heap_t* heap; |
| 5464 | ulint* offsets; |
| 5465 | #ifdef UNIV_DEBUG |
| 5466 | bool parent_latched = false; |
| 5467 | #endif /* UNIV_DEBUG */ |
| 5468 | |
| 5469 | block = btr_cur_get_block(cursor); |
| 5470 | page = buf_block_get_frame(block); |
| 5471 | index = btr_cur_get_index(cursor); |
| 5472 | |
| 5473 | ut_ad(flags == 0 || flags == BTR_CREATE_FLAG); |
| 5474 | ut_ad(!dict_index_is_online_ddl(index) |
| 5475 | || dict_index_is_clust(index) |
| 5476 | || (flags & BTR_CREATE_FLAG)); |
| 5477 | ut_ad(mtr_memo_contains_flagged(mtr, dict_index_get_lock(index), |
| 5478 | MTR_MEMO_X_LOCK |
| 5479 | | MTR_MEMO_SX_LOCK)); |
| 5480 | ut_ad(mtr_memo_contains(mtr, block, MTR_MEMO_PAGE_X_FIX)); |
| 5481 | ut_ad(mtr->is_named_space(index->table->space)); |
| 5482 | |
| 5483 | if (!has_reserved_extents) { |
| 5484 | /* First reserve enough free space for the file segments |
| 5485 | of the index tree, so that the node pointer updates will |
| 5486 | not fail because of lack of space */ |
| 5487 | |
| 5488 | ulint n_extents = cursor->tree_height / 32 + 1; |
| 5489 | |
| 5490 | success = fsp_reserve_free_extents(&n_reserved, |
| 5491 | index->table->space, |
| 5492 | n_extents, |
| 5493 | FSP_CLEANING, mtr); |
| 5494 | if (!success) { |
| 5495 | *err = DB_OUT_OF_FILE_SPACE; |
| 5496 | |
| 5497 | return(FALSE); |
| 5498 | } |
| 5499 | } |
| 5500 | |
| 5501 | heap = mem_heap_create(1024); |
| 5502 | rec = btr_cur_get_rec(cursor); |
| 5503 | page_zip = buf_block_get_page_zip(block); |
| 5504 | #ifdef UNIV_ZIP_DEBUG |
| 5505 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 5506 | #endif /* UNIV_ZIP_DEBUG */ |
| 5507 | |
| 5508 | offsets = rec_get_offsets(rec, index, NULL, page_is_leaf(page), |
| 5509 | ULINT_UNDEFINED, &heap); |
| 5510 | |
| 5511 | if (rec_offs_any_extern(offsets)) { |
| 5512 | btr_rec_free_externally_stored_fields(index, |
| 5513 | rec, offsets, page_zip, |
| 5514 | rollback, mtr); |
| 5515 | #ifdef UNIV_ZIP_DEBUG |
| 5516 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 5517 | #endif /* UNIV_ZIP_DEBUG */ |
| 5518 | } |
| 5519 | |
| 5520 | if (page_is_leaf(page)) { |
| 5521 | const bool is_default_row = rec_get_info_bits( |
| 5522 | rec, page_rec_is_comp(rec)) & REC_INFO_MIN_REC_FLAG; |
| 5523 | if (UNIV_UNLIKELY(is_default_row)) { |
| 5524 | /* This should be rolling back instant ADD COLUMN. |
| 5525 | If this is a recovered transaction, then |
| 5526 | index->is_instant() will hold until the |
| 5527 | insert into SYS_COLUMNS is rolled back. */ |
| 5528 | ut_ad(rollback); |
| 5529 | ut_ad(index->table->supports_instant()); |
| 5530 | ut_ad(index->is_primary()); |
| 5531 | } else if (flags == 0) { |
| 5532 | lock_update_delete(block, rec); |
| 5533 | } |
| 5534 | |
| 5535 | if (!page_is_root(page)) { |
| 5536 | if (page_get_n_recs(page) < 2) { |
| 5537 | goto discard_page; |
| 5538 | } |
| 5539 | } else if (page_get_n_recs(page) == 1 |
| 5540 | + (index->is_instant() |
| 5541 | && !rec_is_default_row(rec, index))) { |
| 5542 | /* The whole index (and table) becomes logically empty. |
| 5543 | Empty the whole page. That is, if we are deleting the |
| 5544 | only user record, also delete the 'default row' record |
| 5545 | if one exists (it exists if and only if is_instant()). |
| 5546 | If we are deleting the 'default row' record and the |
| 5547 | table becomes empty, clean up the whole page. */ |
| 5548 | ut_ad(!index->is_instant() |
| 5549 | || rec_is_default_row( |
| 5550 | page_rec_get_next_const( |
| 5551 | page_get_infimum_rec(page)), |
| 5552 | index)); |
| 5553 | btr_page_empty(block, page_zip, index, 0, mtr); |
| 5554 | page_cur_set_after_last(block, |
| 5555 | btr_cur_get_page_cur(cursor)); |
| 5556 | if (index->is_primary()) { |
| 5557 | /* Concurrent access is prevented by |
| 5558 | index->lock and root_block->lock |
| 5559 | X-latch, so this should be safe. */ |
| 5560 | index->remove_instant(); |
| 5561 | } |
| 5562 | ret = TRUE; |
| 5563 | goto return_after_reservations; |
| 5564 | } |
| 5565 | |
| 5566 | if (UNIV_LIKELY(!is_default_row)) { |
| 5567 | btr_search_update_hash_on_delete(cursor); |
| 5568 | } else { |
| 5569 | page_cur_delete_rec(btr_cur_get_page_cur(cursor), |
| 5570 | index, offsets, mtr); |
| 5571 | /* We must empty the PAGE_FREE list, because |
| 5572 | after rollback, this deleted 'default row' record |
| 5573 | would carry too many fields, and we would be |
| 5574 | unable to know the size of the freed record. */ |
| 5575 | btr_page_reorganize(btr_cur_get_page_cur(cursor), |
| 5576 | index, mtr); |
| 5577 | ut_ad(!ret); |
| 5578 | goto return_after_reservations; |
| 5579 | } |
| 5580 | } else if (UNIV_UNLIKELY(page_rec_is_first(rec, page))) { |
| 5581 | if (page_rec_is_last(rec, page)) { |
| 5582 | discard_page: |
| 5583 | ut_ad(page_get_n_recs(page) == 1); |
| 5584 | /* If there is only one record, drop |
| 5585 | the whole page. */ |
| 5586 | |
| 5587 | btr_discard_page(cursor, mtr); |
| 5588 | |
| 5589 | ret = TRUE; |
| 5590 | goto return_after_reservations; |
| 5591 | } |
| 5592 | |
| 5593 | rec_t* next_rec = page_rec_get_next(rec); |
| 5594 | |
| 5595 | if (!page_has_prev(page)) { |
| 5596 | |
| 5597 | /* If we delete the leftmost node pointer on a |
| 5598 | non-leaf level, we must mark the new leftmost node |
| 5599 | pointer as the predefined minimum record */ |
| 5600 | |
| 5601 | /* This will make page_zip_validate() fail until |
| 5602 | page_cur_delete_rec() completes. This is harmless, |
| 5603 | because everything will take place within a single |
| 5604 | mini-transaction and because writing to the redo log |
| 5605 | is an atomic operation (performed by mtr_commit()). */ |
| 5606 | btr_set_min_rec_mark(next_rec, mtr); |
| 5607 | } else if (dict_index_is_spatial(index)) { |
| 5608 | /* For rtree, if delete the leftmost node pointer, |
| 5609 | we need to update parent page. */ |
| 5610 | rtr_mbr_t father_mbr; |
| 5611 | rec_t* father_rec; |
| 5612 | btr_cur_t father_cursor; |
| 5613 | ulint* offsets; |
| 5614 | bool upd_ret; |
| 5615 | ulint len; |
| 5616 | |
| 5617 | rtr_page_get_father_block(NULL, heap, index, |
| 5618 | block, mtr, NULL, |
| 5619 | &father_cursor); |
| 5620 | offsets = rec_get_offsets( |
| 5621 | btr_cur_get_rec(&father_cursor), index, NULL, |
| 5622 | false, ULINT_UNDEFINED, &heap); |
| 5623 | |
| 5624 | father_rec = btr_cur_get_rec(&father_cursor); |
| 5625 | rtr_read_mbr(rec_get_nth_field( |
| 5626 | father_rec, offsets, 0, &len), &father_mbr); |
| 5627 | |
| 5628 | upd_ret = rtr_update_mbr_field(&father_cursor, offsets, |
| 5629 | NULL, page, &father_mbr, |
| 5630 | next_rec, mtr); |
| 5631 | |
| 5632 | if (!upd_ret) { |
| 5633 | *err = DB_ERROR; |
| 5634 | |
| 5635 | mem_heap_free(heap); |
| 5636 | return(FALSE); |
| 5637 | } |
| 5638 | |
| 5639 | ut_d(parent_latched = true); |
| 5640 | } else { |
| 5641 | /* Otherwise, if we delete the leftmost node pointer |
| 5642 | on a page, we have to change the parent node pointer |
| 5643 | so that it is equal to the new leftmost node pointer |
| 5644 | on the page */ |
| 5645 | |
| 5646 | btr_node_ptr_delete(index, block, mtr); |
| 5647 | const ulint level = btr_page_get_level(page); |
| 5648 | |
| 5649 | dtuple_t* node_ptr = dict_index_build_node_ptr( |
| 5650 | index, next_rec, block->page.id.page_no(), |
| 5651 | heap, level); |
| 5652 | |
| 5653 | btr_insert_on_non_leaf_level( |
| 5654 | flags, index, level + 1, node_ptr, mtr); |
| 5655 | |
| 5656 | ut_d(parent_latched = true); |
| 5657 | } |
| 5658 | } |
| 5659 | |
| 5660 | page_cur_delete_rec(btr_cur_get_page_cur(cursor), index, offsets, mtr); |
| 5661 | #ifdef UNIV_ZIP_DEBUG |
| 5662 | ut_a(!page_zip || page_zip_validate(page_zip, page, index)); |
| 5663 | #endif /* UNIV_ZIP_DEBUG */ |
| 5664 | |
| 5665 | return_after_reservations: |
| 5666 | /* btr_check_node_ptr() needs parent block latched */ |
| 5667 | ut_ad(!parent_latched || btr_check_node_ptr(index, block, mtr)); |
| 5668 | |
| 5669 | *err = DB_SUCCESS; |
| 5670 | |
| 5671 | mem_heap_free(heap); |
| 5672 | |
| 5673 | if (ret == FALSE) { |
| 5674 | ret = btr_cur_compress_if_useful(cursor, FALSE, mtr); |
| 5675 | } |
| 5676 | |
| 5677 | if (!srv_read_only_mode |
| 5678 | && page_is_leaf(page) |
| 5679 | && !dict_index_is_online_ddl(index)) { |
| 5680 | |
| 5681 | mtr_memo_release(mtr, dict_index_get_lock(index), |
| 5682 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK); |
| 5683 | |
| 5684 | /* NOTE: We cannot release root block latch here, because it |
| 5685 | has segment header and already modified in most of cases.*/ |
| 5686 | } |
| 5687 | |
| 5688 | index->table->space->release_free_extents(n_reserved); |
| 5689 | return(ret); |
| 5690 | } |
| 5691 | |
| 5692 | /*******************************************************************//** |
| 5693 | Adds path information to the cursor for the current page, for which |
| 5694 | the binary search has been performed. */ |
| 5695 | static |
| 5696 | void |
| 5697 | btr_cur_add_path_info( |
| 5698 | /*==================*/ |
| 5699 | btr_cur_t* cursor, /*!< in: cursor positioned on a page */ |
| 5700 | ulint height, /*!< in: height of the page in tree; |
| 5701 | 0 means leaf node */ |
| 5702 | ulint root_height) /*!< in: root node height in tree */ |
| 5703 | { |
| 5704 | btr_path_t* slot; |
| 5705 | const rec_t* rec; |
| 5706 | const page_t* page; |
| 5707 | |
| 5708 | ut_a(cursor->path_arr); |
| 5709 | |
| 5710 | if (root_height >= BTR_PATH_ARRAY_N_SLOTS - 1) { |
| 5711 | /* Do nothing; return empty path */ |
| 5712 | |
| 5713 | slot = cursor->path_arr; |
| 5714 | slot->nth_rec = ULINT_UNDEFINED; |
| 5715 | |
| 5716 | return; |
| 5717 | } |
| 5718 | |
| 5719 | if (height == 0) { |
| 5720 | /* Mark end of slots for path */ |
| 5721 | slot = cursor->path_arr + root_height + 1; |
| 5722 | slot->nth_rec = ULINT_UNDEFINED; |
| 5723 | } |
| 5724 | |
| 5725 | rec = btr_cur_get_rec(cursor); |
| 5726 | |
| 5727 | slot = cursor->path_arr + (root_height - height); |
| 5728 | |
| 5729 | page = page_align(rec); |
| 5730 | |
| 5731 | slot->nth_rec = page_rec_get_n_recs_before(rec); |
| 5732 | slot->n_recs = page_get_n_recs(page); |
| 5733 | slot->page_no = page_get_page_no(page); |
| 5734 | slot->page_level = btr_page_get_level(page); |
| 5735 | } |
| 5736 | |
| 5737 | /*******************************************************************//** |
| 5738 | Estimate the number of rows between slot1 and slot2 for any level on a |
| 5739 | B-tree. This function starts from slot1->page and reads a few pages to |
| 5740 | the right, counting their records. If we reach slot2->page quickly then |
| 5741 | we know exactly how many records there are between slot1 and slot2 and |
| 5742 | we set is_n_rows_exact to TRUE. If we cannot reach slot2->page quickly |
| 5743 | then we calculate the average number of records in the pages scanned |
| 5744 | so far and assume that all pages that we did not scan up to slot2->page |
| 5745 | contain the same number of records, then we multiply that average to |
| 5746 | the number of pages between slot1->page and slot2->page (which is |
| 5747 | n_rows_on_prev_level). In this case we set is_n_rows_exact to FALSE. |
| 5748 | @return number of rows, not including the borders (exact or estimated) */ |
| 5749 | static |
| 5750 | ha_rows |
| 5751 | btr_estimate_n_rows_in_range_on_level( |
| 5752 | /*==================================*/ |
| 5753 | dict_index_t* index, /*!< in: index */ |
| 5754 | btr_path_t* slot1, /*!< in: left border */ |
| 5755 | btr_path_t* slot2, /*!< in: right border */ |
| 5756 | ha_rows n_rows_on_prev_level, /*!< in: number of rows |
| 5757 | on the previous level for the |
| 5758 | same descend paths; used to |
| 5759 | determine the number of pages |
| 5760 | on this level */ |
| 5761 | bool* is_n_rows_exact) /*!< out: TRUE if the returned |
| 5762 | value is exact i.e. not an |
| 5763 | estimation */ |
| 5764 | { |
| 5765 | ha_rows n_rows = 0; |
| 5766 | uint n_pages_read = 0; |
| 5767 | ulint level; |
| 5768 | |
| 5769 | /* Assume by default that we will scan all pages between |
| 5770 | slot1->page_no and slot2->page_no. */ |
| 5771 | *is_n_rows_exact = true; |
| 5772 | |
| 5773 | /* Add records from slot1->page_no which are to the right of |
| 5774 | the record which serves as a left border of the range, if any |
| 5775 | (we don't include the record itself in this count). */ |
| 5776 | if (slot1->nth_rec <= slot1->n_recs) { |
| 5777 | n_rows += slot1->n_recs - slot1->nth_rec; |
| 5778 | } |
| 5779 | |
| 5780 | /* Add records from slot2->page_no which are to the left of |
| 5781 | the record which servers as a right border of the range, if any |
| 5782 | (we don't include the record itself in this count). */ |
| 5783 | if (slot2->nth_rec > 1) { |
| 5784 | n_rows += slot2->nth_rec - 1; |
| 5785 | } |
| 5786 | |
| 5787 | /* Count the records in the pages between slot1->page_no and |
| 5788 | slot2->page_no (non inclusive), if any. */ |
| 5789 | |
| 5790 | /* Do not read more than this number of pages in order not to hurt |
| 5791 | performance with this code which is just an estimation. If we read |
| 5792 | this many pages before reaching slot2->page_no then we estimate the |
| 5793 | average from the pages scanned so far. */ |
| 5794 | # define N_PAGES_READ_LIMIT 10 |
| 5795 | |
| 5796 | const fil_space_t* space = index->table->space; |
| 5797 | page_id_t page_id(space->id, slot1->page_no); |
| 5798 | const page_size_t page_size(space->flags); |
| 5799 | |
| 5800 | level = slot1->page_level; |
| 5801 | |
| 5802 | do { |
| 5803 | mtr_t mtr; |
| 5804 | page_t* page; |
| 5805 | buf_block_t* block; |
| 5806 | dberr_t err=DB_SUCCESS; |
| 5807 | |
| 5808 | mtr_start(&mtr); |
| 5809 | |
| 5810 | /* Fetch the page. Because we are not holding the |
| 5811 | index->lock, the tree may have changed and we may be |
| 5812 | attempting to read a page that is no longer part of |
| 5813 | the B-tree. We pass BUF_GET_POSSIBLY_FREED in order to |
| 5814 | silence a debug assertion about this. */ |
| 5815 | block = buf_page_get_gen(page_id, page_size, RW_S_LATCH, |
| 5816 | NULL, BUF_GET_POSSIBLY_FREED, |
| 5817 | __FILE__, __LINE__, &mtr, &err); |
| 5818 | |
| 5819 | ut_ad((block != NULL) == (err == DB_SUCCESS)); |
| 5820 | |
| 5821 | if (err != DB_SUCCESS) { |
| 5822 | if (err == DB_DECRYPTION_FAILED) { |
| 5823 | ib_push_warning((void *)NULL, |
| 5824 | DB_DECRYPTION_FAILED, |
| 5825 | "Table %s is encrypted but encryption service or" |
| 5826 | " used key_id is not available. " |
| 5827 | " Can't continue reading table.", |
| 5828 | index->table->name); |
| 5829 | index->table->file_unreadable = true; |
| 5830 | } |
| 5831 | |
| 5832 | mtr_commit(&mtr); |
| 5833 | goto inexact; |
| 5834 | } |
| 5835 | |
| 5836 | page = buf_block_get_frame(block); |
| 5837 | |
| 5838 | /* It is possible that the tree has been reorganized in the |
| 5839 | meantime and this is a different page. If this happens the |
| 5840 | calculated estimate will be bogus, which is not fatal as |
| 5841 | this is only an estimate. We are sure that a page with |
| 5842 | page_no exists because InnoDB never frees pages, only |
| 5843 | reuses them. */ |
| 5844 | if (!fil_page_index_page_check(page) |
| 5845 | || btr_page_get_index_id(page) != index->id |
| 5846 | || btr_page_get_level(page) != level) { |
| 5847 | |
| 5848 | /* The page got reused for something else */ |
| 5849 | mtr_commit(&mtr); |
| 5850 | goto inexact; |
| 5851 | } |
| 5852 | |
| 5853 | /* It is possible but highly unlikely that the page was |
| 5854 | originally written by an old version of InnoDB that did |
| 5855 | not initialize FIL_PAGE_TYPE on other than B-tree pages. |
| 5856 | For example, this could be an almost-empty BLOB page |
| 5857 | that happens to contain the magic values in the fields |
| 5858 | that we checked above. */ |
| 5859 | |
| 5860 | n_pages_read++; |
| 5861 | |
| 5862 | if (page_id.page_no() != slot1->page_no) { |
| 5863 | /* Do not count the records on slot1->page_no, |
| 5864 | we already counted them before this loop. */ |
| 5865 | n_rows += page_get_n_recs(page); |
| 5866 | } |
| 5867 | |
| 5868 | page_id.set_page_no(btr_page_get_next(page, &mtr)); |
| 5869 | |
| 5870 | mtr_commit(&mtr); |
| 5871 | |
| 5872 | if (n_pages_read == N_PAGES_READ_LIMIT |
| 5873 | || page_id.page_no() == FIL_NULL) { |
| 5874 | /* Either we read too many pages or |
| 5875 | we reached the end of the level without passing |
| 5876 | through slot2->page_no, the tree must have changed |
| 5877 | in the meantime */ |
| 5878 | goto inexact; |
| 5879 | } |
| 5880 | |
| 5881 | } while (page_id.page_no() != slot2->page_no); |
| 5882 | |
| 5883 | return(n_rows); |
| 5884 | |
| 5885 | inexact: |
| 5886 | |
| 5887 | *is_n_rows_exact = false; |
| 5888 | |
| 5889 | /* We did interrupt before reaching slot2->page */ |
| 5890 | |
| 5891 | if (n_pages_read > 0) { |
| 5892 | /* The number of pages on this level is |
| 5893 | n_rows_on_prev_level, multiply it by the |
| 5894 | average number of recs per page so far */ |
| 5895 | n_rows = n_rows_on_prev_level * n_rows / n_pages_read; |
| 5896 | } else { |
| 5897 | /* The tree changed before we could even |
| 5898 | start with slot1->page_no */ |
| 5899 | n_rows = 10; |
| 5900 | } |
| 5901 | |
| 5902 | return(n_rows); |
| 5903 | } |
| 5904 | |
| 5905 | /** If the tree gets changed too much between the two dives for the left |
| 5906 | and right boundary then btr_estimate_n_rows_in_range_low() will retry |
| 5907 | that many times before giving up and returning the value stored in |
| 5908 | rows_in_range_arbitrary_ret_val. */ |
| 5909 | static const unsigned rows_in_range_max_retries = 4; |
| 5910 | |
| 5911 | /** We pretend that a range has that many records if the tree keeps changing |
| 5912 | for rows_in_range_max_retries retries while we try to estimate the records |
| 5913 | in a given range. */ |
| 5914 | static const ha_rows rows_in_range_arbitrary_ret_val = 10; |
| 5915 | |
| 5916 | /** Estimates the number of rows in a given index range. |
| 5917 | @param[in] index index |
| 5918 | @param[in] tuple1 range start, may also be empty tuple |
| 5919 | @param[in] mode1 search mode for range start |
| 5920 | @param[in] tuple2 range end, may also be empty tuple |
| 5921 | @param[in] mode2 search mode for range end |
| 5922 | @param[in] nth_attempt if the tree gets modified too much while |
| 5923 | we are trying to analyze it, then we will retry (this function will call |
| 5924 | itself, incrementing this parameter) |
| 5925 | @return estimated number of rows; if after rows_in_range_max_retries |
| 5926 | retries the tree keeps changing, then we will just return |
| 5927 | rows_in_range_arbitrary_ret_val as a result (if |
| 5928 | nth_attempt >= rows_in_range_max_retries and the tree is modified between |
| 5929 | the two dives). */ |
| 5930 | static |
| 5931 | ha_rows |
| 5932 | btr_estimate_n_rows_in_range_low( |
| 5933 | dict_index_t* index, |
| 5934 | const dtuple_t* tuple1, |
| 5935 | page_cur_mode_t mode1, |
| 5936 | const dtuple_t* tuple2, |
| 5937 | page_cur_mode_t mode2, |
| 5938 | unsigned nth_attempt) |
| 5939 | { |
| 5940 | btr_path_t path1[BTR_PATH_ARRAY_N_SLOTS]; |
| 5941 | btr_path_t path2[BTR_PATH_ARRAY_N_SLOTS]; |
| 5942 | btr_cur_t cursor; |
| 5943 | btr_path_t* slot1; |
| 5944 | btr_path_t* slot2; |
| 5945 | bool diverged; |
| 5946 | bool diverged_lot; |
| 5947 | ulint divergence_level; |
| 5948 | ha_rows n_rows; |
| 5949 | bool is_n_rows_exact; |
| 5950 | ulint i; |
| 5951 | mtr_t mtr; |
| 5952 | ha_rows table_n_rows; |
| 5953 | |
| 5954 | table_n_rows = dict_table_get_n_rows(index->table); |
| 5955 | |
| 5956 | /* Below we dive to the two records specified by tuple1 and tuple2 and |
| 5957 | we remember the entire dive paths from the tree root. The place where |
| 5958 | the tuple1 path ends on the leaf level we call "left border" of our |
| 5959 | interval and the place where the tuple2 path ends on the leaf level - |
| 5960 | "right border". We take care to either include or exclude the interval |
| 5961 | boundaries depending on whether <, <=, > or >= was specified. For |
| 5962 | example if "5 < x AND x <= 10" then we should not include the left |
| 5963 | boundary, but should include the right one. */ |
| 5964 | |
| 5965 | mtr_start(&mtr); |
| 5966 | |
| 5967 | cursor.path_arr = path1; |
| 5968 | |
| 5969 | bool should_count_the_left_border; |
| 5970 | |
| 5971 | if (dtuple_get_n_fields(tuple1) > 0) { |
| 5972 | |
| 5973 | btr_cur_search_to_nth_level(index, 0, tuple1, mode1, |
| 5974 | BTR_SEARCH_LEAF | BTR_ESTIMATE, |
| 5975 | &cursor, 0, |
| 5976 | __FILE__, __LINE__, &mtr); |
| 5977 | |
| 5978 | ut_ad(!page_rec_is_infimum(btr_cur_get_rec(&cursor))); |
| 5979 | |
| 5980 | /* We should count the border if there are any records to |
| 5981 | match the criteria, i.e. if the maximum record on the tree is |
| 5982 | 5 and x > 3 is specified then the cursor will be positioned at |
| 5983 | 5 and we should count the border, but if x > 7 is specified, |
| 5984 | then the cursor will be positioned at 'sup' on the rightmost |
| 5985 | leaf page in the tree and we should not count the border. */ |
| 5986 | should_count_the_left_border |
| 5987 | = !page_rec_is_supremum(btr_cur_get_rec(&cursor)); |
| 5988 | } else { |
| 5989 | dberr_t err = DB_SUCCESS; |
| 5990 | |
| 5991 | err = btr_cur_open_at_index_side(true, index, |
| 5992 | BTR_SEARCH_LEAF | BTR_ESTIMATE, |
| 5993 | &cursor, 0, &mtr); |
| 5994 | |
| 5995 | if (err != DB_SUCCESS) { |
| 5996 | ib::warn() << " Error code: "<< err |
| 5997 | << " btr_estimate_n_rows_in_range_low " |
| 5998 | << " called from file: " |
| 5999 | << __FILE__ << " line: "<< __LINE__ |
| 6000 | << " table: "<< index->table->name |
| 6001 | << " index: "<< index->name; |
| 6002 | } |
| 6003 | |
| 6004 | ut_ad(page_rec_is_infimum(btr_cur_get_rec(&cursor))); |
| 6005 | |
| 6006 | /* The range specified is wihout a left border, just |
| 6007 | 'x < 123' or 'x <= 123' and btr_cur_open_at_index_side() |
| 6008 | positioned the cursor on the infimum record on the leftmost |
| 6009 | page, which must not be counted. */ |
| 6010 | should_count_the_left_border = false; |
| 6011 | } |
| 6012 | |
| 6013 | mtr_commit(&mtr); |
| 6014 | |
| 6015 | if (!index->is_readable()) { |
| 6016 | return 0; |
| 6017 | } |
| 6018 | |
| 6019 | mtr_start(&mtr); |
| 6020 | |
| 6021 | cursor.path_arr = path2; |
| 6022 | |
| 6023 | bool should_count_the_right_border; |
| 6024 | |
| 6025 | if (dtuple_get_n_fields(tuple2) > 0) { |
| 6026 | |
| 6027 | btr_cur_search_to_nth_level(index, 0, tuple2, mode2, |
| 6028 | BTR_SEARCH_LEAF | BTR_ESTIMATE, |
| 6029 | &cursor, 0, |
| 6030 | __FILE__, __LINE__, &mtr); |
| 6031 | |
| 6032 | const rec_t* rec = btr_cur_get_rec(&cursor); |
| 6033 | |
| 6034 | ut_ad(!(mode2 == PAGE_CUR_L && page_rec_is_supremum(rec))); |
| 6035 | |
| 6036 | should_count_the_right_border |
| 6037 | = (mode2 == PAGE_CUR_LE /* if the range is '<=' */ |
| 6038 | /* and the record was found */ |
| 6039 | && cursor.low_match >= dtuple_get_n_fields(tuple2)) |
| 6040 | || (mode2 == PAGE_CUR_L /* or if the range is '<' */ |
| 6041 | /* and there are any records to match the criteria, |
| 6042 | i.e. if the minimum record on the tree is 5 and |
| 6043 | x < 7 is specified then the cursor will be |
| 6044 | positioned at 5 and we should count the border, but |
| 6045 | if x < 2 is specified, then the cursor will be |
| 6046 | positioned at 'inf' and we should not count the |
| 6047 | border */ |
| 6048 | && !page_rec_is_infimum(rec)); |
| 6049 | /* Notice that for "WHERE col <= 'foo'" MySQL passes to |
| 6050 | ha_innobase::records_in_range(): |
| 6051 | min_key=NULL (left-unbounded) which is expected |
| 6052 | max_key='foo' flag=HA_READ_AFTER_KEY (PAGE_CUR_G), which is |
| 6053 | unexpected - one would expect |
| 6054 | flag=HA_READ_KEY_OR_PREV (PAGE_CUR_LE). In this case the |
| 6055 | cursor will be positioned on the first record to the right of |
| 6056 | the requested one (can also be positioned on the 'sup') and |
| 6057 | we should not count the right border. */ |
| 6058 | } else { |
| 6059 | dberr_t err = DB_SUCCESS; |
| 6060 | |
| 6061 | err = btr_cur_open_at_index_side(false, index, |
| 6062 | BTR_SEARCH_LEAF | BTR_ESTIMATE, |
| 6063 | &cursor, 0, &mtr); |
| 6064 | |
| 6065 | if (err != DB_SUCCESS) { |
| 6066 | ib::warn() << " Error code: "<< err |
| 6067 | << " btr_estimate_n_rows_in_range_low " |
| 6068 | << " called from file: " |
| 6069 | << __FILE__ << " line: "<< __LINE__ |
| 6070 | << " table: "<< index->table->name |
| 6071 | << " index: "<< index->name; |
| 6072 | } |
| 6073 | |
| 6074 | ut_ad(page_rec_is_supremum(btr_cur_get_rec(&cursor))); |
| 6075 | |
| 6076 | /* The range specified is wihout a right border, just |
| 6077 | 'x > 123' or 'x >= 123' and btr_cur_open_at_index_side() |
| 6078 | positioned the cursor on the supremum record on the rightmost |
| 6079 | page, which must not be counted. */ |
| 6080 | should_count_the_right_border = false; |
| 6081 | } |
| 6082 | |
| 6083 | mtr_commit(&mtr); |
| 6084 | |
| 6085 | /* We have the path information for the range in path1 and path2 */ |
| 6086 | |
| 6087 | n_rows = 0; |
| 6088 | is_n_rows_exact = true; |
| 6089 | |
| 6090 | /* This becomes true when the two paths do not pass through the |
| 6091 | same pages anymore. */ |
| 6092 | diverged = false; |
| 6093 | |
| 6094 | /* This becomes true when the paths are not the same or adjacent |
| 6095 | any more. This means that they pass through the same or |
| 6096 | neighboring-on-the-same-level pages only. */ |
| 6097 | diverged_lot = false; |
| 6098 | |
| 6099 | /* This is the level where paths diverged a lot. */ |
| 6100 | divergence_level = 1000000; |
| 6101 | |
| 6102 | for (i = 0; ; i++) { |
| 6103 | ut_ad(i < BTR_PATH_ARRAY_N_SLOTS); |
| 6104 | |
| 6105 | slot1 = path1 + i; |
| 6106 | slot2 = path2 + i; |
| 6107 | |
| 6108 | if (slot1->nth_rec == ULINT_UNDEFINED |
| 6109 | || slot2->nth_rec == ULINT_UNDEFINED) { |
| 6110 | |
| 6111 | /* Here none of the borders were counted. For example, |
| 6112 | if on the leaf level we descended to: |
| 6113 | (inf, a, b, c, d, e, f, sup) |
| 6114 | ^ ^ |
| 6115 | path1 path2 |
| 6116 | then n_rows will be 2 (c and d). */ |
| 6117 | |
| 6118 | if (is_n_rows_exact) { |
| 6119 | /* Only fiddle to adjust this off-by-one |
| 6120 | if the number is exact, otherwise we do |
| 6121 | much grosser adjustments below. */ |
| 6122 | |
| 6123 | btr_path_t* last1 = &path1[i - 1]; |
| 6124 | btr_path_t* last2 = &path2[i - 1]; |
| 6125 | |
| 6126 | /* If both paths end up on the same record on |
| 6127 | the leaf level. */ |
| 6128 | if (last1->page_no == last2->page_no |
| 6129 | && last1->nth_rec == last2->nth_rec) { |
| 6130 | |
| 6131 | /* n_rows can be > 0 here if the paths |
| 6132 | were first different and then converged |
| 6133 | to the same record on the leaf level. |
| 6134 | For example: |
| 6135 | SELECT ... LIKE 'wait/synch/rwlock%' |
| 6136 | mode1=PAGE_CUR_GE, |
| 6137 | tuple1="wait/synch/rwlock" |
| 6138 | path1[0]={nth_rec=58, n_recs=58, |
| 6139 | page_no=3, page_level=1} |
| 6140 | path1[1]={nth_rec=56, n_recs=55, |
| 6141 | page_no=119, page_level=0} |
| 6142 | |
| 6143 | mode2=PAGE_CUR_G |
| 6144 | tuple2="wait/synch/rwlock" |
| 6145 | path2[0]={nth_rec=57, n_recs=57, |
| 6146 | page_no=3, page_level=1} |
| 6147 | path2[1]={nth_rec=56, n_recs=55, |
| 6148 | page_no=119, page_level=0} */ |
| 6149 | |
| 6150 | /* If the range is such that we should |
| 6151 | count both borders, then avoid |
| 6152 | counting that record twice - once as a |
| 6153 | left border and once as a right |
| 6154 | border. */ |
| 6155 | if (should_count_the_left_border |
| 6156 | && should_count_the_right_border) { |
| 6157 | |
| 6158 | n_rows = 1; |
| 6159 | } else { |
| 6160 | /* Some of the borders should |
| 6161 | not be counted, e.g. [3,3). */ |
| 6162 | n_rows = 0; |
| 6163 | } |
| 6164 | } else { |
| 6165 | if (should_count_the_left_border) { |
| 6166 | n_rows++; |
| 6167 | } |
| 6168 | |
| 6169 | if (should_count_the_right_border) { |
| 6170 | n_rows++; |
| 6171 | } |
| 6172 | } |
| 6173 | } |
| 6174 | |
| 6175 | if (i > divergence_level + 1 && !is_n_rows_exact) { |
| 6176 | /* In trees whose height is > 1 our algorithm |
| 6177 | tends to underestimate: multiply the estimate |
| 6178 | by 2: */ |
| 6179 | |
| 6180 | n_rows = n_rows * 2; |
| 6181 | } |
| 6182 | |
| 6183 | DBUG_EXECUTE_IF("bug14007649", return(n_rows);); |
| 6184 | |
| 6185 | /* Do not estimate the number of rows in the range |
| 6186 | to over 1 / 2 of the estimated rows in the whole |
| 6187 | table */ |
| 6188 | |
| 6189 | if (n_rows > table_n_rows / 2 && !is_n_rows_exact) { |
| 6190 | |
| 6191 | n_rows = table_n_rows / 2; |
| 6192 | |
| 6193 | /* If there are just 0 or 1 rows in the table, |
| 6194 | then we estimate all rows are in the range */ |
| 6195 | |
| 6196 | if (n_rows == 0) { |
| 6197 | n_rows = table_n_rows; |
| 6198 | } |
| 6199 | } |
| 6200 | |
| 6201 | return(n_rows); |
| 6202 | } |
| 6203 | |
| 6204 | if (!diverged && slot1->nth_rec != slot2->nth_rec) { |
| 6205 | |
| 6206 | /* If both slots do not point to the same page, |
| 6207 | this means that the tree must have changed between |
| 6208 | the dive for slot1 and the dive for slot2 at the |
| 6209 | beginning of this function. */ |
| 6210 | if (slot1->page_no != slot2->page_no |
| 6211 | || slot1->page_level != slot2->page_level) { |
| 6212 | |
| 6213 | /* If the tree keeps changing even after a |
| 6214 | few attempts, then just return some arbitrary |
| 6215 | number. */ |
| 6216 | if (nth_attempt >= rows_in_range_max_retries) { |
| 6217 | return(rows_in_range_arbitrary_ret_val); |
| 6218 | } |
| 6219 | |
| 6220 | return btr_estimate_n_rows_in_range_low( |
| 6221 | index, tuple1, mode1, |
| 6222 | tuple2, mode2, nth_attempt + 1); |
| 6223 | } |
| 6224 | |
| 6225 | diverged = true; |
| 6226 | |
| 6227 | if (slot1->nth_rec < slot2->nth_rec) { |
| 6228 | /* We do not count the borders (nor the left |
| 6229 | nor the right one), thus "- 1". */ |
| 6230 | n_rows = slot2->nth_rec - slot1->nth_rec - 1; |
| 6231 | |
| 6232 | if (n_rows > 0) { |
| 6233 | /* There is at least one row between |
| 6234 | the two borders pointed to by slot1 |
| 6235 | and slot2, so on the level below the |
| 6236 | slots will point to non-adjacent |
| 6237 | pages. */ |
| 6238 | diverged_lot = true; |
| 6239 | divergence_level = i; |
| 6240 | } |
| 6241 | } else { |
| 6242 | /* It is possible that |
| 6243 | slot1->nth_rec >= slot2->nth_rec |
| 6244 | if, for example, we have a single page |
| 6245 | tree which contains (inf, 5, 6, supr) |
| 6246 | and we select where x > 20 and x < 30; |
| 6247 | in this case slot1->nth_rec will point |
| 6248 | to the supr record and slot2->nth_rec |
| 6249 | will point to 6. */ |
| 6250 | n_rows = 0; |
| 6251 | should_count_the_left_border = false; |
| 6252 | should_count_the_right_border = false; |
| 6253 | } |
| 6254 | |
| 6255 | } else if (diverged && !diverged_lot) { |
| 6256 | |
| 6257 | if (slot1->nth_rec < slot1->n_recs |
| 6258 | || slot2->nth_rec > 1) { |
| 6259 | |
| 6260 | diverged_lot = true; |
| 6261 | divergence_level = i; |
| 6262 | |
| 6263 | n_rows = 0; |
| 6264 | |
| 6265 | if (slot1->nth_rec < slot1->n_recs) { |
| 6266 | n_rows += slot1->n_recs |
| 6267 | - slot1->nth_rec; |
| 6268 | } |
| 6269 | |
| 6270 | if (slot2->nth_rec > 1) { |
| 6271 | n_rows += slot2->nth_rec - 1; |
| 6272 | } |
| 6273 | } |
| 6274 | } else if (diverged_lot) { |
| 6275 | |
| 6276 | n_rows = btr_estimate_n_rows_in_range_on_level( |
| 6277 | index, slot1, slot2, n_rows, |
| 6278 | &is_n_rows_exact); |
| 6279 | } |
| 6280 | } |
| 6281 | } |
| 6282 | |
| 6283 | /** Estimates the number of rows in a given index range. |
| 6284 | @param[in] index index |
| 6285 | @param[in] tuple1 range start, may also be empty tuple |
| 6286 | @param[in] mode1 search mode for range start |
| 6287 | @param[in] tuple2 range end, may also be empty tuple |
| 6288 | @param[in] mode2 search mode for range end |
| 6289 | @return estimated number of rows */ |
| 6290 | ha_rows |
| 6291 | btr_estimate_n_rows_in_range( |
| 6292 | dict_index_t* index, |
| 6293 | const dtuple_t* tuple1, |
| 6294 | page_cur_mode_t mode1, |
| 6295 | const dtuple_t* tuple2, |
| 6296 | page_cur_mode_t mode2) |
| 6297 | { |
| 6298 | return btr_estimate_n_rows_in_range_low( |
| 6299 | index, tuple1, mode1, tuple2, mode2, 1); |
| 6300 | } |
| 6301 | |
| 6302 | /*******************************************************************//** |
| 6303 | Record the number of non_null key values in a given index for |
| 6304 | each n-column prefix of the index where 1 <= n <= dict_index_get_n_unique(index). |
| 6305 | The estimates are eventually stored in the array: |
| 6306 | index->stat_n_non_null_key_vals[], which is indexed from 0 to n-1. */ |
| 6307 | static |
| 6308 | void |
| 6309 | btr_record_not_null_field_in_rec( |
| 6310 | /*=============================*/ |
| 6311 | ulint n_unique, /*!< in: dict_index_get_n_unique(index), |
| 6312 | number of columns uniquely determine |
| 6313 | an index entry */ |
| 6314 | const ulint* offsets, /*!< in: rec_get_offsets(rec, index), |
| 6315 | its size could be for all fields or |
| 6316 | that of "n_unique" */ |
| 6317 | ib_uint64_t* n_not_null) /*!< in/out: array to record number of |
| 6318 | not null rows for n-column prefix */ |
| 6319 | { |
| 6320 | ulint i; |
| 6321 | |
| 6322 | ut_ad(rec_offs_n_fields(offsets) >= n_unique); |
| 6323 | |
| 6324 | if (n_not_null == NULL) { |
| 6325 | return; |
| 6326 | } |
| 6327 | |
| 6328 | for (i = 0; i < n_unique; i++) { |
| 6329 | if (rec_offs_nth_sql_null(offsets, i)) { |
| 6330 | break; |
| 6331 | } |
| 6332 | |
| 6333 | n_not_null[i]++; |
| 6334 | } |
| 6335 | } |
| 6336 | |
| 6337 | /*******************************************************************//** |
| 6338 | Estimates the number of different key values in a given index, for |
| 6339 | each n-column prefix of the index where 1 <= n <= dict_index_get_n_unique(index). |
| 6340 | The estimates are stored in the array index->stat_n_diff_key_vals[] (indexed |
| 6341 | 0..n_uniq-1) and the number of pages that were sampled is saved in |
| 6342 | index->stat_n_sample_sizes[]. |
| 6343 | If innodb_stats_method is nulls_ignored, we also record the number of |
| 6344 | non-null values for each prefix and stored the estimates in |
| 6345 | array index->stat_n_non_null_key_vals. |
| 6346 | @return true if the index is available and we get the estimated numbers, |
| 6347 | false if the index is unavailable. */ |
| 6348 | bool |
| 6349 | btr_estimate_number_of_different_key_vals( |
| 6350 | /*======================================*/ |
| 6351 | dict_index_t* index) /*!< in: index */ |
| 6352 | { |
| 6353 | btr_cur_t cursor; |
| 6354 | page_t* page; |
| 6355 | rec_t* rec; |
| 6356 | ulint n_cols; |
| 6357 | ib_uint64_t* n_diff; |
| 6358 | ib_uint64_t* n_not_null; |
| 6359 | ibool stats_null_not_equal; |
| 6360 | uintmax_t n_sample_pages=1; /* number of pages to sample */ |
| 6361 | ulint not_empty_flag = 0; |
| 6362 | ulint total_external_size = 0; |
| 6363 | ulint i; |
| 6364 | ulint j; |
| 6365 | uintmax_t add_on; |
| 6366 | mtr_t mtr; |
| 6367 | mem_heap_t* heap = NULL; |
| 6368 | ulint* offsets_rec = NULL; |
| 6369 | ulint* offsets_next_rec = NULL; |
| 6370 | |
| 6371 | /* For spatial index, there is no such stats can be |
| 6372 | fetched. */ |
| 6373 | if (dict_index_is_spatial(index)) { |
| 6374 | return(false); |
| 6375 | } |
| 6376 | |
| 6377 | n_cols = dict_index_get_n_unique(index); |
| 6378 | |
| 6379 | heap = mem_heap_create((sizeof *n_diff + sizeof *n_not_null) |
| 6380 | * n_cols |
| 6381 | + dict_index_get_n_fields(index) |
| 6382 | * (sizeof *offsets_rec |
| 6383 | + sizeof *offsets_next_rec)); |
| 6384 | |
| 6385 | n_diff = (ib_uint64_t*) mem_heap_zalloc( |
| 6386 | heap, n_cols * sizeof(n_diff[0])); |
| 6387 | |
| 6388 | n_not_null = NULL; |
| 6389 | |
| 6390 | /* Check srv_innodb_stats_method setting, and decide whether we |
| 6391 | need to record non-null value and also decide if NULL is |
| 6392 | considered equal (by setting stats_null_not_equal value) */ |
| 6393 | switch (srv_innodb_stats_method) { |
| 6394 | case SRV_STATS_NULLS_IGNORED: |
| 6395 | n_not_null = (ib_uint64_t*) mem_heap_zalloc( |
| 6396 | heap, n_cols * sizeof *n_not_null); |
| 6397 | /* fall through */ |
| 6398 | |
| 6399 | case SRV_STATS_NULLS_UNEQUAL: |
| 6400 | /* for both SRV_STATS_NULLS_IGNORED and SRV_STATS_NULLS_UNEQUAL |
| 6401 | case, we will treat NULLs as unequal value */ |
| 6402 | stats_null_not_equal = TRUE; |
| 6403 | break; |
| 6404 | |
| 6405 | case SRV_STATS_NULLS_EQUAL: |
| 6406 | stats_null_not_equal = FALSE; |
| 6407 | break; |
| 6408 | |
| 6409 | default: |
| 6410 | ut_error; |
| 6411 | } |
| 6412 | |
| 6413 | if (srv_stats_sample_traditional) { |
| 6414 | /* It makes no sense to test more pages than are contained |
| 6415 | in the index, thus we lower the number if it is too high */ |
| 6416 | if (srv_stats_transient_sample_pages > index->stat_index_size) { |
| 6417 | if (index->stat_index_size > 0) { |
| 6418 | n_sample_pages = index->stat_index_size; |
| 6419 | } |
| 6420 | } else { |
| 6421 | n_sample_pages = srv_stats_transient_sample_pages; |
| 6422 | } |
| 6423 | } else { |
| 6424 | /* New logaritmic number of pages that are estimated. |
| 6425 | Number of pages estimated should be between 1 and |
| 6426 | index->stat_index_size. |
| 6427 | |
| 6428 | If we have only 0 or 1 index pages then we can only take 1 |
| 6429 | sample. We have already initialized n_sample_pages to 1. |
| 6430 | |
| 6431 | So taking index size as I and sample as S and log(I)*S as L |
| 6432 | |
| 6433 | requirement 1) we want the out limit of the expression to not exceed I; |
| 6434 | requirement 2) we want the ideal pages to be at least S; |
| 6435 | so the current expression is min(I, max( min(S,I), L) |
| 6436 | |
| 6437 | looking for simplifications: |
| 6438 | |
| 6439 | case 1: assume S < I |
| 6440 | min(I, max( min(S,I), L) -> min(I , max( S, L)) |
| 6441 | |
| 6442 | but since L=LOG2(I)*S and log2(I) >=1 L>S always so max(S,L) = L. |
| 6443 | |
| 6444 | so we have: min(I , L) |
| 6445 | |
| 6446 | case 2: assume I < S |
| 6447 | min(I, max( min(S,I), L) -> min(I, max( I, L)) |
| 6448 | |
| 6449 | case 2a: L > I |
| 6450 | min(I, max( I, L)) -> min(I, L) -> I |
| 6451 | |
| 6452 | case 2b: when L < I |
| 6453 | min(I, max( I, L)) -> min(I, I ) -> I |
| 6454 | |
| 6455 | so taking all case2 paths is I, our expression is: |
| 6456 | n_pages = S < I? min(I,L) : I |
| 6457 | */ |
| 6458 | if (index->stat_index_size > 1) { |
| 6459 | n_sample_pages = (srv_stats_transient_sample_pages < index->stat_index_size) ? |
| 6460 | ut_min(static_cast<ulint>(index->stat_index_size), |
| 6461 | static_cast<ulint>(log2(index->stat_index_size)*srv_stats_transient_sample_pages)) |
| 6462 | : index->stat_index_size; |
| 6463 | |
| 6464 | } |
| 6465 | } |
| 6466 | |
| 6467 | /* Sanity check */ |
| 6468 | ut_ad(n_sample_pages > 0 && n_sample_pages <= (index->stat_index_size <= 1 ? 1 : index->stat_index_size)); |
| 6469 | |
| 6470 | /* We sample some pages in the index to get an estimate */ |
| 6471 | |
| 6472 | for (i = 0; i < n_sample_pages; i++) { |
| 6473 | mtr_start(&mtr); |
| 6474 | |
| 6475 | bool available; |
| 6476 | |
| 6477 | available = btr_cur_open_at_rnd_pos(index, BTR_SEARCH_LEAF, |
| 6478 | &cursor, &mtr); |
| 6479 | |
| 6480 | if (!available) { |
| 6481 | mtr_commit(&mtr); |
| 6482 | mem_heap_free(heap); |
| 6483 | |
| 6484 | return(false); |
| 6485 | } |
| 6486 | |
| 6487 | /* Count the number of different key values for each prefix of |
| 6488 | the key on this index page. If the prefix does not determine |
| 6489 | the index record uniquely in the B-tree, then we subtract one |
| 6490 | because otherwise our algorithm would give a wrong estimate |
| 6491 | for an index where there is just one key value. */ |
| 6492 | |
| 6493 | if (!index->is_readable()) { |
| 6494 | mtr_commit(&mtr); |
| 6495 | goto exit_loop; |
| 6496 | } |
| 6497 | |
| 6498 | page = btr_cur_get_page(&cursor); |
| 6499 | |
| 6500 | rec = page_rec_get_next(page_get_infimum_rec(page)); |
| 6501 | const bool is_leaf = page_is_leaf(page); |
| 6502 | |
| 6503 | if (!page_rec_is_supremum(rec)) { |
| 6504 | not_empty_flag = 1; |
| 6505 | offsets_rec = rec_get_offsets(rec, index, offsets_rec, |
| 6506 | is_leaf, |
| 6507 | ULINT_UNDEFINED, &heap); |
| 6508 | |
| 6509 | if (n_not_null != NULL) { |
| 6510 | btr_record_not_null_field_in_rec( |
| 6511 | n_cols, offsets_rec, n_not_null); |
| 6512 | } |
| 6513 | } |
| 6514 | |
| 6515 | while (!page_rec_is_supremum(rec)) { |
| 6516 | ulint matched_fields; |
| 6517 | rec_t* next_rec = page_rec_get_next(rec); |
| 6518 | if (page_rec_is_supremum(next_rec)) { |
| 6519 | total_external_size += |
| 6520 | btr_rec_get_externally_stored_len( |
| 6521 | rec, offsets_rec); |
| 6522 | break; |
| 6523 | } |
| 6524 | |
| 6525 | offsets_next_rec = rec_get_offsets(next_rec, index, |
| 6526 | offsets_next_rec, |
| 6527 | is_leaf, |
| 6528 | ULINT_UNDEFINED, |
| 6529 | &heap); |
| 6530 | |
| 6531 | cmp_rec_rec_with_match(rec, next_rec, |
| 6532 | offsets_rec, offsets_next_rec, |
| 6533 | index, stats_null_not_equal, |
| 6534 | &matched_fields); |
| 6535 | |
| 6536 | for (j = matched_fields; j < n_cols; j++) { |
| 6537 | /* We add one if this index record has |
| 6538 | a different prefix from the previous */ |
| 6539 | |
| 6540 | n_diff[j]++; |
| 6541 | } |
| 6542 | |
| 6543 | if (n_not_null != NULL) { |
| 6544 | btr_record_not_null_field_in_rec( |
| 6545 | n_cols, offsets_next_rec, n_not_null); |
| 6546 | } |
| 6547 | |
| 6548 | total_external_size |
| 6549 | += btr_rec_get_externally_stored_len( |
| 6550 | rec, offsets_rec); |
| 6551 | |
| 6552 | rec = next_rec; |
| 6553 | /* Initialize offsets_rec for the next round |
| 6554 | and assign the old offsets_rec buffer to |
| 6555 | offsets_next_rec. */ |
| 6556 | { |
| 6557 | ulint* offsets_tmp = offsets_rec; |
| 6558 | offsets_rec = offsets_next_rec; |
| 6559 | offsets_next_rec = offsets_tmp; |
| 6560 | } |
| 6561 | } |
| 6562 | |
| 6563 | if (n_cols == dict_index_get_n_unique_in_tree(index) |
| 6564 | && page_has_siblings(page)) { |
| 6565 | |
| 6566 | /* If there is more than one leaf page in the tree, |
| 6567 | we add one because we know that the first record |
| 6568 | on the page certainly had a different prefix than the |
| 6569 | last record on the previous index page in the |
| 6570 | alphabetical order. Before this fix, if there was |
| 6571 | just one big record on each clustered index page, the |
| 6572 | algorithm grossly underestimated the number of rows |
| 6573 | in the table. */ |
| 6574 | |
| 6575 | n_diff[n_cols - 1]++; |
| 6576 | } |
| 6577 | |
| 6578 | mtr_commit(&mtr); |
| 6579 | } |
| 6580 | |
| 6581 | exit_loop: |
| 6582 | /* If we saw k borders between different key values on |
| 6583 | n_sample_pages leaf pages, we can estimate how many |
| 6584 | there will be in index->stat_n_leaf_pages */ |
| 6585 | |
| 6586 | /* We must take into account that our sample actually represents |
| 6587 | also the pages used for external storage of fields (those pages are |
| 6588 | included in index->stat_n_leaf_pages) */ |
| 6589 | |
| 6590 | for (j = 0; j < n_cols; j++) { |
| 6591 | index->stat_n_diff_key_vals[j] |
| 6592 | = BTR_TABLE_STATS_FROM_SAMPLE( |
| 6593 | n_diff[j], index, n_sample_pages, |
| 6594 | total_external_size, not_empty_flag); |
| 6595 | |
| 6596 | /* If the tree is small, smaller than |
| 6597 | 10 * n_sample_pages + total_external_size, then |
| 6598 | the above estimate is ok. For bigger trees it is common that we |
| 6599 | do not see any borders between key values in the few pages |
| 6600 | we pick. But still there may be n_sample_pages |
| 6601 | different key values, or even more. Let us try to approximate |
| 6602 | that: */ |
| 6603 | |
| 6604 | add_on = index->stat_n_leaf_pages |
| 6605 | / (10 * (n_sample_pages |
| 6606 | + total_external_size)); |
| 6607 | |
| 6608 | if (add_on > n_sample_pages) { |
| 6609 | add_on = n_sample_pages; |
| 6610 | } |
| 6611 | |
| 6612 | index->stat_n_diff_key_vals[j] += add_on; |
| 6613 | |
| 6614 | index->stat_n_sample_sizes[j] = n_sample_pages; |
| 6615 | |
| 6616 | /* Update the stat_n_non_null_key_vals[] with our |
| 6617 | sampled result. stat_n_non_null_key_vals[] is created |
| 6618 | and initialized to zero in dict_index_add_to_cache(), |
| 6619 | along with stat_n_diff_key_vals[] array */ |
| 6620 | if (n_not_null != NULL) { |
| 6621 | index->stat_n_non_null_key_vals[j] = |
| 6622 | BTR_TABLE_STATS_FROM_SAMPLE( |
| 6623 | n_not_null[j], index, n_sample_pages, |
| 6624 | total_external_size, not_empty_flag); |
| 6625 | } |
| 6626 | } |
| 6627 | |
| 6628 | mem_heap_free(heap); |
| 6629 | |
| 6630 | return(true); |
| 6631 | } |
| 6632 | |
| 6633 | /*================== EXTERNAL STORAGE OF BIG FIELDS ===================*/ |
| 6634 | |
| 6635 | /***********************************************************//** |
| 6636 | Gets the offset of the pointer to the externally stored part of a field. |
| 6637 | @return offset of the pointer to the externally stored part */ |
| 6638 | static |
| 6639 | ulint |
| 6640 | btr_rec_get_field_ref_offs( |
| 6641 | /*=======================*/ |
| 6642 | const ulint* offsets,/*!< in: array returned by rec_get_offsets() */ |
| 6643 | ulint n) /*!< in: index of the external field */ |
| 6644 | { |
| 6645 | ulint field_ref_offs; |
| 6646 | ulint local_len; |
| 6647 | |
| 6648 | ut_a(rec_offs_nth_extern(offsets, n)); |
| 6649 | field_ref_offs = rec_get_nth_field_offs(offsets, n, &local_len); |
| 6650 | ut_a(len_is_stored(local_len)); |
| 6651 | ut_a(local_len >= BTR_EXTERN_FIELD_REF_SIZE); |
| 6652 | |
| 6653 | return(field_ref_offs + local_len - BTR_EXTERN_FIELD_REF_SIZE); |
| 6654 | } |
| 6655 | |
| 6656 | /** Gets a pointer to the externally stored part of a field. |
| 6657 | @param rec record |
| 6658 | @param offsets rec_get_offsets(rec) |
| 6659 | @param n index of the externally stored field |
| 6660 | @return pointer to the externally stored part */ |
| 6661 | #define btr_rec_get_field_ref(rec, offsets, n) \ |
| 6662 | ((rec) + btr_rec_get_field_ref_offs(offsets, n)) |
| 6663 | |
| 6664 | /** Gets the externally stored size of a record, in units of a database page. |
| 6665 | @param[in] rec record |
| 6666 | @param[in] offsets array returned by rec_get_offsets() |
| 6667 | @return externally stored part, in units of a database page */ |
| 6668 | ulint |
| 6669 | btr_rec_get_externally_stored_len( |
| 6670 | const rec_t* rec, |
| 6671 | const ulint* offsets) |
| 6672 | { |
| 6673 | ulint n_fields; |
| 6674 | ulint total_extern_len = 0; |
| 6675 | ulint i; |
| 6676 | |
| 6677 | ut_ad(!rec_offs_comp(offsets) || !rec_get_node_ptr_flag(rec)); |
| 6678 | |
| 6679 | if (!rec_offs_any_extern(offsets)) { |
| 6680 | return(0); |
| 6681 | } |
| 6682 | |
| 6683 | n_fields = rec_offs_n_fields(offsets); |
| 6684 | |
| 6685 | for (i = 0; i < n_fields; i++) { |
| 6686 | if (rec_offs_nth_extern(offsets, i)) { |
| 6687 | |
| 6688 | ulint extern_len = mach_read_from_4( |
| 6689 | btr_rec_get_field_ref(rec, offsets, i) |
| 6690 | + BTR_EXTERN_LEN + 4); |
| 6691 | |
| 6692 | total_extern_len += ut_calc_align( |
| 6693 | extern_len, ulint(srv_page_size)); |
| 6694 | } |
| 6695 | } |
| 6696 | |
| 6697 | return total_extern_len >> srv_page_size_shift; |
| 6698 | } |
| 6699 | |
| 6700 | /*******************************************************************//** |
| 6701 | Sets the ownership bit of an externally stored field in a record. */ |
| 6702 | static |
| 6703 | void |
| 6704 | btr_cur_set_ownership_of_extern_field( |
| 6705 | /*==================================*/ |
| 6706 | page_zip_des_t* page_zip,/*!< in/out: compressed page whose uncompressed |
| 6707 | part will be updated, or NULL */ |
| 6708 | rec_t* rec, /*!< in/out: clustered index record */ |
| 6709 | dict_index_t* index, /*!< in: index of the page */ |
| 6710 | const ulint* offsets,/*!< in: array returned by rec_get_offsets() */ |
| 6711 | ulint i, /*!< in: field number */ |
| 6712 | ibool val, /*!< in: value to set */ |
| 6713 | mtr_t* mtr) /*!< in: mtr, or NULL if not logged */ |
| 6714 | { |
| 6715 | byte* data; |
| 6716 | ulint local_len; |
| 6717 | ulint byte_val; |
| 6718 | |
| 6719 | data = rec_get_nth_field(rec, offsets, i, &local_len); |
| 6720 | ut_ad(rec_offs_nth_extern(offsets, i)); |
| 6721 | ut_a(local_len >= BTR_EXTERN_FIELD_REF_SIZE); |
| 6722 | |
| 6723 | local_len -= BTR_EXTERN_FIELD_REF_SIZE; |
| 6724 | |
| 6725 | byte_val = mach_read_from_1(data + local_len + BTR_EXTERN_LEN); |
| 6726 | |
| 6727 | if (val) { |
| 6728 | byte_val &= ~BTR_EXTERN_OWNER_FLAG; |
| 6729 | } else { |
| 6730 | #if defined UNIV_DEBUG || defined UNIV_BLOB_LIGHT_DEBUG |
| 6731 | ut_a(!(byte_val & BTR_EXTERN_OWNER_FLAG)); |
| 6732 | #endif /* UNIV_DEBUG || UNIV_BLOB_LIGHT_DEBUG */ |
| 6733 | byte_val |= BTR_EXTERN_OWNER_FLAG; |
| 6734 | } |
| 6735 | |
| 6736 | if (page_zip) { |
| 6737 | mach_write_to_1(data + local_len + BTR_EXTERN_LEN, byte_val); |
| 6738 | page_zip_write_blob_ptr(page_zip, rec, index, offsets, i, mtr); |
| 6739 | } else if (mtr != NULL) { |
| 6740 | |
| 6741 | mlog_write_ulint(data + local_len + BTR_EXTERN_LEN, byte_val, |
| 6742 | MLOG_1BYTE, mtr); |
| 6743 | } else { |
| 6744 | mach_write_to_1(data + local_len + BTR_EXTERN_LEN, byte_val); |
| 6745 | } |
| 6746 | } |
| 6747 | |
| 6748 | /*******************************************************************//** |
| 6749 | Marks non-updated off-page fields as disowned by this record. The ownership |
| 6750 | must be transferred to the updated record which is inserted elsewhere in the |
| 6751 | index tree. In purge only the owner of externally stored field is allowed |
| 6752 | to free the field. */ |
| 6753 | void |
| 6754 | btr_cur_disown_inherited_fields( |
| 6755 | /*============================*/ |
| 6756 | page_zip_des_t* page_zip,/*!< in/out: compressed page whose uncompressed |
| 6757 | part will be updated, or NULL */ |
| 6758 | rec_t* rec, /*!< in/out: record in a clustered index */ |
| 6759 | dict_index_t* index, /*!< in: index of the page */ |
| 6760 | const ulint* offsets,/*!< in: array returned by rec_get_offsets() */ |
| 6761 | const upd_t* update, /*!< in: update vector */ |
| 6762 | mtr_t* mtr) /*!< in/out: mini-transaction */ |
| 6763 | { |
| 6764 | ulint i; |
| 6765 | |
| 6766 | ut_ad(rec_offs_validate(rec, index, offsets)); |
| 6767 | ut_ad(!rec_offs_comp(offsets) || !rec_get_node_ptr_flag(rec)); |
| 6768 | ut_ad(rec_offs_any_extern(offsets)); |
| 6769 | |
| 6770 | for (i = 0; i < rec_offs_n_fields(offsets); i++) { |
| 6771 | if (rec_offs_nth_extern(offsets, i) |
| 6772 | && !upd_get_field_by_field_no(update, i, false)) { |
| 6773 | btr_cur_set_ownership_of_extern_field( |
| 6774 | page_zip, rec, index, offsets, i, FALSE, mtr); |
| 6775 | } |
| 6776 | } |
| 6777 | } |
| 6778 | |
| 6779 | /*******************************************************************//** |
| 6780 | Marks all extern fields in a record as owned by the record. This function |
| 6781 | should be called if the delete mark of a record is removed: a not delete |
| 6782 | marked record always owns all its extern fields. */ |
| 6783 | static |
| 6784 | void |
| 6785 | btr_cur_unmark_extern_fields( |
| 6786 | /*=========================*/ |
| 6787 | page_zip_des_t* page_zip,/*!< in/out: compressed page whose uncompressed |
| 6788 | part will be updated, or NULL */ |
| 6789 | rec_t* rec, /*!< in/out: record in a clustered index */ |
| 6790 | dict_index_t* index, /*!< in: index of the page */ |
| 6791 | const ulint* offsets,/*!< in: array returned by rec_get_offsets() */ |
| 6792 | mtr_t* mtr) /*!< in: mtr, or NULL if not logged */ |
| 6793 | { |
| 6794 | ulint n; |
| 6795 | ulint i; |
| 6796 | |
| 6797 | ut_ad(!rec_offs_comp(offsets) || !rec_get_node_ptr_flag(rec)); |
| 6798 | n = rec_offs_n_fields(offsets); |
| 6799 | |
| 6800 | if (!rec_offs_any_extern(offsets)) { |
| 6801 | |
| 6802 | return; |
| 6803 | } |
| 6804 | |
| 6805 | for (i = 0; i < n; i++) { |
| 6806 | if (rec_offs_nth_extern(offsets, i)) { |
| 6807 | |
| 6808 | btr_cur_set_ownership_of_extern_field( |
| 6809 | page_zip, rec, index, offsets, i, TRUE, mtr); |
| 6810 | } |
| 6811 | } |
| 6812 | } |
| 6813 | |
| 6814 | /*******************************************************************//** |
| 6815 | Flags the data tuple fields that are marked as extern storage in the |
| 6816 | update vector. We use this function to remember which fields we must |
| 6817 | mark as extern storage in a record inserted for an update. |
| 6818 | @return number of flagged external columns */ |
| 6819 | ulint |
| 6820 | btr_push_update_extern_fields( |
| 6821 | /*==========================*/ |
| 6822 | dtuple_t* tuple, /*!< in/out: data tuple */ |
| 6823 | const upd_t* update, /*!< in: update vector */ |
| 6824 | mem_heap_t* heap) /*!< in: memory heap */ |
| 6825 | { |
| 6826 | ulint n_pushed = 0; |
| 6827 | ulint n; |
| 6828 | const upd_field_t* uf; |
| 6829 | |
| 6830 | uf = update->fields; |
| 6831 | n = upd_get_n_fields(update); |
| 6832 | |
| 6833 | for (; n--; uf++) { |
| 6834 | if (dfield_is_ext(&uf->new_val)) { |
| 6835 | dfield_t* field |
| 6836 | = dtuple_get_nth_field(tuple, uf->field_no); |
| 6837 | |
| 6838 | if (!dfield_is_ext(field)) { |
| 6839 | dfield_set_ext(field); |
| 6840 | n_pushed++; |
| 6841 | } |
| 6842 | |
| 6843 | switch (uf->orig_len) { |
| 6844 | byte* data; |
| 6845 | ulint len; |
| 6846 | byte* buf; |
| 6847 | case 0: |
| 6848 | break; |
| 6849 | case BTR_EXTERN_FIELD_REF_SIZE: |
| 6850 | /* Restore the original locally stored |
| 6851 | part of the column. In the undo log, |
| 6852 | InnoDB writes a longer prefix of externally |
| 6853 | stored columns, so that column prefixes |
| 6854 | in secondary indexes can be reconstructed. */ |
| 6855 | dfield_set_data(field, |
| 6856 | (byte*) dfield_get_data(field) |
| 6857 | + dfield_get_len(field) |
| 6858 | - BTR_EXTERN_FIELD_REF_SIZE, |
| 6859 | BTR_EXTERN_FIELD_REF_SIZE); |
| 6860 | dfield_set_ext(field); |
| 6861 | break; |
| 6862 | default: |
| 6863 | /* Reconstruct the original locally |
| 6864 | stored part of the column. The data |
| 6865 | will have to be copied. */ |
| 6866 | ut_a(uf->orig_len > BTR_EXTERN_FIELD_REF_SIZE); |
| 6867 | |
| 6868 | data = (byte*) dfield_get_data(field); |
| 6869 | len = dfield_get_len(field); |
| 6870 | |
| 6871 | buf = (byte*) mem_heap_alloc(heap, |
| 6872 | uf->orig_len); |
| 6873 | /* Copy the locally stored prefix. */ |
| 6874 | memcpy(buf, data, |
| 6875 | unsigned(uf->orig_len) |
| 6876 | - BTR_EXTERN_FIELD_REF_SIZE); |
| 6877 | /* Copy the BLOB pointer. */ |
| 6878 | memcpy(buf + unsigned(uf->orig_len) |
| 6879 | - BTR_EXTERN_FIELD_REF_SIZE, |
| 6880 | data + len - BTR_EXTERN_FIELD_REF_SIZE, |
| 6881 | BTR_EXTERN_FIELD_REF_SIZE); |
| 6882 | |
| 6883 | dfield_set_data(field, buf, uf->orig_len); |
| 6884 | dfield_set_ext(field); |
| 6885 | } |
| 6886 | } |
| 6887 | } |
| 6888 | |
| 6889 | return(n_pushed); |
| 6890 | } |
| 6891 | |
| 6892 | /*******************************************************************//** |
| 6893 | Returns the length of a BLOB part stored on the header page. |
| 6894 | @return part length */ |
| 6895 | static |
| 6896 | ulint |
| 6897 | btr_blob_get_part_len( |
| 6898 | /*==================*/ |
| 6899 | const byte* blob_header) /*!< in: blob header */ |
| 6900 | { |
| 6901 | return(mach_read_from_4(blob_header + BTR_BLOB_HDR_PART_LEN)); |
| 6902 | } |
| 6903 | |
| 6904 | /*******************************************************************//** |
| 6905 | Returns the page number where the next BLOB part is stored. |
| 6906 | @return page number or FIL_NULL if no more pages */ |
| 6907 | static |
| 6908 | ulint |
| 6909 | btr_blob_get_next_page_no( |
| 6910 | /*======================*/ |
| 6911 | const byte* blob_header) /*!< in: blob header */ |
| 6912 | { |
| 6913 | return(mach_read_from_4(blob_header + BTR_BLOB_HDR_NEXT_PAGE_NO)); |
| 6914 | } |
| 6915 | |
| 6916 | /*******************************************************************//** |
| 6917 | Deallocate a buffer block that was reserved for a BLOB part. */ |
| 6918 | static |
| 6919 | void |
| 6920 | btr_blob_free( |
| 6921 | /*==========*/ |
| 6922 | buf_block_t* block, /*!< in: buffer block */ |
| 6923 | ibool all, /*!< in: TRUE=remove also the compressed page |
| 6924 | if there is one */ |
| 6925 | mtr_t* mtr) /*!< in: mini-transaction to commit */ |
| 6926 | { |
| 6927 | buf_pool_t* buf_pool = buf_pool_from_block(block); |
| 6928 | ulint space = block->page.id.space(); |
| 6929 | ulint page_no = block->page.id.page_no(); |
| 6930 | |
| 6931 | ut_ad(mtr_memo_contains(mtr, block, MTR_MEMO_PAGE_X_FIX)); |
| 6932 | |
| 6933 | mtr_commit(mtr); |
| 6934 | |
| 6935 | buf_pool_mutex_enter(buf_pool); |
| 6936 | |
| 6937 | /* Only free the block if it is still allocated to |
| 6938 | the same file page. */ |
| 6939 | |
| 6940 | if (buf_block_get_state(block) |
| 6941 | == BUF_BLOCK_FILE_PAGE |
| 6942 | && block->page.id.space() == space |
| 6943 | && block->page.id.page_no() == page_no) { |
| 6944 | |
| 6945 | if (!buf_LRU_free_page(&block->page, all) |
| 6946 | && all && block->page.zip.data) { |
| 6947 | /* Attempt to deallocate the uncompressed page |
| 6948 | if the whole block cannot be deallocted. */ |
| 6949 | |
| 6950 | buf_LRU_free_page(&block->page, false); |
| 6951 | } |
| 6952 | } |
| 6953 | |
| 6954 | buf_pool_mutex_exit(buf_pool); |
| 6955 | } |
| 6956 | |
| 6957 | /** Helper class used while writing blob pages, during insert or update. */ |
| 6958 | struct btr_blob_log_check_t { |
| 6959 | /** Persistent cursor on a clusterex index record with blobs. */ |
| 6960 | btr_pcur_t* m_pcur; |
| 6961 | /** Mini transaction holding the latches for m_pcur */ |
| 6962 | mtr_t* m_mtr; |
| 6963 | /** rec_get_offsets(rec, index); offset of clust_rec */ |
| 6964 | const ulint* m_offsets; |
| 6965 | /** The block containing clustered record */ |
| 6966 | buf_block_t** m_block; |
| 6967 | /** The clustered record pointer */ |
| 6968 | rec_t** m_rec; |
| 6969 | /** The blob operation code */ |
| 6970 | enum blob_op m_op; |
| 6971 | |
| 6972 | /** Constructor |
| 6973 | @param[in] pcur persistent cursor on a clustered |
| 6974 | index record with blobs. |
| 6975 | @param[in] mtr mini-transaction holding latches for |
| 6976 | pcur. |
| 6977 | @param[in] offsets offsets of the clust_rec |
| 6978 | @param[in,out] block record block containing pcur record |
| 6979 | @param[in,out] rec the clustered record pointer |
| 6980 | @param[in] op the blob operation code */ |
| 6981 | btr_blob_log_check_t( |
| 6982 | btr_pcur_t* pcur, |
| 6983 | mtr_t* mtr, |
| 6984 | const ulint* offsets, |
| 6985 | buf_block_t** block, |
| 6986 | rec_t** rec, |
| 6987 | enum blob_op op) |
| 6988 | : m_pcur(pcur), |
| 6989 | m_mtr(mtr), |
| 6990 | m_offsets(offsets), |
| 6991 | m_block(block), |
| 6992 | m_rec(rec), |
| 6993 | m_op(op) |
| 6994 | { |
| 6995 | ut_ad(rec_offs_validate(*m_rec, m_pcur->index(), m_offsets)); |
| 6996 | ut_ad((*m_block)->frame == page_align(*m_rec)); |
| 6997 | ut_ad(*m_rec == btr_pcur_get_rec(m_pcur)); |
| 6998 | } |
| 6999 | |
| 7000 | /** Check if there is enough space in log file. Commit and re-start the |
| 7001 | mini transaction. */ |
| 7002 | void check() |
| 7003 | { |
| 7004 | dict_index_t* index = m_pcur->index(); |
| 7005 | ulint offs = 0; |
| 7006 | ulint page_no = ULINT_UNDEFINED; |
| 7007 | FlushObserver* observer = m_mtr->get_flush_observer(); |
| 7008 | |
| 7009 | if (m_op == BTR_STORE_INSERT_BULK) { |
| 7010 | offs = page_offset(*m_rec); |
| 7011 | page_no = page_get_page_no( |
| 7012 | buf_block_get_frame(*m_block)); |
| 7013 | |
| 7014 | buf_block_buf_fix_inc(*m_block, __FILE__, __LINE__); |
| 7015 | } else { |
| 7016 | btr_pcur_store_position(m_pcur, m_mtr); |
| 7017 | } |
| 7018 | m_mtr->commit(); |
| 7019 | |
| 7020 | DEBUG_SYNC_C("blob_write_middle"); |
| 7021 | |
| 7022 | log_free_check(); |
| 7023 | |
| 7024 | DEBUG_SYNC_C("blob_write_middle_after_check"); |
| 7025 | |
| 7026 | const mtr_log_t log_mode = m_mtr->get_log_mode(); |
| 7027 | m_mtr->start(); |
| 7028 | m_mtr->set_log_mode(log_mode); |
| 7029 | index->set_modified(*m_mtr); |
| 7030 | m_mtr->set_flush_observer(observer); |
| 7031 | |
| 7032 | if (m_op == BTR_STORE_INSERT_BULK) { |
| 7033 | mtr_x_lock(dict_index_get_lock(index), m_mtr); |
| 7034 | m_pcur->btr_cur.page_cur.block = btr_block_get( |
| 7035 | page_id_t(index->table->space->id, page_no), |
| 7036 | page_size_t(index->table->space->flags), |
| 7037 | RW_X_LATCH, index, m_mtr); |
| 7038 | m_pcur->btr_cur.page_cur.rec |
| 7039 | = m_pcur->btr_cur.page_cur.block->frame |
| 7040 | + offs; |
| 7041 | |
| 7042 | buf_block_buf_fix_dec(m_pcur->btr_cur.page_cur.block); |
| 7043 | } else { |
| 7044 | ut_ad(m_pcur->rel_pos == BTR_PCUR_ON); |
| 7045 | bool ret = btr_pcur_restore_position( |
| 7046 | BTR_MODIFY_LEAF | BTR_MODIFY_EXTERNAL, |
| 7047 | m_pcur, m_mtr); |
| 7048 | |
| 7049 | ut_a(ret); |
| 7050 | } |
| 7051 | |
| 7052 | *m_block = btr_pcur_get_block(m_pcur); |
| 7053 | *m_rec = btr_pcur_get_rec(m_pcur); |
| 7054 | |
| 7055 | rec_offs_make_valid(*m_rec, index, true, |
| 7056 | const_cast<ulint*>(m_offsets)); |
| 7057 | |
| 7058 | ut_ad(m_mtr->memo_contains_page_flagged( |
| 7059 | *m_rec, |
| 7060 | MTR_MEMO_PAGE_X_FIX | MTR_MEMO_PAGE_SX_FIX)); |
| 7061 | |
| 7062 | ut_ad(mtr_memo_contains_flagged(m_mtr, |
| 7063 | dict_index_get_lock(index), |
| 7064 | MTR_MEMO_SX_LOCK | MTR_MEMO_X_LOCK)); |
| 7065 | } |
| 7066 | }; |
| 7067 | |
| 7068 | /*******************************************************************//** |
| 7069 | Stores the fields in big_rec_vec to the tablespace and puts pointers to |
| 7070 | them in rec. The extern flags in rec will have to be set beforehand. |
| 7071 | The fields are stored on pages allocated from leaf node |
| 7072 | file segment of the index tree. |
| 7073 | |
| 7074 | TODO: If the allocation extends the tablespace, it will not be redo logged, in |
| 7075 | any mini-transaction. Tablespace extension should be redo-logged, so that |
| 7076 | recovery will not fail when the big_rec was written to the extended portion of |
| 7077 | the file, in case the file was somehow truncated in the crash. |
| 7078 | |
| 7079 | @return DB_SUCCESS or DB_OUT_OF_FILE_SPACE */ |
| 7080 | dberr_t |
| 7081 | btr_store_big_rec_extern_fields( |
| 7082 | /*============================*/ |
| 7083 | btr_pcur_t* pcur, /*!< in/out: a persistent cursor. if |
| 7084 | btr_mtr is restarted, then this can |
| 7085 | be repositioned. */ |
| 7086 | ulint* offsets, /*!< in/out: rec_get_offsets() on |
| 7087 | pcur. the "external storage" flags |
| 7088 | in offsets will correctly correspond |
| 7089 | to rec when this function returns */ |
| 7090 | const big_rec_t*big_rec_vec, /*!< in: vector containing fields |
| 7091 | to be stored externally */ |
| 7092 | mtr_t* btr_mtr, /*!< in/out: mtr containing the |
| 7093 | latches to the clustered index. can be |
| 7094 | committed and restarted. */ |
| 7095 | enum blob_op op) /*! in: operation code */ |
| 7096 | { |
| 7097 | ulint rec_page_no; |
| 7098 | byte* field_ref; |
| 7099 | ulint extern_len; |
| 7100 | ulint store_len; |
| 7101 | ulint page_no; |
| 7102 | ulint space_id; |
| 7103 | ulint prev_page_no; |
| 7104 | ulint hint_page_no; |
| 7105 | ulint i; |
| 7106 | mtr_t mtr; |
| 7107 | mtr_t mtr_bulk; |
| 7108 | mem_heap_t* heap = NULL; |
| 7109 | page_zip_des_t* page_zip; |
| 7110 | z_stream c_stream; |
| 7111 | dberr_t error = DB_SUCCESS; |
| 7112 | dict_index_t* index = pcur->index(); |
| 7113 | buf_block_t* rec_block = btr_pcur_get_block(pcur); |
| 7114 | rec_t* rec = btr_pcur_get_rec(pcur); |
| 7115 | |
| 7116 | ut_ad(rec_offs_validate(rec, index, offsets)); |
| 7117 | ut_ad(rec_offs_any_extern(offsets)); |
| 7118 | ut_ad(mtr_memo_contains_flagged(btr_mtr, dict_index_get_lock(index), |
| 7119 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK)); |
| 7120 | ut_ad(mtr_memo_contains(btr_mtr, rec_block, MTR_MEMO_PAGE_X_FIX)); |
| 7121 | ut_ad(buf_block_get_frame(rec_block) == page_align(rec)); |
| 7122 | ut_a(dict_index_is_clust(index)); |
| 7123 | |
| 7124 | ut_a(dict_table_page_size(index->table) |
| 7125 | .equals_to(rec_block->page.size)); |
| 7126 | |
| 7127 | btr_blob_log_check_t redo_log(pcur, btr_mtr, offsets, &rec_block, |
| 7128 | &rec, op); |
| 7129 | page_zip = buf_block_get_page_zip(rec_block); |
| 7130 | space_id = rec_block->page.id.space(); |
| 7131 | rec_page_no = rec_block->page.id.page_no(); |
| 7132 | ut_a(fil_page_index_page_check(page_align(rec)) |
| 7133 | || op == BTR_STORE_INSERT_BULK); |
| 7134 | |
| 7135 | if (page_zip) { |
| 7136 | int err; |
| 7137 | |
| 7138 | /* Zlib deflate needs 128 kilobytes for the default |
| 7139 | window size, plus 512 << memLevel, plus a few |
| 7140 | kilobytes for small objects. We use reduced memLevel |
| 7141 | to limit the memory consumption, and preallocate the |
| 7142 | heap, hoping to avoid memory fragmentation. */ |
| 7143 | heap = mem_heap_create(250000); |
| 7144 | page_zip_set_alloc(&c_stream, heap); |
| 7145 | |
| 7146 | err = deflateInit2(&c_stream, int(page_zip_level), |
| 7147 | Z_DEFLATED, 15, 7, Z_DEFAULT_STRATEGY); |
| 7148 | ut_a(err == Z_OK); |
| 7149 | } |
| 7150 | |
| 7151 | #if defined UNIV_DEBUG || defined UNIV_BLOB_LIGHT_DEBUG |
| 7152 | /* All pointers to externally stored columns in the record |
| 7153 | must either be zero or they must be pointers to inherited |
| 7154 | columns, owned by this record or an earlier record version. */ |
| 7155 | for (i = 0; i < big_rec_vec->n_fields; i++) { |
| 7156 | field_ref = btr_rec_get_field_ref( |
| 7157 | rec, offsets, big_rec_vec->fields[i].field_no); |
| 7158 | |
| 7159 | ut_a(!(field_ref[BTR_EXTERN_LEN] & BTR_EXTERN_OWNER_FLAG)); |
| 7160 | /* Either this must be an update in place, |
| 7161 | or the BLOB must be inherited, or the BLOB pointer |
| 7162 | must be zero (will be written in this function). */ |
| 7163 | ut_a(op == BTR_STORE_UPDATE |
| 7164 | || (field_ref[BTR_EXTERN_LEN] & BTR_EXTERN_INHERITED_FLAG) |
| 7165 | || !memcmp(field_ref, field_ref_zero, |
| 7166 | BTR_EXTERN_FIELD_REF_SIZE)); |
| 7167 | } |
| 7168 | #endif /* UNIV_DEBUG || UNIV_BLOB_LIGHT_DEBUG */ |
| 7169 | |
| 7170 | const page_size_t page_size(dict_table_page_size(index->table)); |
| 7171 | |
| 7172 | /* Space available in compressed page to carry blob data */ |
| 7173 | const ulint payload_size_zip = page_size.physical() |
| 7174 | - FIL_PAGE_DATA; |
| 7175 | |
| 7176 | /* Space available in uncompressed page to carry blob data */ |
| 7177 | const ulint payload_size = page_size.physical() |
| 7178 | - FIL_PAGE_DATA - BTR_BLOB_HDR_SIZE - FIL_PAGE_DATA_END; |
| 7179 | |
| 7180 | /* We have to create a file segment to the tablespace |
| 7181 | for each field and put the pointer to the field in rec */ |
| 7182 | |
| 7183 | for (i = 0; i < big_rec_vec->n_fields; i++) { |
| 7184 | const ulint field_no = big_rec_vec->fields[i].field_no; |
| 7185 | |
| 7186 | field_ref = btr_rec_get_field_ref(rec, offsets, field_no); |
| 7187 | #if defined UNIV_DEBUG || defined UNIV_BLOB_LIGHT_DEBUG |
| 7188 | /* A zero BLOB pointer should have been initially inserted. */ |
| 7189 | ut_a(!memcmp(field_ref, field_ref_zero, |
| 7190 | BTR_EXTERN_FIELD_REF_SIZE)); |
| 7191 | #endif /* UNIV_DEBUG || UNIV_BLOB_LIGHT_DEBUG */ |
| 7192 | extern_len = big_rec_vec->fields[i].len; |
| 7193 | UNIV_MEM_ASSERT_RW(big_rec_vec->fields[i].data, |
| 7194 | extern_len); |
| 7195 | |
| 7196 | ut_a(extern_len > 0); |
| 7197 | |
| 7198 | prev_page_no = FIL_NULL; |
| 7199 | |
| 7200 | if (page_zip) { |
| 7201 | int err = deflateReset(&c_stream); |
| 7202 | ut_a(err == Z_OK); |
| 7203 | |
| 7204 | c_stream.next_in = (Bytef*) |
| 7205 | big_rec_vec->fields[i].data; |
| 7206 | c_stream.avail_in = static_cast<uInt>(extern_len); |
| 7207 | } |
| 7208 | |
| 7209 | for (ulint blob_npages = 0;; ++blob_npages) { |
| 7210 | buf_block_t* block; |
| 7211 | page_t* page; |
| 7212 | const ulint commit_freq = 4; |
| 7213 | ulint r_extents; |
| 7214 | |
| 7215 | ut_ad(page_align(field_ref) == page_align(rec)); |
| 7216 | |
| 7217 | if (!(blob_npages % commit_freq)) { |
| 7218 | |
| 7219 | redo_log.check(); |
| 7220 | |
| 7221 | field_ref = btr_rec_get_field_ref( |
| 7222 | rec, offsets, field_no); |
| 7223 | |
| 7224 | page_zip = buf_block_get_page_zip(rec_block); |
| 7225 | rec_page_no = rec_block->page.id.page_no(); |
| 7226 | } |
| 7227 | |
| 7228 | mtr.start(); |
| 7229 | index->set_modified(mtr); |
| 7230 | mtr.set_log_mode(btr_mtr->get_log_mode()); |
| 7231 | mtr.set_flush_observer(btr_mtr->get_flush_observer()); |
| 7232 | |
| 7233 | buf_page_get(rec_block->page.id, |
| 7234 | rec_block->page.size, RW_X_LATCH, &mtr); |
| 7235 | |
| 7236 | if (prev_page_no == FIL_NULL) { |
| 7237 | hint_page_no = 1 + rec_page_no; |
| 7238 | } else { |
| 7239 | hint_page_no = prev_page_no + 1; |
| 7240 | } |
| 7241 | |
| 7242 | mtr_t *alloc_mtr; |
| 7243 | |
| 7244 | if (op == BTR_STORE_INSERT_BULK) { |
| 7245 | mtr_bulk.start(); |
| 7246 | mtr_bulk.set_spaces(mtr); |
| 7247 | alloc_mtr = &mtr_bulk; |
| 7248 | } else { |
| 7249 | alloc_mtr = &mtr; |
| 7250 | } |
| 7251 | |
| 7252 | if (!fsp_reserve_free_extents(&r_extents, |
| 7253 | index->table->space, 1, |
| 7254 | FSP_BLOB, alloc_mtr, |
| 7255 | 1)) { |
| 7256 | |
| 7257 | alloc_mtr->commit(); |
| 7258 | error = DB_OUT_OF_FILE_SPACE; |
| 7259 | goto func_exit; |
| 7260 | } |
| 7261 | |
| 7262 | block = btr_page_alloc(index, hint_page_no, FSP_NO_DIR, |
| 7263 | 0, alloc_mtr, &mtr); |
| 7264 | |
| 7265 | index->table->space->release_free_extents(r_extents); |
| 7266 | |
| 7267 | if (op == BTR_STORE_INSERT_BULK) { |
| 7268 | mtr_bulk.commit(); |
| 7269 | } |
| 7270 | |
| 7271 | ut_a(block != NULL); |
| 7272 | |
| 7273 | page_no = block->page.id.page_no(); |
| 7274 | page = buf_block_get_frame(block); |
| 7275 | |
| 7276 | if (prev_page_no != FIL_NULL) { |
| 7277 | buf_block_t* prev_block; |
| 7278 | page_t* prev_page; |
| 7279 | |
| 7280 | prev_block = buf_page_get( |
| 7281 | page_id_t(space_id, prev_page_no), |
| 7282 | rec_block->page.size, |
| 7283 | RW_X_LATCH, &mtr); |
| 7284 | |
| 7285 | buf_block_dbg_add_level(prev_block, |
| 7286 | SYNC_EXTERN_STORAGE); |
| 7287 | prev_page = buf_block_get_frame(prev_block); |
| 7288 | |
| 7289 | if (page_zip) { |
| 7290 | mlog_write_ulint( |
| 7291 | prev_page + FIL_PAGE_NEXT, |
| 7292 | page_no, MLOG_4BYTES, &mtr); |
| 7293 | memcpy(buf_block_get_page_zip( |
| 7294 | prev_block) |
| 7295 | ->data + FIL_PAGE_NEXT, |
| 7296 | prev_page + FIL_PAGE_NEXT, 4); |
| 7297 | } else { |
| 7298 | mlog_write_ulint( |
| 7299 | prev_page + FIL_PAGE_DATA |
| 7300 | + BTR_BLOB_HDR_NEXT_PAGE_NO, |
| 7301 | page_no, MLOG_4BYTES, &mtr); |
| 7302 | } |
| 7303 | |
| 7304 | } else if (dict_index_is_online_ddl(index)) { |
| 7305 | row_log_table_blob_alloc(index, page_no); |
| 7306 | } |
| 7307 | |
| 7308 | if (page_zip) { |
| 7309 | int err; |
| 7310 | page_zip_des_t* blob_page_zip; |
| 7311 | |
| 7312 | /* Write FIL_PAGE_TYPE to the redo log |
| 7313 | separately, before logging any other |
| 7314 | changes to the page, so that the debug |
| 7315 | assertions in |
| 7316 | recv_parse_or_apply_log_rec_body() can |
| 7317 | be made simpler. Before InnoDB Plugin |
| 7318 | 1.0.4, the initialization of |
| 7319 | FIL_PAGE_TYPE was logged as part of |
| 7320 | the mlog_log_string() below. */ |
| 7321 | |
| 7322 | mlog_write_ulint(page + FIL_PAGE_TYPE, |
| 7323 | prev_page_no == FIL_NULL |
| 7324 | ? FIL_PAGE_TYPE_ZBLOB |
| 7325 | : FIL_PAGE_TYPE_ZBLOB2, |
| 7326 | MLOG_2BYTES, &mtr); |
| 7327 | |
| 7328 | c_stream.next_out = page |
| 7329 | + FIL_PAGE_DATA; |
| 7330 | c_stream.avail_out = static_cast<uInt>( |
| 7331 | payload_size_zip); |
| 7332 | |
| 7333 | err = deflate(&c_stream, Z_FINISH); |
| 7334 | ut_a(err == Z_OK || err == Z_STREAM_END); |
| 7335 | ut_a(err == Z_STREAM_END |
| 7336 | || c_stream.avail_out == 0); |
| 7337 | |
| 7338 | /* Write the "next BLOB page" pointer */ |
| 7339 | mlog_write_ulint(page + FIL_PAGE_NEXT, |
| 7340 | FIL_NULL, MLOG_4BYTES, &mtr); |
| 7341 | /* Initialize the unused "prev page" pointer */ |
| 7342 | mlog_write_ulint(page + FIL_PAGE_PREV, |
| 7343 | FIL_NULL, MLOG_4BYTES, &mtr); |
| 7344 | /* Write a back pointer to the record |
| 7345 | into the otherwise unused area. This |
| 7346 | information could be useful in |
| 7347 | debugging. Later, we might want to |
| 7348 | implement the possibility to relocate |
| 7349 | BLOB pages. Then, we would need to be |
| 7350 | able to adjust the BLOB pointer in the |
| 7351 | record. We do not store the heap |
| 7352 | number of the record, because it can |
| 7353 | change in page_zip_reorganize() or |
| 7354 | btr_page_reorganize(). However, also |
| 7355 | the page number of the record may |
| 7356 | change when B-tree nodes are split or |
| 7357 | merged. |
| 7358 | NOTE: FIL_PAGE_FILE_FLUSH_LSN space is |
| 7359 | used by R-tree index for a Split Sequence |
| 7360 | Number */ |
| 7361 | ut_ad(!dict_index_is_spatial(index)); |
| 7362 | |
| 7363 | mlog_write_ulint(page |
| 7364 | + FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION, |
| 7365 | space_id, |
| 7366 | MLOG_4BYTES, &mtr); |
| 7367 | mlog_write_ulint(page |
| 7368 | + FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION + 4, |
| 7369 | rec_page_no, |
| 7370 | MLOG_4BYTES, &mtr); |
| 7371 | |
| 7372 | /* Zero out the unused part of the page. */ |
| 7373 | memset(page + page_zip_get_size(page_zip) |
| 7374 | - c_stream.avail_out, |
| 7375 | 0, c_stream.avail_out); |
| 7376 | mlog_log_string(page |
| 7377 | + FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION, |
| 7378 | page_zip_get_size(page_zip) |
| 7379 | - FIL_PAGE_FILE_FLUSH_LSN_OR_KEY_VERSION, |
| 7380 | &mtr); |
| 7381 | /* Copy the page to compressed storage, |
| 7382 | because it will be flushed to disk |
| 7383 | from there. */ |
| 7384 | blob_page_zip = buf_block_get_page_zip(block); |
| 7385 | ut_ad(blob_page_zip); |
| 7386 | ut_ad(page_zip_get_size(blob_page_zip) |
| 7387 | == page_zip_get_size(page_zip)); |
| 7388 | memcpy(blob_page_zip->data, page, |
| 7389 | page_zip_get_size(page_zip)); |
| 7390 | |
| 7391 | if (err == Z_OK && prev_page_no != FIL_NULL) { |
| 7392 | |
| 7393 | goto next_zip_page; |
| 7394 | } |
| 7395 | |
| 7396 | if (err == Z_STREAM_END) { |
| 7397 | mach_write_to_4(field_ref |
| 7398 | + BTR_EXTERN_LEN, 0); |
| 7399 | mach_write_to_4(field_ref |
| 7400 | + BTR_EXTERN_LEN + 4, |
| 7401 | c_stream.total_in); |
| 7402 | } else { |
| 7403 | memset(field_ref + BTR_EXTERN_LEN, |
| 7404 | 0, 8); |
| 7405 | } |
| 7406 | |
| 7407 | if (prev_page_no == FIL_NULL) { |
| 7408 | ut_ad(blob_npages == 0); |
| 7409 | mach_write_to_4(field_ref |
| 7410 | + BTR_EXTERN_SPACE_ID, |
| 7411 | space_id); |
| 7412 | |
| 7413 | mach_write_to_4(field_ref |
| 7414 | + BTR_EXTERN_PAGE_NO, |
| 7415 | page_no); |
| 7416 | |
| 7417 | mach_write_to_4(field_ref |
| 7418 | + BTR_EXTERN_OFFSET, |
| 7419 | FIL_PAGE_NEXT); |
| 7420 | } |
| 7421 | |
| 7422 | /* We compress a page when finish bulk insert.*/ |
| 7423 | if (op != BTR_STORE_INSERT_BULK) { |
| 7424 | page_zip_write_blob_ptr( |
| 7425 | page_zip, rec, index, offsets, |
| 7426 | field_no, &mtr); |
| 7427 | } |
| 7428 | |
| 7429 | next_zip_page: |
| 7430 | prev_page_no = page_no; |
| 7431 | |
| 7432 | /* Commit mtr and release the |
| 7433 | uncompressed page frame to save memory. */ |
| 7434 | btr_blob_free(block, FALSE, &mtr); |
| 7435 | |
| 7436 | if (err == Z_STREAM_END) { |
| 7437 | break; |
| 7438 | } |
| 7439 | } else { |
| 7440 | mlog_write_ulint(page + FIL_PAGE_TYPE, |
| 7441 | FIL_PAGE_TYPE_BLOB, |
| 7442 | MLOG_2BYTES, &mtr); |
| 7443 | |
| 7444 | if (extern_len > payload_size) { |
| 7445 | store_len = payload_size; |
| 7446 | } else { |
| 7447 | store_len = extern_len; |
| 7448 | } |
| 7449 | |
| 7450 | mlog_write_string(page + FIL_PAGE_DATA |
| 7451 | + BTR_BLOB_HDR_SIZE, |
| 7452 | (const byte*) |
| 7453 | big_rec_vec->fields[i].data |
| 7454 | + big_rec_vec->fields[i].len |
| 7455 | - extern_len, |
| 7456 | store_len, &mtr); |
| 7457 | mlog_write_ulint(page + FIL_PAGE_DATA |
| 7458 | + BTR_BLOB_HDR_PART_LEN, |
| 7459 | store_len, MLOG_4BYTES, &mtr); |
| 7460 | mlog_write_ulint(page + FIL_PAGE_DATA |
| 7461 | + BTR_BLOB_HDR_NEXT_PAGE_NO, |
| 7462 | FIL_NULL, MLOG_4BYTES, &mtr); |
| 7463 | |
| 7464 | extern_len -= store_len; |
| 7465 | |
| 7466 | mlog_write_ulint(field_ref + BTR_EXTERN_LEN, 0, |
| 7467 | MLOG_4BYTES, &mtr); |
| 7468 | mlog_write_ulint(field_ref |
| 7469 | + BTR_EXTERN_LEN + 4, |
| 7470 | big_rec_vec->fields[i].len |
| 7471 | - extern_len, |
| 7472 | MLOG_4BYTES, &mtr); |
| 7473 | |
| 7474 | if (prev_page_no == FIL_NULL) { |
| 7475 | ut_ad(blob_npages == 0); |
| 7476 | mlog_write_ulint(field_ref |
| 7477 | + BTR_EXTERN_SPACE_ID, |
| 7478 | space_id, MLOG_4BYTES, |
| 7479 | &mtr); |
| 7480 | |
| 7481 | mlog_write_ulint(field_ref |
| 7482 | + BTR_EXTERN_PAGE_NO, |
| 7483 | page_no, MLOG_4BYTES, |
| 7484 | &mtr); |
| 7485 | |
| 7486 | mlog_write_ulint(field_ref |
| 7487 | + BTR_EXTERN_OFFSET, |
| 7488 | FIL_PAGE_DATA, |
| 7489 | MLOG_4BYTES, |
| 7490 | &mtr); |
| 7491 | } |
| 7492 | |
| 7493 | prev_page_no = page_no; |
| 7494 | |
| 7495 | mtr.commit(); |
| 7496 | |
| 7497 | if (extern_len == 0) { |
| 7498 | break; |
| 7499 | } |
| 7500 | } |
| 7501 | } |
| 7502 | |
| 7503 | DBUG_EXECUTE_IF("btr_store_big_rec_extern", |
| 7504 | error = DB_OUT_OF_FILE_SPACE; |
| 7505 | goto func_exit;); |
| 7506 | |
| 7507 | rec_offs_make_nth_extern(offsets, field_no); |
| 7508 | } |
| 7509 | |
| 7510 | func_exit: |
| 7511 | if (page_zip) { |
| 7512 | deflateEnd(&c_stream); |
| 7513 | } |
| 7514 | |
| 7515 | if (heap != NULL) { |
| 7516 | mem_heap_free(heap); |
| 7517 | } |
| 7518 | |
| 7519 | #if defined UNIV_DEBUG || defined UNIV_BLOB_LIGHT_DEBUG |
| 7520 | /* All pointers to externally stored columns in the record |
| 7521 | must be valid. */ |
| 7522 | for (i = 0; i < rec_offs_n_fields(offsets); i++) { |
| 7523 | if (!rec_offs_nth_extern(offsets, i)) { |
| 7524 | continue; |
| 7525 | } |
| 7526 | |
| 7527 | field_ref = btr_rec_get_field_ref(rec, offsets, i); |
| 7528 | |
| 7529 | /* The pointer must not be zero if the operation |
| 7530 | succeeded. */ |
| 7531 | ut_a(0 != memcmp(field_ref, field_ref_zero, |
| 7532 | BTR_EXTERN_FIELD_REF_SIZE) |
| 7533 | || error != DB_SUCCESS); |
| 7534 | /* The column must not be disowned by this record. */ |
| 7535 | ut_a(!(field_ref[BTR_EXTERN_LEN] & BTR_EXTERN_OWNER_FLAG)); |
| 7536 | } |
| 7537 | #endif /* UNIV_DEBUG || UNIV_BLOB_LIGHT_DEBUG */ |
| 7538 | return(error); |
| 7539 | } |
| 7540 | |
| 7541 | /*******************************************************************//** |
| 7542 | Check the FIL_PAGE_TYPE on an uncompressed BLOB page. */ |
| 7543 | static |
| 7544 | void |
| 7545 | btr_check_blob_fil_page_type( |
| 7546 | /*=========================*/ |
| 7547 | ulint space_id, /*!< in: space id */ |
| 7548 | ulint page_no, /*!< in: page number */ |
| 7549 | const page_t* page, /*!< in: page */ |
| 7550 | ibool read) /*!< in: TRUE=read, FALSE=purge */ |
| 7551 | { |
| 7552 | ulint type = fil_page_get_type(page); |
| 7553 | |
| 7554 | ut_a(space_id == page_get_space_id(page)); |
| 7555 | ut_a(page_no == page_get_page_no(page)); |
| 7556 | |
| 7557 | if (UNIV_UNLIKELY(type != FIL_PAGE_TYPE_BLOB)) { |
| 7558 | ulint flags = fil_space_get_flags(space_id); |
| 7559 | |
| 7560 | #ifndef UNIV_DEBUG /* Improve debug test coverage */ |
| 7561 | if (!DICT_TF_HAS_ATOMIC_BLOBS(flags)) { |
| 7562 | /* Old versions of InnoDB did not initialize |
| 7563 | FIL_PAGE_TYPE on BLOB pages. Do not print |
| 7564 | anything about the type mismatch when reading |
| 7565 | a BLOB page that may be from old versions. */ |
| 7566 | return; |
| 7567 | } |
| 7568 | #endif /* !UNIV_DEBUG */ |
| 7569 | |
| 7570 | ib::fatal() << "FIL_PAGE_TYPE="<< type |
| 7571 | << " on BLOB "<< (read ? "read": "purge") |
| 7572 | << " space "<< space_id << " page "<< page_no |
| 7573 | << " flags "<< flags; |
| 7574 | } |
| 7575 | } |
| 7576 | |
| 7577 | /*******************************************************************//** |
| 7578 | Frees the space in an externally stored field to the file space |
| 7579 | management if the field in data is owned by the externally stored field, |
| 7580 | in a rollback we may have the additional condition that the field must |
| 7581 | not be inherited. */ |
| 7582 | void |
| 7583 | btr_free_externally_stored_field( |
| 7584 | /*=============================*/ |
| 7585 | dict_index_t* index, /*!< in: index of the data, the index |
| 7586 | tree MUST be X-latched; if the tree |
| 7587 | height is 1, then also the root page |
| 7588 | must be X-latched! (this is relevant |
| 7589 | in the case this function is called |
| 7590 | from purge where 'data' is located on |
| 7591 | an undo log page, not an index |
| 7592 | page) */ |
| 7593 | byte* field_ref, /*!< in/out: field reference */ |
| 7594 | const rec_t* rec, /*!< in: record containing field_ref, for |
| 7595 | page_zip_write_blob_ptr(), or NULL */ |
| 7596 | const ulint* offsets, /*!< in: rec_get_offsets(rec, index), |
| 7597 | or NULL */ |
| 7598 | page_zip_des_t* page_zip, /*!< in: compressed page corresponding |
| 7599 | to rec, or NULL if rec == NULL */ |
| 7600 | ulint i, /*!< in: field number of field_ref; |
| 7601 | ignored if rec == NULL */ |
| 7602 | bool rollback, /*!< in: performing rollback? */ |
| 7603 | mtr_t* local_mtr) /*!< in: mtr |
| 7604 | containing the latch to data an an |
| 7605 | X-latch to the index tree */ |
| 7606 | { |
| 7607 | page_t* page; |
| 7608 | const ulint space_id = mach_read_from_4( |
| 7609 | field_ref + BTR_EXTERN_SPACE_ID); |
| 7610 | const ulint start_page = mach_read_from_4( |
| 7611 | field_ref + BTR_EXTERN_PAGE_NO); |
| 7612 | ulint page_no; |
| 7613 | ulint next_page_no; |
| 7614 | mtr_t mtr; |
| 7615 | |
| 7616 | ut_ad(index->is_primary()); |
| 7617 | ut_ad(mtr_memo_contains_flagged(local_mtr, dict_index_get_lock(index), |
| 7618 | MTR_MEMO_X_LOCK | MTR_MEMO_SX_LOCK)); |
| 7619 | ut_ad(mtr_memo_contains_page(local_mtr, field_ref, |
| 7620 | MTR_MEMO_PAGE_X_FIX)); |
| 7621 | ut_ad(!rec || rec_offs_validate(rec, index, offsets)); |
| 7622 | ut_ad(!rec || field_ref == btr_rec_get_field_ref(rec, offsets, i)); |
| 7623 | ut_ad(local_mtr->is_named_space( |
| 7624 | page_get_space_id(page_align(field_ref)))); |
| 7625 | |
| 7626 | if (UNIV_UNLIKELY(!memcmp(field_ref, field_ref_zero, |
| 7627 | BTR_EXTERN_FIELD_REF_SIZE))) { |
| 7628 | /* In the rollback, we may encounter a clustered index |
| 7629 | record with some unwritten off-page columns. There is |
| 7630 | nothing to free then. */ |
| 7631 | ut_a(rollback); |
| 7632 | return; |
| 7633 | } |
| 7634 | |
| 7635 | ut_ad(!(mach_read_from_4(field_ref + BTR_EXTERN_LEN) |
| 7636 | & ~((BTR_EXTERN_OWNER_FLAG |
| 7637 | | BTR_EXTERN_INHERITED_FLAG) << 24))); |
| 7638 | ut_ad(space_id == index->table->space->id); |
| 7639 | |
| 7640 | const page_size_t ext_page_size(dict_table_page_size(index->table)); |
| 7641 | const page_size_t& rec_page_size(rec == NULL |
| 7642 | ? univ_page_size |
| 7643 | : ext_page_size); |
| 7644 | if (rec == NULL) { |
| 7645 | /* This is a call from row_purge_upd_exist_or_extern(). */ |
| 7646 | ut_ad(!page_zip); |
| 7647 | } |
| 7648 | |
| 7649 | for (;;) { |
| 7650 | #ifdef UNIV_DEBUG |
| 7651 | buf_block_t* rec_block; |
| 7652 | #endif /* UNIV_DEBUG */ |
| 7653 | buf_block_t* ext_block; |
| 7654 | |
| 7655 | mtr_start(&mtr); |
| 7656 | mtr.set_spaces(*local_mtr); |
| 7657 | mtr.set_log_mode(local_mtr->get_log_mode()); |
| 7658 | |
| 7659 | ut_ad(!index->table->is_temporary() |
| 7660 | || local_mtr->get_log_mode() == MTR_LOG_NO_REDO); |
| 7661 | |
| 7662 | const page_t* p = page_align(field_ref); |
| 7663 | |
| 7664 | const page_id_t page_id(page_get_space_id(p), |
| 7665 | page_get_page_no(p)); |
| 7666 | |
| 7667 | #ifdef UNIV_DEBUG |
| 7668 | rec_block = |
| 7669 | #endif /* UNIV_DEBUG */ |
| 7670 | buf_page_get(page_id, rec_page_size, RW_X_LATCH, &mtr); |
| 7671 | |
| 7672 | buf_block_dbg_add_level(rec_block, SYNC_NO_ORDER_CHECK); |
| 7673 | page_no = mach_read_from_4(field_ref + BTR_EXTERN_PAGE_NO); |
| 7674 | |
| 7675 | if (/* There is no external storage data */ |
| 7676 | page_no == FIL_NULL |
| 7677 | /* This field does not own the externally stored field */ |
| 7678 | || (mach_read_from_1(field_ref + BTR_EXTERN_LEN) |
| 7679 | & BTR_EXTERN_OWNER_FLAG) |
| 7680 | /* Rollback and inherited field */ |
| 7681 | || (rollback |
| 7682 | && (mach_read_from_1(field_ref + BTR_EXTERN_LEN) |
| 7683 | & BTR_EXTERN_INHERITED_FLAG))) { |
| 7684 | |
| 7685 | /* Do not free */ |
| 7686 | mtr_commit(&mtr); |
| 7687 | |
| 7688 | return; |
| 7689 | } |
| 7690 | |
| 7691 | if (page_no == start_page && dict_index_is_online_ddl(index)) { |
| 7692 | row_log_table_blob_free(index, start_page); |
| 7693 | } |
| 7694 | |
| 7695 | ext_block = buf_page_get( |
| 7696 | page_id_t(space_id, page_no), ext_page_size, |
| 7697 | RW_X_LATCH, &mtr); |
| 7698 | |
| 7699 | buf_block_dbg_add_level(ext_block, SYNC_EXTERN_STORAGE); |
| 7700 | page = buf_block_get_frame(ext_block); |
| 7701 | |
| 7702 | if (ext_page_size.is_compressed()) { |
| 7703 | /* Note that page_zip will be NULL |
| 7704 | in row_purge_upd_exist_or_extern(). */ |
| 7705 | switch (fil_page_get_type(page)) { |
| 7706 | case FIL_PAGE_TYPE_ZBLOB: |
| 7707 | case FIL_PAGE_TYPE_ZBLOB2: |
| 7708 | break; |
| 7709 | default: |
| 7710 | ut_error; |
| 7711 | } |
| 7712 | next_page_no = mach_read_from_4(page + FIL_PAGE_NEXT); |
| 7713 | |
| 7714 | btr_page_free_low(index, ext_block, 0, |
| 7715 | true, &mtr); |
| 7716 | |
| 7717 | if (page_zip != NULL) { |
| 7718 | mach_write_to_4(field_ref + BTR_EXTERN_PAGE_NO, |
| 7719 | next_page_no); |
| 7720 | mach_write_to_4(field_ref + BTR_EXTERN_LEN + 4, |
| 7721 | 0); |
| 7722 | page_zip_write_blob_ptr(page_zip, rec, index, |
| 7723 | offsets, i, &mtr); |
| 7724 | } else { |
| 7725 | mlog_write_ulint(field_ref |
| 7726 | + BTR_EXTERN_PAGE_NO, |
| 7727 | next_page_no, |
| 7728 | MLOG_4BYTES, &mtr); |
| 7729 | mlog_write_ulint(field_ref |
| 7730 | + BTR_EXTERN_LEN + 4, 0, |
| 7731 | MLOG_4BYTES, &mtr); |
| 7732 | } |
| 7733 | } else { |
| 7734 | ut_a(!page_zip); |
| 7735 | btr_check_blob_fil_page_type(space_id, page_no, page, |
| 7736 | FALSE); |
| 7737 | |
| 7738 | next_page_no = mach_read_from_4( |
| 7739 | page + FIL_PAGE_DATA |
| 7740 | + BTR_BLOB_HDR_NEXT_PAGE_NO); |
| 7741 | |
| 7742 | /* We must supply the page level (= 0) as an argument |
| 7743 | because we did not store it on the page (we save the |
| 7744 | space overhead from an index page header. */ |
| 7745 | btr_page_free_low(index, ext_block, 0, |
| 7746 | true, &mtr); |
| 7747 | |
| 7748 | mlog_write_ulint(field_ref + BTR_EXTERN_PAGE_NO, |
| 7749 | next_page_no, |
| 7750 | MLOG_4BYTES, &mtr); |
| 7751 | /* Zero out the BLOB length. If the server |
| 7752 | crashes during the execution of this function, |
| 7753 | trx_rollback_all_recovered() could |
| 7754 | dereference the half-deleted BLOB, fetching a |
| 7755 | wrong prefix for the BLOB. */ |
| 7756 | mlog_write_ulint(field_ref + BTR_EXTERN_LEN + 4, |
| 7757 | 0, |
| 7758 | MLOG_4BYTES, &mtr); |
| 7759 | } |
| 7760 | |
| 7761 | /* Commit mtr and release the BLOB block to save memory. */ |
| 7762 | btr_blob_free(ext_block, TRUE, &mtr); |
| 7763 | } |
| 7764 | } |
| 7765 | |
| 7766 | /***********************************************************//** |
| 7767 | Frees the externally stored fields for a record. */ |
| 7768 | static |
| 7769 | void |
| 7770 | btr_rec_free_externally_stored_fields( |
| 7771 | /*==================================*/ |
| 7772 | dict_index_t* index, /*!< in: index of the data, the index |
| 7773 | tree MUST be X-latched */ |
| 7774 | rec_t* rec, /*!< in/out: record */ |
| 7775 | const ulint* offsets,/*!< in: rec_get_offsets(rec, index) */ |
| 7776 | page_zip_des_t* page_zip,/*!< in: compressed page whose uncompressed |
| 7777 | part will be updated, or NULL */ |
| 7778 | bool rollback,/*!< in: performing rollback? */ |
| 7779 | mtr_t* mtr) /*!< in: mini-transaction handle which contains |
| 7780 | an X-latch to record page and to the index |
| 7781 | tree */ |
| 7782 | { |
| 7783 | ulint n_fields; |
| 7784 | ulint i; |
| 7785 | |
| 7786 | ut_ad(rec_offs_validate(rec, index, offsets)); |
| 7787 | ut_ad(mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_X_FIX)); |
| 7788 | ut_ad(index->is_primary()); |
| 7789 | ut_ad(page_rec_is_leaf(rec)); |
| 7790 | /* Free possible externally stored fields in the record */ |
| 7791 | |
| 7792 | ut_ad(dict_table_is_comp(index->table) == !!rec_offs_comp(offsets)); |
| 7793 | n_fields = rec_offs_n_fields(offsets); |
| 7794 | |
| 7795 | for (i = 0; i < n_fields; i++) { |
| 7796 | if (rec_offs_nth_extern(offsets, i)) { |
| 7797 | btr_free_externally_stored_field( |
| 7798 | index, btr_rec_get_field_ref(rec, offsets, i), |
| 7799 | rec, offsets, page_zip, i, rollback, mtr); |
| 7800 | } |
| 7801 | } |
| 7802 | } |
| 7803 | |
| 7804 | /***********************************************************//** |
| 7805 | Frees the externally stored fields for a record, if the field is mentioned |
| 7806 | in the update vector. */ |
| 7807 | static |
| 7808 | void |
| 7809 | btr_rec_free_updated_extern_fields( |
| 7810 | /*===============================*/ |
| 7811 | dict_index_t* index, /*!< in: index of rec; the index tree MUST be |
| 7812 | X-latched */ |
| 7813 | rec_t* rec, /*!< in/out: record */ |
| 7814 | page_zip_des_t* page_zip,/*!< in: compressed page whose uncompressed |
| 7815 | part will be updated, or NULL */ |
| 7816 | const ulint* offsets,/*!< in: rec_get_offsets(rec, index) */ |
| 7817 | const upd_t* update, /*!< in: update vector */ |
| 7818 | bool rollback,/*!< in: performing rollback? */ |
| 7819 | mtr_t* mtr) /*!< in: mini-transaction handle which contains |
| 7820 | an X-latch to record page and to the tree */ |
| 7821 | { |
| 7822 | ulint n_fields; |
| 7823 | ulint i; |
| 7824 | |
| 7825 | ut_ad(rec_offs_validate(rec, index, offsets)); |
| 7826 | ut_ad(mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_X_FIX)); |
| 7827 | |
| 7828 | /* Free possible externally stored fields in the record */ |
| 7829 | |
| 7830 | n_fields = upd_get_n_fields(update); |
| 7831 | |
| 7832 | for (i = 0; i < n_fields; i++) { |
| 7833 | const upd_field_t* ufield = upd_get_nth_field(update, i); |
| 7834 | |
| 7835 | if (rec_offs_nth_extern(offsets, ufield->field_no)) { |
| 7836 | ulint len; |
| 7837 | byte* data = rec_get_nth_field( |
| 7838 | rec, offsets, ufield->field_no, &len); |
| 7839 | ut_a(len >= BTR_EXTERN_FIELD_REF_SIZE); |
| 7840 | |
| 7841 | btr_free_externally_stored_field( |
| 7842 | index, data + len - BTR_EXTERN_FIELD_REF_SIZE, |
| 7843 | rec, offsets, page_zip, |
| 7844 | ufield->field_no, rollback, mtr); |
| 7845 | } |
| 7846 | } |
| 7847 | } |
| 7848 | |
| 7849 | /*******************************************************************//** |
| 7850 | Copies the prefix of an uncompressed BLOB. The clustered index record |
| 7851 | that points to this BLOB must be protected by a lock or a page latch. |
| 7852 | @return number of bytes written to buf */ |
| 7853 | static |
| 7854 | ulint |
| 7855 | btr_copy_blob_prefix( |
| 7856 | /*=================*/ |
| 7857 | byte* buf, /*!< out: the externally stored part of |
| 7858 | the field, or a prefix of it */ |
| 7859 | ulint len, /*!< in: length of buf, in bytes */ |
| 7860 | ulint space_id,/*!< in: space id of the BLOB pages */ |
| 7861 | ulint page_no,/*!< in: page number of the first BLOB page */ |
| 7862 | ulint offset) /*!< in: offset on the first BLOB page */ |
| 7863 | { |
| 7864 | ulint copied_len = 0; |
| 7865 | |
| 7866 | for (;;) { |
| 7867 | mtr_t mtr; |
| 7868 | buf_block_t* block; |
| 7869 | const page_t* page; |
| 7870 | const byte* blob_header; |
| 7871 | ulint part_len; |
| 7872 | ulint copy_len; |
| 7873 | |
| 7874 | mtr_start(&mtr); |
| 7875 | |
| 7876 | block = buf_page_get(page_id_t(space_id, page_no), |
| 7877 | univ_page_size, RW_S_LATCH, &mtr); |
| 7878 | buf_block_dbg_add_level(block, SYNC_EXTERN_STORAGE); |
| 7879 | page = buf_block_get_frame(block); |
| 7880 | |
| 7881 | btr_check_blob_fil_page_type(space_id, page_no, page, TRUE); |
| 7882 | |
| 7883 | blob_header = page + offset; |
| 7884 | part_len = btr_blob_get_part_len(blob_header); |
| 7885 | copy_len = ut_min(part_len, len - copied_len); |
| 7886 | |
| 7887 | memcpy(buf + copied_len, |
| 7888 | blob_header + BTR_BLOB_HDR_SIZE, copy_len); |
| 7889 | copied_len += copy_len; |
| 7890 | |
| 7891 | page_no = btr_blob_get_next_page_no(blob_header); |
| 7892 | |
| 7893 | mtr_commit(&mtr); |
| 7894 | |
| 7895 | if (page_no == FIL_NULL || copy_len != part_len) { |
| 7896 | UNIV_MEM_ASSERT_RW(buf, copied_len); |
| 7897 | return(copied_len); |
| 7898 | } |
| 7899 | |
| 7900 | /* On other BLOB pages except the first the BLOB header |
| 7901 | always is at the page data start: */ |
| 7902 | |
| 7903 | offset = FIL_PAGE_DATA; |
| 7904 | |
| 7905 | ut_ad(copied_len <= len); |
| 7906 | } |
| 7907 | } |
| 7908 | |
| 7909 | /** Copies the prefix of a compressed BLOB. |
| 7910 | The clustered index record that points to this BLOB must be protected |
| 7911 | by a lock or a page latch. |
| 7912 | @param[out] buf the externally stored part of the field, |
| 7913 | or a prefix of it |
| 7914 | @param[in] len length of buf, in bytes |
| 7915 | @param[in] page_size compressed BLOB page size |
| 7916 | @param[in] space_id space id of the BLOB pages |
| 7917 | @param[in] offset offset on the first BLOB page |
| 7918 | @return number of bytes written to buf */ |
| 7919 | static |
| 7920 | ulint |
| 7921 | btr_copy_zblob_prefix( |
| 7922 | byte* buf, |
| 7923 | ulint len, |
| 7924 | const page_size_t& page_size, |
| 7925 | ulint space_id, |
| 7926 | ulint page_no, |
| 7927 | ulint offset) |
| 7928 | { |
| 7929 | ulint page_type = FIL_PAGE_TYPE_ZBLOB; |
| 7930 | mem_heap_t* heap; |
| 7931 | int err; |
| 7932 | z_stream d_stream; |
| 7933 | |
| 7934 | d_stream.next_out = buf; |
| 7935 | d_stream.avail_out = static_cast<uInt>(len); |
| 7936 | d_stream.next_in = Z_NULL; |
| 7937 | d_stream.avail_in = 0; |
| 7938 | |
| 7939 | /* Zlib inflate needs 32 kilobytes for the default |
| 7940 | window size, plus a few kilobytes for small objects. */ |
| 7941 | heap = mem_heap_create(40000); |
| 7942 | page_zip_set_alloc(&d_stream, heap); |
| 7943 | |
| 7944 | ut_ad(page_size.is_compressed()); |
| 7945 | ut_ad(space_id); |
| 7946 | |
| 7947 | err = inflateInit(&d_stream); |
| 7948 | ut_a(err == Z_OK); |
| 7949 | |
| 7950 | for (;;) { |
| 7951 | buf_page_t* bpage; |
| 7952 | ulint next_page_no; |
| 7953 | |
| 7954 | /* There is no latch on bpage directly. Instead, |
| 7955 | bpage is protected by the B-tree page latch that |
| 7956 | is being held on the clustered index record, or, |
| 7957 | in row_merge_copy_blobs(), by an exclusive table lock. */ |
| 7958 | bpage = buf_page_get_zip(page_id_t(space_id, page_no), |
| 7959 | page_size); |
| 7960 | |
| 7961 | if (UNIV_UNLIKELY(!bpage)) { |
| 7962 | ib::error() << "Cannot load compressed BLOB " |
| 7963 | << page_id_t(space_id, page_no); |
| 7964 | goto func_exit; |
| 7965 | } |
| 7966 | |
| 7967 | if (UNIV_UNLIKELY |
| 7968 | (fil_page_get_type(bpage->zip.data) != page_type)) { |
| 7969 | |
| 7970 | ib::error() << "Unexpected type " |
| 7971 | << fil_page_get_type(bpage->zip.data) |
| 7972 | << " of compressed BLOB page " |
| 7973 | << page_id_t(space_id, page_no); |
| 7974 | |
| 7975 | ut_ad(0); |
| 7976 | goto end_of_blob; |
| 7977 | } |
| 7978 | |
| 7979 | next_page_no = mach_read_from_4(bpage->zip.data + offset); |
| 7980 | |
| 7981 | if (UNIV_LIKELY(offset == FIL_PAGE_NEXT)) { |
| 7982 | /* When the BLOB begins at page header, |
| 7983 | the compressed data payload does not |
| 7984 | immediately follow the next page pointer. */ |
| 7985 | offset = FIL_PAGE_DATA; |
| 7986 | } else { |
| 7987 | offset += 4; |
| 7988 | } |
| 7989 | |
| 7990 | d_stream.next_in = bpage->zip.data + offset; |
| 7991 | d_stream.avail_in = static_cast<uInt>(page_size.physical() |
| 7992 | - offset); |
| 7993 | |
| 7994 | err = inflate(&d_stream, Z_NO_FLUSH); |
| 7995 | switch (err) { |
| 7996 | case Z_OK: |
| 7997 | if (!d_stream.avail_out) { |
| 7998 | goto end_of_blob; |
| 7999 | } |
| 8000 | break; |
| 8001 | case Z_STREAM_END: |
| 8002 | if (next_page_no == FIL_NULL) { |
| 8003 | goto end_of_blob; |
| 8004 | } |
| 8005 | /* fall through */ |
| 8006 | default: |
| 8007 | inflate_error: |
| 8008 | ib::error() << "inflate() of compressed BLOB page " |
| 8009 | << page_id_t(space_id, page_no) |
| 8010 | << " returned "<< err |
| 8011 | << " ("<< d_stream.msg << ")"; |
| 8012 | |
| 8013 | case Z_BUF_ERROR: |
| 8014 | goto end_of_blob; |
| 8015 | } |
| 8016 | |
| 8017 | if (next_page_no == FIL_NULL) { |
| 8018 | if (!d_stream.avail_in) { |
| 8019 | ib::error() |
| 8020 | << "Unexpected end of compressed " |
| 8021 | << "BLOB page " |
| 8022 | << page_id_t(space_id, page_no); |
| 8023 | } else { |
| 8024 | err = inflate(&d_stream, Z_FINISH); |
| 8025 | switch (err) { |
| 8026 | case Z_STREAM_END: |
| 8027 | case Z_BUF_ERROR: |
| 8028 | break; |
| 8029 | default: |
| 8030 | goto inflate_error; |
| 8031 | } |
| 8032 | } |
| 8033 | |
| 8034 | end_of_blob: |
| 8035 | buf_page_release_zip(bpage); |
| 8036 | goto func_exit; |
| 8037 | } |
| 8038 | |
| 8039 | buf_page_release_zip(bpage); |
| 8040 | |
| 8041 | /* On other BLOB pages except the first |
| 8042 | the BLOB header always is at the page header: */ |
| 8043 | |
| 8044 | page_no = next_page_no; |
| 8045 | offset = FIL_PAGE_NEXT; |
| 8046 | page_type = FIL_PAGE_TYPE_ZBLOB2; |
| 8047 | } |
| 8048 | |
| 8049 | func_exit: |
| 8050 | inflateEnd(&d_stream); |
| 8051 | mem_heap_free(heap); |
| 8052 | UNIV_MEM_ASSERT_RW(buf, d_stream.total_out); |
| 8053 | return(d_stream.total_out); |
| 8054 | } |
| 8055 | |
| 8056 | /** Copies the prefix of an externally stored field of a record. |
| 8057 | The clustered index record that points to this BLOB must be protected |
| 8058 | by a lock or a page latch. |
| 8059 | @param[out] buf the externally stored part of the |
| 8060 | field, or a prefix of it |
| 8061 | @param[in] len length of buf, in bytes |
| 8062 | @param[in] page_size BLOB page size |
| 8063 | @param[in] space_id space id of the first BLOB page |
| 8064 | @param[in] page_no page number of the first BLOB page |
| 8065 | @param[in] offset offset on the first BLOB page |
| 8066 | @return number of bytes written to buf */ |
| 8067 | static |
| 8068 | ulint |
| 8069 | btr_copy_externally_stored_field_prefix_low( |
| 8070 | byte* buf, |
| 8071 | ulint len, |
| 8072 | const page_size_t& page_size, |
| 8073 | ulint space_id, |
| 8074 | ulint page_no, |
| 8075 | ulint offset) |
| 8076 | { |
| 8077 | if (len == 0) { |
| 8078 | return(0); |
| 8079 | } |
| 8080 | |
| 8081 | if (page_size.is_compressed()) { |
| 8082 | return(btr_copy_zblob_prefix(buf, len, page_size, |
| 8083 | space_id, page_no, offset)); |
| 8084 | } else { |
| 8085 | ut_ad(page_size.equals_to(univ_page_size)); |
| 8086 | return(btr_copy_blob_prefix(buf, len, space_id, |
| 8087 | page_no, offset)); |
| 8088 | } |
| 8089 | } |
| 8090 | |
| 8091 | /** Copies the prefix of an externally stored field of a record. |
| 8092 | The clustered index record must be protected by a lock or a page latch. |
| 8093 | @param[out] buf the field, or a prefix of it |
| 8094 | @param[in] len length of buf, in bytes |
| 8095 | @param[in] page_size BLOB page size |
| 8096 | @param[in] data 'internally' stored part of the field |
| 8097 | containing also the reference to the external part; must be protected by |
| 8098 | a lock or a page latch |
| 8099 | @param[in] local_len length of data, in bytes |
| 8100 | @return the length of the copied field, or 0 if the column was being |
| 8101 | or has been deleted */ |
| 8102 | ulint |
| 8103 | btr_copy_externally_stored_field_prefix( |
| 8104 | byte* buf, |
| 8105 | ulint len, |
| 8106 | const page_size_t& page_size, |
| 8107 | const byte* data, |
| 8108 | ulint local_len) |
| 8109 | { |
| 8110 | ulint space_id; |
| 8111 | ulint page_no; |
| 8112 | ulint offset; |
| 8113 | |
| 8114 | ut_a(local_len >= BTR_EXTERN_FIELD_REF_SIZE); |
| 8115 | |
| 8116 | local_len -= BTR_EXTERN_FIELD_REF_SIZE; |
| 8117 | |
| 8118 | if (UNIV_UNLIKELY(local_len >= len)) { |
| 8119 | memcpy(buf, data, len); |
| 8120 | return(len); |
| 8121 | } |
| 8122 | |
| 8123 | memcpy(buf, data, local_len); |
| 8124 | data += local_len; |
| 8125 | |
| 8126 | ut_a(memcmp(data, field_ref_zero, BTR_EXTERN_FIELD_REF_SIZE)); |
| 8127 | |
| 8128 | if (!mach_read_from_4(data + BTR_EXTERN_LEN + 4)) { |
| 8129 | /* The externally stored part of the column has been |
| 8130 | (partially) deleted. Signal the half-deleted BLOB |
| 8131 | to the caller. */ |
| 8132 | |
| 8133 | return(0); |
| 8134 | } |
| 8135 | |
| 8136 | space_id = mach_read_from_4(data + BTR_EXTERN_SPACE_ID); |
| 8137 | |
| 8138 | page_no = mach_read_from_4(data + BTR_EXTERN_PAGE_NO); |
| 8139 | |
| 8140 | offset = mach_read_from_4(data + BTR_EXTERN_OFFSET); |
| 8141 | |
| 8142 | return(local_len |
| 8143 | + btr_copy_externally_stored_field_prefix_low(buf + local_len, |
| 8144 | len - local_len, |
| 8145 | page_size, |
| 8146 | space_id, page_no, |
| 8147 | offset)); |
| 8148 | } |
| 8149 | |
| 8150 | /** Copies an externally stored field of a record to mem heap. |
| 8151 | The clustered index record must be protected by a lock or a page latch. |
| 8152 | @param[out] len length of the whole field |
| 8153 | @param[in] data 'internally' stored part of the field |
| 8154 | containing also the reference to the external part; must be protected by |
| 8155 | a lock or a page latch |
| 8156 | @param[in] page_size BLOB page size |
| 8157 | @param[in] local_len length of data |
| 8158 | @param[in,out] heap mem heap |
| 8159 | @return the whole field copied to heap */ |
| 8160 | byte* |
| 8161 | btr_copy_externally_stored_field( |
| 8162 | ulint* len, |
| 8163 | const byte* data, |
| 8164 | const page_size_t& page_size, |
| 8165 | ulint local_len, |
| 8166 | mem_heap_t* heap) |
| 8167 | { |
| 8168 | ulint space_id; |
| 8169 | ulint page_no; |
| 8170 | ulint offset; |
| 8171 | ulint extern_len; |
| 8172 | byte* buf; |
| 8173 | |
| 8174 | ut_a(local_len >= BTR_EXTERN_FIELD_REF_SIZE); |
| 8175 | |
| 8176 | local_len -= BTR_EXTERN_FIELD_REF_SIZE; |
| 8177 | |
| 8178 | space_id = mach_read_from_4(data + local_len + BTR_EXTERN_SPACE_ID); |
| 8179 | |
| 8180 | page_no = mach_read_from_4(data + local_len + BTR_EXTERN_PAGE_NO); |
| 8181 | |
| 8182 | offset = mach_read_from_4(data + local_len + BTR_EXTERN_OFFSET); |
| 8183 | |
| 8184 | /* Currently a BLOB cannot be bigger than 4 GB; we |
| 8185 | leave the 4 upper bytes in the length field unused */ |
| 8186 | |
| 8187 | extern_len = mach_read_from_4(data + local_len + BTR_EXTERN_LEN + 4); |
| 8188 | |
| 8189 | buf = (byte*) mem_heap_alloc(heap, local_len + extern_len); |
| 8190 | |
| 8191 | memcpy(buf, data, local_len); |
| 8192 | *len = local_len |
| 8193 | + btr_copy_externally_stored_field_prefix_low(buf + local_len, |
| 8194 | extern_len, |
| 8195 | page_size, |
| 8196 | space_id, |
| 8197 | page_no, offset); |
| 8198 | |
| 8199 | return(buf); |
| 8200 | } |
| 8201 | |
| 8202 | /** Copies an externally stored field of a record to mem heap. |
| 8203 | @param[in] rec record in a clustered index; must be |
| 8204 | protected by a lock or a page latch |
| 8205 | @param[in] offset array returned by rec_get_offsets() |
| 8206 | @param[in] page_size BLOB page size |
| 8207 | @param[in] no field number |
| 8208 | @param[out] len length of the field |
| 8209 | @param[in,out] heap mem heap |
| 8210 | @return the field copied to heap, or NULL if the field is incomplete */ |
| 8211 | byte* |
| 8212 | btr_rec_copy_externally_stored_field( |
| 8213 | const rec_t* rec, |
| 8214 | const ulint* offsets, |
| 8215 | const page_size_t& page_size, |
| 8216 | ulint no, |
| 8217 | ulint* len, |
| 8218 | mem_heap_t* heap) |
| 8219 | { |
| 8220 | ulint local_len; |
| 8221 | const byte* data; |
| 8222 | |
| 8223 | ut_a(rec_offs_nth_extern(offsets, no)); |
| 8224 | |
| 8225 | /* An externally stored field can contain some initial |
| 8226 | data from the field, and in the last 20 bytes it has the |
| 8227 | space id, page number, and offset where the rest of the |
| 8228 | field data is stored, and the data length in addition to |
| 8229 | the data stored locally. We may need to store some data |
| 8230 | locally to get the local record length above the 128 byte |
| 8231 | limit so that field offsets are stored in two bytes, and |
| 8232 | the extern bit is available in those two bytes. */ |
| 8233 | |
| 8234 | data = rec_get_nth_field(rec, offsets, no, &local_len); |
| 8235 | |
| 8236 | ut_a(local_len >= BTR_EXTERN_FIELD_REF_SIZE); |
| 8237 | |
| 8238 | if (UNIV_UNLIKELY |
| 8239 | (!memcmp(data + local_len - BTR_EXTERN_FIELD_REF_SIZE, |
| 8240 | field_ref_zero, BTR_EXTERN_FIELD_REF_SIZE))) { |
| 8241 | /* The externally stored field was not written yet. |
| 8242 | This record should only be seen by |
| 8243 | recv_recovery_rollback_active() or any |
| 8244 | TRX_ISO_READ_UNCOMMITTED transactions. */ |
| 8245 | return(NULL); |
| 8246 | } |
| 8247 | |
| 8248 | return(btr_copy_externally_stored_field(len, data, |
| 8249 | page_size, local_len, heap)); |
| 8250 | } |
| 8251 |
Generated on 2018-May-24 from project MariaDB revision 20180524
Powered by 
Code Browser 2.1
Generator usage only permitted with license.