1 | /***************************************************************************** |
2 | |
3 | Copyright (c) 1995, 2017, Oracle and/or its affiliates. All Rights Reserved. |
4 | Copyright (c) 2015, 2017, MariaDB Corporation. |
5 | |
6 | This program is free software; you can redistribute it and/or modify it under |
7 | the terms of the GNU General Public License as published by the Free Software |
8 | Foundation; version 2 of the License. |
9 | |
10 | This program is distributed in the hope that it will be useful, but WITHOUT |
11 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS |
12 | FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. |
13 | |
14 | You should have received a copy of the GNU General Public License along with |
15 | this program; if not, write to the Free Software Foundation, Inc., |
16 | 51 Franklin Street, Suite 500, Boston, MA 02110-1335 USA |
17 | |
18 | *****************************************************************************/ |
19 | |
20 | /**************************************************//** |
21 | @file buf/buf0rea.cc |
22 | The database buffer read |
23 | |
24 | Created 11/5/1995 Heikki Tuuri |
25 | *******************************************************/ |
26 | |
27 | #include "ha_prototypes.h" |
28 | #include <mysql/service_thd_wait.h> |
29 | |
30 | #include "buf0rea.h" |
31 | #include "fil0fil.h" |
32 | #include "mtr0mtr.h" |
33 | #include "buf0buf.h" |
34 | #include "buf0flu.h" |
35 | #include "buf0lru.h" |
36 | #include "buf0dblwr.h" |
37 | #include "ibuf0ibuf.h" |
38 | #include "log0recv.h" |
39 | #include "trx0sys.h" |
40 | #include "os0file.h" |
41 | #include "srv0start.h" |
42 | #include "srv0srv.h" |
43 | |
44 | /** There must be at least this many pages in buf_pool in the area to start |
45 | a random read-ahead */ |
46 | #define BUF_READ_AHEAD_RANDOM_THRESHOLD(b) \ |
47 | (5 + BUF_READ_AHEAD_AREA(b) / 8) |
48 | |
49 | /** If there are buf_pool->curr_size per the number below pending reads, then |
50 | read-ahead is not done: this is to prevent flooding the buffer pool with |
51 | i/o-fixed buffer blocks */ |
52 | #define BUF_READ_AHEAD_PEND_LIMIT 2 |
53 | |
54 | /********************************************************************//** |
55 | Unfixes the pages, unlatches the page, |
56 | removes it from page_hash and removes it from LRU. */ |
57 | static |
58 | void |
59 | buf_read_page_handle_error( |
60 | /*=======================*/ |
61 | buf_page_t* bpage) /*!< in: pointer to the block */ |
62 | { |
63 | buf_pool_t* buf_pool = buf_pool_from_bpage(bpage); |
64 | const bool uncompressed = (buf_page_get_state(bpage) |
65 | == BUF_BLOCK_FILE_PAGE); |
66 | |
67 | /* First unfix and release lock on the bpage */ |
68 | buf_pool_mutex_enter(buf_pool); |
69 | mutex_enter(buf_page_get_mutex(bpage)); |
70 | ut_ad(buf_page_get_io_fix(bpage) == BUF_IO_READ); |
71 | ut_ad(bpage->buf_fix_count == 0); |
72 | |
73 | /* Set BUF_IO_NONE before we remove the block from LRU list */ |
74 | buf_page_set_io_fix(bpage, BUF_IO_NONE); |
75 | |
76 | if (uncompressed) { |
77 | rw_lock_x_unlock_gen( |
78 | &((buf_block_t*) bpage)->lock, |
79 | BUF_IO_READ); |
80 | } |
81 | |
82 | mutex_exit(buf_page_get_mutex(bpage)); |
83 | |
84 | /* remove the block from LRU list */ |
85 | buf_LRU_free_one_page(bpage); |
86 | |
87 | ut_ad(buf_pool->n_pend_reads > 0); |
88 | buf_pool->n_pend_reads--; |
89 | |
90 | buf_pool_mutex_exit(buf_pool); |
91 | } |
92 | |
93 | /** Low-level function which reads a page asynchronously from a file to the |
94 | buffer buf_pool if it is not already there, in which case does nothing. |
95 | Sets the io_fix flag and sets an exclusive lock on the buffer frame. The |
96 | flag is cleared and the x-lock released by an i/o-handler thread. |
97 | |
98 | @param[out] err DB_SUCCESS, DB_TABLESPACE_DELETED or |
99 | DB_TABLESPACE_TRUNCATED if we are trying |
100 | to read from a non-existent tablespace, a |
101 | tablespace which is just now being dropped, |
102 | or a tablespace which is truncated |
103 | @param[in] sync true if synchronous aio is desired |
104 | @param[in] type IO type, SIMULATED, IGNORE_MISSING |
105 | @param[in] mode BUF_READ_IBUF_PAGES_ONLY, ..., |
106 | @param[in] page_id page id |
107 | @param[in] unzip true=request uncompressed page |
108 | @param[in] ignore_missing_space true=ignore missing space when reading |
109 | @return 1 if a read request was queued, 0 if the page already resided |
110 | in buf_pool, or if the page is in the doublewrite buffer blocks in |
111 | which case it is never read into the pool, or if the tablespace does |
112 | not exist or is being dropped */ |
113 | static |
114 | ulint |
115 | buf_read_page_low( |
116 | dberr_t* err, |
117 | bool sync, |
118 | ulint type, |
119 | ulint mode, |
120 | const page_id_t& page_id, |
121 | const page_size_t& page_size, |
122 | bool unzip, |
123 | bool ignore_missing_space = false) |
124 | { |
125 | buf_page_t* bpage; |
126 | |
127 | *err = DB_SUCCESS; |
128 | |
129 | if (page_id.space() == TRX_SYS_SPACE |
130 | && buf_dblwr_page_inside(page_id.page_no())) { |
131 | |
132 | ib::error() << "Trying to read doublewrite buffer page " |
133 | << page_id; |
134 | return(0); |
135 | } |
136 | |
137 | if (ibuf_bitmap_page(page_id, page_size) || trx_sys_hdr_page(page_id)) { |
138 | |
139 | /* Trx sys header is so low in the latching order that we play |
140 | safe and do not leave the i/o-completion to an asynchronous |
141 | i/o-thread. Ibuf bitmap pages must always be read with |
142 | syncronous i/o, to make sure they do not get involved in |
143 | thread deadlocks. */ |
144 | |
145 | sync = true; |
146 | } |
147 | |
148 | /* The following call will also check if the tablespace does not exist |
149 | or is being dropped; if we succeed in initing the page in the buffer |
150 | pool for read, then DISCARD cannot proceed until the read has |
151 | completed */ |
152 | bpage = buf_page_init_for_read(err, mode, page_id, page_size, unzip); |
153 | |
154 | if (bpage == NULL) { |
155 | |
156 | return(0); |
157 | } |
158 | |
159 | DBUG_LOG("ib_buf" , |
160 | "read page " << page_id << " size=" << page_size.physical() |
161 | << " unzip=" << unzip << ',' << (sync ? "sync" : "async" )); |
162 | |
163 | ut_ad(buf_page_in_file(bpage)); |
164 | |
165 | if (sync) { |
166 | thd_wait_begin(NULL, THD_WAIT_DISKIO); |
167 | } |
168 | |
169 | void* dst; |
170 | |
171 | if (page_size.is_compressed()) { |
172 | dst = bpage->zip.data; |
173 | } else { |
174 | ut_a(buf_page_get_state(bpage) == BUF_BLOCK_FILE_PAGE); |
175 | |
176 | dst = ((buf_block_t*) bpage)->frame; |
177 | } |
178 | |
179 | DBUG_EXECUTE_IF( |
180 | "innodb_invalid_read_after_truncate" , |
181 | if (fil_space_t* space = fil_space_acquire(page_id.space())) { |
182 | if (!strcmp(space->name, "test/t1" ) |
183 | && page_id.page_no() == space->size - 1) { |
184 | type = 0; |
185 | sync = true; |
186 | } |
187 | space->release(); |
188 | }); |
189 | |
190 | IORequest request(type | IORequest::READ); |
191 | |
192 | *err = fil_io( |
193 | request, sync, page_id, page_size, 0, page_size.physical(), |
194 | dst, bpage, ignore_missing_space); |
195 | |
196 | if (sync) { |
197 | thd_wait_end(NULL); |
198 | } |
199 | |
200 | if (*err != DB_SUCCESS) { |
201 | if (*err == DB_TABLESPACE_TRUNCATED) { |
202 | /* Remove the page which is outside the |
203 | truncated tablespace bounds when recovering |
204 | from a crash happened during a truncation */ |
205 | buf_read_page_handle_error(bpage); |
206 | if (recv_recovery_on) { |
207 | mutex_enter(&recv_sys->mutex); |
208 | ut_ad(recv_sys->n_addrs > 0); |
209 | recv_sys->n_addrs--; |
210 | mutex_exit(&recv_sys->mutex); |
211 | } |
212 | return(0); |
213 | } else if (IORequest::ignore_missing(type) |
214 | || *err == DB_TABLESPACE_DELETED) { |
215 | buf_read_page_handle_error(bpage); |
216 | return(0); |
217 | } |
218 | |
219 | ut_error; |
220 | } |
221 | |
222 | if (sync) { |
223 | /* The i/o is already completed when we arrive from |
224 | fil_read */ |
225 | *err = buf_page_io_complete(bpage); |
226 | |
227 | if (*err != DB_SUCCESS) { |
228 | return(0); |
229 | } |
230 | } |
231 | |
232 | return(1); |
233 | } |
234 | |
235 | /** Applies a random read-ahead in buf_pool if there are at least a threshold |
236 | value of accessed pages from the random read-ahead area. Does not read any |
237 | page, not even the one at the position (space, offset), if the read-ahead |
238 | mechanism is not activated. NOTE 1: the calling thread may own latches on |
239 | pages: to avoid deadlocks this function must be written such that it cannot |
240 | end up waiting for these latches! NOTE 2: the calling thread must want |
241 | access to the page given: this rule is set to prevent unintended read-aheads |
242 | performed by ibuf routines, a situation which could result in a deadlock if |
243 | the OS does not support asynchronous i/o. |
244 | @param[in] page_id page id of a page which the current thread |
245 | wants to access |
246 | @param[in] page_size page size |
247 | @param[in] inside_ibuf TRUE if we are inside ibuf routine |
248 | @return number of page read requests issued; NOTE that if we read ibuf |
249 | pages, it may happen that the page at the given page number does not |
250 | get read even if we return a positive value! */ |
251 | ulint |
252 | buf_read_ahead_random( |
253 | const page_id_t& page_id, |
254 | const page_size_t& page_size, |
255 | ibool inside_ibuf) |
256 | { |
257 | buf_pool_t* buf_pool = buf_pool_get(page_id); |
258 | ulint recent_blocks = 0; |
259 | ulint ibuf_mode; |
260 | ulint count; |
261 | ulint low, high; |
262 | dberr_t err = DB_SUCCESS; |
263 | ulint i; |
264 | const ulint buf_read_ahead_random_area |
265 | = BUF_READ_AHEAD_AREA(buf_pool); |
266 | |
267 | if (!srv_random_read_ahead) { |
268 | /* Disabled by user */ |
269 | return(0); |
270 | } |
271 | |
272 | if (srv_startup_is_before_trx_rollback_phase) { |
273 | /* No read-ahead to avoid thread deadlocks */ |
274 | return(0); |
275 | } |
276 | |
277 | if (ibuf_bitmap_page(page_id, page_size) || trx_sys_hdr_page(page_id)) { |
278 | |
279 | /* If it is an ibuf bitmap page or trx sys hdr, we do |
280 | no read-ahead, as that could break the ibuf page access |
281 | order */ |
282 | |
283 | return(0); |
284 | } |
285 | |
286 | low = (page_id.page_no() / buf_read_ahead_random_area) |
287 | * buf_read_ahead_random_area; |
288 | |
289 | high = (page_id.page_no() / buf_read_ahead_random_area + 1) |
290 | * buf_read_ahead_random_area; |
291 | |
292 | /* Remember the tablespace version before we ask the tablespace size |
293 | below: if DISCARD + IMPORT changes the actual .ibd file meanwhile, we |
294 | do not try to read outside the bounds of the tablespace! */ |
295 | if (fil_space_t* space = fil_space_acquire(page_id.space())) { |
296 | |
297 | #ifdef UNIV_DEBUG |
298 | if (srv_file_per_table) { |
299 | ulint size = 0; |
300 | |
301 | for (const fil_node_t* node = |
302 | UT_LIST_GET_FIRST(space->chain); |
303 | node != NULL; |
304 | node = UT_LIST_GET_NEXT(chain, node)) { |
305 | |
306 | size += ulint(os_file_get_size(node->handle) |
307 | / page_size.physical()); |
308 | } |
309 | |
310 | ut_ad(size == space->size); |
311 | } |
312 | #endif /* UNIV_DEBUG */ |
313 | |
314 | if (high > space->size) { |
315 | high = space->size; |
316 | } |
317 | space->release(); |
318 | } else { |
319 | return(0); |
320 | } |
321 | |
322 | buf_pool_mutex_enter(buf_pool); |
323 | |
324 | if (buf_pool->n_pend_reads |
325 | > buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) { |
326 | buf_pool_mutex_exit(buf_pool); |
327 | |
328 | return(0); |
329 | } |
330 | |
331 | /* Count how many blocks in the area have been recently accessed, |
332 | that is, reside near the start of the LRU list. */ |
333 | |
334 | for (i = low; i < high; i++) { |
335 | DBUG_EXECUTE_IF( |
336 | "innodb_invalid_read_after_truncate" , |
337 | if (fil_space_t* space = fil_space_acquire( |
338 | page_id.space())) { |
339 | bool skip = !strcmp(space->name, "test/t1" ); |
340 | space->release(); |
341 | if (skip) { |
342 | high = space->size; |
343 | buf_pool_mutex_exit(buf_pool); |
344 | goto read_ahead; |
345 | } |
346 | }); |
347 | |
348 | const buf_page_t* bpage = buf_page_hash_get( |
349 | buf_pool, page_id_t(page_id.space(), i)); |
350 | |
351 | if (bpage != NULL |
352 | && buf_page_is_accessed(bpage) |
353 | && buf_page_peek_if_young(bpage)) { |
354 | |
355 | recent_blocks++; |
356 | |
357 | if (recent_blocks |
358 | >= BUF_READ_AHEAD_RANDOM_THRESHOLD(buf_pool)) { |
359 | |
360 | buf_pool_mutex_exit(buf_pool); |
361 | goto read_ahead; |
362 | } |
363 | } |
364 | } |
365 | |
366 | buf_pool_mutex_exit(buf_pool); |
367 | /* Do nothing */ |
368 | return(0); |
369 | |
370 | read_ahead: |
371 | /* Read all the suitable blocks within the area */ |
372 | |
373 | if (inside_ibuf) { |
374 | ibuf_mode = BUF_READ_IBUF_PAGES_ONLY; |
375 | } else { |
376 | ibuf_mode = BUF_READ_ANY_PAGE; |
377 | } |
378 | |
379 | count = 0; |
380 | |
381 | for (i = low; i < high; i++) { |
382 | /* It is only sensible to do read-ahead in the non-sync aio |
383 | mode: hence FALSE as the first parameter */ |
384 | |
385 | const page_id_t cur_page_id(page_id.space(), i); |
386 | |
387 | if (!ibuf_bitmap_page(cur_page_id, page_size)) { |
388 | count += buf_read_page_low( |
389 | &err, false, |
390 | IORequest::DO_NOT_WAKE, |
391 | ibuf_mode, |
392 | cur_page_id, page_size, false); |
393 | |
394 | switch (err) { |
395 | case DB_SUCCESS: |
396 | case DB_TABLESPACE_TRUNCATED: |
397 | case DB_ERROR: |
398 | break; |
399 | case DB_TABLESPACE_DELETED: |
400 | ib::info() << "Random readahead trying to" |
401 | " access page " << cur_page_id |
402 | << " in nonexisting or" |
403 | " being-dropped tablespace" ; |
404 | break; |
405 | default: |
406 | ut_error; |
407 | } |
408 | } |
409 | } |
410 | |
411 | /* In simulated aio we wake the aio handler threads only after |
412 | queuing all aio requests, in native aio the following call does |
413 | nothing: */ |
414 | |
415 | os_aio_simulated_wake_handler_threads(); |
416 | |
417 | if (count) { |
418 | DBUG_PRINT("ib_buf" , ("random read-ahead %u pages, %u:%u" , |
419 | (unsigned) count, |
420 | (unsigned) page_id.space(), |
421 | (unsigned) page_id.page_no())); |
422 | } |
423 | |
424 | /* Read ahead is considered one I/O operation for the purpose of |
425 | LRU policy decision. */ |
426 | buf_LRU_stat_inc_io(); |
427 | |
428 | buf_pool->stat.n_ra_pages_read_rnd += count; |
429 | srv_stats.buf_pool_reads.add(count); |
430 | return(count); |
431 | } |
432 | |
433 | /** High-level function which reads a page asynchronously from a file to the |
434 | buffer buf_pool if it is not already there. Sets the io_fix flag and sets |
435 | an exclusive lock on the buffer frame. The flag is cleared and the x-lock |
436 | released by the i/o-handler thread. |
437 | @param[in] page_id page id |
438 | @param[in] page_size page size |
439 | @retval DB_SUCCESS if the page was read and is not corrupted, |
440 | @retval DB_PAGE_CORRUPTED if page based on checksum check is corrupted, |
441 | @retval DB_DECRYPTION_FAILED if page post encryption checksum matches but |
442 | after decryption normal page checksum does not match. |
443 | @retval DB_TABLESPACE_DELETED if tablespace .ibd file is missing */ |
444 | dberr_t |
445 | ( |
446 | const page_id_t& page_id, |
447 | const page_size_t& page_size) |
448 | { |
449 | ulint count; |
450 | dberr_t err = DB_SUCCESS; |
451 | |
452 | /* We do synchronous IO because our AIO completion code |
453 | is sub-optimal. See buf_page_io_complete(), we have to |
454 | acquire the buffer pool mutex before acquiring the block |
455 | mutex, required for updating the page state. The acquire |
456 | of the buffer pool mutex becomes an expensive bottleneck. */ |
457 | |
458 | count = buf_read_page_low( |
459 | &err, true, |
460 | 0, BUF_READ_ANY_PAGE, page_id, page_size, false); |
461 | |
462 | srv_stats.buf_pool_reads.add(count); |
463 | |
464 | if (err == DB_TABLESPACE_DELETED) { |
465 | ib::info() << "trying to read page " << page_id |
466 | << " in nonexisting or being-dropped tablespace" ; |
467 | } |
468 | |
469 | /* Increment number of I/O operations used for LRU policy. */ |
470 | buf_LRU_stat_inc_io(); |
471 | |
472 | return(err); |
473 | } |
474 | |
475 | /** High-level function which reads a page asynchronously from a file to the |
476 | buffer buf_pool if it is not already there. Sets the io_fix flag and sets |
477 | an exclusive lock on the buffer frame. The flag is cleared and the x-lock |
478 | released by the i/o-handler thread. |
479 | @param[in] page_id page id |
480 | @param[in] page_size page size |
481 | @param[in] sync true if synchronous aio is desired */ |
482 | void |
483 | buf_read_page_background( |
484 | const page_id_t& page_id, |
485 | const page_size_t& page_size, |
486 | bool sync) |
487 | { |
488 | ulint count; |
489 | dberr_t err; |
490 | |
491 | count = buf_read_page_low( |
492 | &err, sync, |
493 | IORequest::DO_NOT_WAKE | IORequest::IGNORE_MISSING, |
494 | BUF_READ_ANY_PAGE, |
495 | page_id, page_size, false); |
496 | |
497 | switch (err) { |
498 | case DB_SUCCESS: |
499 | case DB_TABLESPACE_TRUNCATED: |
500 | case DB_ERROR: |
501 | break; |
502 | case DB_TABLESPACE_DELETED: |
503 | ib::info() << "trying to read page " << page_id |
504 | << " in the background" |
505 | " in a non-existing or being-dropped tablespace" ; |
506 | break; |
507 | case DB_PAGE_CORRUPTED: |
508 | case DB_DECRYPTION_FAILED: |
509 | ib::error() |
510 | << "Background Page read failed to " |
511 | "read or decrypt " << page_id; |
512 | break; |
513 | default: |
514 | ib::fatal() << "Error " << err << " in background read of " |
515 | << page_id; |
516 | } |
517 | |
518 | srv_stats.buf_pool_reads.add(count); |
519 | |
520 | /* We do not increment number of I/O operations used for LRU policy |
521 | here (buf_LRU_stat_inc_io()). We use this in heuristics to decide |
522 | about evicting uncompressed version of compressed pages from the |
523 | buffer pool. Since this function is called from buffer pool load |
524 | these IOs are deliberate and are not part of normal workload we can |
525 | ignore these in our heuristics. */ |
526 | } |
527 | |
528 | /** Applies linear read-ahead if in the buf_pool the page is a border page of |
529 | a linear read-ahead area and all the pages in the area have been accessed. |
530 | Does not read any page if the read-ahead mechanism is not activated. Note |
531 | that the algorithm looks at the 'natural' adjacent successor and |
532 | predecessor of the page, which on the leaf level of a B-tree are the next |
533 | and previous page in the chain of leaves. To know these, the page specified |
534 | in (space, offset) must already be present in the buf_pool. Thus, the |
535 | natural way to use this function is to call it when a page in the buf_pool |
536 | is accessed the first time, calling this function just after it has been |
537 | bufferfixed. |
538 | NOTE 1: as this function looks at the natural predecessor and successor |
539 | fields on the page, what happens, if these are not initialized to any |
540 | sensible value? No problem, before applying read-ahead we check that the |
541 | area to read is within the span of the space, if not, read-ahead is not |
542 | applied. An uninitialized value may result in a useless read operation, but |
543 | only very improbably. |
544 | NOTE 2: the calling thread may own latches on pages: to avoid deadlocks this |
545 | function must be written such that it cannot end up waiting for these |
546 | latches! |
547 | NOTE 3: the calling thread must want access to the page given: this rule is |
548 | set to prevent unintended read-aheads performed by ibuf routines, a situation |
549 | which could result in a deadlock if the OS does not support asynchronous io. |
550 | @param[in] page_id page id; see NOTE 3 above |
551 | @param[in] page_size page size |
552 | @param[in] inside_ibuf TRUE if we are inside ibuf routine |
553 | @return number of page read requests issued */ |
554 | ulint |
555 | buf_read_ahead_linear( |
556 | const page_id_t& page_id, |
557 | const page_size_t& page_size, |
558 | ibool inside_ibuf) |
559 | { |
560 | buf_pool_t* buf_pool = buf_pool_get(page_id); |
561 | buf_page_t* bpage; |
562 | buf_frame_t* frame; |
563 | buf_page_t* pred_bpage = NULL; |
564 | ulint pred_offset; |
565 | ulint succ_offset; |
566 | int asc_or_desc; |
567 | ulint new_offset; |
568 | ulint fail_count; |
569 | ulint low, high; |
570 | dberr_t err = DB_SUCCESS; |
571 | ulint i; |
572 | const ulint buf_read_ahead_linear_area |
573 | = BUF_READ_AHEAD_AREA(buf_pool); |
574 | ulint threshold; |
575 | |
576 | /* check if readahead is disabled */ |
577 | if (!srv_read_ahead_threshold) { |
578 | return(0); |
579 | } |
580 | |
581 | if (srv_startup_is_before_trx_rollback_phase) { |
582 | /* No read-ahead to avoid thread deadlocks */ |
583 | return(0); |
584 | } |
585 | |
586 | low = (page_id.page_no() / buf_read_ahead_linear_area) |
587 | * buf_read_ahead_linear_area; |
588 | high = (page_id.page_no() / buf_read_ahead_linear_area + 1) |
589 | * buf_read_ahead_linear_area; |
590 | |
591 | if ((page_id.page_no() != low) && (page_id.page_no() != high - 1)) { |
592 | /* This is not a border page of the area: return */ |
593 | |
594 | return(0); |
595 | } |
596 | |
597 | if (ibuf_bitmap_page(page_id, page_size) || trx_sys_hdr_page(page_id)) { |
598 | |
599 | /* If it is an ibuf bitmap page or trx sys hdr, we do |
600 | no read-ahead, as that could break the ibuf page access |
601 | order */ |
602 | |
603 | return(0); |
604 | } |
605 | |
606 | /* Remember the tablespace version before we ask te tablespace size |
607 | below: if DISCARD + IMPORT changes the actual .ibd file meanwhile, we |
608 | do not try to read outside the bounds of the tablespace! */ |
609 | ulint space_size; |
610 | |
611 | if (fil_space_t* space = fil_space_acquire(page_id.space())) { |
612 | space_size = space->size; |
613 | space->release(); |
614 | |
615 | if (high > space_size) { |
616 | /* The area is not whole */ |
617 | return(0); |
618 | } |
619 | } else { |
620 | return(0); |
621 | } |
622 | |
623 | buf_pool_mutex_enter(buf_pool); |
624 | |
625 | if (buf_pool->n_pend_reads |
626 | > buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) { |
627 | buf_pool_mutex_exit(buf_pool); |
628 | |
629 | return(0); |
630 | } |
631 | |
632 | /* Check that almost all pages in the area have been accessed; if |
633 | offset == low, the accesses must be in a descending order, otherwise, |
634 | in an ascending order. */ |
635 | |
636 | asc_or_desc = 1; |
637 | |
638 | if (page_id.page_no() == low) { |
639 | asc_or_desc = -1; |
640 | } |
641 | |
642 | /* How many out of order accessed pages can we ignore |
643 | when working out the access pattern for linear readahead */ |
644 | threshold = ut_min(static_cast<ulint>(64 - srv_read_ahead_threshold), |
645 | BUF_READ_AHEAD_AREA(buf_pool)); |
646 | |
647 | fail_count = 0; |
648 | |
649 | for (i = low; i < high; i++) { |
650 | bpage = buf_page_hash_get(buf_pool, |
651 | page_id_t(page_id.space(), i)); |
652 | |
653 | if (bpage == NULL || !buf_page_is_accessed(bpage)) { |
654 | /* Not accessed */ |
655 | fail_count++; |
656 | |
657 | } else if (pred_bpage) { |
658 | /* Note that buf_page_is_accessed() returns |
659 | the time of the first access. If some blocks |
660 | of the extent existed in the buffer pool at |
661 | the time of a linear access pattern, the first |
662 | access times may be nonmonotonic, even though |
663 | the latest access times were linear. The |
664 | threshold (srv_read_ahead_factor) should help |
665 | a little against this. */ |
666 | int res = ut_ulint_cmp( |
667 | buf_page_is_accessed(bpage), |
668 | buf_page_is_accessed(pred_bpage)); |
669 | /* Accesses not in the right order */ |
670 | if (res != 0 && res != asc_or_desc) { |
671 | fail_count++; |
672 | } |
673 | } |
674 | |
675 | if (fail_count > threshold) { |
676 | /* Too many failures: return */ |
677 | buf_pool_mutex_exit(buf_pool); |
678 | return(0); |
679 | } |
680 | |
681 | if (bpage && buf_page_is_accessed(bpage)) { |
682 | pred_bpage = bpage; |
683 | } |
684 | } |
685 | |
686 | /* If we got this far, we know that enough pages in the area have |
687 | been accessed in the right order: linear read-ahead can be sensible */ |
688 | |
689 | bpage = buf_page_hash_get(buf_pool, page_id); |
690 | |
691 | if (bpage == NULL) { |
692 | buf_pool_mutex_exit(buf_pool); |
693 | |
694 | return(0); |
695 | } |
696 | |
697 | switch (buf_page_get_state(bpage)) { |
698 | case BUF_BLOCK_ZIP_PAGE: |
699 | frame = bpage->zip.data; |
700 | break; |
701 | case BUF_BLOCK_FILE_PAGE: |
702 | frame = ((buf_block_t*) bpage)->frame; |
703 | break; |
704 | default: |
705 | ut_error; |
706 | break; |
707 | } |
708 | |
709 | /* Read the natural predecessor and successor page addresses from |
710 | the page; NOTE that because the calling thread may have an x-latch |
711 | on the page, we do not acquire an s-latch on the page, this is to |
712 | prevent deadlocks. Even if we read values which are nonsense, the |
713 | algorithm will work. */ |
714 | |
715 | pred_offset = fil_page_get_prev(frame); |
716 | succ_offset = fil_page_get_next(frame); |
717 | |
718 | buf_pool_mutex_exit(buf_pool); |
719 | |
720 | if ((page_id.page_no() == low) |
721 | && (succ_offset == page_id.page_no() + 1)) { |
722 | |
723 | /* This is ok, we can continue */ |
724 | new_offset = pred_offset; |
725 | |
726 | } else if ((page_id.page_no() == high - 1) |
727 | && (pred_offset == page_id.page_no() - 1)) { |
728 | |
729 | /* This is ok, we can continue */ |
730 | new_offset = succ_offset; |
731 | } else { |
732 | /* Successor or predecessor not in the right order */ |
733 | |
734 | return(0); |
735 | } |
736 | |
737 | low = (new_offset / buf_read_ahead_linear_area) |
738 | * buf_read_ahead_linear_area; |
739 | high = (new_offset / buf_read_ahead_linear_area + 1) |
740 | * buf_read_ahead_linear_area; |
741 | |
742 | if ((new_offset != low) && (new_offset != high - 1)) { |
743 | /* This is not a border page of the area: return */ |
744 | |
745 | return(0); |
746 | } |
747 | |
748 | if (high > space_size) { |
749 | /* The area is not whole, return */ |
750 | |
751 | return(0); |
752 | } |
753 | |
754 | ulint count = 0; |
755 | |
756 | /* If we got this far, read-ahead can be sensible: do it */ |
757 | |
758 | ulint ibuf_mode; |
759 | |
760 | ibuf_mode = inside_ibuf ? BUF_READ_IBUF_PAGES_ONLY : BUF_READ_ANY_PAGE; |
761 | |
762 | /* Since Windows XP seems to schedule the i/o handler thread |
763 | very eagerly, and consequently it does not wait for the |
764 | full read batch to be posted, we use special heuristics here */ |
765 | |
766 | os_aio_simulated_put_read_threads_to_sleep(); |
767 | |
768 | for (i = low; i < high; i++) { |
769 | /* It is only sensible to do read-ahead in the non-sync |
770 | aio mode: hence FALSE as the first parameter */ |
771 | |
772 | const page_id_t cur_page_id(page_id.space(), i); |
773 | |
774 | if (!ibuf_bitmap_page(cur_page_id, page_size)) { |
775 | count += buf_read_page_low( |
776 | &err, false, |
777 | IORequest::DO_NOT_WAKE, |
778 | ibuf_mode, cur_page_id, page_size, false); |
779 | |
780 | switch (err) { |
781 | case DB_SUCCESS: |
782 | case DB_TABLESPACE_TRUNCATED: |
783 | case DB_TABLESPACE_DELETED: |
784 | case DB_ERROR: |
785 | break; |
786 | case DB_PAGE_CORRUPTED: |
787 | case DB_DECRYPTION_FAILED: |
788 | ib::error() << "linear readahead failed to" |
789 | " read or decrypt " |
790 | << page_id_t(page_id.space(), i); |
791 | break; |
792 | default: |
793 | ut_error; |
794 | } |
795 | } |
796 | } |
797 | |
798 | /* In simulated aio we wake the aio handler threads only after |
799 | queuing all aio requests, in native aio the following call does |
800 | nothing: */ |
801 | |
802 | os_aio_simulated_wake_handler_threads(); |
803 | |
804 | if (count) { |
805 | DBUG_PRINT("ib_buf" , ("linear read-ahead " ULINTPF " pages, " |
806 | "%u:%u" , |
807 | count, |
808 | page_id.space(), |
809 | page_id.page_no())); |
810 | } |
811 | |
812 | /* Read ahead is considered one I/O operation for the purpose of |
813 | LRU policy decision. */ |
814 | buf_LRU_stat_inc_io(); |
815 | |
816 | buf_pool->stat.n_ra_pages_read += count; |
817 | return(count); |
818 | } |
819 | |
820 | /********************************************************************//** |
821 | Issues read requests for pages which the ibuf module wants to read in, in |
822 | order to contract the insert buffer tree. Technically, this function is like |
823 | a read-ahead function. */ |
824 | void |
825 | buf_read_ibuf_merge_pages( |
826 | /*======================*/ |
827 | bool sync, /*!< in: true if the caller |
828 | wants this function to wait |
829 | for the highest address page |
830 | to get read in, before this |
831 | function returns */ |
832 | const ulint* space_ids, /*!< in: array of space ids */ |
833 | const ulint* page_nos, /*!< in: array of page numbers |
834 | to read, with the highest page |
835 | number the last in the |
836 | array */ |
837 | ulint n_stored) /*!< in: number of elements |
838 | in the arrays */ |
839 | { |
840 | #ifdef UNIV_IBUF_DEBUG |
841 | ut_a(n_stored < srv_page_size); |
842 | #endif |
843 | |
844 | for (ulint i = 0; i < n_stored; i++) { |
845 | bool found; |
846 | const page_size_t page_size(fil_space_get_page_size( |
847 | space_ids[i], &found)); |
848 | |
849 | if (!found) { |
850 | tablespace_deleted: |
851 | /* The tablespace was not found: remove all |
852 | entries for it */ |
853 | ibuf_delete_for_discarded_space(space_ids[i]); |
854 | while (i + 1 < n_stored |
855 | && space_ids[i + 1] == space_ids[i]) { |
856 | i++; |
857 | } |
858 | continue; |
859 | } |
860 | |
861 | const page_id_t page_id(space_ids[i], page_nos[i]); |
862 | |
863 | buf_pool_t* buf_pool = buf_pool_get(page_id); |
864 | |
865 | while (buf_pool->n_pend_reads |
866 | > buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) { |
867 | os_thread_sleep(500000); |
868 | } |
869 | |
870 | dberr_t err; |
871 | |
872 | buf_read_page_low(&err, |
873 | sync && (i + 1 == n_stored), |
874 | 0, |
875 | BUF_READ_ANY_PAGE, page_id, page_size, |
876 | true, true /* ignore_missing_space */); |
877 | |
878 | switch(err) { |
879 | case DB_SUCCESS: |
880 | case DB_TABLESPACE_TRUNCATED: |
881 | case DB_ERROR: |
882 | break; |
883 | case DB_TABLESPACE_DELETED: |
884 | goto tablespace_deleted; |
885 | case DB_PAGE_CORRUPTED: |
886 | case DB_DECRYPTION_FAILED: |
887 | ib::error() << "Failed to read or decrypt " << page_id |
888 | << " for change buffer merge" ; |
889 | break; |
890 | default: |
891 | ut_error; |
892 | } |
893 | } |
894 | |
895 | os_aio_simulated_wake_handler_threads(); |
896 | |
897 | if (n_stored) { |
898 | DBUG_PRINT("ib_buf" , |
899 | ("ibuf merge read-ahead %u pages, space %u" , |
900 | unsigned(n_stored), unsigned(space_ids[0]))); |
901 | } |
902 | } |
903 | |
904 | /** Issues read requests for pages which recovery wants to read in. |
905 | @param[in] sync true if the caller wants this function to wait |
906 | for the highest address page to get read in, before this function returns |
907 | @param[in] space_id tablespace id |
908 | @param[in] page_nos array of page numbers to read, with the |
909 | highest page number the last in the array |
910 | @param[in] n_stored number of page numbers in the array */ |
911 | void |
912 | buf_read_recv_pages( |
913 | bool sync, |
914 | ulint space_id, |
915 | const ulint* page_nos, |
916 | ulint n_stored) |
917 | { |
918 | fil_space_t* space = fil_space_get(space_id); |
919 | |
920 | if (space == NULL) { |
921 | /* The tablespace is missing: do nothing */ |
922 | return; |
923 | } |
924 | |
925 | fil_space_open_if_needed(space); |
926 | |
927 | const page_size_t page_size(space->flags); |
928 | |
929 | for (ulint i = 0; i < n_stored; i++) { |
930 | buf_pool_t* buf_pool; |
931 | const page_id_t cur_page_id(space_id, page_nos[i]); |
932 | |
933 | ulint count = 0; |
934 | |
935 | buf_pool = buf_pool_get(cur_page_id); |
936 | while (buf_pool->n_pend_reads >= recv_n_pool_free_frames / 2) { |
937 | |
938 | os_aio_simulated_wake_handler_threads(); |
939 | os_thread_sleep(10000); |
940 | |
941 | count++; |
942 | |
943 | if (!(count % 1000)) { |
944 | |
945 | ib::error() |
946 | << "Waited for " << count / 100 |
947 | << " seconds for " |
948 | << buf_pool->n_pend_reads |
949 | << " pending reads" ; |
950 | } |
951 | } |
952 | |
953 | dberr_t err; |
954 | |
955 | if (sync && i + 1 == n_stored) { |
956 | buf_read_page_low( |
957 | &err, true, |
958 | 0, |
959 | BUF_READ_ANY_PAGE, |
960 | cur_page_id, page_size, true); |
961 | } else { |
962 | buf_read_page_low( |
963 | &err, false, |
964 | IORequest::DO_NOT_WAKE, |
965 | BUF_READ_ANY_PAGE, |
966 | cur_page_id, page_size, true); |
967 | } |
968 | |
969 | if (err == DB_DECRYPTION_FAILED || err == DB_PAGE_CORRUPTED) { |
970 | ib::error() << "Recovery failed to read or decrypt " |
971 | << cur_page_id; |
972 | } |
973 | } |
974 | |
975 | os_aio_simulated_wake_handler_threads(); |
976 | |
977 | DBUG_PRINT("ib_buf" , ("recovery read-ahead (%u pages)" , |
978 | unsigned(n_stored))); |
979 | } |
980 | |