1
by brian
clean slate |
1 |
/******************************************************
|
2 |
The database buffer read
|
|
3 |
||
4 |
(c) 1995 Innobase Oy
|
|
5 |
||
6 |
Created 11/5/1995 Heikki Tuuri
|
|
7 |
*******************************************************/
|
|
8 |
||
9 |
#include "buf0rea.h" |
|
10 |
||
11 |
#include "fil0fil.h" |
|
12 |
#include "mtr0mtr.h" |
|
13 |
||
14 |
#include "buf0buf.h" |
|
15 |
#include "buf0flu.h" |
|
16 |
#include "buf0lru.h" |
|
17 |
#include "ibuf0ibuf.h" |
|
18 |
#include "log0recv.h" |
|
19 |
#include "trx0sys.h" |
|
20 |
#include "os0file.h" |
|
21 |
#include "srv0start.h" |
|
22 |
||
23 |
extern ulint srv_read_ahead_rnd; |
|
24 |
extern ulint srv_read_ahead_seq; |
|
25 |
extern ulint srv_buf_pool_reads; |
|
26 |
||
27 |
/* The size in blocks of the area where the random read-ahead algorithm counts
|
|
28 |
the accessed pages when deciding whether to read-ahead */
|
|
29 |
#define BUF_READ_AHEAD_RANDOM_AREA BUF_READ_AHEAD_AREA
|
|
30 |
||
31 |
/* There must be at least this many pages in buf_pool in the area to start
|
|
32 |
a random read-ahead */
|
|
33 |
#define BUF_READ_AHEAD_RANDOM_THRESHOLD (5 + BUF_READ_AHEAD_RANDOM_AREA / 8)
|
|
34 |
||
35 |
/* The linear read-ahead area size */
|
|
36 |
#define BUF_READ_AHEAD_LINEAR_AREA BUF_READ_AHEAD_AREA
|
|
37 |
||
38 |
/* The linear read-ahead threshold */
|
|
39 |
#define BUF_READ_AHEAD_LINEAR_THRESHOLD (3 * BUF_READ_AHEAD_LINEAR_AREA / 8)
|
|
40 |
||
41 |
/* If there are buf_pool->curr_size per the number below pending reads, then
|
|
42 |
read-ahead is not done: this is to prevent flooding the buffer pool with
|
|
43 |
i/o-fixed buffer blocks */
|
|
44 |
#define BUF_READ_AHEAD_PEND_LIMIT 2
|
|
45 |
||
46 |
/************************************************************************
|
|
47 |
Low-level function which reads a page asynchronously from a file to the
|
|
48 |
buffer buf_pool if it is not already there, in which case does nothing.
|
|
49 |
Sets the io_fix flag and sets an exclusive lock on the buffer frame. The
|
|
50 |
flag is cleared and the x-lock released by an i/o-handler thread. */
|
|
51 |
static
|
|
52 |
ulint
|
|
53 |
buf_read_page_low( |
|
54 |
/*==============*/
|
|
55 |
/* out: 1 if a read request was queued, 0 if the page
|
|
56 |
already resided in buf_pool, or if the page is in
|
|
57 |
the doublewrite buffer blocks in which case it is never
|
|
58 |
read into the pool, or if the tablespace does not
|
|
59 |
exist or is being dropped */
|
|
60 |
ulint* err, /* out: DB_SUCCESS or DB_TABLESPACE_DELETED if we are |
|
61 |
trying to read from a non-existent tablespace, or a
|
|
62 |
tablespace which is just now being dropped */
|
|
63 |
ibool sync, /* in: TRUE if synchronous aio is desired */ |
|
64 |
ulint mode, /* in: BUF_READ_IBUF_PAGES_ONLY, ..., |
|
65 |
ORed to OS_AIO_SIMULATED_WAKE_LATER (see below
|
|
66 |
at read-ahead functions) */
|
|
67 |
ulint space, /* in: space id */ |
|
68 |
ib_longlong tablespace_version, /* in: if the space memory object has |
|
69 |
this timestamp different from what we are giving here,
|
|
70 |
treat the tablespace as dropped; this is a timestamp we
|
|
71 |
use to stop dangling page reads from a tablespace
|
|
72 |
which we have DISCARDed + IMPORTed back */
|
|
73 |
ulint offset) /* in: page number */ |
|
74 |
{
|
|
75 |
buf_block_t* block; |
|
76 |
ulint wake_later; |
|
77 |
||
78 |
*err = DB_SUCCESS; |
|
79 |
||
80 |
wake_later = mode & OS_AIO_SIMULATED_WAKE_LATER; |
|
81 |
mode = mode & ~OS_AIO_SIMULATED_WAKE_LATER; |
|
82 |
||
83 |
if (trx_doublewrite && space == TRX_SYS_SPACE |
|
84 |
&& ( (offset >= trx_doublewrite->block1 |
|
85 |
&& offset < trx_doublewrite->block1 |
|
86 |
+ TRX_SYS_DOUBLEWRITE_BLOCK_SIZE) |
|
87 |
|| (offset >= trx_doublewrite->block2 |
|
88 |
&& offset < trx_doublewrite->block2 |
|
89 |
+ TRX_SYS_DOUBLEWRITE_BLOCK_SIZE))) { |
|
90 |
ut_print_timestamp(stderr); |
|
91 |
fprintf(stderr, |
|
92 |
" InnoDB: Warning: trying to read"
|
|
93 |
" doublewrite buffer page %lu\n", |
|
94 |
(ulong) offset); |
|
95 |
||
96 |
return(0); |
|
97 |
}
|
|
98 |
||
99 |
if (ibuf_bitmap_page(offset) || trx_sys_hdr_page(space, offset)) { |
|
100 |
||
101 |
/* Trx sys header is so low in the latching order that we play
|
|
102 |
safe and do not leave the i/o-completion to an asynchronous
|
|
103 |
i/o-thread. Ibuf bitmap pages must always be read with
|
|
104 |
syncronous i/o, to make sure they do not get involved in
|
|
105 |
thread deadlocks. */
|
|
106 |
||
107 |
sync = TRUE; |
|
108 |
}
|
|
109 |
||
110 |
/* The following call will also check if the tablespace does not exist
|
|
111 |
or is being dropped; if we succeed in initing the page in the buffer
|
|
112 |
pool for read, then DISCARD cannot proceed until the read has
|
|
113 |
completed */
|
|
114 |
block = buf_page_init_for_read(err, mode, space, tablespace_version, |
|
115 |
offset); |
|
116 |
if (block == NULL) { |
|
117 |
||
118 |
return(0); |
|
119 |
}
|
|
120 |
||
121 |
#ifdef UNIV_DEBUG
|
|
122 |
if (buf_debug_prints) { |
|
123 |
fprintf(stderr, |
|
124 |
"Posting read request for page %lu, sync %lu\n", |
|
125 |
(ulong) offset, |
|
126 |
(ulong) sync); |
|
127 |
}
|
|
128 |
#endif
|
|
129 |
||
130 |
ut_a(block->state == BUF_BLOCK_FILE_PAGE); |
|
131 |
||
132 |
*err = fil_io(OS_FILE_READ | wake_later, |
|
133 |
sync, space, |
|
134 |
offset, 0, UNIV_PAGE_SIZE, |
|
135 |
(void*)block->frame, (void*)block); |
|
136 |
ut_a(*err == DB_SUCCESS); |
|
137 |
||
138 |
if (sync) { |
|
139 |
/* The i/o is already completed when we arrive from
|
|
140 |
fil_read */
|
|
141 |
buf_page_io_complete(block); |
|
142 |
}
|
|
143 |
||
144 |
return(1); |
|
145 |
}
|
|
146 |
||
147 |
/************************************************************************
|
|
148 |
Applies a random read-ahead in buf_pool if there are at least a threshold
|
|
149 |
value of accessed pages from the random read-ahead area. Does not read any
|
|
150 |
page, not even the one at the position (space, offset), if the read-ahead
|
|
151 |
mechanism is not activated. NOTE 1: the calling thread may own latches on
|
|
152 |
pages: to avoid deadlocks this function must be written such that it cannot
|
|
153 |
end up waiting for these latches! NOTE 2: the calling thread must want
|
|
154 |
access to the page given: this rule is set to prevent unintended read-aheads
|
|
155 |
performed by ibuf routines, a situation which could result in a deadlock if
|
|
156 |
the OS does not support asynchronous i/o. */
|
|
157 |
static
|
|
158 |
ulint
|
|
159 |
buf_read_ahead_random( |
|
160 |
/*==================*/
|
|
161 |
/* out: number of page read requests issued; NOTE
|
|
162 |
that if we read ibuf pages, it may happen that
|
|
163 |
the page at the given page number does not get
|
|
164 |
read even if we return a value > 0! */
|
|
165 |
ulint space, /* in: space id */ |
|
166 |
ulint offset) /* in: page number of a page which the current thread |
|
167 |
wants to access */
|
|
168 |
{
|
|
169 |
ib_longlong tablespace_version; |
|
170 |
buf_block_t* block; |
|
171 |
ulint recent_blocks = 0; |
|
172 |
ulint count; |
|
173 |
ulint LRU_recent_limit; |
|
174 |
ulint ibuf_mode; |
|
175 |
ulint low, high; |
|
176 |
ulint err; |
|
177 |
ulint i; |
|
178 |
||
179 |
if (srv_startup_is_before_trx_rollback_phase) { |
|
180 |
/* No read-ahead to avoid thread deadlocks */
|
|
181 |
return(0); |
|
182 |
}
|
|
183 |
||
184 |
if (ibuf_bitmap_page(offset) || trx_sys_hdr_page(space, offset)) { |
|
185 |
||
186 |
/* If it is an ibuf bitmap page or trx sys hdr, we do
|
|
187 |
no read-ahead, as that could break the ibuf page access
|
|
188 |
order */
|
|
189 |
||
190 |
return(0); |
|
191 |
}
|
|
192 |
||
193 |
/* Remember the tablespace version before we ask te tablespace size
|
|
194 |
below: if DISCARD + IMPORT changes the actual .ibd file meanwhile, we
|
|
195 |
do not try to read outside the bounds of the tablespace! */
|
|
196 |
||
197 |
tablespace_version = fil_space_get_version(space); |
|
198 |
||
199 |
low = (offset / BUF_READ_AHEAD_RANDOM_AREA) |
|
200 |
* BUF_READ_AHEAD_RANDOM_AREA; |
|
201 |
high = (offset / BUF_READ_AHEAD_RANDOM_AREA + 1) |
|
202 |
* BUF_READ_AHEAD_RANDOM_AREA; |
|
203 |
if (high > fil_space_get_size(space)) { |
|
204 |
||
205 |
high = fil_space_get_size(space); |
|
206 |
}
|
|
207 |
||
208 |
/* Get the minimum LRU_position field value for an initial segment
|
|
209 |
of the LRU list, to determine which blocks have recently been added
|
|
210 |
to the start of the list. */
|
|
211 |
||
212 |
LRU_recent_limit = buf_LRU_get_recent_limit(); |
|
213 |
||
214 |
mutex_enter(&(buf_pool->mutex)); |
|
215 |
||
216 |
if (buf_pool->n_pend_reads |
|
217 |
> buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) { |
|
218 |
mutex_exit(&(buf_pool->mutex)); |
|
219 |
||
220 |
return(0); |
|
221 |
}
|
|
222 |
||
223 |
/* Count how many blocks in the area have been recently accessed,
|
|
224 |
that is, reside near the start of the LRU list. */
|
|
225 |
||
226 |
for (i = low; i < high; i++) { |
|
227 |
block = buf_page_hash_get(space, i); |
|
228 |
||
229 |
if ((block) |
|
230 |
&& (block->LRU_position > LRU_recent_limit) |
|
231 |
&& block->accessed) { |
|
232 |
||
233 |
recent_blocks++; |
|
234 |
}
|
|
235 |
}
|
|
236 |
||
237 |
mutex_exit(&(buf_pool->mutex)); |
|
238 |
||
239 |
if (recent_blocks < BUF_READ_AHEAD_RANDOM_THRESHOLD) { |
|
240 |
/* Do nothing */
|
|
241 |
||
242 |
return(0); |
|
243 |
}
|
|
244 |
||
245 |
/* Read all the suitable blocks within the area */
|
|
246 |
||
247 |
if (ibuf_inside()) { |
|
248 |
ibuf_mode = BUF_READ_IBUF_PAGES_ONLY; |
|
249 |
} else { |
|
250 |
ibuf_mode = BUF_READ_ANY_PAGE; |
|
251 |
}
|
|
252 |
||
253 |
count = 0; |
|
254 |
||
255 |
for (i = low; i < high; i++) { |
|
256 |
/* It is only sensible to do read-ahead in the non-sync aio
|
|
257 |
mode: hence FALSE as the first parameter */
|
|
258 |
||
259 |
if (!ibuf_bitmap_page(i)) { |
|
260 |
count += buf_read_page_low( |
|
261 |
&err, FALSE, |
|
262 |
ibuf_mode | OS_AIO_SIMULATED_WAKE_LATER, |
|
263 |
space, tablespace_version, i); |
|
264 |
if (err == DB_TABLESPACE_DELETED) { |
|
265 |
ut_print_timestamp(stderr); |
|
266 |
fprintf(stderr, |
|
267 |
" InnoDB: Warning: in random"
|
|
268 |
" readahead trying to access\n" |
|
269 |
"InnoDB: tablespace %lu page %lu,\n" |
|
270 |
"InnoDB: but the tablespace does not"
|
|
271 |
" exist or is just being dropped.\n", |
|
272 |
(ulong) space, (ulong) i); |
|
273 |
}
|
|
274 |
}
|
|
275 |
}
|
|
276 |
||
277 |
/* In simulated aio we wake the aio handler threads only after
|
|
278 |
queuing all aio requests, in native aio the following call does
|
|
279 |
nothing: */
|
|
280 |
||
281 |
os_aio_simulated_wake_handler_threads(); |
|
282 |
||
283 |
#ifdef UNIV_DEBUG
|
|
284 |
if (buf_debug_prints && (count > 0)) { |
|
285 |
fprintf(stderr, |
|
286 |
"Random read-ahead space %lu offset %lu pages %lu\n", |
|
287 |
(ulong) space, (ulong) offset, |
|
288 |
(ulong) count); |
|
289 |
}
|
|
290 |
#endif /* UNIV_DEBUG */ |
|
291 |
||
292 |
++srv_read_ahead_rnd; |
|
293 |
return(count); |
|
294 |
}
|
|
295 |
||
296 |
/************************************************************************
|
|
297 |
High-level function which reads a page asynchronously from a file to the
|
|
298 |
buffer buf_pool if it is not already there. Sets the io_fix flag and sets
|
|
299 |
an exclusive lock on the buffer frame. The flag is cleared and the x-lock
|
|
300 |
released by the i/o-handler thread. Does a random read-ahead if it seems
|
|
301 |
sensible. */
|
|
302 |
||
303 |
ulint
|
|
304 |
buf_read_page( |
|
305 |
/*==========*/
|
|
306 |
/* out: number of page read requests issued: this can
|
|
307 |
be > 1 if read-ahead occurred */
|
|
308 |
ulint space, /* in: space id */ |
|
309 |
ulint offset) /* in: page number */ |
|
310 |
{
|
|
311 |
ib_longlong tablespace_version; |
|
312 |
ulint count; |
|
313 |
ulint count2; |
|
314 |
ulint err; |
|
315 |
||
316 |
tablespace_version = fil_space_get_version(space); |
|
317 |
||
318 |
count = buf_read_ahead_random(space, offset); |
|
319 |
||
320 |
/* We do the i/o in the synchronous aio mode to save thread
|
|
321 |
switches: hence TRUE */
|
|
322 |
||
323 |
count2 = buf_read_page_low(&err, TRUE, BUF_READ_ANY_PAGE, space, |
|
324 |
tablespace_version, offset); |
|
325 |
srv_buf_pool_reads+= count2; |
|
326 |
if (err == DB_TABLESPACE_DELETED) { |
|
327 |
ut_print_timestamp(stderr); |
|
328 |
fprintf(stderr, |
|
329 |
" InnoDB: Error: trying to access"
|
|
330 |
" tablespace %lu page no. %lu,\n" |
|
331 |
"InnoDB: but the tablespace does not exist"
|
|
332 |
" or is just being dropped.\n", |
|
333 |
(ulong) space, (ulong) offset); |
|
334 |
}
|
|
335 |
||
336 |
/* Flush pages from the end of the LRU list if necessary */
|
|
337 |
buf_flush_free_margin(); |
|
338 |
||
339 |
return(count + count2); |
|
340 |
}
|
|
341 |
||
342 |
/************************************************************************
|
|
343 |
Applies linear read-ahead if in the buf_pool the page is a border page of
|
|
344 |
a linear read-ahead area and all the pages in the area have been accessed.
|
|
345 |
Does not read any page if the read-ahead mechanism is not activated. Note
|
|
346 |
that the the algorithm looks at the 'natural' adjacent successor and
|
|
347 |
predecessor of the page, which on the leaf level of a B-tree are the next
|
|
348 |
and previous page in the chain of leaves. To know these, the page specified
|
|
349 |
in (space, offset) must already be present in the buf_pool. Thus, the
|
|
350 |
natural way to use this function is to call it when a page in the buf_pool
|
|
351 |
is accessed the first time, calling this function just after it has been
|
|
352 |
bufferfixed.
|
|
353 |
NOTE 1: as this function looks at the natural predecessor and successor
|
|
354 |
fields on the page, what happens, if these are not initialized to any
|
|
355 |
sensible value? No problem, before applying read-ahead we check that the
|
|
356 |
area to read is within the span of the space, if not, read-ahead is not
|
|
357 |
applied. An uninitialized value may result in a useless read operation, but
|
|
358 |
only very improbably.
|
|
359 |
NOTE 2: the calling thread may own latches on pages: to avoid deadlocks this
|
|
360 |
function must be written such that it cannot end up waiting for these
|
|
361 |
latches!
|
|
362 |
NOTE 3: the calling thread must want access to the page given: this rule is
|
|
363 |
set to prevent unintended read-aheads performed by ibuf routines, a situation
|
|
364 |
which could result in a deadlock if the OS does not support asynchronous io. */
|
|
365 |
||
366 |
ulint
|
|
367 |
buf_read_ahead_linear( |
|
368 |
/*==================*/
|
|
369 |
/* out: number of page read requests issued */
|
|
370 |
ulint space, /* in: space id */ |
|
371 |
ulint offset) /* in: page number of a page; NOTE: the current thread |
|
372 |
must want access to this page (see NOTE 3 above) */
|
|
373 |
{
|
|
374 |
ib_longlong tablespace_version; |
|
375 |
buf_block_t* block; |
|
376 |
buf_frame_t* frame; |
|
377 |
buf_block_t* pred_block = NULL; |
|
378 |
ulint pred_offset; |
|
379 |
ulint succ_offset; |
|
380 |
ulint count; |
|
381 |
int asc_or_desc; |
|
382 |
ulint new_offset; |
|
383 |
ulint fail_count; |
|
384 |
ulint ibuf_mode; |
|
385 |
ulint low, high; |
|
386 |
ulint err; |
|
387 |
ulint i; |
|
388 |
||
389 |
if (srv_startup_is_before_trx_rollback_phase) { |
|
390 |
/* No read-ahead to avoid thread deadlocks */
|
|
391 |
return(0); |
|
392 |
}
|
|
393 |
||
394 |
if (ibuf_bitmap_page(offset) || trx_sys_hdr_page(space, offset)) { |
|
395 |
||
396 |
/* If it is an ibuf bitmap page or trx sys hdr, we do
|
|
397 |
no read-ahead, as that could break the ibuf page access
|
|
398 |
order */
|
|
399 |
||
400 |
return(0); |
|
401 |
}
|
|
402 |
||
403 |
low = (offset / BUF_READ_AHEAD_LINEAR_AREA) |
|
404 |
* BUF_READ_AHEAD_LINEAR_AREA; |
|
405 |
high = (offset / BUF_READ_AHEAD_LINEAR_AREA + 1) |
|
406 |
* BUF_READ_AHEAD_LINEAR_AREA; |
|
407 |
||
408 |
if ((offset != low) && (offset != high - 1)) { |
|
409 |
/* This is not a border page of the area: return */
|
|
410 |
||
411 |
return(0); |
|
412 |
}
|
|
413 |
||
414 |
/* Remember the tablespace version before we ask te tablespace size
|
|
415 |
below: if DISCARD + IMPORT changes the actual .ibd file meanwhile, we
|
|
416 |
do not try to read outside the bounds of the tablespace! */
|
|
417 |
||
418 |
tablespace_version = fil_space_get_version(space); |
|
419 |
||
420 |
mutex_enter(&(buf_pool->mutex)); |
|
421 |
||
422 |
if (high > fil_space_get_size(space)) { |
|
423 |
mutex_exit(&(buf_pool->mutex)); |
|
424 |
/* The area is not whole, return */
|
|
425 |
||
426 |
return(0); |
|
427 |
}
|
|
428 |
||
429 |
if (buf_pool->n_pend_reads |
|
430 |
> buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) { |
|
431 |
mutex_exit(&(buf_pool->mutex)); |
|
432 |
||
433 |
return(0); |
|
434 |
}
|
|
435 |
||
436 |
/* Check that almost all pages in the area have been accessed; if
|
|
437 |
offset == low, the accesses must be in a descending order, otherwise,
|
|
438 |
in an ascending order. */
|
|
439 |
||
440 |
asc_or_desc = 1; |
|
441 |
||
442 |
if (offset == low) { |
|
443 |
asc_or_desc = -1; |
|
444 |
}
|
|
445 |
||
446 |
fail_count = 0; |
|
447 |
||
448 |
for (i = low; i < high; i++) { |
|
449 |
block = buf_page_hash_get(space, i); |
|
450 |
||
451 |
if ((block == NULL) || !block->accessed) { |
|
452 |
/* Not accessed */
|
|
453 |
fail_count++; |
|
454 |
||
455 |
} else if (pred_block |
|
456 |
&& (ut_ulint_cmp(block->LRU_position, |
|
457 |
pred_block->LRU_position) |
|
458 |
!= asc_or_desc)) { |
|
459 |
/* Accesses not in the right order */
|
|
460 |
||
461 |
fail_count++; |
|
462 |
pred_block = block; |
|
463 |
}
|
|
464 |
}
|
|
465 |
||
466 |
if (fail_count > BUF_READ_AHEAD_LINEAR_AREA |
|
467 |
- BUF_READ_AHEAD_LINEAR_THRESHOLD) { |
|
468 |
/* Too many failures: return */
|
|
469 |
||
470 |
mutex_exit(&(buf_pool->mutex)); |
|
471 |
||
472 |
return(0); |
|
473 |
}
|
|
474 |
||
475 |
/* If we got this far, we know that enough pages in the area have
|
|
476 |
been accessed in the right order: linear read-ahead can be sensible */
|
|
477 |
||
478 |
block = buf_page_hash_get(space, offset); |
|
479 |
||
480 |
if (block == NULL) { |
|
481 |
mutex_exit(&(buf_pool->mutex)); |
|
482 |
||
483 |
return(0); |
|
484 |
}
|
|
485 |
||
486 |
frame = block->frame; |
|
487 |
||
488 |
/* Read the natural predecessor and successor page addresses from
|
|
489 |
the page; NOTE that because the calling thread may have an x-latch
|
|
490 |
on the page, we do not acquire an s-latch on the page, this is to
|
|
491 |
prevent deadlocks. Even if we read values which are nonsense, the
|
|
492 |
algorithm will work. */
|
|
493 |
||
494 |
pred_offset = fil_page_get_prev(frame); |
|
495 |
succ_offset = fil_page_get_next(frame); |
|
496 |
||
497 |
mutex_exit(&(buf_pool->mutex)); |
|
498 |
||
499 |
if ((offset == low) && (succ_offset == offset + 1)) { |
|
500 |
||
501 |
/* This is ok, we can continue */
|
|
502 |
new_offset = pred_offset; |
|
503 |
||
504 |
} else if ((offset == high - 1) && (pred_offset == offset - 1)) { |
|
505 |
||
506 |
/* This is ok, we can continue */
|
|
507 |
new_offset = succ_offset; |
|
508 |
} else { |
|
509 |
/* Successor or predecessor not in the right order */
|
|
510 |
||
511 |
return(0); |
|
512 |
}
|
|
513 |
||
514 |
low = (new_offset / BUF_READ_AHEAD_LINEAR_AREA) |
|
515 |
* BUF_READ_AHEAD_LINEAR_AREA; |
|
516 |
high = (new_offset / BUF_READ_AHEAD_LINEAR_AREA + 1) |
|
517 |
* BUF_READ_AHEAD_LINEAR_AREA; |
|
518 |
||
519 |
if ((new_offset != low) && (new_offset != high - 1)) { |
|
520 |
/* This is not a border page of the area: return */
|
|
521 |
||
522 |
return(0); |
|
523 |
}
|
|
524 |
||
525 |
if (high > fil_space_get_size(space)) { |
|
526 |
/* The area is not whole, return */
|
|
527 |
||
528 |
return(0); |
|
529 |
}
|
|
530 |
||
531 |
/* If we got this far, read-ahead can be sensible: do it */
|
|
532 |
||
533 |
if (ibuf_inside()) { |
|
534 |
ibuf_mode = BUF_READ_IBUF_PAGES_ONLY; |
|
535 |
} else { |
|
536 |
ibuf_mode = BUF_READ_ANY_PAGE; |
|
537 |
}
|
|
538 |
||
539 |
count = 0; |
|
540 |
||
541 |
/* Since Windows XP seems to schedule the i/o handler thread
|
|
542 |
very eagerly, and consequently it does not wait for the
|
|
543 |
full read batch to be posted, we use special heuristics here */
|
|
544 |
||
545 |
os_aio_simulated_put_read_threads_to_sleep(); |
|
546 |
||
547 |
for (i = low; i < high; i++) { |
|
548 |
/* It is only sensible to do read-ahead in the non-sync
|
|
549 |
aio mode: hence FALSE as the first parameter */
|
|
550 |
||
551 |
if (!ibuf_bitmap_page(i)) { |
|
552 |
count += buf_read_page_low( |
|
553 |
&err, FALSE, |
|
554 |
ibuf_mode | OS_AIO_SIMULATED_WAKE_LATER, |
|
555 |
space, tablespace_version, i); |
|
556 |
if (err == DB_TABLESPACE_DELETED) { |
|
557 |
ut_print_timestamp(stderr); |
|
558 |
fprintf(stderr, |
|
559 |
" InnoDB: Warning: in"
|
|
560 |
" linear readahead trying to access\n" |
|
561 |
"InnoDB: tablespace %lu page %lu,\n" |
|
562 |
"InnoDB: but the tablespace does not"
|
|
563 |
" exist or is just being dropped.\n", |
|
564 |
(ulong) space, (ulong) i); |
|
565 |
}
|
|
566 |
}
|
|
567 |
}
|
|
568 |
||
569 |
/* In simulated aio we wake the aio handler threads only after
|
|
570 |
queuing all aio requests, in native aio the following call does
|
|
571 |
nothing: */
|
|
572 |
||
573 |
os_aio_simulated_wake_handler_threads(); |
|
574 |
||
575 |
/* Flush pages from the end of the LRU list if necessary */
|
|
576 |
buf_flush_free_margin(); |
|
577 |
||
578 |
#ifdef UNIV_DEBUG
|
|
579 |
if (buf_debug_prints && (count > 0)) { |
|
580 |
fprintf(stderr, |
|
581 |
"LINEAR read-ahead space %lu offset %lu pages %lu\n", |
|
582 |
(ulong) space, (ulong) offset, (ulong) count); |
|
583 |
}
|
|
584 |
#endif /* UNIV_DEBUG */ |
|
585 |
||
586 |
++srv_read_ahead_seq; |
|
587 |
return(count); |
|
588 |
}
|
|
589 |
||
590 |
/************************************************************************
|
|
591 |
Issues read requests for pages which the ibuf module wants to read in, in
|
|
592 |
order to contract the insert buffer tree. Technically, this function is like
|
|
593 |
a read-ahead function. */
|
|
594 |
||
595 |
void
|
|
596 |
buf_read_ibuf_merge_pages( |
|
597 |
/*======================*/
|
|
598 |
ibool sync, /* in: TRUE if the caller wants this function |
|
599 |
to wait for the highest address page to get
|
|
600 |
read in, before this function returns */
|
|
601 |
ulint* space_ids, /* in: array of space ids */ |
|
602 |
ib_longlong* space_versions,/* in: the spaces must have this version |
|
603 |
number (timestamp), otherwise we discard the
|
|
604 |
read; we use this to cancel reads if
|
|
605 |
DISCARD + IMPORT may have changed the
|
|
606 |
tablespace size */
|
|
607 |
ulint* page_nos, /* in: array of page numbers to read, with the |
|
608 |
highest page number the last in the array */
|
|
609 |
ulint n_stored) /* in: number of page numbers in the array */ |
|
610 |
{
|
|
611 |
ulint err; |
|
612 |
ulint i; |
|
613 |
||
614 |
ut_ad(!ibuf_inside()); |
|
615 |
#ifdef UNIV_IBUF_DEBUG
|
|
616 |
ut_a(n_stored < UNIV_PAGE_SIZE); |
|
617 |
#endif
|
|
618 |
while (buf_pool->n_pend_reads |
|
619 |
> buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) { |
|
620 |
os_thread_sleep(500000); |
|
621 |
}
|
|
622 |
||
623 |
for (i = 0; i < n_stored; i++) { |
|
624 |
buf_read_page_low(&err, |
|
625 |
(i + 1 == n_stored) && sync, |
|
626 |
BUF_READ_ANY_PAGE, |
|
627 |
space_ids[i], space_versions[i], |
|
628 |
page_nos[i]); |
|
629 |
||
630 |
if (err == DB_TABLESPACE_DELETED) { |
|
631 |
/* We have deleted or are deleting the single-table
|
|
632 |
tablespace: remove the entries for that page */
|
|
633 |
||
634 |
ibuf_merge_or_delete_for_page(NULL, space_ids[i], |
|
635 |
page_nos[i], FALSE); |
|
636 |
}
|
|
637 |
}
|
|
638 |
||
639 |
os_aio_simulated_wake_handler_threads(); |
|
640 |
||
641 |
/* Flush pages from the end of the LRU list if necessary */
|
|
642 |
buf_flush_free_margin(); |
|
643 |
||
644 |
#ifdef UNIV_DEBUG
|
|
645 |
if (buf_debug_prints) { |
|
646 |
fprintf(stderr, |
|
647 |
"Ibuf merge read-ahead space %lu pages %lu\n", |
|
648 |
(ulong) space_ids[0], (ulong) n_stored); |
|
649 |
}
|
|
650 |
#endif /* UNIV_DEBUG */ |
|
651 |
}
|
|
652 |
||
653 |
/************************************************************************
|
|
654 |
Issues read requests for pages which recovery wants to read in. */
|
|
655 |
||
656 |
void
|
|
657 |
buf_read_recv_pages( |
|
658 |
/*================*/
|
|
659 |
ibool sync, /* in: TRUE if the caller wants this function |
|
660 |
to wait for the highest address page to get
|
|
661 |
read in, before this function returns */
|
|
662 |
ulint space, /* in: space id */ |
|
663 |
ulint* page_nos, /* in: array of page numbers to read, with the |
|
664 |
highest page number the last in the array */
|
|
665 |
ulint n_stored) /* in: number of page numbers in the array */ |
|
666 |
{
|
|
667 |
ib_longlong tablespace_version; |
|
668 |
ulint count; |
|
669 |
ulint err; |
|
670 |
ulint i; |
|
671 |
||
672 |
tablespace_version = fil_space_get_version(space); |
|
673 |
||
674 |
for (i = 0; i < n_stored; i++) { |
|
675 |
||
676 |
count = 0; |
|
677 |
||
678 |
os_aio_print_debug = FALSE; |
|
679 |
||
680 |
while (buf_pool->n_pend_reads >= recv_n_pool_free_frames / 2) { |
|
681 |
||
682 |
os_aio_simulated_wake_handler_threads(); |
|
683 |
os_thread_sleep(500000); |
|
684 |
||
685 |
count++; |
|
686 |
||
687 |
if (count > 100) { |
|
688 |
fprintf(stderr, |
|
689 |
"InnoDB: Error: InnoDB has waited for"
|
|
690 |
" 50 seconds for pending\n" |
|
691 |
"InnoDB: reads to the buffer pool to"
|
|
692 |
" be finished.\n" |
|
693 |
"InnoDB: Number of pending reads %lu,"
|
|
694 |
" pending pread calls %lu\n", |
|
695 |
(ulong) buf_pool->n_pend_reads, |
|
696 |
(ulong)os_file_n_pending_preads); |
|
697 |
||
698 |
os_aio_print_debug = TRUE; |
|
699 |
}
|
|
700 |
}
|
|
701 |
||
702 |
os_aio_print_debug = FALSE; |
|
703 |
||
704 |
if ((i + 1 == n_stored) && sync) { |
|
705 |
buf_read_page_low(&err, TRUE, BUF_READ_ANY_PAGE, |
|
706 |
space, tablespace_version, |
|
707 |
page_nos[i]); |
|
708 |
} else { |
|
709 |
buf_read_page_low(&err, FALSE, BUF_READ_ANY_PAGE |
|
710 |
| OS_AIO_SIMULATED_WAKE_LATER, |
|
711 |
space, tablespace_version, |
|
712 |
page_nos[i]); |
|
713 |
}
|
|
714 |
}
|
|
715 |
||
716 |
os_aio_simulated_wake_handler_threads(); |
|
717 |
||
718 |
/* Flush pages from the end of the LRU list if necessary */
|
|
719 |
buf_flush_free_margin(); |
|
720 |
||
721 |
#ifdef UNIV_DEBUG
|
|
722 |
if (buf_debug_prints) { |
|
723 |
fprintf(stderr, |
|
724 |
"Recovery applies read-ahead pages %lu\n", |
|
725 |
(ulong) n_stored); |
|
726 |
}
|
|
727 |
#endif /* UNIV_DEBUG */ |
|
728 |
}
|