buf0buf.c

这是linux下运行的mysql软件包,可用于linux 下安装 php + mysql + apach 的网络配置

/*   Innobase relational database engine; Copyright (C) 2001 Innobase Oy
     
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License 2
     as published by the Free Software Foundation in June 1991.
     
     This program is distributed in the hope that it will be useful,
     but WITHOUT ANY WARRANTY; without even the implied warranty of
     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
     GNU General Public License for more details.
     
     You should have received a copy of the GNU General Public License 2
     along with this program (in file COPYING); if not, write to the Free
     Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. */
/******************************************************
The database buffer buf_pool

(c) 1995 Innobase Oy

Created 11/5/1995 Heikki Tuuri
*******************************************************/

#include "buf0buf.h"

#ifdef UNIV_NONINL
#include "buf0buf.ic"
#endif

#include "mem0mem.h"
#include "btr0btr.h"
#include "fil0fil.h"
#include "lock0lock.h"
#include "btr0sea.h"
#include "ibuf0ibuf.h"
#include "dict0dict.h"
#include "log0recv.h"
#include "log0log.h"
#include "trx0undo.h"
#include "srv0srv.h"

/*
		IMPLEMENTATION OF THE BUFFER POOL
		=================================

Performance improvement: 
------------------------
Thread scheduling in NT may be so slow that the OS wait mechanism should
not be used even in waiting for disk reads to complete.
Rather, we should put waiting query threads to the queue of
waiting jobs, and let the OS thread do something useful while the i/o
is processed. In this way we could remove most OS thread switches in
an i/o-intensive benchmark like TPC-C.

A possibility is to put a user space thread library between the database
and NT. User space thread libraries might be very fast.

SQL Server 7.0 can be configured to use 'fibers' which are lightweight
threads in NT. These should be studied.

		Buffer frames and blocks
		------------------------
Following the terminology of Gray and Reuter, we call the memory
blocks where file pages are loaded buffer frames. For each buffer
frame there is a control block, or shortly, a block, in the buffer
control array. The control info which does not need to be stored
in the file along with the file page, resides in the control block.

		Buffer pool struct
		------------------
The buffer buf_pool contains a single mutex which protects all the
control data structures of the buf_pool. The content of a buffer frame is
protected by a separate read-write lock in its control block, though.
These locks can be locked and unlocked without owning the buf_pool mutex.
The OS events in the buf_pool struct can be waited for without owning the
buf_pool mutex.

The buf_pool mutex is a hot-spot in main memory, causing a lot of
memory bus traffic on multiprocessor systems when processors
alternately access the mutex. On our Pentium, the mutex is accessed
maybe every 10 microseconds. We gave up the solution to have mutexes
for each control block, for instance, because it seemed to be
complicated.

A solution to reduce mutex contention of the buf_pool mutex is to
create a separate mutex for the page hash table. On Pentium,
accessing the hash table takes 2 microseconds, about half
of the total buf_pool mutex hold time.

		Control blocks
		--------------

The control block contains, for instance, the bufferfix count
which is incremented when a thread wants a file page to be fixed
in a buffer frame. The bufferfix operation does not lock the
contents of the frame, however. For this purpose, the control
block contains a read-write lock.

The buffer frames have to be aligned so that the start memory
address of a frame is divisible by the universal page size, which
is a power of two.

We intend to make the buffer buf_pool size on-line reconfigurable,
that is, the buf_pool size can be changed without closing the database.
Then the database administarator may adjust it to be bigger
at night, for example. The control block array must
contain enough control blocks for the maximum buffer buf_pool size
which is used in the particular database.
If the buf_pool size is cut, we exploit the virtual memory mechanism of
the OS, and just refrain from using frames at high addresses. Then the OS
can swap them to disk.

The control blocks containing file pages are put to a hash table
according to the file address of the page.
We could speed up the access to an individual page by using
"pointer swizzling": we could replace the page references on
non-leaf index pages by direct pointers to the page, if it exists
in the buf_pool. We could make a separate hash table where we could
chain all the page references in non-leaf pages residing in the buf_pool,
using the page reference as the hash key,
and at the time of reading of a page update the pointers accordingly.
Drawbacks of this solution are added complexity and,
possibly, extra space required on non-leaf pages for memory pointers.
A simpler solution is just to speed up the hash table mechanism
in the database, using tables whose size is a power of 2.

		Lists of blocks
		---------------

There are several lists of control blocks. The free list contains
blocks which are currently not used.

The LRU-list contains all the blocks holding a file page
except those for which the bufferfix count is non-zero.
The pages are in the LRU list roughly in the order of the last
access to the page, so that the oldest pages are at the end of the
list. We also keep a pointer to near the end of the LRU list,
which we can use when we want to artificially age a page in the
buf_pool. This is used if we know that some page is not needed
again for some time: we insert the block right after the pointer,
causing it to be replaced sooner than would noramlly be the case.
Currently this aging mechanism is used for read-ahead mechanism
of pages, and it can also be used when there is a scan of a full
table which cannot fit in the memory. Putting the pages near the
of the LRU list, we make sure that most of the buf_pool stays in the
main memory, undisturbed.

The chain of modified blocks contains the blocks
holding file pages that have been modified in the memory
but not written to disk yet. The block with the oldest modification
which has not yet been written to disk is at the end of the chain.

		Loading a file page
		-------------------

First, a victim block for replacement has to be found in the
buf_pool. It is taken from the free list or searched for from the
end of the LRU-list. An exclusive lock is reserved for the frame,
the io_fix field is set in the block fixing the block in buf_pool,
and the io-operation for loading the page is queued. The io-handler thread
releases the X-lock on the frame and resets the io_fix field
when the io operation completes.

A thread may request the above operation using the buf_page_get-
function. It may then continue to request a lock on the frame.
The lock is granted when the io-handler releases the x-lock.

		Read-ahead
		----------

The read-ahead mechanism is intended to be intelligent and
isolated from the semantically higher levels of the database
index management. From the higher level we only need the
information if a file page has a natural successor or
predecessor page. On the leaf level of a B-tree index,
these are the next and previous pages in the natural
order of the pages.

Let us first explain the read-ahead mechanism when the leafs
of a B-tree are scanned in an ascending or descending order.
When a read page is the first time referenced in the buf_pool,
the buffer manager checks if it is at the border of a so-called
linear read-ahead area. The tablespace is divided into these
areas of size 64 blocks, for example. So if the page is at the
border of such an area, the read-ahead mechanism checks if
all the other blocks in the area have been accessed in an
ascending or descending order. If this is the case, the system
looks at the natural successor or predecessor of the page,
checks if that is at the border of another area, and in this case
issues read-requests for all the pages in that area. Maybe
we could relax the condition that all the pages in the area
have to be accessed: if data is deleted from a table, there may
appear holes of unused pages in the area.

A different read-ahead mechanism is used when there appears
to be a random access pattern to a file.
If a new page is referenced in the buf_pool, and several pages
of its random access area (for instance, 32 consecutive pages
in a tablespace) have recently been referenced, we may predict
that the whole area may be needed in the near future, and issue
the read requests for the whole area.

		AWE implementation
		------------------

By a 'block' we mean the buffer header of type buf_block_t. By a 'page'
we mean the physical 16 kB memory area allocated from RAM for that block.
By a 'frame' we mean a 16 kB area in the virtual address space of the
process, in the frame_mem of buf_pool.

We can map pages to the frames of the buffer pool.

1) A buffer block allocated to use as a non-data page, e.g., to the lock
table, is always mapped to a frame.
2) A bufferfixed or io-fixed data page is always mapped to a frame.
3) When we need to map a block to frame, we look from the list
awe_LRU_free_mapped and try to unmap its last block, but note that
bufferfixed or io-fixed pages cannot be unmapped.
4) For every frame in the buffer pool there is always a block whose page is
mapped to it. When we create the buffer pool, we map the first elements
in the free list to the frames.
5) When we have AWE enabled, we disable adaptive hash indexes.
*/

buf_pool_t*	buf_pool = NULL; /* The buffer buf_pool of the database */

#ifdef UNIV_DEBUG
ulint		buf_dbg_counter	= 0; /* This is used to insert validation
					operations in excution in the
					debug version */
ibool		buf_debug_prints = FALSE; /* If this is set TRUE,
					the program prints info whenever
					read-ahead or flush occurs */
#endif /* UNIV_DEBUG */
/************************************************************************
Calculates a page checksum which is stored to the page when it is written
to a file. Note that we must be careful to calculate the same value on
32-bit and 64-bit architectures. */

ulint
buf_calc_page_new_checksum(
/*=======================*/
		       /* out: checksum */
	byte*    page) /* in: buffer page */
{
  	ulint checksum;

        /* Since the field FIL_PAGE_FILE_FLUSH_LSN, and in versions <= 4.1.x
        ..._ARCH_LOG_NO, are written outside the buffer pool to the first
        pages of data files, we have to skip them in the page checksum
        calculation.
	We must also skip the field FIL_PAGE_SPACE_OR_CHKSUM where the
	checksum is stored, and also the last 8 bytes of page because
	there we store the old formula checksum. */
  	
  	checksum = ut_fold_binary(page + FIL_PAGE_OFFSET,
				 FIL_PAGE_FILE_FLUSH_LSN - FIL_PAGE_OFFSET)
  		   + ut_fold_binary(page + FIL_PAGE_DATA, 
				           UNIV_PAGE_SIZE - FIL_PAGE_DATA
				           - FIL_PAGE_END_LSN_OLD_CHKSUM);
  	checksum = checksum & 0xFFFFFFFFUL;

  	return(checksum);
}

/************************************************************************
In versions < 4.0.14 and < 4.1.1 there was a bug that the checksum only
looked at the first few bytes of the page. This calculates that old
checksum. 
NOTE: we must first store the new formula checksum to
FIL_PAGE_SPACE_OR_CHKSUM before calculating and storing this old checksum
because this takes that field as an input! */

ulint
buf_calc_page_old_checksum(
/*=======================*/
		       /* out: checksum */
	byte*    page) /* in: buffer page */
{
  	ulint checksum;
  	
  	checksum = ut_fold_binary(page, FIL_PAGE_FILE_FLUSH_LSN);

  	checksum = checksum & 0xFFFFFFFFUL;

  	return(checksum);
}

/************************************************************************
Checks if a page is corrupt. */

ibool
buf_page_is_corrupted(
/*==================*/
				/* out: TRUE if corrupted */
	byte*	read_buf)	/* in: a database page */
{
	ulint	checksum;
	ulint	old_checksum;
	ulint	checksum_field;
	ulint	old_checksum_field;
#ifndef UNIV_HOTBACKUP
	dulint	current_lsn;
#endif
	if (mach_read_from_4(read_buf + FIL_PAGE_LSN + 4)
	     != mach_read_from_4(read_buf + UNIV_PAGE_SIZE
				- FIL_PAGE_END_LSN_OLD_CHKSUM + 4)) {

		/* Stored log sequence numbers at the start and the end
		of page do not match */

		return(TRUE);
	}

#ifndef UNIV_HOTBACKUP
	if (recv_lsn_checks_on && log_peek_lsn(&current_lsn)) {
		if (ut_dulint_cmp(current_lsn,
				  mach_read_from_8(read_buf + FIL_PAGE_LSN))
				 < 0) {
			ut_print_timestamp(stderr);

			fprintf(stderr,
"  InnoDB: Error: page %lu log sequence number %lu %lu\n"
"InnoDB: is in the future! Current system log sequence number %lu %lu.\n"
"InnoDB: Your database may be corrupt or you may have copied the InnoDB\n"
"InnoDB: tablespace but not the InnoDB log files. See\n"
"http://dev.mysql.com/doc/mysql/en/backing-up.html for more information.\n",
		        (ulong) mach_read_from_4(read_buf + FIL_PAGE_OFFSET),
			(ulong) ut_dulint_get_high(
				mach_read_from_8(read_buf + FIL_PAGE_LSN)),
			(ulong) ut_dulint_get_low(
				mach_read_from_8(read_buf + FIL_PAGE_LSN)),
			(ulong) ut_dulint_get_high(current_lsn),
			(ulong) ut_dulint_get_low(current_lsn));
		}
	}
#endif
  
  /* If we use checksums validation, make additional check before returning
  TRUE to ensure that the checksum is not equal to BUF_NO_CHECKSUM_MAGIC which
  might be stored by InnoDB with checksums disabled.
     Otherwise, skip checksum calculation and return FALSE */
  
  if (srv_use_checksums) {
    old_checksum = buf_calc_page_old_checksum(read_buf); 

    old_checksum_field = mach_read_from_4(read_buf + UNIV_PAGE_SIZE
					- FIL_PAGE_END_LSN_OLD_CHKSUM);

    /* There are 2 valid formulas for old_checksum_field:
	  1. Very old versions of InnoDB only stored 8 byte lsn to the start
	  and the end of the page.
	  2. Newer InnoDB versions store the old formula checksum there. */
	
    if (old_checksum_field != mach_read_from_4(read_buf + FIL_PAGE_LSN)
        && old_checksum_field != old_checksum
        && old_checksum_field != BUF_NO_CHECKSUM_MAGIC) {

      return(TRUE);
    }

    checksum = buf_calc_page_new_checksum(read_buf);
    checksum_field = mach_read_from_4(read_buf + FIL_PAGE_SPACE_OR_CHKSUM);

    /* InnoDB versions < 4.0.14 and < 4.1.1 stored the space id
	  (always equal to 0), to FIL_PAGE_SPACE_SPACE_OR_CHKSUM */

    if (checksum_field != 0 && checksum_field != checksum
        && checksum_field != BUF_NO_CHECKSUM_MAGIC) {

      return(TRUE);
    }
  }
  
	return(FALSE);
}

/************************************************************************
Prints a page to stderr. */

void
buf_page_print(
/*===========*/
	byte*	read_buf)	/* in: a database page */
{
	dict_index_t*	index;
	ulint		checksum;
	ulint		old_checksum;

	ut_print_timestamp(stderr);
	fprintf(stderr, "  InnoDB: Page dump in ascii and hex (%lu bytes):\n",
		(ulint)UNIV_PAGE_SIZE);
	ut_print_buf(stderr, read_buf, UNIV_PAGE_SIZE);
	fputs("InnoDB: End of page dump\n", stderr);

	checksum = srv_use_checksums ?
    buf_calc_page_new_checksum(read_buf) : BUF_NO_CHECKSUM_MAGIC;
	old_checksum = srv_use_checksums ?
    buf_calc_page_old_checksum(read_buf) : BUF_NO_CHECKSUM_MAGIC;

	ut_print_timestamp(stderr);
	fprintf(stderr, 
"  InnoDB: Page checksum %lu, prior-to-4.0.14-form checksum %lu\n"
"InnoDB: stored checksum %lu, prior-to-4.0.14-form stored checksum %lu\n",
			(ulong) checksum, (ulong) old_checksum,
			(ulong) mach_read_from_4(read_buf + FIL_PAGE_SPACE_OR_CHKSUM),
			(ulong) mach_read_from_4(read_buf + UNIV_PAGE_SIZE
					- FIL_PAGE_END_LSN_OLD_CHKSUM));
	fprintf(stderr,
"InnoDB: Page lsn %lu %lu, low 4 bytes of lsn at page end %lu\n"
"InnoDB: Page number (if stored to page already) %lu,\n"
"InnoDB: space id (if created with >= MySQL-4.1.1 and stored already) %lu\n",
		(ulong) mach_read_from_4(read_buf + FIL_PAGE_LSN),
		(ulong) mach_read_from_4(read_buf + FIL_PAGE_LSN + 4),
		(ulong) mach_read_from_4(read_buf + UNIV_PAGE_SIZE
					- FIL_PAGE_END_LSN_OLD_CHKSUM + 4),
		(ulong) mach_read_from_4(read_buf + FIL_PAGE_OFFSET),
		(ulong) mach_read_from_4(read_buf + FIL_PAGE_ARCH_LOG_NO_OR_SPACE_ID));

	if (mach_read_from_2(read_buf + TRX_UNDO_PAGE_HDR + TRX_UNDO_PAGE_TYPE)
	    == TRX_UNDO_INSERT) {
	    	fprintf(stderr,
			"InnoDB: Page may be an insert undo log page\n");
	} else if (mach_read_from_2(read_buf + TRX_UNDO_PAGE_HDR
						+ TRX_UNDO_PAGE_TYPE)
	    	== TRX_UNDO_UPDATE) {
	    	fprintf(stderr,
			"InnoDB: Page may be an update undo log page\n");
	}

	if (fil_page_get_type(read_buf) == FIL_PAGE_INDEX) {
	    	fprintf(stderr,
"InnoDB: Page may be an index page where index id is %lu %lu\n",
			(ulong) ut_dulint_get_high(btr_page_get_index_id(read_buf)),
			(ulong) ut_dulint_get_low(btr_page_get_index_id(read_buf)));

		/* If the code is in ibbackup, dict_sys may be uninitialized,
		i.e., NULL */

		if (dict_sys != NULL) {

		        index = dict_index_find_on_id_low(
					btr_page_get_index_id(read_buf));
		        if (index) {
				fputs("InnoDB: (", stderr);
				dict_index_name_print(stderr, NULL, index);
				fputs(")\n", stderr);
			}
		}
	} else if (fil_page_get_type(read_buf) == FIL_PAGE_INODE) {
		fputs("InnoDB: Page may be an 'inode' page\n", stderr);
	} else if (fil_page_get_type(read_buf) == FIL_PAGE_IBUF_FREE_LIST) {
		fputs("InnoDB: Page may be an insert buffer free list page\n",
			stderr);
	}
}

/************************************************************************
Initializes a buffer control block when the buf_pool is created. */
static
void
buf_block_init(
/*===========*/
	buf_block_t*	block,	/* in: pointer to control block */
	byte*		frame)	/* in: pointer to buffer frame, or NULL if in
				the case of AWE there is no frame */
{
	block->magic_n = 0;

	block->state = BUF_BLOCK_NOT_USED;
	
	block->frame = frame;

	block->awe_info = NULL;

	block->buf_fix_count = 0;
	block->io_fix = 0;