ibuf0ibuf.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

/******************************************************
Insert buffer

(c) 1997 Innobase Oy

Created 7/19/1997 Heikki Tuuri
*******************************************************/

#include "ibuf0ibuf.h"

#ifdef UNIV_NONINL
#include "ibuf0ibuf.ic"
#endif

#include "buf0buf.h"
#include "buf0rea.h"
#include "fsp0fsp.h"
#include "trx0sys.h"
#include "fil0fil.h"
#include "thr0loc.h"
#include "rem0rec.h"
#include "btr0cur.h"
#include "btr0pcur.h"
#include "btr0btr.h"
#include "sync0sync.h"
#include "dict0boot.h"
#include "fut0lst.h"
#include "lock0lock.h"
#include "log0recv.h"
#include "que0que.h"

/*      STRUCTURE OF AN INSERT BUFFER RECORD

In versions < 4.1.x:

1. The first field is the page number.
2. The second field is an array which stores type info for each subsequent
   field. We store the information which affects the ordering of records, and
   also the physical storage size of an SQL NULL value. E.g., for CHAR(10) it
   is 10 bytes.
3. Next we have the fields of the actual index record.

In versions >= 4.1.x:

Note that contary to what we planned in the 1990's, there will only be one
insert buffer tree, and that is in the system tablespace of InnoDB.

1. The first field is the space id.
2. The second field is a one-byte marker (0) which differentiates records from
   the < 4.1.x storage format.
3. The third field is the page number.
4. The fourth field contains the type info, where we have also added 2 bytes to
   store the charset. In the compressed table format of 5.0.x we must add more
   information here so that we can build a dummy 'index' struct which 5.0.x
   can use in the binary search on the index page in the ibuf merge phase.
5. The rest of the fields contain the fields of the actual index record.

In versions >= 5.0.3:

The first byte of the fourth field is an additional marker (0) if the record
is in the compact format.  The presence of this marker can be detected by
looking at the length of the field modulo DATA_NEW_ORDER_NULL_TYPE_BUF_SIZE.

The high-order bit of the character set field in the type info is the
"nullable" flag for the field. */


/*	PREVENTING DEADLOCKS IN THE INSERT BUFFER SYSTEM

If an OS thread performs any operation that brings in disk pages from
non-system tablespaces into the buffer pool, or creates such a page there,
then the operation may have as a side effect an insert buffer index tree
compression. Thus, the tree latch of the insert buffer tree may be acquired
in the x-mode, and also the file space latch of the system tablespace may
be acquired in the x-mode.

Also, an insert to an index in a non-system tablespace can have the same
effect. How do we know this cannot lead to a deadlock of OS threads? There
is a problem with the i\o-handler threads: they break the latching order
because they own x-latches to pages which are on a lower level than the
insert buffer tree latch, its page latches, and the tablespace latch an
insert buffer operation can reserve.

The solution is the following: Let all the tree and page latches connected
with the insert buffer be later in the latching order than the fsp latch and
fsp page latches.

Insert buffer pages must be such that the insert buffer is never invoked
when these pages are accessed as this would result in a recursion violating
the latching order. We let a special i/o-handler thread take care of i/o to
the insert buffer pages and the ibuf bitmap pages, as well as the fsp bitmap
pages and the first inode page, which contains the inode of the ibuf tree: let
us call all these ibuf pages. To prevent deadlocks, we do not let a read-ahead
access both non-ibuf and ibuf pages.

Then an i/o-handler for the insert buffer never needs to access recursively the
insert buffer tree and thus obeys the latching order. On the other hand, other
i/o-handlers for other tablespaces may require access to the insert buffer,
but because all kinds of latches they need to access there are later in the
latching order, no violation of the latching order occurs in this case,
either.

A problem is how to grow and contract an insert buffer tree. As it is later
in the latching order than the fsp management, we have to reserve the fsp
latch first, before adding or removing pages from the insert buffer tree.
We let the insert buffer tree have its own file space management: a free
list of pages linked to the tree root. To prevent recursive using of the
insert buffer when adding pages to the tree, we must first load these pages
to memory, obtaining a latch on them, and only after that add them to the
free list of the insert buffer tree. More difficult is removing of pages
from the free list. If there is an excess of pages in the free list of the
ibuf tree, they might be needed if some thread reserves the fsp latch,
intending to allocate more file space. So we do the following: if a thread
reserves the fsp latch, we check the writer count field of the latch. If
this field has value 1, it means that the thread did not own the latch
before entering the fsp system, and the mtr of the thread contains no
modifications to the fsp pages. Now we are free to reserve the ibuf latch,
and check if there is an excess of pages in the free list. We can then, in a
separate mini-transaction, take them out of the free list and free them to
the fsp system.

To avoid deadlocks in the ibuf system, we divide file pages into three levels:

(1) non-ibuf pages,
(2) ibuf tree pages and the pages in the ibuf tree free list, and
(3) ibuf bitmap pages.

No OS thread is allowed to access higher level pages if it has latches to
lower level pages; even if the thread owns a B-tree latch it must not access
the B-tree non-leaf pages if it has latches on lower level pages. Read-ahead
is only allowed for level 1 and 2 pages. Dedicated i/o-handler threads handle
exclusively level 1 i/o. A dedicated i/o handler thread handles exclusively
level 2 i/o. However, if an OS thread does the i/o handling for itself, i.e.,
it uses synchronous aio, it can access any pages, as long as it obeys the
access order rules. */

/* Buffer pool size per the maximum insert buffer size */
#define IBUF_POOL_SIZE_PER_MAX_SIZE	2

/* The insert buffer control structure */
ibuf_t*	ibuf			= NULL;

static
ulint	ibuf_rnd		= 986058871;

ulint	ibuf_flush_count	= 0;

/* Dimensions for the ibuf_count array */
#define IBUF_COUNT_N_SPACES	500
#define IBUF_COUNT_N_PAGES	2000

/* Buffered entry counts for file pages, used in debugging */
static ulint*	ibuf_counts[IBUF_COUNT_N_SPACES];

static ibool	ibuf_counts_inited	= FALSE;

/* The start address for an insert buffer bitmap page bitmap */
#define IBUF_BITMAP		PAGE_DATA

/* Offsets in bits for the bits describing a single page in the bitmap */
#define	IBUF_BITMAP_FREE	0
#define IBUF_BITMAP_BUFFERED	2
#define IBUF_BITMAP_IBUF	3	/* TRUE if page is a part of the ibuf
					tree, excluding the root page, or is
					in the free list of the ibuf */

/* Number of bits describing a single page */
#define IBUF_BITS_PER_PAGE	4
#if IBUF_BITS_PER_PAGE % 2
# error "IBUF_BITS_PER_PAGE must be an even number!"
#endif

/* The mutex used to block pessimistic inserts to ibuf trees */
static mutex_t	ibuf_pessimistic_insert_mutex;

/* The mutex protecting the insert buffer structs */
static mutex_t	ibuf_mutex;

/* The mutex protecting the insert buffer bitmaps */
static mutex_t	ibuf_bitmap_mutex;

/* The area in pages from which contract looks for page numbers for merge */
#define	IBUF_MERGE_AREA			8

/* Inside the merge area, pages which have at most 1 per this number less
buffered entries compared to maximum volume that can buffered for a single
page are merged along with the page whose buffer became full */
#define IBUF_MERGE_THRESHOLD		4

/* In ibuf_contract at most this number of pages is read to memory in one
batch, in order to merge the entries for them in the insert buffer */
#define	IBUF_MAX_N_PAGES_MERGED		IBUF_MERGE_AREA

/* If the combined size of the ibuf trees exceeds ibuf->max_size by this
many pages, we start to contract it in connection to inserts there, using
non-synchronous contract */
#define IBUF_CONTRACT_ON_INSERT_NON_SYNC	0

/* Same as above, but use synchronous contract */
#define IBUF_CONTRACT_ON_INSERT_SYNC		5

/* Same as above, but no insert is done, only contract is called */
#define IBUF_CONTRACT_DO_NOT_INSERT		10

/* TODO: how to cope with drop table if there are records in the insert
buffer for the indexes of the table? Is there actually any problem,
because ibuf merge is done to a page when it is read in, and it is
still physically like the index page even if the index would have been
dropped! So, there seems to be no problem. */

/**********************************************************************
Validates the ibuf data structures when the caller owns ibuf_mutex. */

ibool
ibuf_validate_low(void);
/*===================*/
			/* out: TRUE if ok */

/**********************************************************************
Sets the flag in the current OS thread local storage denoting that it is
inside an insert buffer routine. */
UNIV_INLINE
void
ibuf_enter(void)
/*============*/
{
	ibool*	ptr;

	ptr = thr_local_get_in_ibuf_field();

	ut_ad(*ptr == FALSE);

	*ptr = TRUE;
}

/**********************************************************************
Sets the flag in the current OS thread local storage denoting that it is
exiting an insert buffer routine. */
UNIV_INLINE
void
ibuf_exit(void)
/*===========*/
{
	ibool*	ptr;

	ptr = thr_local_get_in_ibuf_field();

	ut_ad(*ptr == TRUE);

	*ptr = FALSE;
}

/**********************************************************************
Returns TRUE if the current OS thread is performing an insert buffer
routine. */

ibool
ibuf_inside(void)
/*=============*/
		/* out: TRUE if inside an insert buffer routine: for instance,
		a read-ahead of non-ibuf pages is then forbidden */
{
	return(*thr_local_get_in_ibuf_field());
}

/**********************************************************************
Gets the ibuf header page and x-latches it. */
static
page_t*
ibuf_header_page_get(
/*=================*/
			/* out: insert buffer header page */
	ulint	space,	/* in: space id */
	mtr_t*	mtr)	/* in: mtr */
{
	page_t*	page;

	ut_a(space == 0);

	ut_ad(!ibuf_inside());

	page = buf_page_get(space, FSP_IBUF_HEADER_PAGE_NO, RW_X_LATCH, mtr);

#ifdef UNIV_SYNC_DEBUG
	buf_page_dbg_add_level(page, SYNC_IBUF_HEADER);
#endif /* UNIV_SYNC_DEBUG */

	return(page);
}

/**********************************************************************
Gets the root page and x-latches it. */
static
page_t*
ibuf_tree_root_get(
/*===============*/
				/* out: insert buffer tree root page */
	ibuf_data_t*	data,	/* in: ibuf data */
	ulint		space,	/* in: space id */
	mtr_t*		mtr)	/* in: mtr */
{
	page_t*	page;

	ut_a(space == 0);
	ut_ad(ibuf_inside());

	mtr_x_lock(dict_tree_get_lock((data->index)->tree), mtr);

	page = buf_page_get(space, FSP_IBUF_TREE_ROOT_PAGE_NO, RW_X_LATCH,
									mtr);
#ifdef UNIV_SYNC_DEBUG
	buf_page_dbg_add_level(page, SYNC_TREE_NODE);
#endif /* UNIV_SYNC_DEBUG */

	return(page);
}

/**********************************************************************
Gets the ibuf count for a given page. */

ulint
ibuf_count_get(
/*===========*/
			/* out: number of entries in the insert buffer
			currently buffered for this page */
	ulint	space,	/* in: space id */
	ulint	page_no)/* in: page number */
{
	ut_ad(space < IBUF_COUNT_N_SPACES);
	ut_ad(page_no < IBUF_COUNT_N_PAGES);

	if (!ibuf_counts_inited) {

		return(0);
	}

	return(*(ibuf_counts[space] + page_no));
}

/**********************************************************************
Sets the ibuf count for a given page. */
#ifdef UNIV_IBUF_DEBUG
static
void
ibuf_count_set(
/*===========*/
	ulint	space,	/* in: space id */
	ulint	page_no,/* in: page number */
	ulint	val)	/* in: value to set */
{
	ut_a(space < IBUF_COUNT_N_SPACES);
	ut_a(page_no < IBUF_COUNT_N_PAGES);
	ut_a(val < UNIV_PAGE_SIZE);

	*(ibuf_counts[space] + page_no) = val;
}
#endif

/**********************************************************************
Creates the insert buffer data structure at a database startup and initializes
the data structures for the insert buffer. */

void
ibuf_init_at_db_start(void)
/*=======================*/
{
	ibuf = mem_alloc(sizeof(ibuf_t));

	/* Note that also a pessimistic delete can sometimes make a B-tree
	grow in size, as the references on the upper levels of the tree can
	change */
	
	ibuf->max_size = buf_pool_get_curr_size() / UNIV_PAGE_SIZE
						/ IBUF_POOL_SIZE_PER_MAX_SIZE;
	ibuf->meter = IBUF_THRESHOLD + 1;

	UT_LIST_INIT(ibuf->data_list);

	ibuf->size = 0;					

#ifdef UNIV_IBUF_DEBUG
	{
		ulint	i, j;

		for (i = 0; i < IBUF_COUNT_N_SPACES; i++) {

			ibuf_counts[i] = mem_alloc(sizeof(ulint)
						* IBUF_COUNT_N_PAGES);
			for (j = 0; j < IBUF_COUNT_N_PAGES; j++) {
				ibuf_count_set(i, j, 0);
			}
		}
	}	
#endif
	mutex_create(&ibuf_pessimistic_insert_mutex);

	mutex_set_level(&ibuf_pessimistic_insert_mutex,
						SYNC_IBUF_PESS_INSERT_MUTEX);
	mutex_create(&ibuf_mutex);

	mutex_set_level(&ibuf_mutex, SYNC_IBUF_MUTEX);

	mutex_create(&ibuf_bitmap_mutex);

	mutex_set_level(&ibuf_bitmap_mutex, SYNC_IBUF_BITMAP_MUTEX);

	fil_ibuf_init_at_db_start();

	ibuf_counts_inited = TRUE;
}

/**********************************************************************
Updates the size information in an ibuf data, assuming the segment size has
not changed. */
static
void
ibuf_data_sizes_update(
/*===================*/
	ibuf_data_t*	data,	/* in: ibuf data struct */
	page_t*		root,	/* in: ibuf tree root */
	mtr_t*		mtr)	/* in: mtr */
{
	ulint	old_size;

#ifdef UNIV_SYNC_DEBUG
	ut_ad(mutex_own(&ibuf_mutex));
#endif /* UNIV_SYNC_DEBUG */

	old_size = data->size;
	
	data->free_list_len = flst_get_len(root + PAGE_HEADER
					   + PAGE_BTR_IBUF_FREE_LIST, mtr);

	data->height = 1 + btr_page_get_level(root, mtr);

	data->size = data->seg_size - (1 + data->free_list_len);
				/* the '1 +' is the ibuf header page */
	ut_ad(data->size < data->seg_size);

	if (page_get_n_recs(root) == 0) {

		data->empty = TRUE;
	} else {
		data->empty = FALSE;
	}

	ut_ad(ibuf->size + data->size >= old_size);

	ibuf->size = ibuf->size + data->size - old_size;

/*	fprintf(stderr, "ibuf size %lu, space ibuf size %lu\n", ibuf->size,
							data->size); */
}

/**********************************************************************
Creates the insert buffer data struct for a single tablespace. Reads the
root page of the insert buffer tree in the tablespace. This function can
be called only after the dictionary system has been initialized, as this
creates also the insert buffer table and index into this tablespace. */

ibuf_data_t*
ibuf_data_init_for_space(
/*=====================*/
			/* out, own: ibuf data struct, linked to the list
			in ibuf control structure */
	ulint	space)	/* in: space id */
{
	ibuf_data_t*	data;
	page_t*		root;
	page_t*		header_page;
	mtr_t		mtr;
	char		buf[50];
	dict_table_t*	table;
	dict_index_t*	index;
	ulint		n_used;
	
	ut_a(space == 0);

#ifdef UNIV_LOG_DEBUG
	if (space % 2 == 1) {

		fputs("No ibuf op in replicate space\n", stderr);

		return(NULL);
	}
#endif
	data = mem_alloc(sizeof(ibuf_data_t));

	data->space = space;

	mtr_start(&mtr);

	mutex_enter(&ibuf_mutex);

	mtr_x_lock(fil_space_get_latch(space), &mtr);

	header_page = ibuf_header_page_get(space, &mtr);

	fseg_n_reserved_pages(header_page + IBUF_HEADER + IBUF_TREE_SEG_HEADER,
								&n_used, &mtr);
	ibuf_enter();