/**********************************************************************//**
Adds a column definition to a table. */
UNIV_INTERN
void
dict_mem_table_add_col(
/*===================*/
dict_table_t* table, /*!< in: table */
mem_heap_t* heap, /*!< in: temporary memory heap, or NULL */
const char* name, /*!< in: column name, or NULL */
ulint mtype, /*!< in: main datatype */
ulint prtype, /*!< in: precise type */
ulint len) /*!< in: precision */
{
dict_col_t* col;
#ifndef UNIV_HOTBACKUP
ulint mbminlen;
ulint mbmaxlen;
#endif /* !UNIV_HOTBACKUP */
ulint i;
ut_ad(table);
ut_ad(table->magic_n == DICT_TABLE_MAGIC_N);
ut_ad(!heap == !name);
i = table->n_def++;
if (name) {
if (UNIV_UNLIKELY(table->n_def == table->n_cols)) {
heap = table->heap;
}
if (UNIV_LIKELY(i) && UNIV_UNLIKELY(!table->col_names)) {
/* All preceding column names are empty. */
char* s = mem_heap_zalloc(heap, table->n_def);
table->col_names = s;
}
table->col_names = dict_add_col_name(table->col_names,
i, name, heap);
}
col = dict_table_get_nth_col(table, i);
col->ind = (unsigned int) i;
col->ord_part = 0;
col->mtype = (unsigned int) mtype;
col->prtype = (unsigned int) prtype;
col->len = (unsigned int) len;
#ifndef UNIV_HOTBACKUP
dtype_get_mblen(mtype, prtype, &mbminlen, &mbmaxlen);
col->mbminlen = (unsigned int) mbminlen;
col->mbmaxlen = (unsigned int) mbmaxlen;
#endif /* !UNIV_HOTBACKUP */
}
/***********************************************************//**
Parses the row reference and other info in a fresh insert undo record. */
static
void
row_undo_ins_parse_undo_rec(
/*========================*/
ib_recovery_t recovery, /*!< in: recovery flag */
undo_node_t* node) /*!< in/out: row undo node */
{
dict_index_t* clust_index;
byte* ptr;
undo_no_t undo_no;
dulint table_id;
ulint type;
ulint dummy;
ibool dummy_extern;
ut_ad(node);
ptr = trx_undo_rec_get_pars(node->undo_rec, &type, &dummy,
&dummy_extern, &undo_no, &table_id);
ut_ad(type == TRX_UNDO_INSERT_REC);
node->rec_type = type;
node->update = NULL;
node->table = dict_table_get_on_id(
srv_force_recovery, table_id, node->trx);
/* Skip the UNDO if we can't find the table or the .ibd file. */
if (UNIV_UNLIKELY(node->table == NULL)) {
} else if (UNIV_UNLIKELY(node->table->ibd_file_missing)) {
node->table = NULL;
} else {
clust_index = dict_table_get_first_index(node->table);
if (clust_index != NULL) {
ptr = trx_undo_rec_get_row_ref(
ptr, clust_index, &node->ref, node->heap);
} else {
ut_print_timestamp(ib_stream);
ib_logger(ib_stream, " InnoDB: table ");
ut_print_name(ib_stream, node->trx, TRUE,
node->table->name);
ib_logger(ib_stream, " has no indexes, "
"ignoring the table\n");
node->table = NULL;
}
}
}
/********************************************************************//**
Fills the column prefix cache of an externally stored column. */
static
void
row_ext_cache_fill(
/*===============*/
row_ext_t* ext, /*!< in/out: column prefix cache */
ulint i, /*!< in: index of ext->ext[] */
ulint zip_size,/*!< compressed page size in bytes, or 0 */
const dfield_t* dfield) /*!< in: data field */
{
const byte* field = dfield_get_data(dfield);
ulint f_len = dfield_get_len(dfield);
byte* buf = ext->buf + i * ext->max_len;
ut_ad(ext->max_len > 0);
ut_ad(i < ext->n_ext);
ut_ad(dfield_is_ext(dfield));
ut_a(f_len >= BTR_EXTERN_FIELD_REF_SIZE);
if (UNIV_UNLIKELY(!memcmp(field_ref_zero,
field + f_len - BTR_EXTERN_FIELD_REF_SIZE,
BTR_EXTERN_FIELD_REF_SIZE))) {
/* The BLOB pointer is not set: we cannot fetch it */
ext->len[i] = 0;
} else {
/* Fetch at most ext->max_len of the column.
The column should be non-empty. However,
trx_rollback_or_clean_all_recovered() may try to
access a half-deleted BLOB if the server previously
crashed during the execution of
btr_free_externally_stored_field(). */
ext->len[i] = btr_copy_externally_stored_field_prefix(
buf, ext->max_len, zip_size, field, f_len);
}
}
/*******************************************************************//**
Drops the index tree associated with a row in SYS_INDEXES table. */
UNIV_INTERN
void
dict_drop_index_tree(
/*=================*/
rec_t* rec, /*!< in/out: record in the clustered index
of SYS_INDEXES table */
mtr_t* mtr) /*!< in: mtr having the latch on the record page */
{
ulint root_page_no;
ulint space;
ulint zip_size;
const byte* ptr;
ulint len;
ut_ad(mutex_own(&(dict_sys->mutex)));
ut_a(!dict_table_is_comp(dict_sys->sys_indexes));
ptr = rec_get_nth_field_old(rec, DICT_SYS_INDEXES_PAGE_NO_FIELD, &len);
ut_ad(len == 4);
root_page_no = mtr_read_ulint(ptr, MLOG_4BYTES, mtr);
if (root_page_no == FIL_NULL) {
/* The tree has already been freed */
return;
}
ptr = rec_get_nth_field_old(rec,
DICT_SYS_INDEXES_SPACE_NO_FIELD, &len);
ut_ad(len == 4);
space = mtr_read_ulint(ptr, MLOG_4BYTES, mtr);
zip_size = fil_space_get_zip_size(space);
if (UNIV_UNLIKELY(zip_size == ULINT_UNDEFINED)) {
/* It is a single table tablespace and the .ibd file is
missing: do nothing */
return;
}
/* We free all the pages but the root page first; this operation
may span several mini-transactions */
btr_free_but_not_root(space, zip_size, root_page_no);
/* Then we free the root page in the same mini-transaction where
we write FIL_NULL to the appropriate field in the SYS_INDEXES
record: this mini-transaction marks the B-tree totally freed */
/* printf("Dropping index tree in space %lu root page %lu\n", space,
root_page_no); */
btr_free_root(space, zip_size, root_page_no, mtr);
page_rec_write_index_page_no(rec,
DICT_SYS_INDEXES_PAGE_NO_FIELD,
FIL_NULL, mtr);
}
/********************************************************//**
Parses a log record written by mlog_write_string.
@return parsed record end, NULL if not a complete record */
UNIV_INTERN
byte*
mlog_parse_string(
/*==============*/
byte* ptr, /*!< in: buffer */
byte* end_ptr,/*!< in: buffer end */
byte* page, /*!< in: page where to apply the log record, or NULL */
void* page_zip)/*!< in/out: compressed page, or NULL */
{
ulint offset;
ulint len;
ut_a(!page || !page_zip || fil_page_get_type(page) != FIL_PAGE_INDEX);
if (end_ptr < ptr + 4) {
return(NULL);
}
offset = mach_read_from_2(ptr);
ptr += 2;
len = mach_read_from_2(ptr);
ptr += 2;
if (UNIV_UNLIKELY(offset >= UNIV_PAGE_SIZE)
|| UNIV_UNLIKELY(len + offset > UNIV_PAGE_SIZE)) {
recv_sys->found_corrupt_log = TRUE;
return(NULL);
}
if (end_ptr < ptr + len) {
return(NULL);
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
memcpy(((page_zip_des_t*) page_zip)->data
+ offset, ptr, len);
}
memcpy(page + offset, ptr, len);
}
return(ptr + len);
}
/***********************************************************//**
Undoes a modify in secondary indexes when undo record type is UPD_DEL.
@return DB_SUCCESS or DB_OUT_OF_FILE_SPACE */
static
ulint
row_undo_mod_upd_del_sec(
/*=====================*/
undo_node_t* node, /*!< in: row undo node */
que_thr_t* thr) /*!< in: query thread */
{
mem_heap_t* heap;
dtuple_t* entry;
dict_index_t* index;
ulint err = DB_SUCCESS;
ut_ad(node->rec_type == TRX_UNDO_UPD_DEL_REC);
heap = mem_heap_create(1024);
while (node->index != NULL) {
/* Skip all corrupted secondary index */
dict_table_skip_corrupt_index(node->index);
if (!node->index) {
break;
}
index = node->index;
entry = row_build_index_entry(node->row, node->ext,
index, heap);
if (UNIV_UNLIKELY(!entry)) {
/* The database must have crashed after
inserting a clustered index record but before
writing all the externally stored columns of
that record. Because secondary index entries
are inserted after the clustered index record,
we may assume that the secondary index record
does not exist. However, this situation may
only occur during the rollback of incomplete
transactions. */
ut_a(thr_is_recv(thr));
} else {
err = row_undo_mod_del_mark_or_remove_sec(
node, thr, index, entry);
if (err != DB_SUCCESS) {
break;
}
}
mem_heap_empty(heap);
node->index = dict_table_get_next_index(node->index);
}
mem_heap_free(heap);
return(err);
}
/***********************************************************//**
Undoes a fresh insert of a row to a table. A fresh insert means that
the same clustered index unique key did not have any record, even delete
marked, at the time of the insert. InnoDB is eager in a rollback:
if it figures out that an index record will be removed in the purge
anyway, it will remove it in the rollback.
@return DB_SUCCESS or DB_OUT_OF_FILE_SPACE */
UNIV_INTERN
ulint
row_undo_ins(
/*=========*/
undo_node_t* node) /*!< in: row undo node */
{
ut_ad(node);
ut_ad(node->state == UNDO_NODE_INSERT);
row_undo_ins_parse_undo_rec(node);
if (!node->table || !row_undo_search_clust_to_pcur(node)) {
trx_undo_rec_release(node->trx, node->undo_no);
return(DB_SUCCESS);
}
/* Iterate over all the indexes and undo the insert.*/
/* Skip the clustered index (the first index) */
node->index = dict_table_get_next_index(
dict_table_get_first_index(node->table));
dict_table_skip_corrupt_index(node->index);
while (node->index != NULL) {
dtuple_t* entry;
ulint err;
entry = row_build_index_entry(node->row, node->ext,
node->index, node->heap);
if (UNIV_UNLIKELY(!entry)) {
/* The database must have crashed after
inserting a clustered index record but before
writing all the externally stored columns of
that record. Because secondary index entries
are inserted after the clustered index record,
we may assume that the secondary index record
does not exist. However, this situation may
only occur during the rollback of incomplete
transactions. */
ut_a(trx_is_recv(node->trx));
} else {
log_free_check();
err = row_undo_ins_remove_sec(node->index, entry);
if (err != DB_SUCCESS) {
return(err);
}
}
dict_table_next_uncorrupted_index(node->index);
}
log_free_check();
return(row_undo_ins_remove_clust_rec(node));
}
/***********************************************************//**
Undoes a fresh insert of a row to a table. A fresh insert means that
the same clustered index unique key did not have any record, even delete
marked, at the time of the insert. InnoDB is eager in a rollback:
if it figures out that an index record will be removed in the purge
anyway, it will remove it in the rollback.
@return DB_SUCCESS or DB_OUT_OF_FILE_SPACE */
UNIV_INTERN
ulint
row_undo_ins(
/*=========*/
undo_node_t* node) /*!< in: row undo node */
{
ut_ad(node);
ut_ad(node->state == UNDO_NODE_INSERT);
row_undo_ins_parse_undo_rec(node);
if (!node->table || !row_undo_search_clust_to_pcur(node)) {
trx_undo_rec_release(node->trx, node->undo_no);
return(DB_SUCCESS);
}
/* Iterate over all the indexes and undo the insert.*/
/* Skip the clustered index (the first index) */
node->index = dict_table_get_next_index(
dict_table_get_first_index(node->table));
while (node->index != NULL) {
dtuple_t* entry;
ulint err;
entry = row_build_index_entry(node->row, node->ext,
node->index, node->heap);
if (UNIV_UNLIKELY(!entry)) {
/* The database must have crashed after
inserting a clustered index record but before
writing all the externally stored columns of
that record, or a statement is being rolled
back because an error occurred while storing
off-page columns.
Because secondary index entries are inserted
after the clustered index record, we may
assume that the secondary index record does
not exist. */
} else {
log_free_check();
err = row_undo_ins_remove_sec(node->index, entry);
if (err != DB_SUCCESS) {
return(err);
}
}
node->index = dict_table_get_next_index(node->index);
}
log_free_check();
return(row_undo_ins_remove_clust_rec(node));
}
void
mlog_write_string(
/*==============*/
byte* ptr, /* in: pointer where to write */
const byte* str, /* in: string to write */
ulint len, /* in: string length */
mtr_t* mtr) /* in: mini-transaction handle */
{
byte* log_ptr;
if (UNIV_UNLIKELY(ptr < buf_pool->frame_zero)
|| UNIV_UNLIKELY(ptr >= buf_pool->high_end)) {
fprintf(stderr,
"InnoDB: Error: trying to write to"
" a stray memory location %p\n", (void*) ptr);
ut_error;
}
ut_ad(ptr && mtr);
ut_a(len < UNIV_PAGE_SIZE);
ut_memcpy(ptr, str, len);
log_ptr = mlog_open(mtr, 30);
/* If no logging is requested, we may return now */
if (log_ptr == NULL) {
return;
}
log_ptr = mlog_write_initial_log_record_fast(ptr, MLOG_WRITE_STRING,
log_ptr, mtr);
mach_write_to_2(log_ptr, ptr - buf_frame_align(ptr));
log_ptr += 2;
mach_write_to_2(log_ptr, len);
log_ptr += 2;
mlog_close(mtr, log_ptr);
mlog_catenate_string(mtr, str, len);
}
void
mlog_write_dulint(
/*==============*/
byte* ptr, /* in: pointer where to write */
dulint val, /* in: value to write */
mtr_t* mtr) /* in: mini-transaction handle */
{
byte* log_ptr;
if (UNIV_UNLIKELY(ptr < buf_pool->frame_zero)
|| UNIV_UNLIKELY(ptr >= buf_pool->high_end)) {
fprintf(stderr,
"InnoDB: Error: trying to write to"
" a stray memory location %p\n", (void*) ptr);
ut_error;
}
ut_ad(ptr && mtr);
mach_write_to_8(ptr, val);
log_ptr = mlog_open(mtr, 11 + 2 + 9);
/* If no logging is requested, we may return now */
if (log_ptr == NULL) {
return;
}
log_ptr = mlog_write_initial_log_record_fast(ptr, MLOG_8BYTES,
log_ptr, mtr);
mach_write_to_2(log_ptr, ptr - buf_frame_align(ptr));
log_ptr += 2;
log_ptr += mach_dulint_write_compressed(log_ptr, val);
mlog_close(mtr, log_ptr);
}
/**********************************************************************//**
Adds a column definition to a table. */
UNIV_INTERN
void
dict_mem_table_add_col(
/*===================*/
dict_table_t* table, /*!< in: table */
mem_heap_t* heap, /*!< in: temporary memory heap, or NULL */
const char* name, /*!< in: column name, or NULL */
ulint mtype, /*!< in: main datatype */
ulint prtype, /*!< in: precise type */
ulint len) /*!< in: precision */
{
dict_col_t* col;
ulint i;
ut_ad(table);
ut_ad(table->magic_n == DICT_TABLE_MAGIC_N);
ut_ad(!heap == !name);
i = table->n_def++;
if (name) {
if (UNIV_UNLIKELY(table->n_def == table->n_cols)) {
heap = table->heap;
}
if (UNIV_LIKELY(i) && UNIV_UNLIKELY(!table->col_names)) {
/* All preceding column names are empty. */
char* s = mem_heap_zalloc(heap, table->n_def);
table->col_names = s;
}
table->col_names = dict_add_col_name(table->col_names,
i, name, heap);
}
col = dict_table_get_nth_col(table, i);
dict_mem_fill_column_struct(col, i, mtype, prtype, len);
}
/**********************************************************************//**
Set the next and previous pointers in the undo page for the undo record
that was written to ptr. Update the first free value by the number of bytes
written for this undo record.
@return offset of the inserted entry on the page if succeeded, 0 if fail */
static
ulint
trx_undo_page_set_next_prev_and_add(
/*================================*/
page_t* undo_page, /*!< in/out: undo log page */
byte* ptr, /*!< in: ptr up to where data has been
written on this undo page. */
mtr_t* mtr) /*!< in: mtr */
{
ulint first_free; /*!< offset within undo_page */
ulint end_of_rec; /*!< offset within undo_page */
byte* ptr_to_first_free;
/* pointer within undo_page
that points to the next free
offset value within undo_page.*/
ut_ad(ptr > undo_page);
ut_ad(ptr < undo_page + UNIV_PAGE_SIZE);
if (UNIV_UNLIKELY(trx_undo_left(undo_page, ptr) < 2)) {
return(0);
}
ptr_to_first_free = undo_page + TRX_UNDO_PAGE_HDR + TRX_UNDO_PAGE_FREE;
first_free = mach_read_from_2(ptr_to_first_free);
/* Write offset of the previous undo log record */
mach_write_to_2(ptr, first_free);
ptr += 2;
end_of_rec = ptr - undo_page;
/* Write offset of the next undo log record */
mach_write_to_2(undo_page + first_free, end_of_rec);
/* Update the offset to first free undo record */
mach_write_to_2(ptr_to_first_free, end_of_rec);
/* Write this log entry to the UNDO log */
trx_undof_page_add_undo_rec_log(undo_page, first_free,
end_of_rec, mtr);
return(first_free);
}
/**********************************************************//**
Frees a mutex object. */
UNIV_INTERN
void
os_fast_mutex_free(
/*===============*/
os_fast_mutex_t* fast_mutex) /*!< in: mutex to free */
{
#ifdef __WIN__
ut_a(fast_mutex);
DeleteCriticalSection((LPCRITICAL_SECTION) fast_mutex);
#else
int ret;
ret = pthread_mutex_destroy(fast_mutex);
if (UNIV_UNLIKELY(ret != 0)) {
ut_print_timestamp(stderr);
fprintf(stderr,
" InnoDB: error: return value %lu when calling\n"
"InnoDB: pthread_mutex_destroy().\n", (ulint)ret);
fprintf(stderr,
"InnoDB: Byte contents of the pthread mutex at %p:\n",
(void*) fast_mutex);
ut_print_buf(stderr, fast_mutex, sizeof(os_fast_mutex_t));
putc('\n', stderr);
}
#endif
if (UNIV_LIKELY(os_sync_mutex_inited)) {
/* When freeing the last mutexes, we have
already freed os_sync_mutex */
os_mutex_enter(os_sync_mutex);
}
ut_ad(os_fast_mutex_count > 0);
os_fast_mutex_count--;
if (UNIV_LIKELY(os_sync_mutex_inited)) {
os_mutex_exit(os_sync_mutex);
}
}
/**************************************************************//**
The position of the cursor is stored by taking an initial segment of the
record the cursor is positioned on, before, or after, and copying it to the
cursor data structure, or just setting a flag if the cursor id before the
first in an EMPTY tree, or after the last in an EMPTY tree. NOTE that the
page where the cursor is positioned must not be empty if the index tree is
not totally empty! */
UNIV_INTERN
void
btr_pcur_store_position(
/*====================*/
btr_pcur_t* cursor, /*!< in: persistent cursor */
mtr_t* mtr) /*!< in: mtr */
{
page_cur_t* page_cursor;
buf_block_t* block;
rec_t* rec;
dict_index_t* index;
page_t* page;
ulint offs;
ut_a(cursor->pos_state == BTR_PCUR_IS_POSITIONED);
ut_ad(cursor->latch_mode != BTR_NO_LATCHES);
block = btr_pcur_get_block(cursor);
if (srv_pass_corrupt_table && !block) {
return;
}
ut_a(block);
index = btr_cur_get_index(btr_pcur_get_btr_cur(cursor));
page_cursor = btr_pcur_get_page_cur(cursor);
rec = page_cur_get_rec(page_cursor);
page = page_align(rec);
offs = page_offset(rec);
ut_ad(mtr_memo_contains(mtr, block, MTR_MEMO_PAGE_S_FIX)
|| mtr_memo_contains(mtr, block, MTR_MEMO_PAGE_X_FIX));
ut_a(cursor->latch_mode != BTR_NO_LATCHES);
if (UNIV_UNLIKELY(page_get_n_recs(page) == 0)) {
/* It must be an empty index tree; NOTE that in this case
we do not store the modify_clock, but always do a search
if we restore the cursor position */
ut_a(btr_page_get_next(page, mtr) == FIL_NULL);
ut_a(btr_page_get_prev(page, mtr) == FIL_NULL);
ut_ad(page_is_leaf(page));
ut_ad(page_get_page_no(page) == index->page);
cursor->old_stored = BTR_PCUR_OLD_STORED;
if (page_rec_is_supremum_low(offs)) {
cursor->rel_pos = BTR_PCUR_AFTER_LAST_IN_TREE;
} else {
cursor->rel_pos = BTR_PCUR_BEFORE_FIRST_IN_TREE;
}
return;
}
if (page_rec_is_supremum_low(offs)) {
rec = page_rec_get_prev(rec);
cursor->rel_pos = BTR_PCUR_AFTER;
} else if (page_rec_is_infimum_low(offs)) {
rec = page_rec_get_next(rec);
cursor->rel_pos = BTR_PCUR_BEFORE;
} else {
cursor->rel_pos = BTR_PCUR_ON;
}
cursor->old_stored = BTR_PCUR_OLD_STORED;
cursor->old_rec = dict_index_copy_rec_order_prefix(
index, rec, &cursor->old_n_fields,
&cursor->old_rec_buf, &cursor->buf_size);
cursor->block_when_stored = block;
cursor->modify_clock = buf_block_get_modify_clock(block);
}
/*******************************************************************//**
Fills the "lock_data" member of i_s_locks_row_t object.
If memory can not be allocated then FALSE is returned.
@return FALSE if allocation fails */
static
ibool
fill_lock_data(
/*===========*/
const char** lock_data,/*!< out: "lock_data" to fill */
const lock_t* lock, /*!< in: lock used to find the data */
ulint heap_no,/*!< in: rec num used to find the data */
trx_i_s_cache_t* cache) /*!< in/out: cache where to store
volatile data */
{
mtr_t mtr;
const buf_block_t* block;
const page_t* page;
const rec_t* rec;
ut_a(lock_get_type(lock) == LOCK_REC);
mtr_start(&mtr);
block = buf_page_try_get(lock_rec_get_space_id(lock),
lock_rec_get_page_no(lock),
&mtr);
if (block == NULL) {
*lock_data = NULL;
mtr_commit(&mtr);
return(TRUE);
}
page = (const page_t*) buf_block_get_frame(block);
rec = page_find_rec_with_heap_no(page, heap_no);
if (page_rec_is_infimum(rec)) {
*lock_data = ha_storage_put_str_memlim(
cache->storage, "infimum pseudo-record",
MAX_ALLOWED_FOR_STORAGE(cache));
} else if (page_rec_is_supremum(rec)) {
*lock_data = ha_storage_put_str_memlim(
cache->storage, "supremum pseudo-record",
MAX_ALLOWED_FOR_STORAGE(cache));
} else {
const dict_index_t* index;
ulint n_fields;
mem_heap_t* heap;
ulint offsets_onstack[REC_OFFS_NORMAL_SIZE];
ulint* offsets;
char buf[TRX_I_S_LOCK_DATA_MAX_LEN];
ulint buf_used;
ulint i;
rec_offs_init(offsets_onstack);
offsets = offsets_onstack;
index = lock_rec_get_index(lock);
n_fields = dict_index_get_n_unique(index);
ut_a(n_fields > 0);
heap = NULL;
offsets = rec_get_offsets(rec, index, offsets, n_fields,
&heap);
/* format and store the data */
buf_used = 0;
for (i = 0; i < n_fields; i++) {
buf_used += put_nth_field(
buf + buf_used, sizeof(buf) - buf_used,
i, index, rec, offsets) - 1;
}
*lock_data = (const char*) ha_storage_put_memlim(
cache->storage, buf, buf_used + 1,
MAX_ALLOWED_FOR_STORAGE(cache));
if (UNIV_UNLIKELY(heap != NULL)) {
/* this means that rec_get_offsets() has created a new
heap and has stored offsets in it; check that this is
really the case and free the heap */
ut_a(offsets != offsets_onstack);
mem_heap_free(heap);
}
}
mtr_commit(&mtr);
if (*lock_data == NULL) {
return(FALSE);
}
return(TRUE);
}
/********************************************************************//**
Frees memory to a pool. */
UNIV_INTERN
void
mem_area_free(
/*==========*/
void* ptr, /*!< in, own: pointer to allocated memory
buffer */
mem_pool_t* pool) /*!< in: memory pool */
{
mem_area_t* area;
mem_area_t* buddy;
void* new_ptr;
ulint size;
ulint n;
if (UNIV_LIKELY(srv_use_sys_malloc)) {
free(ptr);
return;
}
/* It may be that the area was really allocated from the OS with
regular malloc: check if ptr points within our memory pool */
if ((byte*)ptr < pool->buf || (byte*)ptr >= pool->buf + pool->size) {
ut_free(ptr);
return;
}
area = (mem_area_t*) (((byte*)ptr) - MEM_AREA_EXTRA_SIZE);
if (mem_area_get_free(area)) {
fprintf(stderr,
"InnoDB: Error: Freeing element to mem pool"
" free list though the\n"
"InnoDB: element is marked free!\n");
mem_analyze_corruption(area);
ut_error;
}
size = mem_area_get_size(area);
UNIV_MEM_FREE(ptr, size - MEM_AREA_EXTRA_SIZE);
if (size == 0) {
fprintf(stderr,
"InnoDB: Error: Mem area size is 0. Possibly a"
" memory overrun of the\n"
"InnoDB: previous allocated area!\n");
mem_analyze_corruption(area);
ut_error;
}
#ifdef UNIV_LIGHT_MEM_DEBUG
if (((byte*)area) + size < pool->buf + pool->size) {
ulint next_size;
next_size = mem_area_get_size(
(mem_area_t*)(((byte*)area) + size));
if (UNIV_UNLIKELY(!next_size || !ut_is_2pow(next_size))) {
fprintf(stderr,
"InnoDB: Error: Memory area size %lu,"
" next area size %lu not a power of 2!\n"
"InnoDB: Possibly a memory overrun of"
" the buffer being freed here.\n",
(ulong) size, (ulong) next_size);
mem_analyze_corruption(area);
ut_error;
}
}
#endif
buddy = mem_area_get_buddy(area, size, pool);
n = ut_2_log(size);
mem_pool_mutex_enter(pool);
mem_n_threads_inside++;
ut_a(mem_n_threads_inside == 1);
if (buddy && mem_area_get_free(buddy)
&& (size == mem_area_get_size(buddy))) {
/* The buddy is in a free list */
if ((byte*)buddy < (byte*)area) {
new_ptr = ((byte*)buddy) + MEM_AREA_EXTRA_SIZE;
mem_area_set_size(buddy, 2 * size);
mem_area_set_free(buddy, FALSE);
} else {
new_ptr = ptr;
mem_area_set_size(area, 2 * size);
}
/* Remove the buddy from its free list and merge it to area */
UT_LIST_REMOVE(free_list, pool->free_list[n], buddy);
pool->reserved += ut_2_exp(n);
mem_n_threads_inside--;
mem_pool_mutex_exit(pool);
mem_area_free(new_ptr, pool);
return;
} else {
UT_LIST_ADD_FIRST(free_list, pool->free_list[n], area);
mem_area_set_free(area, TRUE);
ut_ad(pool->reserved >= size);
pool->reserved -= size;
}
mem_n_threads_inside--;
mem_pool_mutex_exit(pool);
ut_ad(mem_pool_validate(pool));
}
/**************************************************************//**
Restores the stored position of a persistent cursor bufferfixing the page and
obtaining the specified latches. If the cursor position was saved when the
(1) cursor was positioned on a user record: this function restores the position
to the last record LESS OR EQUAL to the stored record;
(2) cursor was positioned on a page infimum record: restores the position to
the last record LESS than the user record which was the successor of the page
infimum;
(3) cursor was positioned on the page supremum: restores to the first record
GREATER than the user record which was the predecessor of the supremum.
(4) cursor was positioned before the first or after the last in an empty tree:
restores to before first or after the last in the tree.
@return TRUE if the cursor position was stored when it was on a user
record and it can be restored on a user record whose ordering fields
are identical to the ones of the original user record */
UNIV_INTERN
ibool
btr_pcur_restore_position_func(
/*===========================*/
ulint latch_mode, /*!< in: BTR_SEARCH_LEAF, ... */
btr_pcur_t* cursor, /*!< in: detached persistent cursor */
const char* file, /*!< in: file name */
ulint line, /*!< in: line where called */
mtr_t* mtr) /*!< in: mtr */
{
dict_index_t* index;
dtuple_t* tuple;
ulint mode;
ulint old_mode;
mem_heap_t* heap;
ut_ad(mtr);
ut_ad(mtr->state == MTR_ACTIVE);
index = btr_cur_get_index(btr_pcur_get_btr_cur(cursor));
if (UNIV_UNLIKELY(cursor->old_stored != BTR_PCUR_OLD_STORED)
|| UNIV_UNLIKELY(cursor->pos_state != BTR_PCUR_WAS_POSITIONED
&& cursor->pos_state != BTR_PCUR_IS_POSITIONED)) {
ut_print_buf(stderr, cursor, sizeof(btr_pcur_t));
putc('\n', stderr);
if (cursor->trx_if_known) {
trx_print(stderr, cursor->trx_if_known, 0);
}
ut_error;
}
if (UNIV_UNLIKELY
(cursor->rel_pos == BTR_PCUR_AFTER_LAST_IN_TREE
|| cursor->rel_pos == BTR_PCUR_BEFORE_FIRST_IN_TREE)) {
/* In these cases we do not try an optimistic restoration,
but always do a search */
btr_cur_open_at_index_side(
cursor->rel_pos == BTR_PCUR_BEFORE_FIRST_IN_TREE,
index, latch_mode, btr_pcur_get_btr_cur(cursor), mtr);
cursor->latch_mode = latch_mode;
cursor->pos_state = BTR_PCUR_IS_POSITIONED;
cursor->block_when_stored = btr_pcur_get_block(cursor);
return(FALSE);
}
ut_a(cursor->old_rec);
ut_a(cursor->old_n_fields);
if (UNIV_LIKELY(latch_mode == BTR_SEARCH_LEAF)
|| UNIV_LIKELY(latch_mode == BTR_MODIFY_LEAF)) {
/* Try optimistic restoration */
if (UNIV_LIKELY(buf_page_optimistic_get(
latch_mode,
cursor->block_when_stored,
cursor->modify_clock,
file, line, mtr))) {
cursor->pos_state = BTR_PCUR_IS_POSITIONED;
buf_block_dbg_add_level(
btr_pcur_get_block(cursor),
dict_index_is_ibuf(index)
? SYNC_IBUF_TREE_NODE : SYNC_TREE_NODE);
if (cursor->rel_pos == BTR_PCUR_ON) {
#ifdef UNIV_DEBUG
const rec_t* rec;
const ulint* offsets1;
const ulint* offsets2;
#endif /* UNIV_DEBUG */
cursor->latch_mode = latch_mode;
#ifdef UNIV_DEBUG
rec = btr_pcur_get_rec(cursor);
heap = mem_heap_create(256);
offsets1 = rec_get_offsets(
cursor->old_rec, index, NULL,
cursor->old_n_fields, &heap);
offsets2 = rec_get_offsets(
rec, index, NULL,
cursor->old_n_fields, &heap);
ut_ad(!cmp_rec_rec(cursor->old_rec,
rec, offsets1, offsets2,
index));
mem_heap_free(heap);
#endif /* UNIV_DEBUG */
return(TRUE);
}
return(FALSE);
}
}
/* If optimistic restoration did not succeed, open the cursor anew */
heap = mem_heap_create(256);
tuple = dict_index_build_data_tuple(index, cursor->old_rec,
cursor->old_n_fields, heap);
/* Save the old search mode of the cursor */
old_mode = cursor->search_mode;
switch (cursor->rel_pos) {
case BTR_PCUR_ON:
mode = PAGE_CUR_LE;
break;
case BTR_PCUR_AFTER:
mode = PAGE_CUR_G;
break;
case BTR_PCUR_BEFORE:
mode = PAGE_CUR_L;
break;
default:
ut_error;
mode = 0;
}
btr_pcur_open_with_no_init_func(index, tuple, mode, latch_mode,
cursor, 0, file, line, mtr);
/* Restore the old search mode */
cursor->search_mode = old_mode;
switch (cursor->rel_pos) {
case BTR_PCUR_ON:
if (btr_pcur_is_on_user_rec(cursor)
&& !cmp_dtuple_rec(
tuple, btr_pcur_get_rec(cursor),
rec_get_offsets(btr_pcur_get_rec(cursor),
index, NULL,
ULINT_UNDEFINED, &heap))) {
/* We have to store the NEW value for
the modify clock, since the cursor can
now be on a different page! But we can
retain the value of old_rec */
cursor->block_when_stored =
btr_pcur_get_block(cursor);
cursor->modify_clock =
buf_block_get_modify_clock(
cursor->block_when_stored);
cursor->old_stored = BTR_PCUR_OLD_STORED;
mem_heap_free(heap);
return(TRUE);
}
#ifdef UNIV_DEBUG
/* fall through */
case BTR_PCUR_BEFORE:
case BTR_PCUR_AFTER:
break;
default:
ut_error;
#endif /* UNIV_DEBUG */
}
mem_heap_free(heap);
/* We have to store new position information, modify_clock etc.,
to the cursor because it can now be on a different page, the record
under it may have been removed, etc. */
btr_pcur_store_position(cursor, mtr);
return(FALSE);
}
ibool
btr_pcur_restore_position(
/*======================*/
/* out: TRUE if the cursor position
was stored when it was on a user record
and it can be restored on a user record
whose ordering fields are identical to
the ones of the original user record */
ulint latch_mode, /* in: BTR_SEARCH_LEAF, ... */
btr_pcur_t* cursor, /* in: detached persistent cursor */
mtr_t* mtr) /* in: mtr */
{
dict_index_t* index;
page_t* page;
dtuple_t* tuple;
ulint mode;
ulint old_mode;
mem_heap_t* heap;
index = btr_cur_get_index(btr_pcur_get_btr_cur(cursor));
if (UNIV_UNLIKELY(cursor->old_stored != BTR_PCUR_OLD_STORED)
|| UNIV_UNLIKELY(cursor->pos_state != BTR_PCUR_WAS_POSITIONED
&& cursor->pos_state != BTR_PCUR_IS_POSITIONED)) {
ut_print_buf(stderr, cursor, sizeof(btr_pcur_t));
if (cursor->trx_if_known) {
trx_print(stderr, cursor->trx_if_known, 0);
}
ut_error;
}
if (UNIV_UNLIKELY(
cursor->rel_pos == BTR_PCUR_AFTER_LAST_IN_TREE
|| cursor->rel_pos == BTR_PCUR_BEFORE_FIRST_IN_TREE)) {
/* In these cases we do not try an optimistic restoration,
but always do a search */
btr_cur_open_at_index_side(
cursor->rel_pos == BTR_PCUR_BEFORE_FIRST_IN_TREE,
index, latch_mode, btr_pcur_get_btr_cur(cursor), mtr);
cursor->block_when_stored
= buf_block_align(btr_pcur_get_page(cursor));
return(FALSE);
}
ut_a(cursor->old_rec);
ut_a(cursor->old_n_fields);
page = btr_cur_get_page(btr_pcur_get_btr_cur(cursor));
if (UNIV_LIKELY(latch_mode == BTR_SEARCH_LEAF)
|| UNIV_LIKELY(latch_mode == BTR_MODIFY_LEAF)) {
/* Try optimistic restoration */
if (UNIV_LIKELY(buf_page_optimistic_get(
latch_mode,
cursor->block_when_stored, page,
cursor->modify_clock, mtr))) {
cursor->pos_state = BTR_PCUR_IS_POSITIONED;
#ifdef UNIV_SYNC_DEBUG
buf_page_dbg_add_level(page, SYNC_TREE_NODE);
#endif /* UNIV_SYNC_DEBUG */
if (cursor->rel_pos == BTR_PCUR_ON) {
#ifdef UNIV_DEBUG
rec_t* rec;
ulint* offsets1;
ulint* offsets2;
#endif /* UNIV_DEBUG */
cursor->latch_mode = latch_mode;
#ifdef UNIV_DEBUG
rec = btr_pcur_get_rec(cursor);
heap = mem_heap_create(256);
offsets1 = rec_get_offsets(
cursor->old_rec, index, NULL,
cursor->old_n_fields, &heap);
offsets2 = rec_get_offsets(
rec, index, NULL,
cursor->old_n_fields, &heap);
ut_ad(!cmp_rec_rec(cursor->old_rec,
rec, offsets1, offsets2,
index));
mem_heap_free(heap);
#endif /* UNIV_DEBUG */
return(TRUE);
}
return(FALSE);
}
}
/* If optimistic restoration did not succeed, open the cursor anew */
heap = mem_heap_create(256);
tuple = dict_index_build_data_tuple(index, cursor->old_rec,
cursor->old_n_fields, heap);
/* Save the old search mode of the cursor */
old_mode = cursor->search_mode;
switch (cursor->rel_pos) {
case BTR_PCUR_ON:
mode = PAGE_CUR_LE;
break;
case BTR_PCUR_AFTER:
mode = PAGE_CUR_G;
break;
case BTR_PCUR_BEFORE:
mode = PAGE_CUR_L;
break;
default:
ut_error;
mode = 0; /* silence a warning */
}
btr_pcur_open_with_no_init(index, tuple, mode, latch_mode,
cursor, 0, mtr);
/* Restore the old search mode */
cursor->search_mode = old_mode;
if (btr_pcur_is_on_user_rec(cursor, mtr)) {
switch (cursor->rel_pos) {
case BTR_PCUR_ON:
if (!cmp_dtuple_rec(
tuple, btr_pcur_get_rec(cursor),
rec_get_offsets(btr_pcur_get_rec(cursor),
index, NULL,
ULINT_UNDEFINED, &heap))) {
/* We have to store the NEW value for
the modify clock, since the cursor can
now be on a different page! But we can
retain the value of old_rec */
cursor->block_when_stored =
buf_block_align(
btr_pcur_get_page(cursor));
cursor->modify_clock =
buf_block_get_modify_clock(
cursor->block_when_stored);
cursor->old_stored = BTR_PCUR_OLD_STORED;
mem_heap_free(heap);
return(TRUE);
}
break;
case BTR_PCUR_BEFORE:
page_cur_move_to_next(btr_pcur_get_page_cur(cursor));
break;
case BTR_PCUR_AFTER:
page_cur_move_to_prev(btr_pcur_get_page_cur(cursor));
break;
#ifdef UNIV_DEBUG
default:
ut_error;
#endif /* UNIV_DEBUG */
}
}
mem_heap_free(heap);
/* We have to store new position information, modify_clock etc.,
to the cursor because it can now be on a different page, the record
under it may have been removed, etc. */
btr_pcur_store_position(cursor, mtr);
return(FALSE);
}
/********************************************************************//**
Fills the specified free list.
@return TRUE if we were able to insert a block to the free list */
static
ibool
mem_pool_fill_free_list(
/*====================*/
ulint i, /*!< in: free list index */
mem_pool_t* pool) /*!< in: memory pool */
{
mem_area_t* area;
mem_area_t* area2;
ibool ret;
ut_ad(mutex_own(&(pool->mutex)));
if (UNIV_UNLIKELY(i >= 63)) {
/* We come here when we have run out of space in the
memory pool: */
return(FALSE);
}
area = UT_LIST_GET_FIRST(pool->free_list[i + 1]);
if (area == NULL) {
if (UT_LIST_GET_LEN(pool->free_list[i + 1]) > 0) {
ut_print_timestamp(stderr);
fprintf(stderr,
" InnoDB: Error: mem pool free list %lu"
" length is %lu\n"
"InnoDB: though the list is empty!\n",
(ulong) i + 1,
(ulong)
UT_LIST_GET_LEN(pool->free_list[i + 1]));
}
ret = mem_pool_fill_free_list(i + 1, pool);
if (ret == FALSE) {
return(FALSE);
}
area = UT_LIST_GET_FIRST(pool->free_list[i + 1]);
}
if (UNIV_UNLIKELY(UT_LIST_GET_LEN(pool->free_list[i + 1]) == 0)) {
mem_analyze_corruption(area);
ut_error;
}
UT_LIST_REMOVE(free_list, pool->free_list[i + 1], area);
area2 = (mem_area_t*)(((byte*)area) + ut_2_exp(i));
UNIV_MEM_ALLOC(area2, MEM_AREA_EXTRA_SIZE);
mem_area_set_size(area2, ut_2_exp(i));
mem_area_set_free(area2, TRUE);
UT_LIST_ADD_FIRST(free_list, pool->free_list[i], area2);
mem_area_set_size(area, ut_2_exp(i));
UT_LIST_ADD_FIRST(free_list, pool->free_list[i], area);
return(TRUE);
}
/*****************************************************************//**
Constructs the last committed version of a clustered index record,
which should be seen by a semi-consistent read.
@return DB_SUCCESS or DB_MISSING_HISTORY */
UNIV_INTERN
ulint
row_vers_build_for_semi_consistent_read(
/*====================================*/
const rec_t* rec, /*!< in: record in a clustered index; the
caller must have a latch on the page; this
latch locks the top of the stack of versions
of this records */
mtr_t* mtr, /*!< in: mtr holding the latch on rec */
dict_index_t* index, /*!< in: the clustered index */
ulint** offsets,/*!< in/out: offsets returned by
rec_get_offsets(rec, index) */
mem_heap_t** offset_heap,/*!< in/out: memory heap from which
the offsets are allocated */
mem_heap_t* in_heap,/*!< in: memory heap from which the memory for
*old_vers is allocated; memory for possible
intermediate versions is allocated and freed
locally within the function */
const rec_t** old_vers)/*!< out: rec, old version, or NULL if the
record does not exist in the view, that is,
it was freshly inserted afterwards */
{
const rec_t* version;
mem_heap_t* heap = NULL;
byte* buf;
ulint err;
trx_id_t rec_trx_id = ut_dulint_zero;
ut_ad(dict_index_is_clust(index));
ut_ad(mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_X_FIX)
|| mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_S_FIX));
#ifdef UNIV_SYNC_DEBUG
ut_ad(!rw_lock_own(&(purge_sys->latch), RW_LOCK_SHARED));
#endif /* UNIV_SYNC_DEBUG */
ut_ad(rec_offs_validate(rec, index, *offsets));
rw_lock_s_lock(&(purge_sys->latch));
/* The S-latch on purge_sys prevents the purge view from
changing. Thus, if we have an uncommitted transaction at
this point, then purge cannot remove its undo log even if
the transaction could commit now. */
version = rec;
for (;;) {
trx_t* version_trx;
mem_heap_t* heap2;
rec_t* prev_version;
trx_id_t version_trx_id;
version_trx_id = row_get_rec_trx_id(version, index, *offsets);
if (rec == version) {
rec_trx_id = version_trx_id;
}
mutex_enter(&kernel_mutex);
version_trx = trx_get_on_id(version_trx_id);
if (version_trx
&& (version_trx->conc_state == TRX_COMMITTED_IN_MEMORY
|| version_trx->conc_state == TRX_NOT_STARTED)) {
version_trx = NULL;
}
mutex_exit(&kernel_mutex);
if (!version_trx) {
/* We found a version that belongs to a
committed transaction: return it. */
#if defined UNIV_DEBUG || defined UNIV_BLOB_LIGHT_DEBUG
ut_a(!rec_offs_any_null_extern(version, *offsets));
#endif /* UNIV_DEBUG || UNIV_BLOB_LIGHT_DEBUG */
if (rec == version) {
*old_vers = rec;
err = DB_SUCCESS;
break;
}
/* We assume that a rolled-back transaction stays in
TRX_ACTIVE state until all the changes have been
rolled back and the transaction is removed from
the global list of transactions. */
if (!ut_dulint_cmp(rec_trx_id, version_trx_id)) {
/* The transaction was committed while
we searched for earlier versions.
Return the current version as a
semi-consistent read. */
version = rec;
*offsets = rec_get_offsets(version,
index, *offsets,
ULINT_UNDEFINED,
offset_heap);
}
buf = mem_heap_alloc(in_heap, rec_offs_size(*offsets));
*old_vers = rec_copy(buf, version, *offsets);
rec_offs_make_valid(*old_vers, index, *offsets);
err = DB_SUCCESS;
break;
}
heap2 = heap;
heap = mem_heap_create(1024);
err = trx_undo_prev_version_build(rec, mtr, version, index,
*offsets, heap,
&prev_version);
if (heap2) {
mem_heap_free(heap2); /* free version */
}
if (UNIV_UNLIKELY(err != DB_SUCCESS)) {
break;
}
if (prev_version == NULL) {
/* It was a freshly inserted version */
*old_vers = NULL;
err = DB_SUCCESS;
break;
}
version = prev_version;
*offsets = rec_get_offsets(version, index, *offsets,
ULINT_UNDEFINED, offset_heap);
#if defined UNIV_DEBUG || defined UNIV_BLOB_LIGHT_DEBUG
ut_a(!rec_offs_any_null_extern(version, *offsets));
#endif /* UNIV_DEBUG || UNIV_BLOB_LIGHT_DEBUG */
}/* for (;;) */
if (heap) {
mem_heap_free(heap);
}
rw_lock_s_unlock(&(purge_sys->latch));
return(err);
}
/********************************************************************//**
Issues read requests for pages which recovery wants to read in. */
UNIV_INTERN
void
buf_read_recv_pages(
/*================*/
ibool sync, /*!< in: TRUE if the caller
wants this function to wait
for the highest address page
to get read in, before this
function returns */
ulint space, /*!< in: space id */
ulint zip_size, /*!< in: compressed page size in
bytes, or 0 */
const ulint* page_nos, /*!< in: array of page numbers
to read, with the highest page
number the last in the
array */
ulint n_stored) /*!< in: number of page numbers
in the array */
{
ib_int64_t tablespace_version;
ulint count;
ulint err;
ulint i;
zip_size = fil_space_get_zip_size(space);
if (UNIV_UNLIKELY(zip_size == ULINT_UNDEFINED)) {
/* It is a single table tablespace and the .ibd file is
missing: do nothing */
/* the log records should be treated here same reason
for http://bugs.mysql.com/bug.php?id=43948 */
if (recv_recovery_is_on()) {
recv_addr_t* recv_addr;
mutex_enter(&(recv_sys->mutex));
if (recv_sys->apply_log_recs == FALSE) {
mutex_exit(&(recv_sys->mutex));
goto not_to_recover;
}
for (i = 0; i < n_stored; i++) {
/* recv_get_fil_addr_struct() */
recv_addr = HASH_GET_FIRST(recv_sys->addr_hash,
hash_calc_hash(ut_fold_ulint_pair(space, page_nos[i]),
recv_sys->addr_hash));
while (recv_addr) {
if ((recv_addr->space == space)
&& (recv_addr->page_no == page_nos[i])) {
break;
}
recv_addr = HASH_GET_NEXT(addr_hash, recv_addr);
}
if ((recv_addr == NULL)
|| (recv_addr->state == RECV_BEING_PROCESSED)
|| (recv_addr->state == RECV_PROCESSED)) {
continue;
}
recv_addr->state = RECV_PROCESSED;
ut_a(recv_sys->n_addrs);
recv_sys->n_addrs--;
}
mutex_exit(&(recv_sys->mutex));
fprintf(stderr, " (cannot find space: %lu)", space);
}
not_to_recover:
return;
}
tablespace_version = fil_space_get_version(space);
for (i = 0; i < n_stored; i++) {
buf_pool_t* buf_pool;
count = 0;
os_aio_print_debug = FALSE;
buf_pool = buf_pool_get(space, page_nos[i]);
while (buf_pool->n_pend_reads >= recv_n_pool_free_frames / 2) {
os_aio_simulated_wake_handler_threads();
os_thread_sleep(10000);
count++;
if (count > 1000) {
fprintf(stderr,
"InnoDB: Error: InnoDB has waited for"
" 10 seconds for pending\n"
"InnoDB: reads to the buffer pool to"
" be finished.\n"
"InnoDB: Number of pending reads %lu,"
" pending pread calls %lu\n",
(ulong) buf_pool->n_pend_reads,
(ulong)os_file_n_pending_preads);
os_aio_print_debug = TRUE;
}
}
os_aio_print_debug = FALSE;
if ((i + 1 == n_stored) && sync) {
buf_read_page_low(&err, TRUE, BUF_READ_ANY_PAGE, space,
zip_size, TRUE, tablespace_version,
page_nos[i], NULL);
} else {
buf_read_page_low(&err, FALSE, BUF_READ_ANY_PAGE
| OS_AIO_SIMULATED_WAKE_LATER,
space, zip_size, TRUE,
tablespace_version, page_nos[i], NULL);
}
}
os_aio_simulated_wake_handler_threads();
/* Flush pages from the end of all the LRU lists if necessary */
buf_flush_free_margins(FALSE);
#ifdef UNIV_DEBUG
if (buf_debug_prints) {
fprintf(stderr,
"Recovery applies read-ahead pages %lu\n",
(ulong) n_stored);
}
#endif /* UNIV_DEBUG */
}
/********************************************************//**
Parses a log record written by mlog_write_ulint or mlog_write_dulint.
@return parsed record end, NULL if not a complete record or a corrupt record */
UNIV_INTERN
byte*
mlog_parse_nbytes(
/*==============*/
ulint type, /*!< in: log record type: MLOG_1BYTE, ... */
byte* ptr, /*!< in: buffer */
byte* end_ptr,/*!< in: buffer end */
byte* page, /*!< in: page where to apply the log record, or NULL */
void* page_zip)/*!< in/out: compressed page, or NULL */
{
ulint offset;
ulint val;
dulint dval;
ut_a(type <= MLOG_8BYTES);
ut_a(!page || !page_zip || fil_page_get_type(page) != FIL_PAGE_INDEX);
if (end_ptr < ptr + 2) {
return(NULL);
}
offset = mach_read_from_2(ptr);
ptr += 2;
if (offset >= UNIV_PAGE_SIZE) {
recv_sys->found_corrupt_log = TRUE;
return(NULL);
}
if (type == MLOG_8BYTES) {
ptr = mach_dulint_parse_compressed(ptr, end_ptr, &dval);
if (ptr == NULL) {
return(NULL);
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_8
(((page_zip_des_t*) page_zip)->data
+ offset, dval);
}
mach_write_to_8(page + offset, dval);
}
return(ptr);
}
ptr = mach_parse_compressed(ptr, end_ptr, &val);
if (ptr == NULL) {
return(NULL);
}
switch (type) {
case MLOG_1BYTE:
if (UNIV_UNLIKELY(val > 0xFFUL)) {
goto corrupt;
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_1
(((page_zip_des_t*) page_zip)->data
+ offset, val);
}
mach_write_to_1(page + offset, val);
}
break;
case MLOG_2BYTES:
if (UNIV_UNLIKELY(val > 0xFFFFUL)) {
goto corrupt;
}
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_2
(((page_zip_des_t*) page_zip)->data
+ offset, val);
}
mach_write_to_2(page + offset, val);
}
break;
case MLOG_4BYTES:
if (page) {
if (UNIV_LIKELY_NULL(page_zip)) {
mach_write_to_4
(((page_zip_des_t*) page_zip)->data
+ offset, val);
}
mach_write_to_4(page + offset, val);
}
break;
default:
corrupt:
recv_sys->found_corrupt_log = TRUE;
ptr = NULL;
}
return(ptr);
}
/****************************************************************//**
Allocates large pages memory.
@return allocated memory */
UNIV_INTERN
void*
os_mem_alloc_large(
/*===============*/
ulint* n) /*!< in/out: number of bytes */
{
void* ptr;
ulint size;
#if defined HAVE_LARGE_PAGES && defined UNIV_LINUX
int shmid;
struct shmid_ds buf;
if (!os_use_large_pages || !os_large_page_size) {
goto skip;
}
/* Align block size to os_large_page_size */
ut_ad(ut_is_2pow(os_large_page_size));
size = ut_2pow_round(*n + (os_large_page_size - 1),
os_large_page_size);
shmid = shmget(IPC_PRIVATE, (size_t)size, SHM_HUGETLB | SHM_R | SHM_W);
if (shmid < 0) {
fprintf(stderr, "InnoDB: HugeTLB: Warning: Failed to allocate"
" %lu bytes. errno %d\n", size, errno);
ptr = NULL;
} else {
ptr = shmat(shmid, NULL, 0);
if (ptr == (void *)-1) {
fprintf(stderr, "InnoDB: HugeTLB: Warning: Failed to"
" attach shared memory segment, errno %d\n",
errno);
ptr = NULL;
}
/* Remove the shared memory segment so that it will be
automatically freed after memory is detached or
process exits */
shmctl(shmid, IPC_RMID, &buf);
}
if (ptr) {
*n = size;
os_fast_mutex_lock(&ut_list_mutex);
ut_total_allocated_memory += size;
os_fast_mutex_unlock(&ut_list_mutex);
UNIV_MEM_ALLOC(ptr, size);
return(ptr);
}
fprintf(stderr, "InnoDB HugeTLB: Warning: Using conventional"
" memory pool\n");
skip:
#endif /* HAVE_LARGE_PAGES && UNIV_LINUX */
#ifdef __WIN__
SYSTEM_INFO system_info;
GetSystemInfo(&system_info);
/* Align block size to system page size */
ut_ad(ut_is_2pow(system_info.dwPageSize));
/* system_info.dwPageSize is only 32-bit. Casting to ulint is required
on 64-bit Windows. */
size = *n = ut_2pow_round(*n + (system_info.dwPageSize - 1),
(ulint) system_info.dwPageSize);
ptr = VirtualAlloc(NULL, size, MEM_COMMIT | MEM_RESERVE,
PAGE_READWRITE);
if (!ptr) {
fprintf(stderr, "InnoDB: VirtualAlloc(%lu bytes) failed;"
" Windows error %lu\n",
(ulong) size, (ulong) GetLastError());
} else {
os_fast_mutex_lock(&ut_list_mutex);
ut_total_allocated_memory += size;
os_fast_mutex_unlock(&ut_list_mutex);
UNIV_MEM_ALLOC(ptr, size);
}
#elif !defined OS_MAP_ANON
size = *n;
ptr = ut_malloc_low(size, TRUE, FALSE);
#else
# ifdef HAVE_GETPAGESIZE
size = getpagesize();
# else
size = UNIV_PAGE_SIZE;
# endif
/* Align block size to system page size */
ut_ad(ut_is_2pow(size));
size = *n = ut_2pow_round(*n + (size - 1), size);
ptr = mmap(NULL, size, PROT_READ | PROT_WRITE,
MAP_PRIVATE | OS_MAP_ANON, -1, 0);
if (UNIV_UNLIKELY(ptr == (void*) -1)) {
fprintf(stderr, "InnoDB: mmap(%lu bytes) failed;"
" errno %lu\n",
(ulong) size, (ulong) errno);
ptr = NULL;
} else {
os_fast_mutex_lock(&ut_list_mutex);
ut_total_allocated_memory += size;
os_fast_mutex_unlock(&ut_list_mutex);
UNIV_MEM_ALLOC(ptr, size);
}
#endif
return(ptr);
}
/*****************************************************************//**
When an insert or purge to a table is performed, this function builds
the entry to be inserted into or purged from an index on the table.
@return index entry which should be inserted or purged, or NULL if the
externally stored columns in the clustered index record are
unavailable and ext != NULL */
UNIV_INTERN
dtuple_t*
row_build_index_entry(
/*==================*/
const dtuple_t* row, /*!< in: row which should be
inserted or purged */
row_ext_t* ext, /*!< in: externally stored column prefixes,
or NULL */
dict_index_t* index, /*!< in: index on the table */
mem_heap_t* heap) /*!< in: memory heap from which the memory for
the index entry is allocated */
{
dtuple_t* entry;
ulint entry_len;
ulint i;
ut_ad(row && index && heap);
ut_ad(dtuple_check_typed(row));
entry_len = dict_index_get_n_fields(index);
entry = dtuple_create(heap, entry_len);
if (UNIV_UNLIKELY(index->type & DICT_UNIVERSAL)) {
dtuple_set_n_fields_cmp(entry, entry_len);
/* There may only be externally stored columns
in a clustered index B-tree of a user table. */
ut_a(!ext);
} else {
dtuple_set_n_fields_cmp(
entry, dict_index_get_n_unique_in_tree(index));
}
for (i = 0; i < entry_len; i++) {
const dict_field_t* ind_field
= dict_index_get_nth_field(index, i);
const dict_col_t* col
= ind_field->col;
ulint col_no
= dict_col_get_no(col);
dfield_t* dfield
= dtuple_get_nth_field(entry, i);
const dfield_t* dfield2
= dtuple_get_nth_field(row, col_no);
ulint len
= dfield_get_len(dfield2);
dfield_copy(dfield, dfield2);
if (dfield_is_null(dfield)) {
continue;
}
if (ind_field->prefix_len == 0
&& (!dfield_is_ext(dfield)
|| dict_index_is_clust(index))) {
/* The dfield_copy() above suffices for
columns that are stored in-page, or for
clustered index record columns that are not
part of a column prefix in the PRIMARY KEY. */
continue;
}
/* If the column is stored externally (off-page) in
the clustered index, it must be an ordering field in
the secondary index. In the Antelope format, only
prefix-indexed columns may be stored off-page in the
clustered index record. In the Barracuda format, also
fully indexed long CHAR or VARCHAR columns may be
stored off-page. */
ut_ad(col->ord_part);
if (UNIV_LIKELY_NULL(ext)) {
/* See if the column is stored externally. */
const byte* buf = row_ext_lookup(ext, col_no,
&len);
if (UNIV_LIKELY_NULL(buf)) {
if (UNIV_UNLIKELY(buf == field_ref_zero)) {
return(NULL);
}
dfield_set_data(dfield, buf, len);
}
if (ind_field->prefix_len == 0) {
/* In the Barracuda format
(ROW_FORMAT=DYNAMIC or
ROW_FORMAT=COMPRESSED), we can have a
secondary index on an entire column
that is stored off-page in the
clustered index. As this is not a
prefix index (prefix_len == 0),
include the entire off-page column in
the secondary index record. */
continue;
}
} else if (dfield_is_ext(dfield)) {
/* This table is either in Antelope format
(ROW_FORMAT=REDUNDANT or ROW_FORMAT=COMPACT)
or a purge record where the ordered part of
the field is not external.
In Antelope, the maximum column prefix
index length is 767 bytes, and the clustered
index record contains a 768-byte prefix of
each off-page column. */
ut_a(len >= BTR_EXTERN_FIELD_REF_SIZE);
len -= BTR_EXTERN_FIELD_REF_SIZE;
dfield_set_len(dfield, len);
}
/* If a column prefix index, take only the prefix. */
if (ind_field->prefix_len) {
len = dtype_get_at_most_n_mbchars(
col->prtype, col->mbminlen, col->mbmaxlen,
ind_field->prefix_len, len,
dfield_get_data(dfield));
dfield_set_len(dfield, len);
}
}
ut_ad(dtuple_check_typed(entry));
return(entry);
}
/*******************************************************************//**
Builds from a secondary index record a row reference with which we can
search the clustered index record. */
UNIV_INTERN
void
row_build_row_ref_in_tuple(
/*=======================*/
dtuple_t* ref, /*!< in/out: row reference built;
see the NOTE below! */
const rec_t* rec, /*!< in: record in the index;
NOTE: the data fields in ref
will point directly into this
record, therefore, the buffer
page of this record must be at
least s-latched and the latch
held as long as the row
reference is used! */
const dict_index_t* index, /*!< in: secondary index */
ulint* offsets,/*!< in: rec_get_offsets(rec, index)
or NULL */
trx_t* trx) /*!< in: transaction */
{
const dict_index_t* clust_index;
dfield_t* dfield;
const byte* field;
ulint len;
ulint ref_len;
ulint pos;
ulint clust_col_prefix_len;
ulint i;
mem_heap_t* heap = NULL;
ulint offsets_[REC_OFFS_NORMAL_SIZE];
rec_offs_init(offsets_);
ut_a(ref);
ut_a(index);
ut_a(rec);
ut_ad(!dict_index_is_clust(index));
if (UNIV_UNLIKELY(!index->table)) {
fputs("InnoDB: table ", stderr);
notfound:
ut_print_name(stderr, trx, TRUE, index->table_name);
fputs(" for index ", stderr);
ut_print_name(stderr, trx, FALSE, index->name);
fputs(" not found\n", stderr);
ut_error;
}
clust_index = dict_table_get_first_index(index->table);
if (UNIV_UNLIKELY(!clust_index)) {
fputs("InnoDB: clust index for table ", stderr);
goto notfound;
}
if (!offsets) {
offsets = rec_get_offsets(rec, index, offsets_,
ULINT_UNDEFINED, &heap);
} else {
ut_ad(rec_offs_validate(rec, index, offsets));
}
/* Secondary indexes must not contain externally stored columns. */
ut_ad(!rec_offs_any_extern(offsets));
ref_len = dict_index_get_n_unique(clust_index);
ut_ad(ref_len == dtuple_get_n_fields(ref));
dict_index_copy_types(ref, clust_index, ref_len);
for (i = 0; i < ref_len; i++) {
dfield = dtuple_get_nth_field(ref, i);
pos = dict_index_get_nth_field_pos(index, clust_index, i);
ut_a(pos != ULINT_UNDEFINED);
field = rec_get_nth_field(rec, offsets, pos, &len);
dfield_set_data(dfield, field, len);
/* If the primary key contains a column prefix, then the
secondary index may contain a longer prefix of the same
column, or the full column, and we must adjust the length
accordingly. */
clust_col_prefix_len = dict_index_get_nth_field(
clust_index, i)->prefix_len;
if (clust_col_prefix_len > 0) {
if (len != UNIV_SQL_NULL) {
const dtype_t* dtype
= dfield_get_type(dfield);
dfield_set_len(dfield,
dtype_get_at_most_n_mbchars(
dtype->prtype,
dtype->mbminlen,
dtype->mbmaxlen,
clust_col_prefix_len,
len, (char*) field));
}
}
}
ut_ad(dtuple_check_typed(ref));
if (UNIV_LIKELY_NULL(heap)) {
mem_heap_free(heap);
}
}
/**************************************************************//**
Moves parts of long fields in entry to the big record vector so that
the size of tuple drops below the maximum record size allowed in the
database. Moves data only from those fields which are not necessary
to determine uniquely the insertion place of the tuple in the index.
@return own: created big record vector, NULL if we are not able to
shorten the entry enough, i.e., if there are too many fixed-length or
short fields in entry or the index is clustered */
UNIV_INTERN
big_rec_t*
dtuple_convert_big_rec(
/*===================*/
dict_index_t* index, /*!< in: index */
dtuple_t* entry, /*!< in/out: index entry */
ulint* n_ext) /*!< in/out: number of
externally stored columns */
{
mem_heap_t* heap;
big_rec_t* vector;
dfield_t* dfield;
dict_field_t* ifield;
ulint size;
ulint n_fields;
ulint local_len;
ulint local_prefix_len;
if (UNIV_UNLIKELY(!dict_index_is_clust(index))) {
return(NULL);
}
if (dict_table_get_format(index->table) < DICT_TF_FORMAT_ZIP) {
/* up to MySQL 5.1: store a 768-byte prefix locally */
local_len = BTR_EXTERN_FIELD_REF_SIZE + DICT_MAX_INDEX_COL_LEN;
} else {
/* new-format table: do not store any BLOB prefix locally */
local_len = BTR_EXTERN_FIELD_REF_SIZE;
}
ut_a(dtuple_check_typed_no_assert(entry));
size = rec_get_converted_size(index, entry, *n_ext);
if (UNIV_UNLIKELY(size > 1000000000)) {
fprintf(stderr,
"InnoDB: Warning: tuple size very big: %lu\n",
(ulong) size);
fputs("InnoDB: Tuple contents: ", stderr);
dtuple_print(stderr, entry);
putc('\n', stderr);
}
heap = mem_heap_create(size + dtuple_get_n_fields(entry)
* sizeof(big_rec_field_t) + 1000);
vector = mem_heap_alloc(heap, sizeof(big_rec_t));
vector->heap = heap;
vector->fields = mem_heap_alloc(heap, dtuple_get_n_fields(entry)
* sizeof(big_rec_field_t));
/* Decide which fields to shorten: the algorithm is to look for
a variable-length field that yields the biggest savings when
stored externally */
n_fields = 0;
while (page_zip_rec_needs_ext(rec_get_converted_size(index, entry,
*n_ext),
dict_table_is_comp(index->table),
dict_index_get_n_fields(index),
dict_table_zip_size(index->table))) {
ulint i;
ulint longest = 0;
ulint longest_i = ULINT_MAX;
byte* data;
big_rec_field_t* b;
for (i = dict_index_get_n_unique_in_tree(index);
i < dtuple_get_n_fields(entry); i++) {
ulint savings;
dfield = dtuple_get_nth_field(entry, i);
ifield = dict_index_get_nth_field(index, i);
/* Skip fixed-length, NULL, externally stored,
or short columns */
if (ifield->fixed_len
|| dfield_is_null(dfield)
|| dfield_is_ext(dfield)
|| dfield_get_len(dfield) <= local_len
|| dfield_get_len(dfield)
<= BTR_EXTERN_FIELD_REF_SIZE * 2) {
goto skip_field;
}
savings = dfield_get_len(dfield) - local_len;
/* Check that there would be savings */
if (longest >= savings) {
goto skip_field;
}
longest_i = i;
longest = savings;
skip_field:
continue;
}
if (!longest) {
/* Cannot shorten more */
mem_heap_free(heap);
return(NULL);
}
/* Move data from field longest_i to big rec vector.
We store the first bytes locally to the record. Then
we can calculate all ordering fields in all indexes
from locally stored data. */
dfield = dtuple_get_nth_field(entry, longest_i);
ifield = dict_index_get_nth_field(index, longest_i);
local_prefix_len = local_len - BTR_EXTERN_FIELD_REF_SIZE;
b = &vector->fields[n_fields];
b->field_no = longest_i;
b->len = dfield_get_len(dfield) - local_prefix_len;
b->data = (char*) dfield_get_data(dfield) + local_prefix_len;
/* Allocate the locally stored part of the column. */
data = mem_heap_alloc(heap, local_len);
/* Copy the local prefix. */
memcpy(data, dfield_get_data(dfield), local_prefix_len);
/* Clear the extern field reference (BLOB pointer). */
memset(data + local_prefix_len, 0, BTR_EXTERN_FIELD_REF_SIZE);
#if 0
/* The following would fail the Valgrind checks in
page_cur_insert_rec_low() and page_cur_insert_rec_zip().
The BLOB pointers in the record will be initialized after
the record and the BLOBs have been written. */
UNIV_MEM_ALLOC(data + local_prefix_len,
BTR_EXTERN_FIELD_REF_SIZE);
#endif
dfield_set_data(dfield, data, local_len);
dfield_set_ext(dfield);
n_fields++;
(*n_ext)++;
ut_ad(n_fields < dtuple_get_n_fields(entry));
}
vector->n_fields = n_fields;
return(vector);
}
/********************************************************************//**
Adds the update undo log as the first log in the history list. Removes the
update undo log segment from the rseg slot if it is too big for reuse. */
UNIV_INTERN
void
trx_purge_add_update_undo_to_history(
/*=================================*/
trx_t* trx, /*!< in: transaction */
page_t* undo_page, /*!< in: update undo log header page,
x-latched */
mtr_t* mtr) /*!< in: mtr */
{
trx_undo_t* undo;
trx_rsegf_t* rseg_header;
trx_ulogf_t* undo_header;
undo = trx->update_undo;
ut_ad(undo);
ut_ad(mutex_own(&undo->rseg->mutex));
rseg_header = trx_rsegf_get(
undo->rseg->space, undo->rseg->zip_size, undo->rseg->page_no,
mtr);
undo_header = undo_page + undo->hdr_offset;
/* Add the log as the first in the history list */
if (undo->state != TRX_UNDO_CACHED) {
ulint hist_size;
#ifdef UNIV_DEBUG
trx_usegf_t* seg_header = undo_page + TRX_UNDO_SEG_HDR;
#endif /* UNIV_DEBUG */
/* The undo log segment will not be reused */
if (UNIV_UNLIKELY(undo->id >= TRX_RSEG_N_SLOTS)) {
fprintf(stderr,
"InnoDB: Error: undo->id is %lu\n",
(ulong) undo->id);
ut_error;
}
trx_rsegf_set_nth_undo(rseg_header, undo->id, FIL_NULL, mtr);
hist_size = mtr_read_ulint(
rseg_header + TRX_RSEG_HISTORY_SIZE, MLOG_4BYTES, mtr);
ut_ad(undo->size == flst_get_len(
seg_header + TRX_UNDO_PAGE_LIST, mtr));
mlog_write_ulint(
rseg_header + TRX_RSEG_HISTORY_SIZE,
hist_size + undo->size, MLOG_4BYTES, mtr);
}
flst_add_first(
rseg_header + TRX_RSEG_HISTORY,
undo_header + TRX_UNDO_HISTORY_NODE, mtr);
/* Write the trx number to the undo log header */
mlog_write_ull(undo_header + TRX_UNDO_TRX_NO, trx->no, mtr);
/* Write information about delete markings to the undo log header */
if (!undo->del_marks) {
mlog_write_ulint(
undo_header + TRX_UNDO_DEL_MARKS, FALSE,
MLOG_2BYTES, mtr);
}
if (undo->rseg->last_page_no == FIL_NULL) {
undo->rseg->last_trx_no = trx->no;
undo->rseg->last_offset = undo->hdr_offset;
undo->rseg->last_page_no = undo->hdr_page_no;
undo->rseg->last_del_marks = undo->del_marks;
/* FIXME: Add a bin heap validate function to check that
the rseg exists. */
}
mutex_enter(&kernel_mutex);
trx_sys->rseg_history_len++;
mutex_exit(&kernel_mutex);
// if (!(trx_sys->rseg_history_len % srv_purge_batch_size)) { /*should wake up always*/
/* Inform the purge thread that there is work to do. */
srv_wake_purge_thread_if_not_active();
// }
}
/********************************************************************//**
Applies linear read-ahead if in the buf_pool the page is a border page of
a linear read-ahead area and all the pages in the area have been accessed.
Does not read any page if the read-ahead mechanism is not activated. Note
that the algorithm looks at the 'natural' adjacent successor and
predecessor of the page, which on the leaf level of a B-tree are the next
and previous page in the chain of leaves. To know these, the page specified
in (space, offset) must already be present in the buf_pool. Thus, the
natural way to use this function is to call it when a page in the buf_pool
is accessed the first time, calling this function just after it has been
bufferfixed.
NOTE 1: as this function looks at the natural predecessor and successor
fields on the page, what happens, if these are not initialized to any
sensible value? No problem, before applying read-ahead we check that the
area to read is within the span of the space, if not, read-ahead is not
applied. An uninitialized value may result in a useless read operation, but
only very improbably.
NOTE 2: the calling thread may own latches on pages: to avoid deadlocks this
function must be written such that it cannot end up waiting for these
latches!
NOTE 3: the calling thread must want access to the page given: this rule is
set to prevent unintended read-aheads performed by ibuf routines, a situation
which could result in a deadlock if the OS does not support asynchronous io.
@return number of page read requests issued */
UNIV_INTERN
ulint
buf_read_ahead_linear(
/*==================*/
ulint space, /*!< in: space id */
ulint zip_size, /*!< in: compressed page size in bytes, or 0 */
ulint offset, /*!< in: page number; see NOTE 3 above */
ibool inside_ibuf, /*!< in: TRUE if we are inside ibuf routine */
trx_t* trx)
{
buf_pool_t* buf_pool = buf_pool_get(space, offset);
ib_int64_t tablespace_version;
buf_page_t* bpage;
buf_frame_t* frame;
buf_page_t* pred_bpage = NULL;
ulint pred_offset;
ulint succ_offset;
ulint count;
int asc_or_desc;
ulint new_offset;
ulint fail_count;
ulint ibuf_mode;
ulint low, high;
ulint err;
ulint i;
const ulint buf_read_ahead_linear_area
= BUF_READ_AHEAD_AREA(buf_pool);
ulint threshold;
if (!(srv_read_ahead & 2)) {
return(0);
}
if (UNIV_UNLIKELY(srv_startup_is_before_trx_rollback_phase)) {
/* No read-ahead to avoid thread deadlocks */
return(0);
}
low = (offset / buf_read_ahead_linear_area)
* buf_read_ahead_linear_area;
high = (offset / buf_read_ahead_linear_area + 1)
* buf_read_ahead_linear_area;
if ((offset != low) && (offset != high - 1)) {
/* This is not a border page of the area: return */
return(0);
}
if (ibuf_bitmap_page(zip_size, offset)
|| trx_sys_hdr_page(space, offset)) {
/* If it is an ibuf bitmap page or trx sys hdr, we do
no read-ahead, as that could break the ibuf page access
order */
return(0);
}
/* Remember the tablespace version before we ask te tablespace size
below: if DISCARD + IMPORT changes the actual .ibd file meanwhile, we
do not try to read outside the bounds of the tablespace! */
tablespace_version = fil_space_get_version(space);
buf_pool_mutex_enter(buf_pool);
if (high > fil_space_get_size(space)) {
buf_pool_mutex_exit(buf_pool);
/* The area is not whole, return */
return(0);
}
if (buf_pool->n_pend_reads
> buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) {
buf_pool_mutex_exit(buf_pool);
return(0);
}
buf_pool_mutex_exit(buf_pool);
/* Check that almost all pages in the area have been accessed; if
offset == low, the accesses must be in a descending order, otherwise,
in an ascending order. */
asc_or_desc = 1;
if (offset == low) {
asc_or_desc = -1;
}
/* How many out of order accessed pages can we ignore
when working out the access pattern for linear readahead */
threshold = ut_min((64 - srv_read_ahead_threshold),
BUF_READ_AHEAD_AREA(buf_pool));
fail_count = 0;
rw_lock_s_lock(&buf_pool->page_hash_latch);
for (i = low; i < high; i++) {
bpage = buf_page_hash_get(buf_pool, space, i);
if (bpage == NULL || !buf_page_is_accessed(bpage)) {
/* Not accessed */
fail_count++;
} else if (pred_bpage) {
/* Note that buf_page_is_accessed() returns
the time of the first access. If some blocks
of the extent existed in the buffer pool at
the time of a linear access pattern, the first
access times may be nonmonotonic, even though
the latest access times were linear. The
threshold (srv_read_ahead_factor) should help
a little against this. */
int res = ut_ulint_cmp(
buf_page_is_accessed(bpage),
buf_page_is_accessed(pred_bpage));
/* Accesses not in the right order */
if (res != 0 && res != asc_or_desc) {
fail_count++;
}
}
if (fail_count > threshold) {
/* Too many failures: return */
//buf_pool_mutex_exit(buf_pool);
rw_lock_s_unlock(&buf_pool->page_hash_latch);
return(0);
}
if (bpage && buf_page_is_accessed(bpage)) {
pred_bpage = bpage;
}
}
/* If we got this far, we know that enough pages in the area have
been accessed in the right order: linear read-ahead can be sensible */
bpage = buf_page_hash_get(buf_pool, space, offset);
if (bpage == NULL) {
//buf_pool_mutex_exit(buf_pool);
rw_lock_s_unlock(&buf_pool->page_hash_latch);
return(0);
}
switch (buf_page_get_state(bpage)) {
case BUF_BLOCK_ZIP_PAGE:
frame = bpage->zip.data;
break;
case BUF_BLOCK_FILE_PAGE:
frame = ((buf_block_t*) bpage)->frame;
break;
default:
ut_error;
break;
}
/* Read the natural predecessor and successor page addresses from
the page; NOTE that because the calling thread may have an x-latch
on the page, we do not acquire an s-latch on the page, this is to
prevent deadlocks. Even if we read values which are nonsense, the
algorithm will work. */
pred_offset = fil_page_get_prev(frame);
succ_offset = fil_page_get_next(frame);
//buf_pool_mutex_exit(buf_pool);
rw_lock_s_unlock(&buf_pool->page_hash_latch);
if ((offset == low) && (succ_offset == offset + 1)) {
/* This is ok, we can continue */
new_offset = pred_offset;
} else if ((offset == high - 1) && (pred_offset == offset - 1)) {
/* This is ok, we can continue */
new_offset = succ_offset;
} else {
/* Successor or predecessor not in the right order */
return(0);
}
low = (new_offset / buf_read_ahead_linear_area)
* buf_read_ahead_linear_area;
high = (new_offset / buf_read_ahead_linear_area + 1)
* buf_read_ahead_linear_area;
if ((new_offset != low) && (new_offset != high - 1)) {
/* This is not a border page of the area: return */
return(0);
}
if (high > fil_space_get_size(space)) {
/* The area is not whole, return */
return(0);
}
/* If we got this far, read-ahead can be sensible: do it */
ibuf_mode = inside_ibuf
? BUF_READ_IBUF_PAGES_ONLY | OS_AIO_SIMULATED_WAKE_LATER
: BUF_READ_ANY_PAGE | OS_AIO_SIMULATED_WAKE_LATER;
count = 0;
/* Since Windows XP seems to schedule the i/o handler thread
very eagerly, and consequently it does not wait for the
full read batch to be posted, we use special heuristics here */
os_aio_simulated_put_read_threads_to_sleep();
for (i = low; i < high; i++) {
/* It is only sensible to do read-ahead in the non-sync
aio mode: hence FALSE as the first parameter */
if (!ibuf_bitmap_page(zip_size, i)) {
count += buf_read_page_low(
&err, FALSE,
ibuf_mode,
space, zip_size, FALSE, tablespace_version, i, trx);
if (err == DB_TABLESPACE_DELETED) {
ut_print_timestamp(stderr);
fprintf(stderr,
" InnoDB: Warning: in"
" linear readahead trying to access\n"
"InnoDB: tablespace %lu page %lu,\n"
"InnoDB: but the tablespace does not"
" exist or is just being dropped.\n",
(ulong) space, (ulong) i);
}
}
}
/* In simulated aio we wake the aio handler threads only after
queuing all aio requests, in native aio the following call does
nothing: */
os_aio_simulated_wake_handler_threads();
/* Flush pages from the end of the LRU list if necessary */
buf_flush_free_margin(buf_pool, TRUE);
#ifdef UNIV_DEBUG
if (buf_debug_prints && (count > 0)) {
fprintf(stderr,
"LINEAR read-ahead space %lu offset %lu pages %lu\n",
(ulong) space, (ulong) offset, (ulong) count);
}
#endif /* UNIV_DEBUG */
/* Read ahead is considered one I/O operation for the purpose of
LRU policy decision. */
buf_LRU_stat_inc_io();
buf_pool->stat.n_ra_pages_read += count;
return(count);
}
big_rec_t*
dtuple_convert_big_rec(
/*===================*/
/* out, own: created big record vector,
NULL if we are not able to shorten
the entry enough, i.e., if there are
too many short fields in entry */
dict_index_t* index, /* in: index */
dtuple_t* entry, /* in: index entry */
ulint* ext_vec,/* in: array of externally stored fields,
or NULL: if a field already is externally
stored, then we cannot move it to the vector
this function returns */
ulint n_ext_vec)/* in: number of elements is ext_vec */
{
mem_heap_t* heap;
big_rec_t* vector;
dfield_t* dfield;
ulint size;
ulint n_fields;
ulint longest;
ulint longest_i = ULINT_MAX;
ibool is_externally_stored;
ulint i;
ulint j;
ut_a(dtuple_check_typed_no_assert(entry));
size = rec_get_converted_size(index, entry);
if (UNIV_UNLIKELY(size > 1000000000)) {
fprintf(stderr,
"InnoDB: Warning: tuple size very big: %lu\n", (ulong) size);
fputs("InnoDB: Tuple contents: ", stderr);
dtuple_print(stderr, entry);
putc('\n', stderr);
}
heap = mem_heap_create(size + dtuple_get_n_fields(entry)
* sizeof(big_rec_field_t) + 1000);
vector = mem_heap_alloc(heap, sizeof(big_rec_t));
vector->heap = heap;
vector->fields = mem_heap_alloc(heap, dtuple_get_n_fields(entry)
* sizeof(big_rec_field_t));
/* Decide which fields to shorten: the algorithm is to look for
the longest field whose type is DATA_BLOB */
n_fields = 0;
while (rec_get_converted_size(index, entry)
>= ut_min(page_get_free_space_of_empty(
index->table->comp) / 2,
REC_MAX_DATA_SIZE)) {
longest = 0;
for (i = dict_index_get_n_unique_in_tree(index);
i < dtuple_get_n_fields(entry); i++) {
/* Skip over fields which already are externally
stored */
is_externally_stored = FALSE;
if (ext_vec) {
for (j = 0; j < n_ext_vec; j++) {
if (ext_vec[j] == i) {
is_externally_stored = TRUE;
}
}
}
if (!is_externally_stored) {
dfield = dtuple_get_nth_field(entry, i);
if (dfield->len != UNIV_SQL_NULL &&
dfield->len > longest) {
longest = dfield->len;
longest_i = i;
}
}
}
/* We do not store externally fields which are smaller than
DICT_MAX_INDEX_COL_LEN */
ut_a(DICT_MAX_INDEX_COL_LEN > REC_1BYTE_OFFS_LIMIT);
if (longest < BTR_EXTERN_FIELD_REF_SIZE + 10
+ DICT_MAX_INDEX_COL_LEN) {
/* Cannot shorten more */
mem_heap_free(heap);
return(NULL);
}
/* Move data from field longest_i to big rec vector;
we do not let data size of the remaining entry
drop below 128 which is the limit for the 2-byte
offset storage format in a physical record. This
we accomplish by storing 128 bytes of data in entry
itself, and only the remaining part to big rec vec.
We store the first bytes locally to the record. Then
we can calculate all ordering fields in all indexes
from locally stored data. */
dfield = dtuple_get_nth_field(entry, longest_i);
vector->fields[n_fields].field_no = longest_i;
ut_a(dfield->len > DICT_MAX_INDEX_COL_LEN);
vector->fields[n_fields].len = dfield->len
- DICT_MAX_INDEX_COL_LEN;
vector->fields[n_fields].data = mem_heap_alloc(heap,
vector->fields[n_fields].len);
/* Copy data (from the end of field) to big rec vector */
ut_memcpy(vector->fields[n_fields].data,
((byte*)dfield->data) + dfield->len
- vector->fields[n_fields].len,
vector->fields[n_fields].len);
dfield->len = dfield->len - vector->fields[n_fields].len
+ BTR_EXTERN_FIELD_REF_SIZE;
/* Set the extern field reference in dfield to zero */
memset(((byte*)dfield->data)
+ dfield->len - BTR_EXTERN_FIELD_REF_SIZE,
0, BTR_EXTERN_FIELD_REF_SIZE);
n_fields++;
}
vector->n_fields = n_fields;
return(vector);
}
/********************************************************************//**
Issues read requests for pages which the ibuf module wants to read in, in
order to contract the insert buffer tree. Technically, this function is like
a read-ahead function. */
UNIV_INTERN
void
buf_read_ibuf_merge_pages(
/*======================*/
ibool sync, /*!< in: TRUE if the caller
wants this function to wait
for the highest address page
to get read in, before this
function returns */
const ulint* space_ids, /*!< in: array of space ids */
const ib_int64_t* space_versions,/*!< in: the spaces must have
this version number
(timestamp), otherwise we
discard the read; we use this
to cancel reads if DISCARD +
IMPORT may have changed the
tablespace size */
const ulint* page_nos, /*!< in: array of page numbers
to read, with the highest page
number the last in the
array */
ulint n_stored) /*!< in: number of elements
in the arrays */
{
ulint i;
#ifdef UNIV_IBUF_DEBUG
ut_a(n_stored < UNIV_PAGE_SIZE);
#endif
for (i = 0; i < n_stored; i++) {
ulint err;
buf_pool_t* buf_pool;
ulint zip_size = fil_space_get_zip_size(space_ids[i]);
buf_pool = buf_pool_get(space_ids[i], page_nos[i]);
while (buf_pool->n_pend_reads
> buf_pool->curr_size / BUF_READ_AHEAD_PEND_LIMIT) {
os_thread_sleep(500000);
}
if (UNIV_UNLIKELY(zip_size == ULINT_UNDEFINED)) {
goto tablespace_deleted;
}
buf_read_page_low(&err, sync && (i + 1 == n_stored),
BUF_READ_ANY_PAGE, space_ids[i],
zip_size, TRUE, space_versions[i],
page_nos[i], NULL);
if (UNIV_UNLIKELY(err == DB_TABLESPACE_DELETED)) {
tablespace_deleted:
/* We have deleted or are deleting the single-table
tablespace: remove the entries for that page */
ibuf_merge_or_delete_for_page(NULL, space_ids[i],
page_nos[i],
zip_size, FALSE);
}
}
os_aio_simulated_wake_handler_threads();
/* Flush pages from the end of all the LRU lists if necessary */
buf_flush_free_margins(FALSE);
#ifdef UNIV_DEBUG
if (buf_debug_prints) {
fprintf(stderr,
"Ibuf merge read-ahead space %lu pages %lu\n",
(ulong) space_ids[0], (ulong) n_stored);
}
#endif /* UNIV_DEBUG */
}
