/***********************************************************//**
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));
}
/***************************************************************//**
Removes a secondary index entry if found.
@return DB_SUCCESS, DB_FAIL, or DB_OUT_OF_FILE_SPACE */
static
ulint
row_undo_ins_remove_sec_low(
/*========================*/
ulint mode, /*!< in: BTR_MODIFY_LEAF or BTR_MODIFY_TREE,
depending on whether we wish optimistic or
pessimistic descent down the index tree */
dict_index_t* index, /*!< in: index */
dtuple_t* entry) /*!< in: index entry to remove */
{
btr_pcur_t pcur;
btr_cur_t* btr_cur;
ibool found;
ibool success;
ulint err;
mtr_t mtr;
log_free_check();
mtr_start(&mtr);
found = row_search_index_entry(index, entry, mode, &pcur, &mtr);
btr_cur = btr_pcur_get_btr_cur(&pcur);
if (!found) {
/* Not found */
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(DB_SUCCESS);
}
if (mode == BTR_MODIFY_LEAF) {
success = btr_cur_optimistic_delete(btr_cur, &mtr);
if (success) {
err = DB_SUCCESS;
} else {
err = DB_FAIL;
}
} else {
ut_ad(mode == BTR_MODIFY_TREE);
/* No need to distinguish RB_RECOVERY here, because we
are deleting a secondary index record: the distinction
between RB_NORMAL and RB_RECOVERY only matters when
deleting a record that contains externally stored
columns. */
ut_ad(!dict_index_is_clust(index));
btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
RB_NORMAL, &mtr);
}
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(err);
}
/*******************************************************************
Removes a secondary index entry if found. */
static
ulint
row_undo_ins_remove_sec_low(
/*========================*/
/* out: DB_SUCCESS, DB_FAIL, or
DB_OUT_OF_FILE_SPACE */
ulint mode, /* in: BTR_MODIFY_LEAF or BTR_MODIFY_TREE,
depending on whether we wish optimistic or
pessimistic descent down the index tree */
dict_index_t* index, /* in: index */
dtuple_t* entry) /* in: index entry to remove */
{
btr_pcur_t pcur;
btr_cur_t* btr_cur;
ibool found;
ibool success;
ulint err;
mtr_t mtr;
log_free_check();
mtr_start(&mtr);
found = row_search_index_entry(index, entry, mode, &pcur, &mtr);
btr_cur = btr_pcur_get_btr_cur(&pcur);
if (!found) {
/* Not found */
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(DB_SUCCESS);
}
if (mode == BTR_MODIFY_LEAF) {
success = btr_cur_optimistic_delete(btr_cur, &mtr);
if (success) {
err = DB_SUCCESS;
} else {
err = DB_FAIL;
}
} else {
ut_ad(mode == BTR_MODIFY_TREE);
btr_cur_pessimistic_delete(&err, FALSE, btr_cur, TRUE, &mtr);
}
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(err);
}
/*******************************************************************
Removes a secondary index entry if found. */
static
ulint
row_undo_ins_remove_sec_low(
/*========================*/
/* out: DB_SUCCESS, DB_FAIL, or
DB_OUT_OF_FILE_SPACE */
ulint mode, /* in: BTR_MODIFY_LEAF or BTR_MODIFY_TREE,
depending on whether we wish optimistic or
pessimistic descent down the index tree */
dict_index_t* index, /* in: index */
dtuple_t* entry, /* in: index entry to remove */
que_thr_t* thr) /* in: query thread */
{
btr_pcur_t pcur;
btr_cur_t* btr_cur;
ibool found;
ibool success;
ulint err;
mtr_t mtr;
UT_NOT_USED(thr);
log_free_check();
mtr_start(&mtr);
found = row_search_index_entry(index, entry, mode, &pcur, &mtr);
btr_cur = btr_pcur_get_btr_cur(&pcur);
if (!found) {
/* Not found */
/* FIXME: remove printfs in the final version */
/* printf(
"--UNDO INS: Record not found from page %lu index %s\n",
buf_frame_get_page_no(btr_cur_get_rec(btr_cur)),
index->name); */
/* ibuf_print(); */
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(DB_SUCCESS);
}
if (mode == BTR_MODIFY_LEAF) {
success = btr_cur_optimistic_delete(btr_cur, &mtr);
if (success) {
err = DB_SUCCESS;
} else {
err = DB_FAIL;
}
} else {
ut_ad(mode == BTR_MODIFY_TREE);
btr_cur_pessimistic_delete(&err, FALSE, btr_cur, TRUE, &mtr);
}
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(err);
}
/***************************************************************
Updates a secondary index entry of a row. */
static
ulint
row_upd_sec_index_entry(
/*====================*/
/* out: DB_SUCCESS if operation successfully
completed, else error code or DB_LOCK_WAIT */
upd_node_t* node, /* in: row update node */
que_thr_t* thr) /* in: query thread */
{
ibool check_ref;
ibool found;
dict_index_t* index;
dtuple_t* entry;
btr_pcur_t pcur;
btr_cur_t* btr_cur;
mem_heap_t* heap;
rec_t* rec;
ulint err = DB_SUCCESS;
mtr_t mtr;
char err_buf[1000];
index = node->index;
check_ref = row_upd_index_is_referenced(index);
heap = mem_heap_create(1024);
/* Build old index entry */
entry = row_build_index_entry(node->row, index, heap);
log_free_check();
mtr_start(&mtr);
found = row_search_index_entry(index, entry, BTR_MODIFY_LEAF, &pcur,
&mtr);
btr_cur = btr_pcur_get_btr_cur(&pcur);
rec = btr_cur_get_rec(btr_cur);
if (!found) {
fprintf(stderr, "InnoDB: error in sec index entry update in\n"
"InnoDB: index %s table %s\n", index->name,
index->table->name);
dtuple_sprintf(err_buf, 900, entry);
fprintf(stderr, "InnoDB: tuple %s\n", err_buf);
rec_sprintf(err_buf, 900, rec);
fprintf(stderr, "InnoDB: record %s\n", err_buf);
fprintf(stderr,
"InnoDB: Make a detailed bug report and send it\n");
fprintf(stderr, "InnoDB: to [email protected]\n");
trx_print(thr_get_trx(thr));
} else {
/* Delete mark the old index record; it can already be
delete marked if we return after a lock wait in
row_ins_index_entry below */
if (!rec_get_deleted_flag(rec)) {
err = btr_cur_del_mark_set_sec_rec(0, btr_cur, TRUE,
thr, &mtr);
if (err == DB_SUCCESS && check_ref) {
/* NOTE that the following call loses
the position of pcur ! */
err = row_upd_check_references_constraints(
&pcur, index->table,
index, thr, &mtr);
if (err != DB_SUCCESS) {
goto close_cur;
}
}
}
}
close_cur:
btr_pcur_close(&pcur);
mtr_commit(&mtr);
if (node->is_delete || err != DB_SUCCESS) {
mem_heap_free(heap);
return(err);
}
/* Build a new index entry */
row_upd_index_replace_new_col_vals(entry, index, node->update);
/* Insert new index entry */
err = row_ins_index_entry(index, entry, NULL, 0, thr);
mem_heap_free(heap);
return(err);
}
/***************************************************************
Removes a secondary index entry without modifying the index tree,
if possible.
@return TRUE if success or if not found */
static
ibool
row_purge_remove_sec_if_poss_leaf(
/*==============================*/
purge_node_t* node, /*!< in: row purge node */
dict_index_t* index, /*!< in: index */
const dtuple_t* entry) /*!< in: index entry */
{
mtr_t mtr;
btr_pcur_t pcur;
enum row_search_result search_result;
log_free_check();
mtr_start(&mtr);
/* Set the purge node for the call to row_purge_poss_sec(). */
pcur.btr_cur.purge_node = node;
/* Set the query thread, so that ibuf_insert_low() will be
able to invoke thd_get_trx(). */
pcur.btr_cur.thr = que_node_get_parent(node);
search_result = row_search_index_entry(
index, entry, BTR_MODIFY_LEAF | BTR_DELETE, &pcur, &mtr);
switch (search_result) {
ibool success;
case ROW_FOUND:
/* Before attempting to purge a record, check
if it is safe to do so. */
if (row_purge_poss_sec(node, index, entry)) {
btr_cur_t* btr_cur = btr_pcur_get_btr_cur(&pcur);
/* Only delete-marked records should be purged. */
ut_ad(REC_INFO_DELETED_FLAG
& rec_get_info_bits(
btr_cur_get_rec(btr_cur),
dict_table_is_comp(index->table)));
if (!btr_cur_optimistic_delete(btr_cur, &mtr)) {
/* The index entry could not be deleted. */
success = FALSE;
goto func_exit;
}
}
/* fall through (the index entry is still needed,
or the deletion succeeded) */
case ROW_NOT_DELETED_REF:
/* The index entry is still needed. */
case ROW_BUFFERED:
/* The deletion was buffered. */
case ROW_NOT_FOUND:
/* The index entry does not exist, nothing to do. */
success = TRUE;
func_exit:
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(success);
}
ut_error;
return(FALSE);
}
/***************************************************************
Removes a secondary index entry if possible, by modifying the
index tree. Does not try to buffer the delete.
@return TRUE if success or if not found */
static
ibool
row_purge_remove_sec_if_poss_tree(
/*==============================*/
purge_node_t* node, /*!< in: row purge node */
dict_index_t* index, /*!< in: index */
const dtuple_t* entry) /*!< in: index entry */
{
btr_pcur_t pcur;
btr_cur_t* btr_cur;
ibool success = TRUE;
ulint err;
mtr_t mtr;
enum row_search_result search_result;
log_free_check();
mtr_start(&mtr);
search_result = row_search_index_entry(index, entry, BTR_MODIFY_TREE,
&pcur, &mtr);
switch (search_result) {
case ROW_NOT_FOUND:
/* Not found. This is a legitimate condition. In a
rollback, InnoDB will remove secondary recs that would
be purged anyway. Then the actual purge will not find
the secondary index record. Also, the purge itself is
eager: if it comes to consider a secondary index
record, and notices it does not need to exist in the
index, it will remove it. Then if/when the purge
comes to consider the secondary index record a second
time, it will not exist any more in the index. */
/* fputs("PURGE:........sec entry not found\n", stderr); */
/* dtuple_print(stderr, entry); */
goto func_exit;
case ROW_FOUND:
break;
case ROW_BUFFERED:
case ROW_NOT_DELETED_REF:
/* These are invalid outcomes, because the mode passed
to row_search_index_entry() did not include any of the
flags BTR_INSERT, BTR_DELETE, or BTR_DELETE_MARK. */
ut_error;
}
btr_cur = btr_pcur_get_btr_cur(&pcur);
/* We should remove the index record if no later version of the row,
which cannot be purged yet, requires its existence. If some requires,
we should do nothing. */
if (row_purge_poss_sec(node, index, entry)) {
/* Remove the index record, which should have been
marked for deletion. */
ut_ad(REC_INFO_DELETED_FLAG
& rec_get_info_bits(btr_cur_get_rec(btr_cur),
dict_table_is_comp(index->table)));
btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
RB_NONE, &mtr);
switch (UNIV_EXPECT(err, DB_SUCCESS)) {
case DB_SUCCESS:
break;
case DB_OUT_OF_FILE_SPACE:
success = FALSE;
break;
default:
ut_error;
}
}
func_exit:
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(success);
}
/***********************************************************//**
Removes a delete marked clustered index record if possible.
@return TRUE if success, or if not found, or if modified after the
delete marking */
static
ibool
row_purge_remove_clust_if_poss_low(
/*===============================*/
purge_node_t* node, /*!< in: row purge node */
ulint mode) /*!< in: BTR_MODIFY_LEAF or BTR_MODIFY_TREE */
{
dict_index_t* index;
btr_pcur_t* pcur;
btr_cur_t* btr_cur;
ibool success;
ulint err;
mtr_t mtr;
rec_t* rec;
mem_heap_t* heap = NULL;
ulint offsets_[REC_OFFS_NORMAL_SIZE];
rec_offs_init(offsets_);
index = dict_table_get_first_index(node->table);
pcur = &(node->pcur);
btr_cur = btr_pcur_get_btr_cur(pcur);
log_free_check();
mtr_start(&mtr);
success = row_purge_reposition_pcur(mode, node, &mtr);
if (!success) {
/* The record is already removed */
btr_pcur_commit_specify_mtr(pcur, &mtr);
return(TRUE);
}
rec = btr_pcur_get_rec(pcur);
if (node->roll_ptr != row_get_rec_roll_ptr(
rec, index, rec_get_offsets(rec, index, offsets_,
ULINT_UNDEFINED, &heap))) {
if (UNIV_LIKELY_NULL(heap)) {
mem_heap_free(heap);
}
/* Someone else has modified the record later: do not remove */
btr_pcur_commit_specify_mtr(pcur, &mtr);
return(TRUE);
}
if (UNIV_LIKELY_NULL(heap)) {
mem_heap_free(heap);
}
if (mode == BTR_MODIFY_LEAF) {
success = btr_cur_optimistic_delete(btr_cur, &mtr);
} else {
ut_ad(mode == BTR_MODIFY_TREE);
btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
RB_NONE, &mtr);
if (err == DB_SUCCESS) {
success = TRUE;
} else if (err == DB_OUT_OF_FILE_SPACE) {
success = FALSE;
} else {
ut_error;
}
}
btr_pcur_commit_specify_mtr(pcur, &mtr);
return(success);
}
/***************************************************************
Removes a secondary index entry if possible. */
static
ibool
row_purge_remove_sec_if_poss_low(
/*=============================*/
/* out: TRUE if success or if not found */
purge_node_t* node, /* in: row purge node */
que_thr_t* thr, /* in: query thread */
dict_index_t* index, /* in: index */
dtuple_t* entry, /* in: index entry */
ulint mode) /* in: latch mode BTR_MODIFY_LEAF or
BTR_MODIFY_TREE */
{
btr_pcur_t pcur;
btr_cur_t* btr_cur;
ibool success;
ibool old_has;
ibool found;
ulint err;
mtr_t mtr;
mtr_t* mtr_vers;
UT_NOT_USED(thr);
log_free_check();
mtr_start(&mtr);
found = row_search_index_entry(index, entry, mode, &pcur, &mtr);
if (!found) {
/* Not found */
/* printf("PURGE:........sec entry not found\n"); */
/* dtuple_print(entry); */
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(TRUE);
}
btr_cur = btr_pcur_get_btr_cur(&pcur);
/* We should remove the index record if no later version of the row,
which cannot be purged yet, requires its existence. If some requires,
we should do nothing. */
mtr_vers = mem_alloc(sizeof(mtr_t));
mtr_start(mtr_vers);
success = row_purge_reposition_pcur(BTR_SEARCH_LEAF, node, mtr_vers);
if (success) {
old_has = row_vers_old_has_index_entry(TRUE,
btr_pcur_get_rec(&(node->pcur)),
mtr_vers, index, entry);
}
btr_pcur_commit_specify_mtr(&(node->pcur), mtr_vers);
mem_free(mtr_vers);
if (!success || !old_has) {
/* Remove the index record */
if (mode == BTR_MODIFY_LEAF) {
success = btr_cur_optimistic_delete(btr_cur, &mtr);
} else {
ut_ad(mode == BTR_MODIFY_TREE);
btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
FALSE, &mtr);
if (err == DB_SUCCESS) {
success = TRUE;
} else if (err == DB_OUT_OF_FILE_SPACE) {
success = FALSE;
} else {
ut_a(0);
}
}
}
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(success);
}
/***********************************************************//**
Delete unmarks a secondary index entry which must be found. It might not be
delete-marked at the moment, but it does not harm to unmark it anyway. We also
need to update the fields of the secondary index record if we updated its
fields but alphabetically they stayed the same, e.g., 'abc' -> 'aBc'.
@return DB_FAIL or DB_SUCCESS or DB_OUT_OF_FILE_SPACE */
static
ulint
row_undo_mod_del_unmark_sec_and_undo_update(
/*========================================*/
ulint mode, /*!< in: search mode: BTR_MODIFY_LEAF or
BTR_MODIFY_TREE */
que_thr_t* thr, /*!< in: query thread */
dict_index_t* index, /*!< in: index */
const dtuple_t* entry) /*!< in: index entry */
{
mem_heap_t* heap;
btr_pcur_t pcur;
btr_cur_t* btr_cur;
upd_t* update;
ulint err = DB_SUCCESS;
big_rec_t* dummy_big_rec;
mtr_t mtr;
trx_t* trx = thr_get_trx(thr);
enum row_search_result search_result;
/* Ignore indexes that are being created. */
if (UNIV_UNLIKELY(*index->name == TEMP_INDEX_PREFIX)) {
return(DB_SUCCESS);
}
log_free_check();
mtr_start(&mtr);
ut_ad(mode == BTR_MODIFY_TREE || mode == BTR_MODIFY_LEAF);
search_result = row_search_index_entry(index, entry, mode,
&pcur, &mtr);
switch (search_result) {
case ROW_BUFFERED:
case ROW_NOT_DELETED_REF:
/* These are invalid outcomes, because the mode passed
to row_search_index_entry() did not include any of the
flags BTR_INSERT, BTR_DELETE, or BTR_DELETE_MARK. */
ut_error;
case ROW_NOT_FOUND:
fputs("InnoDB: error in sec index entry del undo in\n"
"InnoDB: ", stderr);
dict_index_name_print(stderr, trx, index);
fputs("\n"
"InnoDB: tuple ", stderr);
dtuple_print(stderr, entry);
fputs("\n"
"InnoDB: record ", stderr);
rec_print(stderr, btr_pcur_get_rec(&pcur), index);
putc('\n', stderr);
trx_print(stderr, trx, 0);
fputs("\n"
"InnoDB: Submit a detailed bug report"
" to http://bugs.mysql.com\n", stderr);
break;
case ROW_FOUND:
btr_cur = btr_pcur_get_btr_cur(&pcur);
err = btr_cur_del_mark_set_sec_rec(BTR_NO_LOCKING_FLAG,
btr_cur, FALSE, thr, &mtr);
ut_a(err == DB_SUCCESS);
heap = mem_heap_create(100);
update = row_upd_build_sec_rec_difference_binary(
index, entry, btr_cur_get_rec(btr_cur), trx, heap);
if (upd_get_n_fields(update) == 0) {
/* Do nothing */
} else if (mode == BTR_MODIFY_LEAF) {
/* Try an optimistic updating of the record, keeping
changes within the page */
err = btr_cur_optimistic_update(
BTR_KEEP_SYS_FLAG | BTR_NO_LOCKING_FLAG,
btr_cur, update, 0, thr, &mtr);
switch (err) {
case DB_OVERFLOW:
case DB_UNDERFLOW:
case DB_ZIP_OVERFLOW:
err = DB_FAIL;
}
} else {
ut_a(mode == BTR_MODIFY_TREE);
err = btr_cur_pessimistic_update(
BTR_KEEP_SYS_FLAG | BTR_NO_LOCKING_FLAG,
btr_cur, &heap, &dummy_big_rec,
update, 0, thr, &mtr);
ut_a(!dummy_big_rec);
}
mem_heap_free(heap);
}
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(err);
}
/***********************************************************//**
Delete marks or removes a secondary index entry if found.
@return DB_SUCCESS, DB_FAIL, or DB_OUT_OF_FILE_SPACE */
static
ulint
row_undo_mod_del_mark_or_remove_sec_low(
/*====================================*/
undo_node_t* node, /*!< in: row undo node */
que_thr_t* thr, /*!< in: query thread */
dict_index_t* index, /*!< in: index */
dtuple_t* entry, /*!< in: index entry */
ulint mode) /*!< in: latch mode BTR_MODIFY_LEAF or
BTR_MODIFY_TREE */
{
btr_pcur_t pcur;
btr_cur_t* btr_cur;
ibool success;
ibool old_has;
ulint err;
mtr_t mtr;
mtr_t mtr_vers;
enum row_search_result search_result;
log_free_check();
mtr_start(&mtr);
btr_cur = btr_pcur_get_btr_cur(&pcur);
ut_ad(mode == BTR_MODIFY_TREE || mode == BTR_MODIFY_LEAF);
search_result = row_search_index_entry(index, entry, mode,
&pcur, &mtr);
switch (UNIV_EXPECT(search_result, ROW_FOUND)) {
case ROW_NOT_FOUND:
/* In crash recovery, the secondary index record may
be missing if the UPDATE did not have time to insert
the secondary index records before the crash. When we
are undoing that UPDATE in crash recovery, the record
may be missing.
In normal processing, if an update ends in a deadlock
before it has inserted all updated secondary index
records, then the undo will not find those records. */
err = DB_SUCCESS;
goto func_exit;
case ROW_FOUND:
break;
case ROW_BUFFERED:
case ROW_NOT_DELETED_REF:
/* These are invalid outcomes, because the mode passed
to row_search_index_entry() did not include any of the
flags BTR_INSERT, BTR_DELETE, or BTR_DELETE_MARK. */
ut_error;
}
/* We should remove the index record if no prior version of the row,
which cannot be purged yet, requires its existence. If some requires,
we should delete mark the record. */
mtr_start(&mtr_vers);
success = btr_pcur_restore_position(BTR_SEARCH_LEAF, &(node->pcur),
&mtr_vers);
ut_a(success);
old_has = row_vers_old_has_index_entry(FALSE,
btr_pcur_get_rec(&(node->pcur)),
&mtr_vers, index, entry);
if (old_has) {
err = btr_cur_del_mark_set_sec_rec(BTR_NO_LOCKING_FLAG,
btr_cur, TRUE, thr, &mtr);
ut_ad(err == DB_SUCCESS);
} else {
/* Remove the index record */
if (mode == BTR_MODIFY_LEAF) {
success = btr_cur_optimistic_delete(btr_cur, &mtr);
if (success) {
err = DB_SUCCESS;
} else {
err = DB_FAIL;
}
} else {
ut_ad(mode == BTR_MODIFY_TREE);
/* No need to distinguish RB_RECOVERY_PURGE here,
because we are deleting a secondary index record:
the distinction between RB_NORMAL and
RB_RECOVERY_PURGE only matters when deleting a
record that contains externally stored
columns. */
ut_ad(!dict_index_is_clust(index));
btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
RB_NORMAL, &mtr);
/* The delete operation may fail if we have little
file space left: TODO: easiest to crash the database
and restart with more file space */
}
}
btr_pcur_commit_specify_mtr(&(node->pcur), &mtr_vers);
func_exit:
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(err);
}
/***********************************************************//**
Undoes a modify in a clustered index record. Sets also the node state for the
next round of undo.
@return DB_SUCCESS or error code: we may run out of file space */
static
ulint
row_undo_mod_clust(
/*===============*/
undo_node_t* node, /*!< in: row undo node */
que_thr_t* thr) /*!< in: query thread */
{
btr_pcur_t* pcur;
mtr_t mtr;
ulint err;
ibool success;
ibool more_vers;
undo_no_t new_undo_no;
ut_ad(node && thr);
log_free_check();
/* Check if also the previous version of the clustered index record
should be undone in this same rollback operation */
more_vers = row_undo_mod_undo_also_prev_vers(node, &new_undo_no);
pcur = &(node->pcur);
mtr_start(&mtr);
/* Try optimistic processing of the record, keeping changes within
the index page */
err = row_undo_mod_clust_low(node, thr, &mtr, BTR_MODIFY_LEAF);
if (err != DB_SUCCESS) {
btr_pcur_commit_specify_mtr(pcur, &mtr);
/* We may have to modify tree structure: do a pessimistic
descent down the index tree */
mtr_start(&mtr);
err = row_undo_mod_clust_low(node, thr, &mtr, BTR_MODIFY_TREE);
}
btr_pcur_commit_specify_mtr(pcur, &mtr);
if (err == DB_SUCCESS && node->rec_type == TRX_UNDO_UPD_DEL_REC) {
mtr_start(&mtr);
err = row_undo_mod_remove_clust_low(node, thr, &mtr,
BTR_MODIFY_LEAF);
if (err != DB_SUCCESS) {
btr_pcur_commit_specify_mtr(pcur, &mtr);
/* We may have to modify tree structure: do a
pessimistic descent down the index tree */
mtr_start(&mtr);
err = row_undo_mod_remove_clust_low(node, thr, &mtr,
BTR_MODIFY_TREE);
}
btr_pcur_commit_specify_mtr(pcur, &mtr);
}
node->state = UNDO_NODE_FETCH_NEXT;
trx_undo_rec_release(node->trx, node->undo_no);
if (more_vers && err == DB_SUCCESS) {
/* Reserve the undo log record to the prior version after
committing &mtr: this is necessary to comply with the latching
order, as &mtr may contain the fsp latch which is lower in
the latch hierarchy than trx->undo_mutex. */
success = trx_undo_rec_reserve(node->trx, new_undo_no);
if (success) {
node->state = UNDO_NODE_PREV_VERS;
}
}
return(err);
}