mariadb/row/row0uins.c
marko 467b0a8db0 branches/innodb+: Merge revisions 2799:2835 from branches/zip:
------------------------------------------------------------------------
  r2809 | marko | 2008-10-16 09:41:13 +0300 (Thu, 16 Oct 2008) | 18 lines

  branches/zip: Skip the undo log size check on REDUNDANT and COMPACT tables.

  In ROW_FORMAT=DYNAMIC and ROW_FORMAT=COMPRESSED, column prefix indexes
  require that prefixes of externally stored columns be written to the
  undo log. This may make the undo log record bigger than the record on
  the B-tree page.  The maximum size of an undo log record is the page
  size.  That must be checked for, in dict_index_add_to_cache().

  dict_index_add_to_cache(): Skip the undo log size check for REDUNDANT
  and COMPACT tables. These tables store prefixes of externally stored
  columns locally within the clustered index record.  There are no special
  considerations for the undo log record size.

  innodb-index.test: Ensure that the check exists for ROW_FORMAT=DYNAMIC,
  but not for ROW_FORMAT=COMPACT.

  This fixes issue #99.  rb://28 approved by Sunny.
  ------------------------------------------------------------------------
  r2810 | vasil | 2008-10-16 19:57:58 +0300 (Thu, 16 Oct 2008) | 12 lines

  branches/zip:

  Fix Mantis issue#61:

  In row_undo_ins_parse_undo_rec(): if we find that a table has no indexes
  (dict_table_get_first_index() returns NULL) do not try to call
  trx_undo_rec_get_row_ref() with a NULL pointer because that would lead to
  a crash. Instead, print a warning and set node->table to NULL just like
  it is done if the .ibd file is missing.

  Approved by:    Heikki (via IM)
  ------------------------------------------------------------------------
  r2824 | marko | 2008-10-20 09:58:01 +0300 (Mon, 20 Oct 2008) | 2 lines

  branches/zip: rec_convert_dtuple_to_rec_comp(): Relax a too tight assertion.
  Spotted by Sunny.
  ------------------------------------------------------------------------
  r2825 | vasil | 2008-10-20 13:41:04 +0300 (Mon, 20 Oct 2008) | 6 lines

  branches/zip:

  Print the table name via ut_print_name() and add two spaces before InnoDB.

  Suggested by:   Marko
  ------------------------------------------------------------------------
  r2833 | marko | 2008-10-21 10:16:45 +0300 (Tue, 21 Oct 2008) | 2 lines

  branches/zip: ibuf_insert_low(): Avoid unnecessarily
  acquiring and releasing ibuf_mutex.
  ------------------------------------------------------------------------
  r2834 | marko | 2008-10-21 10:18:57 +0300 (Tue, 21 Oct 2008) | 1 line

  branches/zip: ibuf_delete_rec(): Add debug assertions suggested by Heikki.
  ------------------------------------------------------------------------
  r2835 | marko | 2008-10-21 11:04:06 +0300 (Tue, 21 Oct 2008) | 1 line

  branches/zip: ibuf_insert_low(): Simplify a comparison.
  ------------------------------------------------------------------------
2008-10-21 07:30:53 +00:00

335 lines
8.1 KiB
C

/******************************************************
Fresh insert undo
(c) 1996 Innobase Oy
Created 2/25/1997 Heikki Tuuri
*******************************************************/
#include "row0uins.h"
#ifdef UNIV_NONINL
#include "row0uins.ic"
#endif
#include "dict0dict.h"
#include "dict0boot.h"
#include "dict0crea.h"
#include "trx0undo.h"
#include "trx0roll.h"
#include "btr0btr.h"
#include "mach0data.h"
#include "row0undo.h"
#include "row0vers.h"
#include "trx0trx.h"
#include "trx0rec.h"
#include "row0row.h"
#include "row0upd.h"
#include "que0que.h"
#include "ibuf0ibuf.h"
#include "log0log.h"
/*******************************************************************
Removes a clustered index record. The pcur in node was positioned on the
record, now it is detached. */
static
ulint
row_undo_ins_remove_clust_rec(
/*==========================*/
/* out: DB_SUCCESS or DB_OUT_OF_FILE_SPACE */
undo_node_t* node) /* in: undo node */
{
btr_cur_t* btr_cur;
ibool success;
ulint err;
ulint n_tries = 0;
mtr_t mtr;
mtr_start(&mtr);
success = btr_pcur_restore_position(BTR_MODIFY_LEAF, &(node->pcur),
&mtr);
ut_a(success);
if (ut_dulint_cmp(node->table->id, DICT_INDEXES_ID) == 0) {
ut_ad(node->trx->dict_operation_lock_mode == RW_X_LATCH);
/* Drop the index tree associated with the row in
SYS_INDEXES table: */
dict_drop_index_tree(btr_pcur_get_rec(&(node->pcur)), &mtr);
mtr_commit(&mtr);
mtr_start(&mtr);
success = btr_pcur_restore_position(BTR_MODIFY_LEAF,
&(node->pcur), &mtr);
ut_a(success);
}
btr_cur = btr_pcur_get_btr_cur(&(node->pcur));
success = btr_cur_optimistic_delete(btr_cur, &mtr);
btr_pcur_commit_specify_mtr(&(node->pcur), &mtr);
if (success) {
trx_undo_rec_release(node->trx, node->undo_no);
return(DB_SUCCESS);
}
retry:
/* If did not succeed, try pessimistic descent to tree */
mtr_start(&mtr);
success = btr_pcur_restore_position(BTR_MODIFY_TREE,
&(node->pcur), &mtr);
ut_a(success);
btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
trx_is_recv(node->trx)
? RB_RECOVERY
: 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 */
if (err == DB_OUT_OF_FILE_SPACE
&& n_tries < BTR_CUR_RETRY_DELETE_N_TIMES) {
btr_pcur_commit_specify_mtr(&(node->pcur), &mtr);
n_tries++;
os_thread_sleep(BTR_CUR_RETRY_SLEEP_TIME);
goto retry;
}
btr_pcur_commit_specify_mtr(&(node->pcur), &mtr);
trx_undo_rec_release(node->trx, node->undo_no);
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;
ulint err;
mtr_t mtr;
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 (search_result) {
case ROW_NOT_FOUND:
err = DB_SUCCESS;
goto func_exit;
case ROW_FOUND:
break;
case ROW_BUFFERED:
case ROW_NOT_IN_POOL:
/* These are invalid outcomes, because the mode passed
to row_search_index_entry() did not include any of the
flags BTR_INSERT, BTR_DELETE, BTR_DELETE_MARK, or
BTR_WATCH_LEAF. */
ut_error;
}
if (mode == BTR_MODIFY_LEAF) {
err = btr_cur_optimistic_delete(btr_cur, &mtr)
? DB_SUCCESS : 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);
}
func_exit:
btr_pcur_close(&pcur);
mtr_commit(&mtr);
return(err);
}
/*******************************************************************
Removes a secondary index entry from the index if found. Tries first
optimistic, then pessimistic descent down the tree. */
static
ulint
row_undo_ins_remove_sec(
/*====================*/
/* out: DB_SUCCESS or DB_OUT_OF_FILE_SPACE */
dict_index_t* index, /* in: index */
dtuple_t* entry) /* in: index entry to insert */
{
ulint err;
ulint n_tries = 0;
/* Try first optimistic descent to the B-tree */
err = row_undo_ins_remove_sec_low(BTR_MODIFY_LEAF, index, entry);
if (err == DB_SUCCESS) {
return(err);
}
/* Try then pessimistic descent to the B-tree */
retry:
err = row_undo_ins_remove_sec_low(BTR_MODIFY_TREE, index, entry);
/* The delete operation may fail if we have little
file space left: TODO: easiest to crash the database
and restart with more file space */
if (err != DB_SUCCESS && n_tries < BTR_CUR_RETRY_DELETE_N_TIMES) {
n_tries++;
os_thread_sleep(BTR_CUR_RETRY_SLEEP_TIME);
goto retry;
}
return(err);
}
/***************************************************************
Parses the row reference and other info in a fresh insert undo record. */
static
void
row_undo_ins_parse_undo_rec(
/*========================*/
undo_node_t* node) /* in/out: row undo node */
{
dict_index_t* clust_index;
byte* ptr;
dulint 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(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(stderr);
fprintf(stderr, " InnoDB: table ");
ut_print_name(stderr, node->trx, TRUE,
node->table->name);
fprintf(stderr, " has no indexes, "
"ignoring the table\n");
node->table = NULL;
}
}
}
/***************************************************************
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. */
UNIV_INTERN
ulint
row_undo_ins(
/*=========*/
/* out: DB_SUCCESS or DB_OUT_OF_FILE_SPACE */
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. 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 {
err = row_undo_ins_remove_sec(node->index, entry);
if (err != DB_SUCCESS) {
return(err);
}
}
node->index = dict_table_get_next_index(node->index);
}
return(row_undo_ins_remove_clust_rec(node));
}