/*
* __wt_turtle_update --
* Update the turtle file.
*/
int
__wt_turtle_update(
WT_SESSION_IMPL *session, const char *key, const char *value)
{
FILE *fp;
WT_DECL_RET;
int vmajor, vminor, vpatch;
const char *version;
char *path;
fp = NULL;
path = NULL;
/*
* Create the turtle setup file: we currently re-write it from scratch
* every time.
*/
WT_RET(__wt_filename(session, WT_METADATA_TURTLE_SET, &path));
if ((fp = fopen(path, "w")) == NULL)
ret = __wt_errno();
__wt_free(session, path);
if (fp == NULL)
return (ret);
version = wiredtiger_version(&vmajor, &vminor, &vpatch);
WT_ERR_TEST((fprintf(fp,
"%s\n%s\n%s\n" "major=%d,minor=%d,patch=%d\n%s\n%s\n",
WT_METADATA_VERSION_STR, version,
WT_METADATA_VERSION, vmajor, vminor, vpatch,
key, value) < 0), __wt_errno());
ret = fclose(fp);
fp = NULL;
WT_ERR_TEST(ret == EOF, __wt_errno());
WT_ERR(
__wt_rename(session, WT_METADATA_TURTLE_SET, WT_METADATA_TURTLE));
if (0) {
err: WT_TRET(__wt_remove(session, WT_METADATA_TURTLE_SET));
}
if (fp != NULL)
WT_TRET(fclose(fp) == 0 ? 0 : __wt_errno());
return (ret);
}
/*
* __backup_all --
* Backup all objects in the database.
*/
static int
__backup_all(WT_SESSION_IMPL *session, WT_CURSOR_BACKUP *cb)
{
WT_CONFIG_ITEM cval;
WT_CURSOR *cursor;
WT_DECL_RET;
const char *key, *value;
cursor = NULL;
/*
* Open a cursor on the metadata file and copy all of the entries to
* the hot backup file.
*/
WT_ERR(__wt_metadata_cursor(session, NULL, &cursor));
while ((ret = cursor->next(cursor)) == 0) {
WT_ERR(cursor->get_key(cursor, &key));
WT_ERR(cursor->get_value(cursor, &value));
WT_ERR_TEST((fprintf(
cb->bfp, "%s\n%s\n", key, value) < 0), __wt_errno());
/*
* While reading the metadata file, check there are no "sources"
* or "types" which can't support hot backup. This checks for
* a data source that's non-standard, which can't be backed up,
* but is also sanity checking: if there's an entry backed by
* anything other than a file or lsm entry, we're confused.
*/
if ((ret = __wt_config_getones(
session, value, "type", &cval)) == 0 &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "file") &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "lsm"))
WT_ERR_MSG(session, ENOTSUP,
"hot backup is not supported for objects of "
"type %.*s", (int)cval.len, cval.str);
WT_ERR_NOTFOUND_OK(ret);
if ((ret =__wt_config_getones(
session, value, "source", &cval)) == 0 &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "file:") &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "lsm:"))
WT_ERR_MSG(session, ENOTSUP,
"hot backup is not supported for objects of "
"source %.*s", (int)cval.len, cval.str);
WT_ERR_NOTFOUND_OK(ret);
}
WT_ERR_NOTFOUND_OK(ret);
/* Build a list of the file objects that need to be copied. */
WT_WITH_DHANDLE_LOCK(session,
ret = __wt_meta_btree_apply(
session, __backup_list_all_append, NULL));
err: if (cursor != NULL)
WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_meta_turtle_read --
* Read the turtle file.
*/
int
__wt_meta_turtle_read(
WT_SESSION_IMPL *session, const char *key, const char **valuep)
{
FILE *fp;
WT_DECL_RET;
const char *path;
char *p, line[1024];
fp = NULL;
path = NULL;
/* Open the turtle file. */
WT_RET(__wt_filename(session, WT_METADATA_TURTLE, &path));
WT_ERR_TEST((fp = fopen(path, "r")) == NULL, WT_NOTFOUND);
/* Search for the key. */
ret = WT_NOTFOUND;
while (fgets(line, sizeof(line), fp) != NULL) {
if ((p = strchr(line, '\n')) == NULL)
goto format;
*p = '\0';
if (strcmp(key, line) == 0)
ret = 0;
/* Key matched: read the subsequent line for the value. */
if (fgets(line, sizeof(line), fp) == NULL)
goto format;
if ((p = strchr(line, '\n')) == NULL)
goto format;
*p = '\0';
if (ret == 0)
break;
}
/* Check for an I/O error. */
if (ferror(fp))
WT_ERR(__wt_errno());
WT_ERR(ret);
/* Successful: copy the value for the caller. */
WT_ERR(__wt_strdup(session, line, valuep));
if (0) {
format: return (__wt_illegal_value(session, WT_METADATA_TURTLE));
}
err: if (fp != NULL)
WT_TRET(fclose(fp));
__wt_free(session, path);
return (ret);
}
/*
* __backup_file_create --
* Create the meta-data backup file.
*/
static int
__backup_file_create(WT_SESSION_IMPL *session, WT_CURSOR_BACKUP *cb)
{
WT_DECL_RET;
char *path;
/* Open the hot backup file. */
WT_RET(__wt_filename(session, WT_METADATA_BACKUP, &path));
WT_ERR_TEST((cb->bfp = fopen(path, "w")) == NULL, __wt_errno());
err: __wt_free(session, path);
return (ret);
}
/*
* __wt_rwlock_alloc --
* Allocate and initialize a read/write lock.
*/
int
__wt_rwlock_alloc(
WT_SESSION_IMPL *session, const char *name, WT_RWLOCK **rwlockp)
{
WT_DECL_RET;
WT_RWLOCK *rwlock;
WT_RET(__wt_calloc(session, 1, sizeof(WT_RWLOCK), &rwlock));
WT_ERR_TEST(pthread_rwlock_init(&rwlock->rwlock, NULL), WT_ERROR);
rwlock->name = name;
*rwlockp = rwlock;
WT_VERBOSE_ERR(session, mutex,
"rwlock: alloc %s (%p)", rwlock->name, rwlock);
if (0) {
err: __wt_free(session, rwlock);
}
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, bool truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
bool newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_prev);
WT_STAT_FAST_DATA_INCR(session, cursor_prev);
flags = WT_READ_PREV | WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, false));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_PREV))
__wt_btcur_iterate_setup(cbt);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the previous page, until we reach the start
* of the file.
*/
for (newpage = false;; newpage = true) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (newpage && page != NULL && page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
newpage = true;
}
if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_prev(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
}
/*
* If we saw a lot of deleted records on this page, or we went
* all the way through a page and only saw deleted records, try
* to evict the page when we release it. Otherwise repeatedly
* deleting from the beginning of a tree can have quadratic
* performance. Take care not to force eviction of pages that
* are genuinely empty, in new trees.
*/
if (page != NULL &&
(cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD ||
(newpage && cbt->page_deleted_count > 0)))
__wt_page_evict_soon(page);
cbt->page_deleted_count = 0;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, bool truncating)
{
WT_CURSOR *cursor;
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
bool newpage;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_CONN_INCR(session, cursor_next);
WT_STAT_DATA_INCR(session, cursor_next);
F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET);
WT_RET(__cursor_func_init(cbt, false));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
flags = WT_READ_SKIP_INTL; /* tree walk flags */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
for (newpage = false;; newpage = true) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* Column-store pages may have appended entries. Handle
* it separately from the usual cursor code, it's in a
* simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*
* If we saw a lot of deleted records on this page, or we went
* all the way through a page and only saw deleted records, try
* to evict the page when we release it. Otherwise repeatedly
* deleting from the beginning of a tree can have quadratic
* performance. Take care not to force eviction of pages that
* are genuinely empty, in new trees.
*/
if (page != NULL &&
(cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD ||
(newpage && cbt->page_deleted_count > 0)))
__wt_page_evict_soon(session, cbt->ref);
cbt->page_deleted_count = 0;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
#ifdef HAVE_DIAGNOSTIC
if (ret == 0)
WT_ERR(__wt_cursor_key_order_check(session, cbt, true));
#endif
if (ret == 0)
F_SET(cursor, WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT);
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*将btree cursor移动到下一个记录*/
int __wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
/*btree 扫描标示*/
flags = WT_READ_SKIP_INTL;
if (truncating)
LF_SET(WT_READ_TRUNCATE);
/*激活一个btree cursor*/
WT_RET(__cursor_func_init(cbt, 0));
/*初始化cursor*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*对btree的扫描*/
for (;;){
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
/*column store append方式，在insert header上做扫描*/
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)){
switch (page->type){
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
/*清除掉column store的标记*/
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
}
else if (page != NULL){
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/*找到对应的记录了，直接返回*/
if (ret != WT_NOTFOUND)
break;
/*假如是column store方式，检查是否要扫描insert header list*/
if (page->type != WT_PAGE_ROW_LEAF && (cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*删除的记录太多，对page进行重组，增大page的填充因子*/
if (page != NULL && (cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD || (newpage && cbt->page_deleted_count > 0))){
__wt_page_evict_soon(page);
}
cbt->page_deleted_count = 0;
/*btree cursor跳转到下一个page上*/
WT_ERR(__wt_tree_walk(session, &cbt->ref, NULL, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err:
if (ret != 0) /*失败了，恢复cursor的状态*/
WT_TRET(__cursor_reset(cbt));
return ret;
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_prev);
WT_STAT_FAST_DATA_INCR(session, cursor_prev);
flags = WT_READ_PREV | WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, 0));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_PREV))
__wt_btcur_iterate_setup(cbt, 0);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the previous page, until we reach the start
* of the file.
*/
page = cbt->ref == NULL ? NULL : cbt->ref->page;
for (newpage = 0;; newpage = 1) {
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
newpage = 1;
}
if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_prev(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
}
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
page = cbt->ref->page;
WT_ASSERT(session,
page->type != WT_PAGE_COL_INT &&
page->type != WT_PAGE_ROW_INT);
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
err: if (ret != 0)
WT_TRET(__cursor_error_resolve(cbt));
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, int discard)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_DSTAT_INCR(session, cursor_prev);
flags = WT_TREE_SKIP_INTL | WT_TREE_PREV; /* Tree walk flags. */
if (discard)
LF_SET(WT_TREE_DISCARD);
retry: WT_RET(__cursor_func_init(cbt, 0));
__cursor_position_clear(cbt);
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_PREV))
__wt_btcur_iterate_setup(cbt, 0);
/*
* If this is a modification, we're about to read information from the
* page, save the write generation.
*/
page = cbt->page;
if (discard && page != NULL) {
WT_ERR(__wt_page_modify_init(session, page));
WT_ORDERED_READ(cbt->write_gen, page->modify->write_gen);
}
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the previous page, until we reach the start
* of the file.
*/
for (newpage = 0;; newpage = 1) {
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
newpage = 1;
}
if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_prev(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
}
cbt->page = NULL;
WT_ERR(__wt_tree_walk(session, &page, flags));
WT_ERR_TEST(page == NULL, WT_NOTFOUND);
WT_ASSERT(session,
page->type != WT_PAGE_COL_INT &&
page->type != WT_PAGE_ROW_INT);
cbt->page = page;
/* Initialize the page's modification information */
if (discard) {
WT_ERR(__wt_page_modify_init(session, page));
WT_ORDERED_READ(
cbt->write_gen, page->modify->write_gen);
}
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
err: if (ret == WT_RESTART)
goto retry;
WT_TRET(__cursor_func_resolve(cbt, ret));
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_BSTAT_INCR(session, cursor_read_prev);
__cursor_func_init(cbt, 0);
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_PREV))
__wt_btcur_iterate_setup(cbt, 0);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the previous page, until we reach the start
* of the file.
*/
for (newpage = 0;; newpage = 1) {
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (cbt->page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
newpage = 1;
}
if (cbt->page != NULL) {
switch (cbt->page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_prev(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
}
do {
WT_ERR(__wt_tree_np(session, &cbt->page, 0, 0));
WT_ERR_TEST(cbt->page == NULL, WT_NOTFOUND);
} while (
cbt->page->type == WT_PAGE_COL_INT ||
cbt->page->type == WT_PAGE_ROW_INT);
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (cbt->page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(cbt->page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
err: __cursor_func_resolve(cbt, ret);
return (ret);
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int skipped, newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
flags = WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, 0));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
for (skipped = newpage = 0;; skipped = 0, newpage = 1) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(
cbt, newpage, &skipped);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage, &skipped);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage, &skipped);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*
* If we scanned all the way through a page and only saw
* deleted records, try to evict the page as we release it.
* Otherwise repeatedly deleting from the beginning of a tree
* can have quadratic performance.
*/
if (newpage && skipped)
page->read_gen = WT_READGEN_OLDEST;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
