/*
* __curds_insert --
* WT_CURSOR.insert method for the data-source cursor type.
*/
static int
__curds_insert(WT_CURSOR *cursor)
{
WT_CURSOR *source;
WT_DECL_RET;
WT_SESSION_IMPL *session;
source = ((WT_CURSOR_DATA_SOURCE *)cursor)->source;
CURSOR_UPDATE_API_CALL(cursor, session, insert, NULL);
WT_ERR(__curds_txn_enter(session));
WT_STAT_FAST_CONN_INCR(session, cursor_insert);
WT_STAT_FAST_DATA_INCR(session, cursor_insert);
WT_STAT_FAST_DATA_INCRV(session,
cursor_insert_bytes, cursor->key.size + cursor->value.size);
if (!F_ISSET(cursor, WT_CURSTD_APPEND))
WT_ERR(__curds_key_set(cursor));
WT_ERR(__curds_value_set(cursor));
ret = __curds_cursor_resolve(cursor, source->insert(source));
err: __curds_txn_leave(session);
CURSOR_UPDATE_API_END(session, ret);
return (ret);
}
/*
* __curds_remove --
* WT_CURSOR.remove method for the data-source cursor type.
*/
static int
__curds_remove(WT_CURSOR *cursor)
{
WT_CURSOR *source;
WT_DECL_RET;
WT_SESSION_IMPL *session;
source = ((WT_CURSOR_DATA_SOURCE *)cursor)->source;
CURSOR_UPDATE_API_CALL(cursor, session, remove, NULL);
WT_STAT_FAST_CONN_INCR(session, cursor_remove);
WT_STAT_FAST_DATA_INCR(session, cursor_remove);
WT_STAT_FAST_DATA_INCRV(session, cursor_remove_bytes, cursor->key.size);
WT_ERR(__curds_txn_enter(session));
WT_ERR(__curds_key_set(cursor));
ret = __curds_cursor_resolve(cursor, source->remove(source));
err: __curds_txn_leave(session);
CURSOR_UPDATE_API_END(session, ret);
return (ret);
}
/*
* __wt_btcur_next_random --
* Move to a random record in the tree.
*/
int
__wt_btcur_next_random(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = cbt->btree;
/*
* Only supports row-store: applications can trivially select a random
* value from a column-store, if there were any reason to do so.
*/
if (btree->type != BTREE_ROW)
WT_RET(ENOTSUP);
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
WT_RET(__cursor_func_init(cbt, 1));
WT_ERR(__wt_row_random(session, cbt));
if (__cursor_valid(cbt, &upd))
WT_ERR(__wt_kv_return(session, cbt, upd));
else
WT_ERR(__wt_btcur_search_near(cbt, 0));
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __curlog_search --
* WT_CURSOR.search method for the log cursor type.
*/
static int
__curlog_search(WT_CURSOR *cursor)
{
WT_CURSOR_LOG *cl;
WT_DECL_RET;
WT_LSN key;
WT_SESSION_IMPL *session;
uint32_t counter;
cl = (WT_CURSOR_LOG *)cursor;
CURSOR_API_CALL(cursor, session, search, NULL);
/*
* !!! We are ignoring the counter and only searching based on the LSN.
*/
WT_ERR(__wt_cursor_get_key((WT_CURSOR *)cl,
&key.file, &key.offset, &counter));
WT_ERR(__wt_log_scan(session, &key, WT_LOGSCAN_ONE,
__curlog_logrec, cl));
WT_ERR(__curlog_kv(session, cursor));
WT_STAT_FAST_CONN_INCR(session, cursor_search);
WT_STAT_FAST_DATA_INCR(session, cursor_search);
err: API_END_RET(session, ret);
}
/*
* __ovfl_read --
* Read an overflow item from the disk.
*/
static int
__ovfl_read(WT_SESSION_IMPL *session,
const uint8_t *addr, size_t addr_size, WT_ITEM *store)
{
WT_BTREE *btree;
const WT_PAGE_HEADER *dsk;
btree = S2BT(session);
/*
* Read the overflow item from the block manager, then reference the
* start of the data and set the data's length.
*
* Overflow reads are synchronous. That may bite me at some point, but
* WiredTiger supports large page sizes, overflow items should be rare.
*/
WT_RET(__wt_bt_read(session, store, addr, addr_size));
dsk = store->data;
store->data = WT_PAGE_HEADER_BYTE(btree, dsk);
store->size = dsk->u.datalen;
WT_STAT_FAST_DATA_INCR(session, cache_read_overflow);
return (0);
}
/*
* __hazard_exclusive --
* Request exclusive access to a page.
*/
static int
__hazard_exclusive(WT_SESSION_IMPL *session, WT_REF *ref, int top)
{
/*
* Make sure there is space to track exclusive access so we can unlock
* to clean up.
*/
WT_RET(__wt_realloc_def(session, &session->excl_allocated,
session->excl_next + 1, &session->excl));
/*
* Hazard pointers are acquired down the tree, which means we can't
* deadlock.
*
* Request exclusive access to the page. The top-level page should
* already be in the locked state, lock child pages in memory.
* If another thread already has this page, give up.
*/
if (!top && !WT_ATOMIC_CAS(ref->state, WT_REF_MEM, WT_REF_LOCKED))
return (EBUSY); /* We couldn't change the state. */
WT_ASSERT(session, ref->state == WT_REF_LOCKED);
session->excl[session->excl_next++] = ref;
/* Check for a matching hazard pointer. */
if (__wt_page_hazard_check(session, ref->page) == NULL)
return (0);
WT_STAT_FAST_DATA_INCR(session, cache_eviction_hazard);
WT_STAT_FAST_CONN_INCR(session, cache_eviction_hazard);
WT_VERBOSE_RET(
session, evict, "page %p hazard request failed", ref->page);
return (EBUSY);
}
/*
* __curbulk_insert_fix_bitmap --
* Fixed-length column-store bulk cursor insert for bitmaps.
*/
static int
__curbulk_insert_fix_bitmap(WT_CURSOR *cursor)
{
WT_BTREE *btree;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cbulk = (WT_CURSOR_BULK *)cursor;
btree = cbulk->cbt.btree;
/*
* Bulk cursor inserts are updates, but don't need auto-commit
* transactions because they are single-threaded and not visible
* until the bulk cursor is closed.
*/
CURSOR_API_CALL(cursor, session, insert, btree);
WT_STAT_FAST_DATA_INCR(session, cursor_insert_bulk);
WT_CURSOR_CHECKVALUE(cursor);
/* Insert the current record. */
ret = __wt_bulk_insert_fix_bitmap(session, cbulk);
err: API_END_RET(session, ret);
}
/*
* __wt_btcur_next_random --
* Move to a random record in the tree.
*/
int
__wt_btcur_next_random(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = cbt->btree;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
WT_RET(__cursor_func_init(cbt, 1));
/*
* Only supports row-store: applications can trivially select a random
* value from a column-store, if there were any reason to do so.
*/
WT_ERR(btree->type == BTREE_ROW ?
__wt_row_random(session, cbt) : ENOTSUP);
ret = cbt->compare == 0 ?
__wt_kv_return(session, cbt) : WT_NOTFOUND;
err: if (ret != 0)
WT_TRET(__cursor_error_resolve(cbt));
return (ret);
}
/*
* __curlog_next --
* WT_CURSOR.next method for the step log cursor type.
*/
static int
__curlog_next(WT_CURSOR *cursor)
{
WT_CURSOR_LOG *cl;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cl = (WT_CURSOR_LOG *)cursor;
CURSOR_API_CALL(cursor, session, next, NULL);
/*
* If we don't have a record, or went to the end of the record we
* have, or we are in the zero-fill portion of the record, get a
* new one.
*/
if (cl->stepp == NULL || cl->stepp >= cl->stepp_end || !*cl->stepp) {
cl->txnid = 0;
WT_ERR(__wt_log_scan(session, cl->next_lsn, WT_LOGSCAN_ONE,
__curlog_logrec, cl));
}
WT_ASSERT(session, cl->logrec->data != NULL);
WT_ERR(__curlog_kv(session, cursor));
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
err: API_END_RET(session, ret);
}
/*
* __wt_btcur_reset --
* Invalidate the cursor position.
*/
int
__wt_btcur_reset(WT_CURSOR_BTREE *cbt)
{
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_reset);
WT_STAT_FAST_DATA_INCR(session, cursor_reset);
return (__cursor_reset(cbt));
}
/*
* __cursor_truncate_fix --
* Discard a cursor range from fixed-width column-store tree.
*/
static int
__cursor_truncate_fix(WT_SESSION_IMPL *session,
WT_CURSOR_BTREE *start, WT_CURSOR_BTREE *stop,
int (*rmfunc)(WT_SESSION_IMPL *, WT_CURSOR_BTREE *, int))
{
WT_DECL_RET;
uint8_t *value;
/*
* Handle fixed-length column-store objects separately: for row-store
* and variable-length column-store objects we have "deleted" values
* and so returned objects actually exist: fixed-length column-store
* objects are filled-in if they don't exist, that is, if you create
* record 37, records 1-36 magically appear. Those records can't be
* deleted, which means we have to ignore already "deleted" records.
*
* First, call the standard cursor remove method to do a full search and
* re-position the cursor because we don't have a saved copy of the
* page's write generation information, which we need to remove records.
* Once that's done, we can delete records without a full search, unless
* we encounter a restart error because the page was modified by some
* other thread of control; in that case, repeat the full search to
* refresh the page's modification information.
*/
retry: WT_RET(__wt_btcur_remove(start));
/*
* Reset ret each time through so that we don't loop forever in
* the cursor equals case.
*/
for (ret = 0;;) {
if (stop != NULL && __cursor_equals(start, stop))
break;
if ((ret = __wt_btcur_next(start, 1)) != 0)
break;
start->compare = 0; /* Exact match */
value = (uint8_t *)start->iface.value.data;
if (*value != 0 &&
(ret = rmfunc(session, start, 1)) != 0)
break;
}
if (ret == WT_RESTART) {
WT_STAT_FAST_CONN_INCR(session, cursor_restart);
WT_STAT_FAST_DATA_INCR(session, cursor_restart);
goto retry;
}
WT_RET_NOTFOUND_OK(ret);
return (0);
}
/*
* __cursor_truncate --
* Discard a cursor range from row-store or variable-width column-store
* tree.
*/
static int
__cursor_truncate(WT_SESSION_IMPL *session,
WT_CURSOR_BTREE *start, WT_CURSOR_BTREE *stop,
int (*rmfunc)(WT_SESSION_IMPL *, WT_CURSOR_BTREE *, int))
{
WT_DECL_RET;
/*
* First, call the standard cursor remove method to do a full search and
* re-position the cursor because we don't have a saved copy of the
* page's write generation information, which we need to remove records.
* Once that's done, we can delete records without a full search, unless
* we encounter a restart error because the page was modified by some
* other thread of control; in that case, repeat the full search to
* refresh the page's modification information.
*
* If this is a row-store, we delete leaf pages having no overflow items
* without reading them; for that to work, we have to ensure we read the
* page referenced by the ending cursor, since we may be deleting only a
* partial page at the end of the truncation. Our caller already fully
* instantiated the end cursor, so we know that page is pinned in memory
* and we can proceed without concern.
*/
retry: WT_RET(__wt_btcur_remove(start));
/*
* Reset ret each time through so that we don't loop forever in
* the cursor equals case.
*/
for (ret = 0;;) {
if (stop != NULL && __cursor_equals(start, stop))
break;
if ((ret = __wt_btcur_next(start, 1)) != 0)
break;
start->compare = 0; /* Exact match */
if ((ret = rmfunc(session, start, 1)) != 0)
break;
}
if (ret == WT_RESTART) {
WT_STAT_FAST_CONN_INCR(session, cursor_restart);
WT_STAT_FAST_DATA_INCR(session, cursor_restart);
goto retry;
}
WT_RET_NOTFOUND_OK(ret);
return (0);
}
/*
* __curbulk_insert_fix --
* Fixed-length column-store bulk cursor insert.
*/
static int
__curbulk_insert_fix(WT_CURSOR *cursor)
{
WT_BTREE *btree;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
WT_SESSION_IMPL *session;
uint64_t recno;
cbulk = (WT_CURSOR_BULK *)cursor;
btree = cbulk->cbt.btree;
/*
* Bulk cursor inserts are updates, but don't need auto-commit
* transactions because they are single-threaded and not visible
* until the bulk cursor is closed.
*/
CURSOR_API_CALL(cursor, session, insert, btree);
WT_STAT_FAST_DATA_INCR(session, cursor_insert_bulk);
/*
* If the "append" flag was configured, the application doesn't have to
* supply a key, else require a key.
*/
if (F_ISSET(cursor, WT_CURSTD_APPEND))
recno = cbulk->recno + 1;
else {
WT_CURSOR_CHECKKEY(cursor);
if ((recno = cursor->recno) <= cbulk->recno)
WT_ERR(__bulk_col_keycmp_err(cbulk));
}
WT_CURSOR_CHECKVALUE(cursor);
/*
* Insert any skipped records as deleted records, update the current
* record count.
*/
for (; recno != cbulk->recno + 1; ++cbulk->recno)
WT_ERR(__wt_bulk_insert_fix(session, cbulk, true));
cbulk->recno = recno;
/* Insert the current record. */
ret = __wt_bulk_insert_fix(session, cbulk, false);
err: API_END_RET(session, ret);
}
/*
* __curbulk_insert_row --
* Row-store bulk cursor insert, with key-sort checks.
*/
static int
__curbulk_insert_row(WT_CURSOR *cursor)
{
WT_BTREE *btree;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
WT_SESSION_IMPL *session;
int cmp;
cbulk = (WT_CURSOR_BULK *)cursor;
btree = cbulk->cbt.btree;
/*
* Bulk cursor inserts are updates, but don't need auto-commit
* transactions because they are single-threaded and not visible
* until the bulk cursor is closed.
*/
CURSOR_API_CALL(cursor, session, insert, btree);
WT_STAT_FAST_DATA_INCR(session, cursor_insert_bulk);
WT_CURSOR_CHECKKEY(cursor);
WT_CURSOR_CHECKVALUE(cursor);
/*
* If this isn't the first key inserted, compare it against the last key
* to ensure the application doesn't accidentally corrupt the table.
*/
if (!cbulk->first_insert) {
WT_ERR(__wt_compare(session,
btree->collator, &cursor->key, &cbulk->last, &cmp));
if (cmp <= 0)
WT_ERR(__bulk_row_keycmp_err(cbulk));
} else
cbulk->first_insert = false;
/* Save a copy of the key for the next comparison. */
WT_ERR(__wt_buf_set(session,
&cbulk->last, cursor->key.data, cursor->key.size));
ret = __wt_bulk_insert_row(session, cbulk);
err: API_END_RET(session, ret);
}
/*
* __truncate_dsrc --
* WT_SESSION::truncate for a data-source without a truncate operation.
*/
static int
__truncate_dsrc(WT_SESSION_IMPL *session, const char *uri)
{
WT_CURSOR *cursor;
WT_DECL_RET;
const char *cfg[2];
/* Open a cursor and traverse the object, removing every entry. */
cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
cfg[1] = NULL;
WT_RET(__wt_open_cursor(session, uri, NULL, cfg, &cursor));
while ((ret = cursor->next(cursor)) == 0)
WT_ERR(cursor->remove(cursor));
WT_ERR_NOTFOUND_OK(ret);
WT_STAT_FAST_DATA_INCR(session, cursor_truncate);
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_btcur_search --
* Search for a matching record in the tree.
*/
int
__wt_btcur_search(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_search);
WT_STAT_FAST_DATA_INCR(session, cursor_search);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_func_init(cbt, 1));
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 0) :
__cursor_col_search(session, cbt));
if (cbt->compare == 0 && __cursor_valid(cbt, &upd))
ret = __wt_kv_return(session, cbt, upd);
else if (__cursor_fix_implicit(btree, cbt)) {
/*
* Creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
*/
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
} else
ret = WT_NOTFOUND;
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __curds_reset --
* WT_CURSOR.reset method for the data-source cursor type.
*/
static int
__curds_reset(WT_CURSOR *cursor)
{
WT_CURSOR *source;
WT_DECL_RET;
WT_SESSION_IMPL *session;
source = ((WT_CURSOR_DATA_SOURCE *)cursor)->source;
CURSOR_API_CALL(cursor, session, reset, NULL);
WT_STAT_FAST_CONN_INCR(session, cursor_reset);
WT_STAT_FAST_DATA_INCR(session, cursor_reset);
WT_ERR(source->reset(source));
F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET);
err: API_END_RET(session, ret);
}
/*
* __curds_prev --
* WT_CURSOR.prev method for the data-source cursor type.
*/
static int
__curds_prev(WT_CURSOR *cursor)
{
WT_CURSOR *source;
WT_DECL_RET;
WT_SESSION_IMPL *session;
source = ((WT_CURSOR_DATA_SOURCE *)cursor)->source;
CURSOR_API_CALL(cursor, session, prev, NULL);
WT_STAT_FAST_CONN_INCR(session, cursor_prev);
WT_STAT_FAST_DATA_INCR(session, cursor_prev);
WT_ERR(__curds_txn_enter(session));
F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET);
ret = __curds_cursor_resolve(cursor, source->prev(source));
err: __curds_txn_leave(session);
API_END_RET(session, ret);
}
/*
* __wt_btcur_update_check --
* Check whether an update would conflict.
*
* This can be used to replace WT_CURSOR::insert or WT_CURSOR::update, so
* they only check for conflicts without updating the tree. It is used to
* maintain snapshot isolation for transactions that span multiple chunks
* in an LSM tree.
*/
int
__wt_btcur_update_check(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cursor = &cbt->iface;
btree = cbt->btree;
session = (WT_SESSION_IMPL *)cursor->session;
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, NULL, 1));
/*
* Just check for conflicts.
*/
ret = __curfile_update_check(cbt);
break;
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART) {
WT_STAT_FAST_CONN_INCR(session, cursor_restart);
WT_STAT_FAST_DATA_INCR(session, cursor_restart);
goto retry;
}
WT_TRET(__curfile_leave(cbt));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __curds_search --
* WT_CURSOR.search method for the data-source cursor type.
*/
static int
__curds_search(WT_CURSOR *cursor)
{
WT_CURSOR *source;
WT_DECL_RET;
WT_SESSION_IMPL *session;
source = ((WT_CURSOR_DATA_SOURCE *)cursor)->source;
CURSOR_API_CALL(cursor, session, search, NULL);
WT_STAT_FAST_CONN_INCR(session, cursor_search);
WT_STAT_FAST_DATA_INCR(session, cursor_search);
WT_ERR(__curds_txn_enter(session));
WT_ERR(__curds_key_set(cursor));
ret = __curds_cursor_resolve(cursor, source->search(source));
err: __curds_txn_leave(session);
API_END_RET(session, ret);
}
/*
* __truncate_table --
* WT_SESSION::truncate for a table.
*/
static int
__truncate_table(WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_DECL_RET;
WT_TABLE *table;
u_int i;
WT_RET(__wt_schema_get_table(session, uri, strlen(uri), false, &table));
WT_STAT_FAST_DATA_INCR(session, cursor_truncate);
/* Truncate the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++)
WT_ERR(__wt_schema_truncate(
session, table->cgroups[i]->source, cfg));
/* Truncate the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++)
WT_ERR(__wt_schema_truncate(
session, table->indices[i]->source, cfg));
err: __wt_schema_release_table(session, table);
return (ret);
}
/*
* __wt_curfile_create --
* Open a cursor for a given btree handle.
*/
int
__wt_curfile_create(WT_SESSION_IMPL *session,
WT_CURSOR *owner, const char *cfg[], int bulk, int bitmap,
WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__wt_cursor_set_value, /* set-value */
__curfile_compare, /* compare */
__curfile_equals, /* equals */
__curfile_next, /* next */
__curfile_prev, /* prev */
__curfile_reset, /* reset */
__curfile_search, /* search */
__curfile_search_near, /* search-near */
__curfile_insert, /* insert */
__curfile_update, /* update */
__curfile_remove, /* remove */
__wt_cursor_reconfigure, /* reconfigure */
__curfile_close); /* close */
WT_BTREE *btree;
WT_CONFIG_ITEM cval;
WT_CURSOR *cursor;
WT_CURSOR_BTREE *cbt;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
size_t csize;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_BTREE, iface) == 0);
cbt = NULL;
btree = S2BT(session);
WT_ASSERT(session, btree != NULL);
csize = bulk ? sizeof(WT_CURSOR_BULK) : sizeof(WT_CURSOR_BTREE);
WT_RET(__wt_calloc(session, 1, csize, &cbt));
cursor = &cbt->iface;
*cursor = iface;
cursor->session = &session->iface;
cursor->internal_uri = btree->dhandle->name;
cursor->key_format = btree->key_format;
cursor->value_format = btree->value_format;
cbt->btree = btree;
if (bulk) {
F_SET(cursor, WT_CURSTD_BULK);
cbulk = (WT_CURSOR_BULK *)cbt;
/* Optionally skip the validation of each bulk-loaded key. */
WT_ERR(__wt_config_gets_def(
session, cfg, "skip_sort_check", 0, &cval));
WT_ERR(__wt_curbulk_init(
session, cbulk, bitmap, cval.val == 0 ? 0 : 1));
}
/*
* random_retrieval
* Random retrieval cursors only support next, reset and close.
*/
WT_ERR(__wt_config_gets_def(session, cfg, "next_random", 0, &cval));
if (cval.val != 0) {
__wt_cursor_set_notsup(cursor);
cursor->next = __curfile_next_random;
cursor->reset = __curfile_reset;
}
/* Underlying btree initialization. */
__wt_btcur_open(cbt);
/* __wt_cursor_init is last so we don't have to clean up on error. */
WT_ERR(__wt_cursor_init(
cursor, cursor->internal_uri, owner, cfg, cursorp));
WT_STAT_FAST_CONN_INCR(session, cursor_create);
WT_STAT_FAST_DATA_INCR(session, cursor_create);
if (0) {
err: __wt_free(session, cbt);
}
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_bt_read --
* Read a cookie referenced block into a buffer.
*/
int
__wt_bt_read(WT_SESSION_IMPL *session,
WT_ITEM *buf, const uint8_t *addr, size_t addr_size)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
const WT_PAGE_HEADER *dsk;
size_t result_len;
btree = S2BT(session);
bm = btree->bm;
/*
* If anticipating a compressed block, read into a scratch buffer and
* decompress into the caller's buffer. Else, read directly into the
* caller's buffer.
*/
if (btree->compressor == NULL) {
WT_RET(bm->read(bm, session, buf, addr, addr_size));
dsk = buf->data;
} else {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->read(bm, session, tmp, addr, addr_size));
dsk = tmp->data;
}
/*
* If the block is compressed, copy the skipped bytes of the original
* image into place, then decompress.
*/
if (F_ISSET(dsk, WT_PAGE_COMPRESSED)) {
if (btree->compressor == NULL ||
btree->compressor->decompress == NULL)
WT_ERR_MSG(session, WT_ERROR,
"read compressed block where no compression engine "
"configured");
/*
* We're allocating the exact number of bytes we're expecting
* from decompression.
*/
WT_ERR(__wt_buf_initsize(session, buf, dsk->mem_size));
/*
* Note the source length is NOT the number of compressed bytes,
* it's the length of the block we just read (minus the skipped
* bytes). We don't store the number of compressed bytes: some
* compression engines need that length stored externally, they
* don't have markers in the stream to signal the end of the
* compressed bytes. Those engines must store the compressed
* byte length somehow, see the snappy compression extension for
* an example.
*/
memcpy(buf->mem, tmp->data, WT_BLOCK_COMPRESS_SKIP);
ret = btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)tmp->data + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->mem_size - WT_BLOCK_COMPRESS_SKIP, &result_len);
/*
* If checksums were turned off because we're depending on the
* decompression to fail on any corrupted data, we'll end up
* here after corruption happens. If we're salvaging the file,
* it's OK, otherwise it's really, really bad.
*/
if (ret != 0 ||
result_len != dsk->mem_size - WT_BLOCK_COMPRESS_SKIP)
WT_ERR(
F_ISSET(btree, WT_BTREE_VERIFY) ||
F_ISSET(session, WT_SESSION_SALVAGE_CORRUPT_OK) ?
WT_ERROR :
__wt_illegal_value(session, btree->dhandle->name));
} else
if (btree->compressor == NULL)
buf->size = dsk->mem_size;
else
/*
* We guessed wrong: there was a compressor, but this
* block was not compressed, and now the page is in the
* wrong buffer and the buffer may be of the wrong size.
* This should be rare, but happens with small blocks
* that aren't worth compressing.
*/
WT_ERR(__wt_buf_set(
session, buf, tmp->data, dsk->mem_size));
/* If the handle is a verify handle, verify the physical page. */
if (F_ISSET(btree, WT_BTREE_VERIFY)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->addr_string(bm, session, tmp, addr, addr_size));
WT_ERR(__wt_verify_dsk(session, (const char *)tmp->data, buf));
}
WT_STAT_FAST_CONN_INCR(session, cache_read);
WT_STAT_FAST_DATA_INCR(session, cache_read);
if (F_ISSET(dsk, WT_PAGE_COMPRESSED))
WT_STAT_FAST_DATA_INCR(session, compress_read);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_read, dsk->mem_size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_read, dsk->mem_size);
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_bt_write --
* Write a buffer into a block, returning the block's addr/size and
* checksum.
*/
int
__wt_bt_write(WT_SESSION_IMPL *session, WT_ITEM *buf,
uint8_t *addr, size_t *addr_sizep, bool checkpoint, bool compressed)
{
WT_BM *bm;
WT_BTREE *btree;
WT_ITEM *ip;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t dst_len, len, result_len, size, src_len;
int compression_failed; /* Extension API, so not a bool. */
uint8_t *dst, *src;
bool data_cksum;
btree = S2BT(session);
bm = btree->bm;
/* Checkpoint calls are different than standard calls. */
WT_ASSERT(session,
(!checkpoint && addr != NULL && addr_sizep != NULL) ||
(checkpoint && addr == NULL && addr_sizep == NULL));
#ifdef HAVE_DIAGNOSTIC
/*
* We're passed a table's disk image. Decompress if necessary and
* verify the image. Always check the in-memory length for accuracy.
*/
dsk = buf->mem;
if (compressed) {
WT_ERR(__wt_scr_alloc(session, dsk->mem_size, &tmp));
memcpy(tmp->mem, buf->data, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)buf->data + WT_BLOCK_COMPRESS_SKIP,
buf->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)tmp->data + WT_BLOCK_COMPRESS_SKIP,
tmp->memsize - WT_BLOCK_COMPRESS_SKIP,
&result_len));
WT_ASSERT(session,
dsk->mem_size == result_len + WT_BLOCK_COMPRESS_SKIP);
tmp->size = (uint32_t)result_len + WT_BLOCK_COMPRESS_SKIP;
ip = tmp;
} else {
WT_ASSERT(session, dsk->mem_size == buf->size);
ip = buf;
}
WT_ERR(__wt_verify_dsk(session, "[write-check]", ip));
__wt_scr_free(session, &tmp);
#endif
/*
* Optionally stream-compress the data, but don't compress blocks that
* are already as small as they're going to get.
*/
if (btree->compressor == NULL ||
btree->compressor->compress == NULL || compressed)
ip = buf;
else if (buf->size <= btree->allocsize) {
ip = buf;
WT_STAT_FAST_DATA_INCR(session, compress_write_too_small);
} else {
/* Skip the header bytes of the source data. */
src = (uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP;
src_len = buf->size - WT_BLOCK_COMPRESS_SKIP;
/*
* Compute the size needed for the destination buffer. We only
* allocate enough memory for a copy of the original by default,
* if any compressed version is bigger than the original, we
* won't use it. However, some compression engines (snappy is
* one example), may need more memory because they don't stop
* just because there's no more memory into which to compress.
*/
if (btree->compressor->pre_size == NULL)
len = src_len;
else
WT_ERR(btree->compressor->pre_size(btree->compressor,
&session->iface, src, src_len, &len));
size = len + WT_BLOCK_COMPRESS_SKIP;
WT_ERR(bm->write_size(bm, session, &size));
WT_ERR(__wt_scr_alloc(session, size, &tmp));
/* Skip the header bytes of the destination data. */
dst = (uint8_t *)tmp->mem + WT_BLOCK_COMPRESS_SKIP;
dst_len = len;
compression_failed = 0;
WT_ERR(btree->compressor->compress(btree->compressor,
&session->iface,
src, src_len,
dst, dst_len,
&result_len, &compression_failed));
result_len += WT_BLOCK_COMPRESS_SKIP;
/*
* If compression fails, or doesn't gain us at least one unit of
* allocation, fallback to the original version. This isn't
* unexpected: if compression doesn't work for some chunk of
* data for some reason (noting likely additional format/header
* information which compressed output requires), it just means
* the uncompressed version is as good as it gets, and that's
* what we use.
*/
if (compression_failed ||
buf->size / btree->allocsize <=
result_len / btree->allocsize) {
ip = buf;
WT_STAT_FAST_DATA_INCR(session, compress_write_fail);
} else {
compressed = true;
WT_STAT_FAST_DATA_INCR(session, compress_write);
/*
* Copy in the skipped header bytes, set the final data
* size.
*/
memcpy(tmp->mem, buf->mem, WT_BLOCK_COMPRESS_SKIP);
tmp->size = result_len;
ip = tmp;
}
}
dsk = ip->mem;
/* If the buffer is compressed, set the flag. */
if (compressed)
F_SET(dsk, WT_PAGE_COMPRESSED);
/*
* We increment the block's write generation so it's easy to identify
* newer versions of blocks during salvage. (It's common in WiredTiger,
* at least for the default block manager, for multiple blocks to be
* internally consistent with identical first and last keys, so we need
* a way to know the most recent state of the block. We could check
* which leaf is referenced by a valid internal page, but that implies
* salvaging internal pages, which I don't want to do, and it's not
* as good anyway, because the internal page may not have been written
* after the leaf page was updated. So, write generations it is.
*
* Nothing is locked at this point but two versions of a page with the
* same generation is pretty unlikely, and if we did, they're going to
* be roughly identical for the purposes of salvage, anyway.
*/
dsk->write_gen = ++btree->write_gen;
/*
* Checksum the data if the buffer isn't compressed or checksums are
* configured.
*/
switch (btree->checksum) {
case CKSUM_ON:
data_cksum = true;
break;
case CKSUM_OFF:
data_cksum = false;
break;
case CKSUM_UNCOMPRESSED:
default:
data_cksum = !compressed;
break;
}
/* Call the block manager to write the block. */
WT_ERR(checkpoint ?
bm->checkpoint(bm, session, ip, btree->ckpt, data_cksum) :
bm->write(bm, session, ip, addr, addr_sizep, data_cksum));
WT_STAT_FAST_CONN_INCR(session, cache_write);
WT_STAT_FAST_DATA_INCR(session, cache_write);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_write, dsk->mem_size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_write, dsk->mem_size);
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_btcur_update --
* Update a record in the tree.
*/
int
__wt_btcur_update(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_update);
WT_STAT_FAST_DATA_INCR(session, cursor_update);
WT_STAT_FAST_DATA_INCRV(
session, cursor_update_bytes, cursor->value.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
/*
* The tree is no longer empty: eviction should pay attention to it,
* and it's no longer possible to bulk-load into it.
*/
if (btree->bulk_load_ok) {
btree->bulk_load_ok = 0;
__wt_btree_evictable(session, 1);
}
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ERR(__cursor_col_search(session, cbt));
/*
* If not overwriting, fail if the key doesn't exist. Update
* the record if it exists. Creating a record past the end of
* the tree in a fixed-length column-store implicitly fills the
* gap with empty records. Update the record in that case, the
* record exists.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
(cbt->compare != 0 || !__cursor_valid(cbt, NULL)) &&
!__cursor_fix_implicit(btree, cbt))
WT_ERR(WT_NOTFOUND);
ret = __cursor_col_modify(session, cbt, 0);
break;
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, 1));
/*
* If not overwriting, fail if the key does not exist.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
(cbt->compare != 0 || !__cursor_valid(cbt, NULL)))
WT_ERR(WT_NOTFOUND);
ret = __cursor_row_modify(session, cbt, 0);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART)
goto retry;
/*
* If successful, point the cursor at internal copies of the data. We
* could shuffle memory in the cursor so the key/value pair are in local
* buffer memory, but that's a data copy. We don't want to do another
* search (and we might get a different update structure if we race).
* To make this work, we add a field to the btree cursor to pass back a
* pointer to the modify function's allocated update structure.
*/
if (ret == 0)
WT_TRET(__wt_kv_return(session, cbt, cbt->modify_update));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_remove --
* Remove a record from the tree.
*/
int
__wt_btcur_remove(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_remove);
WT_STAT_FAST_DATA_INCR(session, cursor_remove);
WT_STAT_FAST_DATA_INCRV(session, cursor_remove_bytes, cursor->key.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ERR(__cursor_col_search(session, cbt));
/* Remove the record if it exists. */
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL)) {
if (!__cursor_fix_implicit(btree, cbt))
WT_ERR(WT_NOTFOUND);
/*
* Creating a record past the end of the tree in a
* fixed-length column-store implicitly fills the
* gap with empty records. Return success in that
* case, the record was deleted successfully.
*
* Correct the btree cursor's location: the search
* will have pointed us at the previous/next item,
* and that's not correct.
*/
cbt->recno = cursor->recno;
} else
ret = __cursor_col_modify(session, cbt, 1);
break;
case BTREE_ROW:
/* Remove the record if it exists. */
WT_ERR(__cursor_row_search(session, cbt, 0));
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL))
WT_ERR(WT_NOTFOUND);
ret = __cursor_row_modify(session, cbt, 1);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART)
goto retry;
/*
* If the cursor is configured to overwrite and the record is not
* found, that is exactly what we want.
*/
if (F_ISSET(cursor, WT_CURSTD_OVERWRITE) && ret == WT_NOTFOUND)
ret = 0;
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_insert --
* Insert a record into the tree.
*/
int
__wt_btcur_insert(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_insert);
WT_STAT_FAST_DATA_INCR(session, cursor_insert);
WT_STAT_FAST_DATA_INCRV(session,
cursor_insert_bytes, cursor->key.size + cursor->value.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
/*
* The tree is no longer empty: eviction should pay attention to it,
* and it's no longer possible to bulk-load into it.
*/
if (btree->bulk_load_ok) {
btree->bulk_load_ok = 0;
__wt_btree_evictable(session, 1);
}
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
/*
* If WT_CURSTD_APPEND is set, insert a new record (ignoring
* the application's record number). First we search for the
* maximum possible record number so the search ends on the
* last page. The real record number is assigned by the
* serialized append operation.
*/
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = UINT64_MAX;
WT_ERR(__cursor_col_search(session, cbt));
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = 0;
/*
* If not overwriting, fail if the key exists. Creating a
* record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
* Fail in that case, the record exists.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
((cbt->compare == 0 && __cursor_valid(cbt, NULL)) ||
(cbt->compare != 0 && __cursor_fix_implicit(btree, cbt))))
WT_ERR(WT_DUPLICATE_KEY);
WT_ERR(__cursor_col_modify(session, cbt, 0));
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = cbt->recno;
break;
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, 1));
/*
* If not overwriting, fail if the key exists, else insert the
* key/value pair.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
cbt->compare == 0 && __cursor_valid(cbt, NULL))
WT_ERR(WT_DUPLICATE_KEY);
ret = __cursor_row_modify(session, cbt, 0);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART)
goto retry;
/* Insert doesn't maintain a position across calls, clear resources. */
if (ret == 0)
WT_TRET(__curfile_leave(cbt));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_search_near --
* Search for a record in the tree.
*/
int
__wt_btcur_search_near(WT_CURSOR_BTREE *cbt, int *exactp)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
int exact;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
exact = 0;
WT_STAT_FAST_CONN_INCR(session, cursor_search_near);
WT_STAT_FAST_DATA_INCR(session, cursor_search_near);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_func_init(cbt, 1));
/*
* Set the "insert" flag for the btree row-store search; we may intend
* to position our cursor at the end of the tree, rather than match an
* existing record.
*/
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 1) :
__cursor_col_search(session, cbt));
/*
* If we find an valid key, return it.
*
* Else, creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records. In this
* case, we instantiate the empty record, it's an exact match.
*
* Else, move to the next key in the tree (bias for prefix searches).
* Cursor next skips invalid rows, so we don't have to test for them
* again.
*
* Else, redo the search and move to the previous key in the tree.
* Cursor previous skips invalid rows, so we don't have to test for
* them again.
*
* If that fails, quit, there's no record to return.
*/
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if (__cursor_fix_implicit(btree, cbt)) {
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
exact = 0;
} else if ((ret = __wt_btcur_next(cbt, 0)) != WT_NOTFOUND)
exact = 1;
else {
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 1) :
__cursor_col_search(session, cbt));
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if ((ret = __wt_btcur_prev(cbt, 0)) != WT_NOTFOUND)
exact = -1;
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
if (exactp != NULL && (ret == 0 || ret == WT_NOTFOUND))
*exactp = exact;
return (ret);
}
/*
* __wt_ovfl_cache --
* Handle deletion of an overflow value.
*/
int
__wt_ovfl_cache(WT_SESSION_IMPL *session,
WT_PAGE *page, void *cookie, WT_CELL_UNPACK *vpack)
{
int visible;
/*
* This function solves a problem in reconciliation. The scenario is:
* - reconciling a leaf page that references an overflow item
* - the item is updated and the update committed
* - a checkpoint runs, freeing the backing overflow blocks
* - a snapshot transaction wants the original version of the item
*
* In summary, we may need the original version of an overflow item for
* a snapshot transaction after the item was deleted from a page that's
* subsequently been checkpointed, where the checkpoint must know about
* the freed blocks. We don't have any way to delay a free of the
* underlying blocks until a particular set of transactions exit (and
* this shouldn't be a common scenario), so cache the overflow value in
* memory.
*
* This gets hard because the snapshot transaction reader might:
* - search the WT_UPDATE list and not find an useful entry
* - read the overflow value's address from the on-page cell
* - go to sleep
* - checkpoint runs, caches the overflow value, frees the blocks
* - another thread allocates and overwrites the blocks
* - the reader wakes up and reads the wrong value
*
* Use a read/write lock and the on-page cell to fix the problem: hold
* a write lock when changing the cell type from WT_CELL_VALUE_OVFL to
* WT_CELL_VALUE_OVFL_RM and hold a read lock when reading an overflow
* item.
*
* The read/write lock is per btree, but it could be per page or even
* per overflow item. We don't do any of that because overflow values
* are supposed to be rare and we shouldn't see contention for the lock.
*
* Check for a globally visible update. If there is a globally visible
* update, we don't need to cache the item because it's not possible for
* a running thread to have moved past it.
*/
switch (page->type) {
case WT_PAGE_COL_VAR:
visible = __ovfl_cache_col_visible(session, cookie, vpack);
break;
case WT_PAGE_ROW_LEAF:
visible = __ovfl_cache_row_visible(session, page, cookie);
break;
WT_ILLEGAL_VALUE(session);
}
/*
* If there's no globally visible update, there's a reader in the system
* that might try and read the old value, cache it.
*/
if (!visible) {
WT_RET(__ovfl_cache(session, page, vpack));
WT_STAT_FAST_DATA_INCR(session, cache_overflow_value);
}
/*
* Queue the on-page cell to be set to WT_CELL_VALUE_OVFL_RM and the
* underlying overflow value's blocks to be freed when reconciliation
* completes.
*/
return (__wt_ovfl_discard_add(session, page, vpack->cell));
}
