/*
* __cursor_var_prev --
* Move to the previous, variable-length column-store item.
*/
static inline int
__cursor_var_prev(WT_CURSOR_BTREE *cbt, int newpage)
{
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_COL *cip;
WT_DECL_RET;
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
val = &cbt->iface.value;
/* Initialize for each new page. */
if (newpage) {
cbt->last_standard_recno = __col_last_recno(cbt->page);
if (cbt->last_standard_recno == 0)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, cbt->last_standard_recno);
goto new_page;
}
/* Move to the previous entry and return the item. */
for (;;) {
__cursor_set_recno(cbt, cbt->recno - 1);
new_page: if (cbt->recno < cbt->page->u.col_var.recno)
return (WT_NOTFOUND);
/* Find the matching WT_COL slot. */
if ((cip = __col_var_search(cbt->page, cbt->recno)) == NULL)
return (WT_NOTFOUND);
cbt->slot = WT_COL_SLOT(cbt->page, cip);
/* Check any insert list for a matching record. */
cbt->ins_head = WT_COL_UPDATE_SLOT(cbt->page, cbt->slot);
cbt->ins = __col_insert_search_match(cbt->ins_head, cbt->recno);
upd = cbt->ins == NULL ?
NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd != NULL) {
if (WT_UPDATE_DELETED_ISSET(upd))
continue;
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
return (0);
}
/*
* If we're at the same slot as the last reference and there's
* no matching insert list item, re-use the return information
* (so encoded items with large repeat counts aren't repeatedly
* decoded). Otherwise, unpack the cell and build the return
* information.
*/
if (cbt->cip_saved != cip) {
if ((cell = WT_COL_PTR(cbt->page, cip)) == NULL)
continue;
__wt_cell_unpack(cell, &unpack);
if (unpack.type == WT_CELL_DEL)
continue;
/*
* Restart for a variable-length column-store. We could
* catch restart higher up the call-stack but there's no
* point to it: unlike row-store (where a normal search
* path finds cached overflow values), we have to access
* the page's reconciliation structures, and that's as
* easy here as higher up the stack.
*/
if ((ret = __wt_cell_unpack_ref(
session, &unpack, &cbt->tmp)) == WT_RESTART)
ret = __wt_ovfl_cache_col_restart(
session, cbt->page, &unpack, &cbt->tmp);
WT_RET(ret);
cbt->cip_saved = cip;
}
val->data = cbt->tmp.data;
val->size = cbt->tmp.size;
return (0);
}
/* NOTREACHED */
}
/*
* __verify_dsk_row --
* Walk a WT_PAGE_ROW_INT or WT_PAGE_ROW_LEAF disk page and verify it.
*/
static int
__verify_dsk_row(
WT_SESSION_IMPL *session, const char *addr, WT_PAGE_HEADER *dsk)
{
WT_BM *bm;
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_DECL_ITEM(current);
WT_DECL_ITEM(last_ovfl);
WT_DECL_ITEM(last_pfx);
WT_DECL_RET;
WT_ITEM *last;
enum { FIRST, WAS_KEY, WAS_VALUE } last_cell_type;
void *huffman;
uint32_t cell_num, cell_type, i, prefix;
uint8_t *end;
int cmp;
btree = S2BT(session);
bm = btree->bm;
huffman = btree->huffman_key;
unpack = &_unpack;
WT_ERR(__wt_scr_alloc(session, 0, ¤t));
WT_ERR(__wt_scr_alloc(session, 0, &last_pfx));
WT_ERR(__wt_scr_alloc(session, 0, &last_ovfl));
last = last_ovfl;
end = (uint8_t *)dsk + dsk->mem_size;
last_cell_type = FIRST;
cell_num = 0;
WT_CELL_FOREACH(btree, dsk, cell, unpack, i) {
++cell_num;
/* Carefully unpack the cell. */
if (__wt_cell_unpack_safe(cell, unpack, end) != 0) {
ret = __err_cell_corrupted(session, cell_num, addr);
goto err;
}
/* Check the raw and collapsed cell types. */
WT_ERR(__err_cell_type(
session, cell_num, addr, unpack->raw, dsk->type));
WT_ERR(__err_cell_type(
session, cell_num, addr, unpack->type, dsk->type));
cell_type = unpack->type;
/*
* Check ordering relationships between the WT_CELL entries.
* For row-store internal pages, check for:
* two values in a row,
* two keys in a row,
* a value as the first cell on a page.
* For row-store leaf pages, check for:
* two values in a row,
* a value as the first cell on a page.
*/
switch (cell_type) {
case WT_CELL_KEY:
case WT_CELL_KEY_OVFL:
switch (last_cell_type) {
case FIRST:
case WAS_VALUE:
break;
case WAS_KEY:
if (dsk->type == WT_PAGE_ROW_LEAF)
break;
WT_ERR_VRFY(session,
"cell %" PRIu32 " on page at %s is the "
"first of two adjacent keys",
cell_num - 1, addr);
}
last_cell_type = WAS_KEY;
break;
case WT_CELL_ADDR:
case WT_CELL_VALUE:
case WT_CELL_VALUE_OVFL:
switch (last_cell_type) {
case FIRST:
WT_ERR_VRFY(session,
"page at %s begins with a value", addr);
case WAS_KEY:
break;
case WAS_VALUE:
WT_ERR_VRFY(session,
"cell %" PRIu32 " on page at %s is the "
"first of two adjacent values",
cell_num - 1, addr);
}
last_cell_type = WAS_VALUE;
break;
}
/* Check if any referenced item is entirely in the file. */
switch (cell_type) {
case WT_CELL_ADDR:
case WT_CELL_KEY_OVFL:
case WT_CELL_VALUE_OVFL:
if (!bm->addr_valid(bm,
session, unpack->data, unpack->size))
goto eof;
break;
}
/*
* Remaining checks are for key order and prefix compression.
* If this cell isn't a key, we're done, move to the next cell.
* If this cell is an overflow item, instantiate the key and
* compare it with the last key. Otherwise, we have to deal
* with prefix compression.
*/
switch (cell_type) {
case WT_CELL_KEY:
break;
case WT_CELL_KEY_OVFL:
WT_ERR(__wt_cell_unpack_ref(session, unpack, current));
goto key_compare;
default:
/* Not a key -- continue with the next cell. */
continue;
}
/*
* Prefix compression checks.
*
* Confirm the first non-overflow key on a page has a zero
* prefix compression count.
*/
prefix = unpack->prefix;
if (last_pfx->size == 0 && prefix != 0)
WT_ERR_VRFY(session,
"the %" PRIu32 " key on page at %s is the first "
"non-overflow key on the page and has a non-zero "
"prefix compression value",
cell_num, addr);
/* Confirm the prefix compression count is possible. */
if (cell_num > 1 && prefix > last->size)
WT_ERR_VRFY(session,
"key %" PRIu32 " on page at %s has a prefix "
"compression count of %" PRIu32
", larger than the length of the previous key, %"
PRIu32,
cell_num, addr, prefix, last->size);
/*
* If Huffman decoding required, unpack the cell to build the
* key, then resolve the prefix. Else, we can do it faster
* internally because we don't have to shuffle memory around as
* much.
*/
if (huffman != NULL) {
WT_ERR(__wt_cell_unpack_ref(session, unpack, current));
/*
* If there's a prefix, make sure there's enough buffer
* space, then shift the decoded data past the prefix
* and copy the prefix into place. Take care with the
* pointers: current->data may be pointing inside the
*/
if (prefix != 0) {
WT_ERR(__wt_buf_grow(
session, current, prefix + current->size));
memmove((uint8_t *)current->mem + prefix,
current->data, current->size);
memcpy(current->mem, last->data, prefix);
current->data = current->mem;
current->size += prefix;
}
} else {
/*
* Get the cell's data/length and make sure we have
* enough buffer space.
*/
WT_ERR(__wt_buf_init(
session, current, prefix + unpack->size));
/* Copy the prefix then the data into place. */
if (prefix != 0)
memcpy(current->mem, last->data, prefix);
memcpy((uint8_t *)current->mem + prefix, unpack->data,
unpack->size);
current->size = prefix + unpack->size;
}
key_compare: /*
* Compare the current key against the last key.
*
* Be careful about the 0th key on internal pages: we only store
* the first byte and custom collators may not be able to handle
* truncated keys.
*/
if ((dsk->type == WT_PAGE_ROW_INT && cell_num > 3) ||
(dsk->type != WT_PAGE_ROW_INT && cell_num > 1)) {
WT_ERR(
WT_BTREE_CMP(session, btree, last, current, cmp));
if (cmp >= 0)
WT_ERR_VRFY(session,
"the %" PRIu32 " and %" PRIu32 " keys on "
"page at %s are incorrectly sorted",
cell_num - 2, cell_num, addr);
}
/*
* Swap the buffers: last always references the last key entry,
* last_pfx and last_ovfl reference the last prefix-compressed
* and last overflow key entries. Current gets pointed to the
* buffer we're not using this time around, which is where the
* next key goes.
*/
last = current;
if (cell_type == WT_CELL_KEY) {
current = last_pfx;
last_pfx = last;
} else {
current = last_ovfl;
last_ovfl = last;
}
WT_ASSERT(session, last != current);
}
/*
* __wt_kv_return --
* Return a page referenced key/value pair to the application.
*/
int
__wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_PAGE *page;
WT_ROW *rip;
WT_UPDATE *upd;
uint8_t v;
btree = S2BT(session);
page = cbt->page;
cursor = &cbt->iface;
switch (page->type) {
case WT_PAGE_COL_FIX:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return (__wt_buf_set(session, &cursor->value, &v, 1));
case WT_PAGE_COL_VAR:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
cell = WT_COL_PTR(page, &page->u.col_var.d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->u.row.d[cbt->slot];
/*
* If the cursor references a WT_INSERT item, take the key and
* related WT_UPDATE item. Otherwise, take the key from the
* original page, and the value from any related WT_UPDATE item,
* or the page if the key was never updated.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
} else {
WT_RET(
__wt_row_key(session, page, rip, &cursor->key, 0));
upd = __wt_txn_read(session, WT_ROW_UPDATE(page, rip));
}
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the original cell (which may be empty). */
if ((cell = __wt_row_value(page, rip)) == NULL) {
cursor->value.size = 0;
return (0);
}
break;
WT_ILLEGAL_VALUE(session);
}
/* The value is an on-page cell, unpack and expand it as necessary. */
__wt_cell_unpack(cell, &unpack);
ret = __wt_cell_unpack_ref(session, &unpack, &cursor->value);
/*
* Restart for a variable-length column-store. We could catch restart
* higher up the call-stack but there's no point to it: unlike row-store
* (where a normal search path finds cached overflow values), we have to
* access the page's reconciliation structures, and that's as easy here
* as higher up the stack.
*/
if (ret == WT_RESTART && page->type == WT_PAGE_COL_VAR)
ret = __wt_ovfl_cache_col_restart(
session, page, &unpack, &cursor->value);
return (ret);
}
