/*
* __wt_btcur_update --
* Update a record in the tree.
*/
int
__wt_btcur_update(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_SESSION_IMPL *session;
int ret;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_BSTAT_INCR(session, cursor_updates);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
retry: __cursor_func_init(cbt, 1);
switch (btree->type) {
case BTREE_COL_FIX:
if (cursor->value.size != 1)
WT_RET_MSG(session, EINVAL,
"item size of %" PRIu32 " does not match "
"fixed-length file requirement of 1 byte",
cursor->value.size);
/* FALLTHROUGH */
case BTREE_COL_VAR:
WT_ERR(__wt_col_search(session, cbt, 1));
/*
* 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 ((cbt->compare != 0 || __cursor_invalid(cbt)) &&
!__cursor_fix_implicit(btree, cbt))
ret = WT_NOTFOUND;
else if ((ret = __wt_col_modify(session, cbt, 3)) == WT_RESTART)
goto retry;
break;
case BTREE_ROW:
/* Update the record it it exists. */
WT_ERR(__wt_row_search(session, cbt, 1));
if (cbt->compare != 0 || __cursor_invalid(cbt))
ret = WT_NOTFOUND;
else if ((ret = __wt_row_modify(session, cbt, 0)) == WT_RESTART)
goto retry;
break;
WT_ILLEGAL_VALUE(session);
}
err: __cursor_func_resolve(cbt, ret);
return (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_BSTAT_INCR(session, cursor_resets);
__cursor_func_init(cbt, 1);
__cursor_search_clear(cbt);
return (0);
}
/*
* __wt_btcur_remove --
* Remove a record from the tree.
*/
int
__wt_btcur_remove(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_SESSION_IMPL *session;
int ret;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_BSTAT_INCR(session, cursor_removes);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
retry: __cursor_func_init(cbt, 1);
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ERR(__wt_col_search(session, cbt, 1));
/*
* Remove 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. Return success in that
* case, the record was deleted successfully.
*/
if (cbt->compare != 0 || __cursor_invalid(cbt))
ret =
__cursor_fix_implicit(btree, cbt) ? 0 : WT_NOTFOUND;
else if ((ret = __wt_col_modify(session, cbt, 2)) == WT_RESTART)
goto retry;
break;
case BTREE_ROW:
/* Remove the record if it exists. */
WT_ERR(__wt_row_search(session, cbt, 1));
if (cbt->compare != 0 || __cursor_invalid(cbt))
ret = WT_NOTFOUND;
else if ((ret = __wt_row_modify(session, cbt, 1)) == WT_RESTART)
goto retry;
break;
WT_ILLEGAL_VALUE(session);
}
err: __cursor_func_resolve(cbt, ret);
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_ITEM *val;
WT_SESSION_IMPL *session;
int ret;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_BSTAT_INCR(session, cursor_read);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
__cursor_func_init(cbt, 1);
WT_ERR(btree->type == BTREE_ROW ?
__wt_row_search(session, cbt, 0) :
__wt_col_search(session, cbt, 0));
if (cbt->compare != 0 || __cursor_invalid(cbt)) {
/*
* Creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
*/
if (__cursor_fix_implicit(btree, cbt)) {
cbt->v = 0;
val = &cbt->iface.value;
val->data = &cbt->v;
val->size = 1;
} else
ret = WT_NOTFOUND;
} else
ret = __wt_kv_return(session, cbt, 0);
err: __cursor_func_resolve(cbt, ret);
return (ret);
}
/*
* __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.
*/
if (session->excl_next * sizeof(WT_REF *) == session->excl_allocated)
WT_RET(__wt_realloc(session, &session->excl_allocated,
(session->excl_next + 50) * sizeof(WT_REF *),
&session->excl));
/*
* Hazard references 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 reference. */
if (__wt_page_hazard_check(session, ref->page) == NULL)
return (0);
WT_BSTAT_INCR(session, rec_hazard);
WT_CSTAT_INCR(session, cache_evict_hazard);
WT_VERBOSE_RET(
session, evict, "page %p hazard request failed", ref->page);
return (EBUSY);
}
/*
* __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_insert --
* Insert a record into the tree.
*/
int
__wt_btcur_insert(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_SESSION_IMPL *session;
int ret;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_BSTAT_INCR(session, cursor_inserts);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
retry: __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.
* __wt_col_append_serial_func
*/
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = UINT64_MAX;
WT_ERR(__wt_col_search(session, cbt, 1));
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = 0;
/*
* If WT_CURSTD_OVERWRITE set, insert/update the key/value pair.
*
* If WT_CURSTD_OVERWRITE not set, fail if the key exists, else
* insert the key/value pair. 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_invalid(cbt)) ||
(cbt->compare != 0 && __cursor_fix_implicit(btree, cbt)))) {
ret = WT_DUPLICATE_KEY;
break;
}
if ((ret = __wt_col_modify(session, cbt, 3)) == WT_RESTART)
goto retry;
if (F_ISSET(cursor, WT_CURSTD_APPEND) && ret == 0)
cbt->iface.recno = cbt->recno;
break;
case BTREE_ROW:
/*
* If WT_CURSTD_OVERWRITE not set, fail if the key exists, else
* insert the key/value pair.
*
* If WT_CURSTD_OVERWRITE set, insert/update the key/value pair.
*/
WT_ERR(__wt_row_search(session, cbt, 1));
if (cbt->compare == 0 &&
!__cursor_invalid(cbt) &&
!F_ISSET(cursor, WT_CURSTD_OVERWRITE)) {
ret = WT_DUPLICATE_KEY;
break;
}
if ((ret = __wt_row_modify(session, cbt, 0)) == WT_RESTART)
goto retry;
break;
WT_ILLEGAL_VALUE(session);
}
err: __cursor_func_resolve(cbt, ret);
return (ret);
}
/*
* __wt_btcur_search_near --
* Search for a record in the tree.
*/
int
__wt_btcur_search_near(WT_CURSOR_BTREE *cbt, int *exact)
{
WT_BTREE *btree;
WT_ITEM *val;
WT_CURSOR *cursor;
WT_SESSION_IMPL *session;
int ret;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_BSTAT_INCR(session, cursor_read_near);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
__cursor_func_init(cbt, 1);
WT_ERR(btree->type == BTREE_ROW ?
__wt_row_search(session, cbt, 0) :
__wt_col_search(session, cbt, 0));
/*
* 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, if we find a valid key (one that wasn't deleted), return it.
*
* Else, if we found a deleted key, try to move to the next key in the
* tree (bias for prefix searches). Cursor next skips deleted records,
* so we don't have to test for them again.
*
* Else if there's no larger tree key, redo the search and try and find
* an earlier record. If that fails, quit, there's no record to return.
*/
if (cbt->compare != 0 && __cursor_fix_implicit(btree, cbt)) {
cbt->v = 0;
val = &cbt->iface.value;
val->data = &cbt->v;
val->size = 1;
*exact = 0;
} else if (!__cursor_invalid(cbt)) {
*exact = cbt->compare;
ret = __wt_kv_return(session, cbt, cbt->compare == 0 ? 0 : 1);
} else if ((ret = __wt_btcur_next(cbt)) != WT_NOTFOUND)
*exact = 1;
else {
WT_ERR(btree->type == BTREE_ROW ?
__wt_row_search(session, cbt, 0) :
__wt_col_search(session, cbt, 0));
if (!__cursor_invalid(cbt)) {
*exact = cbt->compare;
ret = __wt_kv_return(
session, cbt, cbt->compare == 0 ? 0 : 1);
} else if ((ret = __wt_btcur_prev(cbt)) != WT_NOTFOUND)
*exact = -1;
}
err: __cursor_func_resolve(cbt, ret);
return (ret);
}
/*
* __wt_block_read_off --
* Read an addr/size pair referenced block into a buffer.
*/
int
__wt_block_read_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, off_t offset, uint32_t size, uint32_t cksum)
{
WT_BLOCK_HEADER *blk;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t result_len;
uint32_t page_cksum;
WT_VERBOSE_RET(session, read,
"off %" PRIuMAX ", size %" PRIu32 ", cksum %" PRIu32,
(uintmax_t)offset, size, cksum);
#ifdef HAVE_DIAGNOSTIC
/*
* In diagnostic mode, verify the block we're about to read isn't on
* either the available or discard lists.
*
* Don't check during salvage, it's possible we're reading an already
* freed overflow page.
*/
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
WT_RET(
__wt_block_misplaced(session, block, "read", offset, size));
#endif
/*
* If we're compressing the file blocks, place the initial read into a
* scratch buffer, we're going to have to re-allocate more memory for
* decompression. Else check the caller's buffer size and grow it as
* necessary, there will only be one buffer.
*/
if (block->compressor == NULL) {
F_SET(buf, WT_ITEM_ALIGNED);
WT_RET(__wt_buf_init(session, buf, size));
buf->size = size;
dsk = buf->mem;
} else {
WT_RET(__wt_scr_alloc(session, size, &tmp));
tmp->size = size;
dsk = tmp->mem;
}
/* Read. */
WT_ERR(__wt_read(session, block->fh, offset, size, dsk));
blk = WT_BLOCK_HEADER_REF(dsk);
/* Validate the checksum. */
if (block->checksum &&
cksum != WT_BLOCK_CHECKSUM_NOT_SET &&
blk->cksum != WT_BLOCK_CHECKSUM_NOT_SET) {
blk->cksum = 0;
page_cksum = __wt_cksum(dsk, size);
if (page_cksum == WT_BLOCK_CHECKSUM_NOT_SET)
++page_cksum;
if (cksum != page_cksum) {
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
__wt_errx(session,
"read checksum error [%"
PRIu32 "B @ %" PRIuMAX ", %"
PRIu32 " != %" PRIu32 "]",
size, (uintmax_t)offset, cksum, page_cksum);
WT_ERR(WT_ERROR);
}
}
/*
* If the in-memory block size is larger than the on-disk block size,
* the block is compressed. Size the user's buffer, copy the skipped
* bytes of the original image into place, then decompress.
*
* If the in-memory block size is less than or equal to the on-disk
* block size, the block is not compressed.
*/
if (blk->disk_size < dsk->size) {
if (block->compressor == NULL)
WT_ERR(__wt_illegal_value(session, block->name));
WT_ERR(__wt_buf_init(session, buf, dsk->size));
buf->size = dsk->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->mem, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(block->compressor->decompress(
block->compressor, &session->iface,
(uint8_t *)tmp->mem + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->size - WT_BLOCK_COMPRESS_SKIP,
&result_len));
if (result_len != dsk->size - WT_BLOCK_COMPRESS_SKIP)
WT_ERR(__wt_illegal_value(session, block->name));
} else
if (block->compressor == NULL)
buf->size = dsk->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, why configure a compressor that
* doesn't work? Allocate a buffer of the right size
* (we used a scratch buffer which might be large), and
* copy the data into place.
*/
WT_ERR(
__wt_buf_set(session, buf, tmp->data, dsk->size));
WT_BSTAT_INCR(session, page_read);
WT_CSTAT_INCR(session, block_read);
err: __wt_scr_free(&tmp);
return (ret);
}
