/*
* __wt_ovfl_read --
* Bring an overflow item into memory.
*/
int
__wt_ovfl_read(WT_SESSION_IMPL *session,
WT_PAGE *page, WT_CELL_UNPACK *unpack, WT_ITEM *store)
{
WT_DECL_RET;
/*
* If no page specified, there's no need to lock and there's no cache
* to search, we don't care about WT_CELL_VALUE_OVFL_RM cells.
*/
if (page == NULL)
return (
__ovfl_read(session, unpack->data, unpack->size, store));
/*
* WT_CELL_VALUE_OVFL_RM cells: If reconciliation deleted an overflow
* value, but there was still a reader in the system that might need it,
* the on-page cell type will have been reset to WT_CELL_VALUE_OVFL_RM
* and we will be passed a page so we can look-aside into the cache of
* such values.
*
* Acquire the overflow lock, and retest the on-page cell's value inside
* the lock.
*/
WT_RET(__wt_readlock(session, S2BT(session)->ovfl_lock));
ret = __wt_cell_type_raw(unpack->cell) == WT_CELL_VALUE_OVFL_RM ?
__wt_ovfl_txnc_search(page, unpack->data, unpack->size, store) :
__ovfl_read(session, unpack->data, unpack->size, store);
WT_TRET(__wt_readunlock(session, S2BT(session)->ovfl_lock));
return (ret);
}
/*
* __wt_ovfl_read --
* Bring an overflow item into memory.
*/
int
__wt_ovfl_read(WT_SESSION_IMPL *session,
WT_PAGE *page, WT_CELL_UNPACK *unpack, WT_ITEM *store, bool *decoded)
{
WT_DECL_RET;
WT_OVFL_TRACK *track;
size_t i;
*decoded = false;
/*
* If no page specified, there's no need to lock and there's no cache
* to search, we don't care about WT_CELL_VALUE_OVFL_RM cells.
*/
if (page == NULL)
return (
__ovfl_read(session, unpack->data, unpack->size, store));
/*
* WT_CELL_VALUE_OVFL_RM cells: If reconciliation deleted an overflow
* value, but there was still a reader in the system that might need it,
* the on-page cell type will have been reset to WT_CELL_VALUE_OVFL_RM
* and we will be passed a page so we can check the on-page cell.
*
* Acquire the overflow lock, and retest the on-page cell's value inside
* the lock.
*/
__wt_readlock(session, &S2BT(session)->ovfl_lock);
if (__wt_cell_type_raw(unpack->cell) == WT_CELL_VALUE_OVFL_RM) {
track = page->modify->ovfl_track;
for (i = 0; i < track->remove_next; ++i)
if (track->remove[i].cell == unpack->cell) {
store->data = track->remove[i].data;
store->size = track->remove[i].size;
break;
}
WT_ASSERT(session, i < track->remove_next);
*decoded = true;
} else
ret = __ovfl_read(session, unpack->data, unpack->size, store);
__wt_readunlock(session, &S2BT(session)->ovfl_lock);
return (ret);
}
/*
* __wt_delete_page --
* If deleting a range, try to delete the page without instantiating it.
*/
int
__wt_delete_page(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
{
WT_DECL_RET;
WT_PAGE *parent;
*skipp = false;
/* If we have a clean page in memory, attempt to evict it. */
if (ref->state == WT_REF_MEM &&
__wt_atomic_casv32(&ref->state, WT_REF_MEM, WT_REF_LOCKED)) {
if (__wt_page_is_modified(ref->page)) {
WT_PUBLISH(ref->state, WT_REF_MEM);
return (0);
}
(void)__wt_atomic_addv32(&S2BT(session)->evict_busy, 1);
ret = __wt_evict_page(session, ref);
(void)__wt_atomic_subv32(&S2BT(session)->evict_busy, 1);
WT_RET_BUSY_OK(ret);
}
/*
* Atomically switch the page's state to lock it. If the page is not
* on-disk, other threads may be using it, no fast delete.
*
* Possible optimization: if the page is already deleted and the delete
* is visible to us (the delete has been committed), we could skip the
* page instead of instantiating it and figuring out there are no rows
* in the page. While that's a huge amount of work to no purpose, it's
* unclear optimizing for overlapping range deletes is worth the effort.
*/
if (ref->state != WT_REF_DISK ||
!__wt_atomic_casv32(&ref->state, WT_REF_DISK, WT_REF_LOCKED))
return (0);
/*
* We cannot fast-delete pages that have overflow key/value items as
* the overflow blocks have to be discarded. The way we figure that
* out is to check the on-page cell type for the page, cells for leaf
* pages that have no overflow items are special.
*
* In some cases, the reference address may not reference an on-page
* cell (for example, some combination of page splits), in which case
* we can't check the original cell value and we fail.
*
* To look at an on-page cell, we need to look at the parent page, and
* that's dangerous, our parent page could change without warning if
* the parent page were to split, deepening the tree. It's safe: the
* page's reference will always point to some valid page, and if we find
* any problems we simply fail the fast-delete optimization.
*
* !!!
* I doubt it's worth the effort, but we could copy the cell's type into
* the reference structure, and then we wouldn't need an on-page cell.
*/
parent = ref->home;
if (__wt_off_page(parent, ref->addr) ||
__wt_cell_type_raw(ref->addr) != WT_CELL_ADDR_LEAF_NO)
goto err;
/*
* This action dirties the parent page: mark it dirty now, there's no
* future reconciliation of the child leaf page that will dirty it as
* we write the tree.
*/
WT_ERR(__wt_page_parent_modify_set(session, ref, false));
/*
* Record the change in the transaction structure and set the change's
* transaction ID.
*/
WT_ERR(__wt_calloc_one(session, &ref->page_del));
ref->page_del->txnid = session->txn.id;
WT_ERR(__wt_txn_modify_ref(session, ref));
*skipp = true;
WT_PUBLISH(ref->state, WT_REF_DELETED);
return (0);
err: __wt_free(session, ref->page_del);
/*
* Restore the page to on-disk status, we'll have to instantiate it.
*/
WT_PUBLISH(ref->state, WT_REF_DISK);
return (ret);
}
/*
* __wt_delete_page --
* If deleting a range, try to delete the page without instantiating it.
*/
int
__wt_delete_page(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
{
WT_DECL_RET;
WT_PAGE *parent;
*skipp = false;
/* If we have a clean page in memory, attempt to evict it. */
if (ref->state == WT_REF_MEM &&
__wt_atomic_casv32(&ref->state, WT_REF_MEM, WT_REF_LOCKED)) {
if (__wt_page_is_modified(ref->page)) {
WT_PUBLISH(ref->state, WT_REF_MEM);
return (0);
}
(void)__wt_atomic_addv32(&S2BT(session)->evict_busy, 1);
ret = __wt_evict(session, ref, false);
(void)__wt_atomic_subv32(&S2BT(session)->evict_busy, 1);
WT_RET_BUSY_OK(ret);
}
/*
* Atomically switch the page's state to lock it. If the page is not
* on-disk, other threads may be using it, no fast delete.
*/
if (ref->state != WT_REF_DISK ||
!__wt_atomic_casv32(&ref->state, WT_REF_DISK, WT_REF_LOCKED))
return (0);
/*
* We cannot fast-delete pages that have overflow key/value items as
* the overflow blocks have to be discarded. The way we figure that
* out is to check the page's cell type, cells for leaf pages without
* overflow items are special.
*
* To look at an on-page cell, we need to look at the parent page, and
* that's dangerous, our parent page could change without warning if
* the parent page were to split, deepening the tree. It's safe: the
* page's reference will always point to some valid page, and if we find
* any problems we simply fail the fast-delete optimization.
*/
parent = ref->home;
if (__wt_off_page(parent, ref->addr) ?
((WT_ADDR *)ref->addr)->type != WT_ADDR_LEAF_NO :
__wt_cell_type_raw(ref->addr) != WT_CELL_ADDR_LEAF_NO)
goto err;
/*
* This action dirties the parent page: mark it dirty now, there's no
* future reconciliation of the child leaf page that will dirty it as
* we write the tree.
*/
WT_ERR(__wt_page_parent_modify_set(session, ref, false));
/*
* Record the change in the transaction structure and set the change's
* transaction ID.
*/
WT_ERR(__wt_calloc_one(session, &ref->page_del));
ref->page_del->txnid = session->txn.id;
WT_ERR(__wt_txn_modify_ref(session, ref));
*skipp = true;
WT_STAT_CONN_INCR(session, rec_page_delete_fast);
WT_STAT_DATA_INCR(session, rec_page_delete_fast);
WT_PUBLISH(ref->state, WT_REF_DELETED);
return (0);
err: __wt_free(session, ref->page_del);
/*
* Restore the page to on-disk status, we'll have to instantiate it.
*/
WT_PUBLISH(ref->state, WT_REF_DISK);
return (ret);
}
