/*
* __merge_switch_page --
* Switch a page from the locked tree into the new tree.
*/
static void
__merge_switch_page(WT_PAGE *parent, WT_REF *ref, WT_VISIT_STATE *state)
{
WT_PAGE *child;
WT_PAGE_MODIFY *modify;
WT_REF *newref;
if (state->split != 0 && state->refcnt++ == state->split) {
state->page = state->second;
state->ref = state->second_ref;
}
newref = state->ref++;
if (ref->addr != NULL)
__merge_transfer_footprint(
state->session, state->page, parent,
(uint32_t)sizeof(WT_ADDR) + ((WT_ADDR *)ref->addr)->size);
if (parent->type == WT_PAGE_ROW_INT)
__merge_transfer_footprint(
state->session, state->page, parent,
(uint32_t)sizeof(WT_IKEY) + ((WT_IKEY *)ref->u.key)->size);
if (ref->state == WT_REF_LOCKED) {
child = ref->page;
/*
* If the child has been split, update the split page to point
* into the new tree. That way, if the split-merge page is
* later swapped into place, it will point to the new parent.
*
* The order here is important: the parent page should point to
* the original child page, so we link that in last.
*/
if ((modify = child->modify) != NULL &&
F_ISSET(modify, WT_PM_REC_SPLIT))
WT_LINK_PAGE(state->page, newref, modify->u.split);
WT_LINK_PAGE(state->page, newref, child);
/*
* If we have a child that is a live internal page, its subtree
* was locked by __rec_review. We're swapping it into the new
* tree, unlock it now.
*/
if (child->type == WT_PAGE_ROW_INT ||
child->type == WT_PAGE_COL_INT)
__merge_unlock(child);
newref->state = WT_REF_MEM;
}
WT_CLEAR(*ref);
}
/*
* __wt_merge_tree --
* Attempt to collapse a stack of split-merge pages in memory into a
* shallow tree. If enough keys are found, create a real internal node
* that can be evicted (and, if necessary, split further).
*
* This code is designed to deal with workloads that otherwise create
* arbitrarily deep (and slow) trees in memory.
*/
int
__wt_merge_tree(WT_SESSION_IMPL *session, WT_PAGE *top)
{
WT_DECL_RET;
WT_PAGE *lchild, *newtop, *rchild;
WT_REF *newref;
WT_VISIT_STATE visit_state;
uint32_t refcnt, split;
int promote;
u_int levels;
uint8_t page_type;
WT_CLEAR(visit_state);
visit_state.session = session;
lchild = newtop = rchild = NULL;
page_type = top->type;
WT_ASSERT(session, __wt_btree_mergeable(top));
WT_ASSERT(session, top->ref->state == WT_REF_LOCKED);
/*
* Walk the subtree, count the references at the bottom level and
* calculate the maximum depth.
*/
WT_RET(__merge_walk(session, top, 1, __merge_count, &visit_state));
/* If there aren't enough useful levels, give up. */
if (visit_state.maxdepth < WT_MERGE_STACK_MIN)
return (EBUSY);
/*
* Don't allow split merges to generate arbitrarily large pages.
* Ideally we would choose a size based on the internal_page_max
* setting for the btree, but we don't have the correct btree handle
* available.
*/
if (visit_state.refcnt > WT_MERGE_MAX_REFS)
return (EBUSY);
/*
* Now we either collapse the internal pages into one split-merge page,
* or if there are "enough" keys, we split into two equal internal
* pages, each of which can be evicted independently.
*
* We set a flag (WT_PM_REC_SPLIT_MERGE) on the created page if it
* isn't big enough to justify the cost of evicting it. If splits
* continue, it will be merged again until it gets over this limit.
*/
promote = 0;
refcnt = (uint32_t)visit_state.refcnt;
if (refcnt >= WT_MERGE_FULL_PAGE && visit_state.seen_live) {
/*
* In the normal case where there are live children spread
* through the subtree, create two child pages.
*
* Handle the case where the only live child is first / last
* specially: put the live child into the top-level page.
*
* Set SPLIT_MERGE on the internal pages if there are any live
* children: they can't be evicted, so there is no point
* permanently deepening the tree.
*/
if (visit_state.first_live == visit_state.last_live &&
(visit_state.first_live == 0 ||
visit_state.first_live == refcnt - 1))
split = (visit_state.first_live == 0) ? 1 : refcnt - 1;
else
split = (refcnt + 1) / 2;
/* Only promote if we can create a real page. */
if (split == 1 || split == refcnt - 1)
promote = 1;
else if (split >= WT_MERGE_FULL_PAGE &&
visit_state.first_live >= split)
promote = 1;
else if (refcnt - split >= WT_MERGE_FULL_PAGE &&
visit_state.last_live < split)
promote = 1;
}
if (promote) {
/* Create a new top-level split-merge page with two entries. */
WT_ERR(__merge_new_page(session, page_type, 2, 1, &newtop));
visit_state.split = split;
/* Left split. */
if (split == 1)
visit_state.first = newtop;
else {
WT_ERR(__merge_new_page(session, page_type, split,
visit_state.first_live < split, &lchild));
visit_state.first = lchild;
}
/* Right split. */
if (split == refcnt - 1) {
visit_state.second = newtop;
visit_state.second_ref = &newtop->u.intl.t[1];
} else {
WT_ERR(__merge_new_page(session, page_type,
refcnt - split, visit_state.last_live >= split,
&rchild));
visit_state.second = rchild;
visit_state.second_ref =
&visit_state.second->u.intl.t[0];
}
} else {
/*
* Create a new split-merge page for small merges, or if the
* page above is a split merge page. When we do a big enough
* merge, we create a real page at the top and don't consider
* it as a merge candidate again. Over time with an insert
* workload the tree will grow deeper, but that's inevitable,
* and this keeps individual merges small.
*/
WT_ERR(__merge_new_page(session, page_type, refcnt,
refcnt < WT_MERGE_FULL_PAGE ||
__wt_btree_mergeable(top->parent),
&newtop));
visit_state.first = newtop;
}
/*
* Copy the references into the new tree, but don't update anything in
* the locked tree in case there is an error and we need to back out.
* We do this in a separate pass so that we can figure out the key for
* the split point: that allocates memory and so it could still fail.
*/
visit_state.page = visit_state.first;
visit_state.ref = visit_state.page->u.intl.t;
visit_state.refcnt = 0;
WT_ERR(__merge_walk(session, top, 0, __merge_copy_ref, &visit_state));
if (promote) {
/* Promote keys into the top-level page. */
if (lchild != NULL) {
newref = &newtop->u.intl.t[0];
WT_LINK_PAGE(newtop, newref, lchild);
newref->state = WT_REF_MEM;
WT_ERR(__merge_promote_key(session, newref));
}
if (rchild != NULL) {
newref = &newtop->u.intl.t[1];
WT_LINK_PAGE(newtop, newref, rchild);
newref->state = WT_REF_MEM;
WT_ERR(__merge_promote_key(session, newref));
}
}
/*
* We have copied everything into place and allocated all of the memory
* we need. Now link all pages into the new tree and unlock them.
*
* The only way this could fail is if a reference state has been
* changed by another thread since they were locked. Panic in that
* case: that should never happen.
*/
visit_state.page = visit_state.first;
visit_state.ref = visit_state.page->u.intl.t;
visit_state.refcnt = 0;
ret = __merge_walk(session, top, 0, __merge_switch_page, &visit_state);
if (ret != 0)
WT_ERR(__wt_illegal_value(session, "__wt_merge_tree"));
newtop->u.intl.recno = top->u.intl.recno;
newtop->parent = top->parent;
newtop->ref = top->ref;
#ifdef HAVE_DIAGNOSTIC
/*
* Before swapping in the new tree, walk the pages we are discarding,
* check that everything looks right.
*/
__merge_check_discard(session, top);
#endif
/*
* Set up the new top-level page as a split so that it will be swapped
* into place by our caller.
*/
top->modify->flags = WT_PM_REC_SPLIT;
top->modify->u.split = newtop;
WT_VERBOSE_ERR(session, evict,
"Successfully %s %" PRIu32
" split-merge pages containing %" PRIu32 " keys\n",
promote ? "promoted" : "merged", visit_state.maxdepth, refcnt);
/* Evict new child pages as soon as possible. */
if (lchild != NULL && !F_ISSET(lchild->modify, WT_PM_REC_SPLIT_MERGE))
lchild->read_gen = WT_READ_GEN_OLDEST;
if (rchild != NULL && !F_ISSET(rchild->modify, WT_PM_REC_SPLIT_MERGE))
rchild->read_gen = WT_READ_GEN_OLDEST;
/* Update statistics. */
WT_CSTAT_INCR(session, cache_eviction_merge);
WT_DSTAT_INCR(session, cache_eviction_merge);
/* How many levels did we remove? */
levels = visit_state.maxdepth - (promote ? 2 : 1);
WT_CSTAT_INCRV(session, cache_eviction_merge_levels, levels);
WT_DSTAT_INCRV(session, cache_eviction_merge_levels, levels);
return (0);
err: WT_VERBOSE_TRET(session, evict,
"Failed to merge %" PRIu32
" split-merge pages containing %" PRIu32 " keys\n",
visit_state.maxdepth, refcnt);
WT_CSTAT_INCR(session, cache_eviction_merge_fail);
WT_DSTAT_INCR(session, cache_eviction_merge_fail);
if (newtop != NULL)
__wt_page_out(session, &newtop);
if (lchild != NULL)
__wt_page_out(session, &lchild);
if (rchild != NULL)
__wt_page_out(session, &rchild);
return (ret);
}
/*
* __wt_page_inmem --
* Build in-memory page information.
*/
int
__wt_page_inmem(
WT_SESSION_IMPL *session, WT_PAGE *parent, WT_REF *parent_ref,
WT_PAGE_HEADER *dsk, int disk_not_alloc, WT_PAGE **pagep)
{
WT_DECL_RET;
WT_PAGE *page;
uint32_t alloc_entries;
size_t size;
alloc_entries = 0;
*pagep = NULL;
/*
* Figure out how many underlying objects the page references so
* we can allocate them along with the page.
*/
switch (dsk->type) {
case WT_PAGE_COL_FIX:
break;
case WT_PAGE_COL_INT:
/*
* Column-store internal page entries map one-to-one to the
* number of physical entries on the page (each physical entry
* is an offset object).
*/
alloc_entries = dsk->u.entries;
break;
case WT_PAGE_COL_VAR:
/*
* Column-store leaf page entries map one-to-one to the number
* of physical entries on the page (each physical entry is a
* data item).
*/
alloc_entries = dsk->u.entries;
break;
case WT_PAGE_ROW_INT:
/*
* Row-store internal page entries map one-to-two to the number
* of physical entries on the page (each in-memory entry is a
* key item and location cookie).
*/
alloc_entries = dsk->u.entries / 2;
break;
case WT_PAGE_ROW_LEAF:
/*
* Row-store leaf page entries map in an indeterminate way to
* the physical entries on the page, we have to walk the page
* to figure it out.
*/
WT_RET(__inmem_row_leaf_entries(session, dsk, &alloc_entries));
break;
WT_ILLEGAL_VALUE(session);
}
/* Allocate and initialize a new WT_PAGE. */
WT_RET(__wt_page_alloc(session, dsk->type, alloc_entries, &page));
page->dsk = dsk;
page->read_gen = WT_READ_GEN_NOTSET;
if (disk_not_alloc)
F_SET_ATOMIC(page, WT_PAGE_DISK_NOT_ALLOC);
/*
* Track the memory allocated to build this page so we can update the
* cache statistics in a single call.
*/
size = disk_not_alloc ? 0 : dsk->mem_size;
switch (page->type) {
case WT_PAGE_COL_FIX:
page->entries = dsk->u.entries;
page->u.col_fix.recno = dsk->recno;
__inmem_col_fix(session, page);
break;
case WT_PAGE_COL_INT:
page->entries = dsk->u.entries;
page->u.intl.recno = dsk->recno;
__inmem_col_int(session, page);
break;
case WT_PAGE_COL_VAR:
page->entries = dsk->u.entries;
page->u.col_var.recno = dsk->recno;
WT_ERR(__inmem_col_var(session, page, &size));
break;
case WT_PAGE_ROW_INT:
page->entries = dsk->u.entries / 2;
WT_ERR(__inmem_row_int(session, page, &size));
break;
case WT_PAGE_ROW_LEAF:
page->entries = alloc_entries;
WT_ERR(__inmem_row_leaf(session, page));
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Update the page's in-memory size and the cache statistics. */
__wt_cache_page_inmem_incr(session, page, size);
/* Link the new page into the parent. */
if (parent_ref != NULL)
WT_LINK_PAGE(parent, parent_ref, page);
*pagep = page;
return (0);
err: __wt_page_out(session, &page);
return (ret);
}
