/*
* __evict_force_check --
* Check if a page matches the criteria for forced eviction.
*/
static int
__evict_force_check(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_BTREE *btree;
btree = S2BT(session);
/* Pages are usually small enough, check that first. */
if (page->memory_footprint < btree->maxmempage)
return (0);
/* Leaf pages only. */
if (WT_PAGE_IS_INTERNAL(page))
return (0);
/*
* It's hard to imagine a page with a huge memory footprint that has
* never been modified, but check to be sure.
*/
if (page->modify == NULL)
return (0);
/* Trigger eviction on the next page release. */
__wt_page_evict_soon(page);
/* Bump the oldest ID, we're about to do some visibility checks. */
__wt_txn_update_oldest(session, 0);
/* If eviction cannot succeed, don't try. */
return (__wt_page_can_evict(session, page, 1, NULL));
}
/*
* __evict_force_check --
* Check if a page matches the criteria for forced eviction.
*/
static int
__evict_force_check(WT_SESSION_IMPL *session, WT_PAGE *page, uint32_t flags)
{
WT_BTREE *btree;
btree = S2BT(session);
/* Pages are usually small enough, check that first. */
if (page->memory_footprint < btree->maxmempage)
return (0);
/* Leaf pages only. */
if (WT_PAGE_IS_INTERNAL(page))
return (0);
/* Eviction may be turned off. */
if (LF_ISSET(WT_READ_NO_EVICT) || F_ISSET(btree, WT_BTREE_NO_EVICTION))
return (0);
/*
* It's hard to imagine a page with a huge memory footprint that has
* never been modified, but check to be sure.
*/
if (page->modify == NULL)
return (0);
/* Trigger eviction on the next page release. */
__wt_page_evict_soon(page);
/* If eviction cannot succeed, don't try. */
return (__wt_page_can_evict(session, page, 1));
}
/*
* __evict_force_check --
* Check if a page matches the criteria for forced eviction.
*/
static bool
__evict_force_check(WT_SESSION_IMPL *session, WT_REF *ref)
{
WT_BTREE *btree;
WT_PAGE *page;
btree = S2BT(session);
page = ref->page;
/* Leaf pages only. */
if (WT_PAGE_IS_INTERNAL(page))
return (false);
/*
* It's hard to imagine a page with a huge memory footprint that has
* never been modified, but check to be sure.
*/
if (page->modify == NULL)
return (false);
/* Pages are usually small enough, check that first. */
if (page->memory_footprint < btree->splitmempage)
return (false);
/*
* If this session has more than one hazard pointer, eviction will fail
* and there is no point trying.
*/
if (__wt_hazard_count(session, page) > 1)
return (false);
/*
* If we have already tried and the transaction state has not moved on,
* eviction is highly likely to fail.
*/
if (page->modify->last_eviction_id == __wt_txn_oldest_id(session))
return (false);
if (page->memory_footprint < btree->maxmempage)
return (__wt_leaf_page_can_split(session, page));
/* Trigger eviction on the next page release. */
__wt_page_evict_soon(session, ref);
/* Bump the oldest ID, we're about to do some visibility checks. */
WT_IGNORE_RET(__wt_txn_update_oldest(session, 0));
/* If eviction cannot succeed, don't try. */
return (__wt_page_can_evict(session, ref, NULL));
}
/*
* __wt_evict_file --
* Discard pages for a specific file.
*/
int
__wt_evict_file(WT_SESSION_IMPL *session, int syncop)
{
WT_DECL_RET;
WT_PAGE *page;
WT_REF *next_ref, *ref;
bool evict_reset;
/*
* We need exclusive access to the file -- disable ordinary eviction
* and drain any blocks already queued.
*/
WT_RET(__wt_evict_file_exclusive_on(session, &evict_reset));
/* Make sure the oldest transaction ID is up-to-date. */
__wt_txn_update_oldest(session, true);
/* Walk the tree, discarding pages. */
next_ref = NULL;
WT_ERR(__wt_tree_walk(session, &next_ref, NULL,
WT_READ_CACHE | WT_READ_NO_EVICT));
while ((ref = next_ref) != NULL) {
page = ref->page;
/*
* Eviction can fail when a page in the evicted page's subtree
* switches state. For example, if we don't evict a page marked
* empty, because we expect it to be merged into its parent, it
* might no longer be empty after it's reconciled, in which case
* eviction of its parent would fail. We can either walk the
* tree multiple times (until it's finally empty), or reconcile
* each page to get it to its final state before considering if
* it's an eviction target or will be merged into its parent.
*
* Don't limit this test to any particular page type, that tends
* to introduce bugs when the reconciliation of other page types
* changes, and there's no advantage to doing so.
*
* Eviction can also fail because an update cannot be written.
* If sessions have disjoint sets of files open, updates in a
* no-longer-referenced file may not yet be globally visible,
* and the write will fail with EBUSY. Our caller handles that
* error, retrying later.
*/
if (syncop == WT_SYNC_CLOSE && __wt_page_is_modified(page))
WT_ERR(__wt_reconcile(session, ref, NULL, WT_EVICTING));
/*
* We can't evict the page just returned to us (it marks our
* place in the tree), so move the walk to one page ahead of
* the page being evicted. Note, we reconciled the returned
* page first: if reconciliation of that page were to change
* the shape of the tree, and we did the next walk call before
* the reconciliation, the next walk call could miss a page in
* the tree.
*/
WT_ERR(__wt_tree_walk(session, &next_ref, NULL,
WT_READ_CACHE | WT_READ_NO_EVICT));
switch (syncop) {
case WT_SYNC_CLOSE:
/*
* Evict the page.
*/
WT_ERR(__wt_evict(session, ref, 1));
break;
case WT_SYNC_DISCARD:
/*
* Dead handles may reference dirty pages; clean the
* page, both to keep statistics correct, and to let
* the page-discard function assert no dirty page is
* ever discarded.
*/
if (F_ISSET(session->dhandle, WT_DHANDLE_DEAD))
__wt_page_modify_clear(session, page);
WT_ASSERT(session,
F_ISSET(session->dhandle, WT_DHANDLE_DEAD) ||
__wt_page_can_evict(session, ref, false, NULL));
__wt_evict_page_clean_update(session, ref, 1);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
}
if (0) {
err: /* On error, clear any left-over tree walk. */
if (next_ref != NULL)
WT_TRET(__wt_page_release(
session, next_ref, WT_READ_NO_EVICT));
}
if (evict_reset)
__wt_evict_file_exclusive_off(session);
return (ret);
}
/*
* __evict_stat_walk --
* Walk all the pages in cache for a dhandle gathering stats information
*/
static void
__evict_stat_walk(WT_SESSION_IMPL *session)
{
WT_BTREE *btree;
WT_CACHE *cache;
WT_PAGE *page;
WT_REF *next_walk;
uint64_t dsk_size, gen_gap, gen_gap_max, gen_gap_sum, max_pagesize;
uint64_t min_written_size, num_memory, num_not_queueable, num_queued;
uint64_t num_smaller_allocsz, pages_clean, pages_dirty, pages_internal;
uint64_t pages_leaf, seen_count, size, visited_count;
uint64_t visited_age_gap_sum, unvisited_count, unvisited_age_gap_sum;
uint64_t walk_count, written_size_cnt, written_size_sum;
btree = S2BT(session);
cache = S2C(session)->cache;
next_walk = NULL;
gen_gap_max = gen_gap_sum = max_pagesize = 0;
num_memory = num_not_queueable = num_queued = 0;
num_smaller_allocsz = pages_clean = pages_dirty = pages_internal = 0;
pages_leaf = seen_count = size = visited_count = 0;
visited_age_gap_sum = unvisited_count = unvisited_age_gap_sum = 0;
walk_count = written_size_cnt = written_size_sum = 0;
min_written_size = UINT64_MAX;
while (__wt_tree_walk_count(session, &next_walk, &walk_count,
WT_READ_CACHE | WT_READ_NO_EVICT |
WT_READ_NO_GEN | WT_READ_NO_WAIT) == 0 &&
next_walk != NULL) {
++seen_count;
page = next_walk->page;
size = page->memory_footprint;
if (__wt_page_is_modified(page))
++pages_dirty;
else
++pages_clean;
if (!__wt_ref_is_root(next_walk) &&
!__wt_page_can_evict(session, next_walk, NULL))
++num_not_queueable;
if (F_ISSET_ATOMIC(page, WT_PAGE_EVICT_LRU))
++num_queued;
if (size > max_pagesize)
max_pagesize = size;
dsk_size = page->dsk != NULL ? page->dsk->mem_size : 0;
if (dsk_size != 0) {
if (dsk_size < btree->allocsize)
++num_smaller_allocsz;
if (dsk_size < min_written_size)
min_written_size = dsk_size;
++written_size_cnt;
written_size_sum += dsk_size;
} else
++num_memory;
if (WT_PAGE_IS_INTERNAL(page))
++pages_internal;
else
++pages_leaf;
/* Skip root pages since they are never considered */
if (__wt_ref_is_root(next_walk))
continue;
if (page->evict_pass_gen == 0) {
unvisited_age_gap_sum +=
(cache->evict_pass_gen - page->cache_create_gen);
++unvisited_count;
} else {
visited_age_gap_sum +=
(cache->evict_pass_gen - page->cache_create_gen);
gen_gap = cache->evict_pass_gen - page->evict_pass_gen;
if (gen_gap > gen_gap_max)
gen_gap_max = gen_gap;
gen_gap_sum += gen_gap;
++visited_count;
}
}
WT_STAT_DATA_SET(session, cache_state_gen_avg_gap,
visited_count == 0 ? 0 : gen_gap_sum / visited_count);
WT_STAT_DATA_SET(session, cache_state_avg_unvisited_age,
unvisited_count == 0 ? 0 : unvisited_age_gap_sum / unvisited_count);
WT_STAT_DATA_SET(session, cache_state_avg_visited_age,
visited_count == 0 ? 0 : visited_age_gap_sum / visited_count);
WT_STAT_DATA_SET(session, cache_state_avg_written_size,
written_size_cnt == 0 ? 0 : written_size_sum / written_size_cnt);
WT_STAT_DATA_SET(session, cache_state_gen_max_gap, gen_gap_max);
WT_STAT_DATA_SET(session, cache_state_max_pagesize, max_pagesize);
WT_STAT_DATA_SET(session,
cache_state_min_written_size, min_written_size);
WT_STAT_DATA_SET(session, cache_state_memory, num_memory);
WT_STAT_DATA_SET(session, cache_state_queued, num_queued);
WT_STAT_DATA_SET(session, cache_state_not_queueable, num_not_queueable);
WT_STAT_DATA_SET(session, cache_state_pages, walk_count);
WT_STAT_DATA_SET(session, cache_state_pages_clean, pages_clean);
WT_STAT_DATA_SET(session, cache_state_pages_dirty, pages_dirty);
WT_STAT_DATA_SET(session, cache_state_pages_internal, pages_internal);
WT_STAT_DATA_SET(session, cache_state_pages_leaf, pages_leaf);
WT_STAT_DATA_SET(session,
cache_state_refs_skipped, walk_count - seen_count);
WT_STAT_DATA_SET(session,
cache_state_smaller_alloc_size, num_smaller_allocsz);
WT_STAT_DATA_SET(session,
cache_state_unvisited_count, unvisited_count);
}
