/*
* __stat_page --
* Stat any Btree page.
*/
static int
__stat_page(WT_SESSION_IMPL *session, WT_PAGE *page, WT_DSRC_STATS **stats)
{
/*
* All internal pages and overflow pages are trivial, all we track is
* a count of the page type.
*/
switch (page->type) {
case WT_PAGE_COL_FIX:
WT_STAT_INCR(session, stats, btree_column_fix);
WT_STAT_INCRV(
session, stats, btree_entries, page->pg_fix_entries);
break;
case WT_PAGE_COL_INT:
WT_STAT_INCR(session, stats, btree_column_internal);
break;
case WT_PAGE_COL_VAR:
__stat_page_col_var(session, page, stats);
break;
case WT_PAGE_ROW_INT:
__stat_page_row_int(session, page, stats);
break;
case WT_PAGE_ROW_LEAF:
__stat_page_row_leaf(session, page, stats);
break;
WT_ILLEGAL_VALUE(session);
}
return (0);
}
/*
* __stat_page_col_var --
* Stat a WT_PAGE_COL_VAR page.
*/
static int
__stat_page_col_var(WT_PAGE *page, WT_DSRC_STATS *stats)
{
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COL *cip;
WT_INSERT *ins;
WT_UPDATE *upd;
uint32_t i;
int orig_deleted;
unpack = &_unpack;
WT_STAT_INCR(stats, btree_column_variable);
/*
* Walk the page, counting regular and overflow data items, and checking
* to be sure any updates weren't deletions. If the item was updated,
* assume it was updated by an item of the same size (it's expensive to
* figure out if it will require the same space or not, especially if
* there's Huffman encoding).
*/
WT_COL_FOREACH(page, cip, i) {
if ((cell = WT_COL_PTR(page, cip)) == NULL) {
orig_deleted = 1;
WT_STAT_INCR(stats, btree_column_deleted);
} else {
orig_deleted = 0;
__wt_cell_unpack(cell, unpack);
WT_STAT_INCRV(
stats, btree_entries, __wt_cell_rle(unpack));
}
/*
* Walk the insert list, checking for changes. For each insert
* we find, correct the original count based on its state.
*/
WT_SKIP_FOREACH(ins, WT_COL_UPDATE(page, cip)) {
upd = ins->upd;
if (WT_UPDATE_DELETED_ISSET(upd)) {
if (orig_deleted)
continue;
WT_STAT_INCR(stats, btree_column_deleted);
WT_STAT_DECR(stats, btree_entries);
} else {
if (!orig_deleted)
continue;
WT_STAT_DECR(stats, btree_column_deleted);
WT_STAT_INCR(stats, btree_entries);
}
}
}
/*
* __stat_page_col_var --
* Stat a WT_PAGE_COL_VAR page.
*/
static void
__stat_page_col_var(
WT_SESSION_IMPL *session, WT_PAGE *page, WT_DSRC_STATS **stats)
{
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COL *cip;
WT_INSERT *ins;
WT_UPDATE *upd;
uint64_t deleted_cnt, entry_cnt, ovfl_cnt, rle_cnt;
uint32_t i;
bool orig_deleted;
unpack = &_unpack;
deleted_cnt = entry_cnt = ovfl_cnt = rle_cnt = 0;
WT_STAT_INCR(session, stats, btree_column_variable);
/*
* Walk the page counting regular items, adjusting if the item has been
* subsequently deleted or not. This is a mess because 10-item RLE might
* have 3 of the items subsequently deleted. Overflow items are harder,
* we can't know if an updated item will be an overflow item or not; do
* our best, and simply count every overflow item (or RLE set of items)
* we see.
*/
WT_COL_FOREACH(page, cip, i) {
if ((cell = WT_COL_PTR(page, cip)) == NULL) {
orig_deleted = true;
++deleted_cnt;
} else {
orig_deleted = false;
__wt_cell_unpack(cell, unpack);
if (unpack->type == WT_CELL_ADDR_DEL)
orig_deleted = true;
else {
entry_cnt += __wt_cell_rle(unpack);
rle_cnt += __wt_cell_rle(unpack) - 1;
}
if (unpack->ovfl)
++ovfl_cnt;
}
/*
* Walk the insert list, checking for changes. For each insert
* we find, correct the original count based on its state.
*/
WT_SKIP_FOREACH(ins, WT_COL_UPDATE(page, cip)) {
upd = ins->upd;
if (WT_UPDATE_DELETED_ISSET(upd)) {
if (!orig_deleted) {
++deleted_cnt;
--entry_cnt;
}
} else
if (orig_deleted) {
--deleted_cnt;
++entry_cnt;
}
}
}
/*
* __rec_review --
* Get exclusive access to the page and review the page and its subtree
* for conditions that would block its eviction.
*
* The ref and page arguments may appear to be redundant, because usually
* ref->page == page and page->ref == ref. However, we need both because
* (a) there are cases where ref == NULL (e.g., for root page or during
* salvage), and (b) we can't safely look at page->ref until we have a
* hazard reference.
*/
static int
__rec_review(WT_SESSION_IMPL *session,
WT_REF *ref, WT_PAGE *page, uint32_t flags, int top)
{
WT_DECL_RET;
WT_PAGE_MODIFY *mod;
WT_TXN *txn;
uint32_t i;
txn = &session->txn;
/*
* Get exclusive access to the page if our caller doesn't have the tree
* locked down.
*/
if (!LF_ISSET(WT_REC_SINGLE))
WT_RET(__hazard_exclusive(session, ref, top));
/*
* Recurse through the page's subtree: this happens first because we
* have to write pages in depth-first order, otherwise we'll dirty
* pages after we've written them.
*/
if (page->type == WT_PAGE_COL_INT || page->type == WT_PAGE_ROW_INT)
WT_REF_FOREACH(page, ref, i)
switch (ref->state) {
case WT_REF_DISK: /* On-disk */
case WT_REF_DELETED: /* On-disk, deleted */
break;
case WT_REF_MEM: /* In-memory */
WT_RET(__rec_review(
session, ref, ref->page, flags, 0));
break;
case WT_REF_EVICT_WALK: /* Walk point */
case WT_REF_LOCKED: /* Being evicted */
case WT_REF_READING: /* Being read */
return (EBUSY);
}
/*
* Check if this page can be evicted:
*
* Fail if the top-level page is a page expected to be removed from the
* tree as part of eviction (an empty page or a split-merge page). Note
* "split" pages are NOT included in this test, because a split page can
* be separately evicted, at which point it's replaced in its parent by
* a reference to a split-merge page. That's a normal part of the leaf
* page life-cycle if it grows too large and must be pushed out of the
* cache. There is also an exception for empty pages, the root page may
* be empty when evicted, but that only happens when the tree is closed.
*
* Fail if any page in the top-level page's subtree can't be merged into
* its parent. You can't evict a page that references such in-memory
* pages, they must be evicted first. The test is necessary but should
* not fire much: the LRU-based eviction code is biased for leaf pages,
* an internal page shouldn't be selected for LRU-based eviction until
* its children have been evicted. Empty, split and split-merge pages
* are all included in this test, they can all be merged into a parent.
*
* We have to write dirty pages to know their final state, a page marked
* empty may have had records added since reconciliation, a page marked
* split may have had records deleted and no longer need to split.
* Split-merge pages are the exception: they can never be change into
* anything other than a split-merge page and are merged regardless of
* being clean or dirty.
*
* Writing the page is expensive, do a cheap test first: if it doesn't
* appear a subtree page can be merged, quit. It's possible the page
* has been emptied since it was last reconciled, and writing it before
* testing might be worthwhile, but it's more probable we're attempting
* to evict an internal page with live children, and that's a waste of
* time.
*
* We don't do a cheap test for the top-level page: we're not called
* to evict split-merge pages, which means the only interesting case
* is an empty page. If the eviction thread picked an "empty" page
* for eviction, it must have had reason, probably the empty page got
* really, really full.
*/
mod = page->modify;
if (!top && (mod == NULL || !F_ISSET(mod,
WT_PM_REC_EMPTY | WT_PM_REC_SPLIT | WT_PM_REC_SPLIT_MERGE)))
return (EBUSY);
/* If the page is dirty, write it so we know the final state. */
if (__wt_page_is_modified(page) &&
!F_ISSET(mod, WT_PM_REC_SPLIT_MERGE)) {
ret = __wt_rec_write(session, page, NULL, flags);
/* If there are unwritten changes on the page, give up. */
if (ret == 0 &&
!LF_ISSET(WT_REC_SINGLE) && __wt_page_is_modified(page))
ret = EBUSY;
if (ret == EBUSY) {
WT_VERBOSE_RET(session, evict,
"page %p written but not clean", page);
if (F_ISSET(txn, TXN_RUNNING) &&
++txn->eviction_fails >= 100) {
txn->eviction_fails = 0;
ret = WT_DEADLOCK;
WT_STAT_INCR(
S2C(session)->stats, txn_fail_cache);
}
/*
* If there aren't multiple cursors active, there
* are no consistency issues: try to bump our snapshot.
*/
if (session->ncursors <= 1) {
__wt_txn_read_last(session);
__wt_txn_read_first(session);
}
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
__wt_col_leaf_obsolete(session, page);
break;
case WT_PAGE_ROW_LEAF:
__wt_row_leaf_obsolete(session, page);
break;
}
}
WT_RET(ret);
txn->eviction_fails = 0;
}
/*
* Repeat the eviction tests.
*
* Fail if the top-level page should be merged into its parent, and it's
* not the root page.
*
* Fail if a page in the top-level page's subtree can't be merged into
* its parent.
*/
if (top) {
/*
* We never get a top-level split-merge page to evict, they are
* ignored by the eviction thread. Check out of sheer paranoia.
*/
if (mod != NULL) {
if (F_ISSET(mod, WT_PM_REC_SPLIT_MERGE))
return (EBUSY);
if (F_ISSET(mod, WT_PM_REC_EMPTY) &&
!WT_PAGE_IS_ROOT(page))
return (EBUSY);
}
} else if (mod == NULL || !F_ISSET(mod,
WT_PM_REC_EMPTY | WT_PM_REC_SPLIT | WT_PM_REC_SPLIT_MERGE))
return (EBUSY);
return (0);
}
/*
* __wt_rec_evict --
* Reconciliation plus eviction.
*/
int
__wt_rec_evict(WT_SESSION_IMPL *session, WT_PAGE *page, uint32_t flags)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
int single;
conn = S2C(session);
WT_VERBOSE_RET(session, evict,
"page %p (%s)", page, __wt_page_type_string(page->type));
WT_ASSERT(session, session->excl_next == 0);
single = LF_ISSET(WT_REC_SINGLE) ? 1 : 0;
/*
* Get exclusive access to the page and review the page and its subtree
* for conditions that would block our eviction of the page. If the
* check fails (for example, we find a child page that can't be merged),
* we're done. We have to make this check for clean pages, too: while
* unlikely eviction would choose an internal page with children, it's
* not disallowed anywhere.
*
* Note that page->ref may be NULL in some cases (e.g., for root pages
* or during salvage). That's OK if WT_REC_SINGLE is set: we won't
* check hazard references in that case.
*/
WT_ERR(__rec_review(session, page->ref, page, flags, 1));
/* Count evictions of internal pages during normal operation. */
if (!single &&
(page->type == WT_PAGE_COL_INT || page->type == WT_PAGE_ROW_INT))
WT_STAT_INCR(conn->stats, cache_evict_internal);
/* Update the parent and discard the page. */
if (page->modify == NULL || !F_ISSET(page->modify, WT_PM_REC_MASK)) {
WT_STAT_INCR(conn->stats, cache_evict_unmodified);
WT_ASSERT(session, single || page->ref->state == WT_REF_LOCKED);
if (WT_PAGE_IS_ROOT(page))
__rec_root_update(session);
else
__rec_page_clean_update(session, page);
/* Discard the page. */
__rec_discard_page(session, page, single);
} else {
WT_STAT_INCR(conn->stats, cache_evict_modified);
if (WT_PAGE_IS_ROOT(page))
__rec_root_update(session);
else
WT_ERR(__rec_page_dirty_update(session, page));
/* Discard the tree rooted in this page. */
__rec_discard_tree(session, page, single);
}
if (0) {
err: /*
* If unable to evict this page, release exclusive reference(s)
* we've acquired.
*/
__rec_excl_clear(session);
}
session->excl_next = 0;
return (ret);
}
/*
* __wt_lsm_stat_init --
* Initialize a LSM statistics structure.
*/
int
__wt_lsm_stat_init(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_CURSOR_STAT *cst, uint32_t flags)
{
WT_CURSOR *stat_cursor;
WT_DECL_ITEM(uribuf);
WT_DECL_RET;
WT_DSRC_STATS *stats;
WT_LSM_CHUNK *chunk;
const char *cfg[] = API_CONF_DEFAULTS(
session, open_cursor, "statistics_fast=on");
const char *disk_cfg[] = API_CONF_DEFAULTS(session,
open_cursor, "checkpoint=WiredTigerCheckpoint,statistics_fast=on");
const char *desc, *pvalue;
uint64_t value;
u_int i;
int locked, stat_key;
WT_UNUSED(flags);
locked = 0;
WT_ERR(__wt_scr_alloc(session, 0, &uribuf));
/* Clear the statistics we are about to recalculate. */
if (cst->stats != NULL)
stats = (WT_DSRC_STATS *)cst->stats;
else {
WT_ERR(__wt_calloc_def(session, 1, &stats));
__wt_stat_init_dsrc_stats(stats);
cst->stats_first = cst->stats = (WT_STATS *)stats;
cst->stats_count = sizeof(*stats) / sizeof(WT_STATS);
}
*stats = lsm_tree->stats;
if (LF_ISSET(WT_STATISTICS_CLEAR))
__wt_stat_clear_dsrc_stats(&lsm_tree->stats);
/* Hold the LSM lock so that we can safely walk through the chunks. */
WT_ERR(__wt_readlock(session, lsm_tree->rwlock));
locked = 1;
/* Set the stats for this run. */
WT_STAT_SET(stats, lsm_chunk_count, lsm_tree->nchunks);
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
if (chunk->generation >
(uint32_t)WT_STAT(stats, lsm_generation_max))
WT_STAT_SET(stats,
lsm_generation_max, chunk->generation);
/*
* LSM chunk reads happen from a checkpoint, so get the
* statistics for a checkpoint if one exists.
*/
WT_ERR(__wt_buf_fmt(
session, uribuf, "statistics:%s", chunk->uri));
ret = __wt_curstat_open(session, uribuf->data,
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) ? disk_cfg : cfg,
&stat_cursor);
/*
* XXX kludge: we may have an empty chunk where no checkpoint
* was written. If so, try to open the ordinary handle on that
* chunk instead.
*/
if (ret == WT_NOTFOUND && F_ISSET(chunk, WT_LSM_CHUNK_ONDISK))
ret = __wt_curstat_open(
session, uribuf->data, cfg, &stat_cursor);
WT_ERR(ret);
while ((ret = stat_cursor->next(stat_cursor)) == 0) {
WT_ERR(stat_cursor->get_key(stat_cursor, &stat_key));
WT_ERR(stat_cursor->get_value(
stat_cursor, &desc, &pvalue, &value));
WT_STAT_INCRKV(stats, stat_key, value);
}
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(stat_cursor->close(stat_cursor));
if (!F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
continue;
WT_STAT_INCR(stats, bloom_count);
WT_STAT_INCRV(stats, bloom_size,
(chunk->count * lsm_tree->bloom_bit_count) / 8);
WT_ERR(__wt_buf_fmt(
session, uribuf, "statistics:%s", chunk->bloom_uri));
WT_ERR(__wt_curstat_open(session, uribuf->data,
cfg, &stat_cursor));
stat_cursor->set_key(
stat_cursor, WT_STAT_DSRC_CACHE_EVICTION_CLEAN);
WT_ERR(stat_cursor->search(stat_cursor));
WT_ERR(stat_cursor->get_value(
stat_cursor, &desc, &pvalue, &value));
WT_STAT_INCRV(stats, cache_eviction_clean, value);
WT_STAT_INCRV(stats, bloom_page_evict, value);
stat_cursor->set_key(
stat_cursor, WT_STAT_DSRC_CACHE_EVICTION_DIRTY);
WT_ERR(stat_cursor->search(stat_cursor));
WT_ERR(stat_cursor->get_value(
stat_cursor, &desc, &pvalue, &value));
WT_STAT_INCRV(stats, cache_eviction_dirty, value);
WT_STAT_INCRV(stats, bloom_page_evict, value);
stat_cursor->set_key(
stat_cursor, WT_STAT_DSRC_CACHE_EVICTION_FAIL);
WT_ERR(stat_cursor->search(stat_cursor));
WT_ERR(stat_cursor->get_value(
stat_cursor, &desc, &pvalue, &value));
WT_STAT_INCRV(stats, cache_eviction_fail, value);
stat_cursor->set_key(stat_cursor, WT_STAT_DSRC_CACHE_READ);
WT_ERR(stat_cursor->search(stat_cursor));
WT_ERR(stat_cursor->get_value(
stat_cursor, &desc, &pvalue, &value));
WT_STAT_INCRV(stats, cache_read, value);
WT_STAT_INCRV(stats, bloom_page_read, value);
stat_cursor->set_key(stat_cursor, WT_STAT_DSRC_CACHE_WRITE);
WT_ERR(stat_cursor->search(stat_cursor));
WT_ERR(stat_cursor->get_value(
stat_cursor, &desc, &pvalue, &value));
WT_STAT_INCRV(stats, cache_write, value);
WT_ERR(stat_cursor->close(stat_cursor));
}
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
__wt_scr_free(&uribuf);
return (ret);
}
