/*
* __stat_tree_walk --
* Gather btree statistics that require traversing the tree.
*/
static int
__stat_tree_walk(WT_SESSION_IMPL *session)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_DSRC_STATS **stats;
WT_REF *next_walk;
btree = S2BT(session);
stats = btree->dhandle->stats;
/*
* Clear the statistics we're about to count.
*/
WT_STAT_SET(session, stats, btree_column_deleted, 0);
WT_STAT_SET(session, stats, btree_column_fix, 0);
WT_STAT_SET(session, stats, btree_column_internal, 0);
WT_STAT_SET(session, stats, btree_column_rle, 0);
WT_STAT_SET(session, stats, btree_column_variable, 0);
WT_STAT_SET(session, stats, btree_entries, 0);
WT_STAT_SET(session, stats, btree_overflow, 0);
WT_STAT_SET(session, stats, btree_row_internal, 0);
WT_STAT_SET(session, stats, btree_row_leaf, 0);
next_walk = NULL;
while ((ret = __wt_tree_walk(
session, &next_walk, 0)) == 0 && next_walk != NULL) {
WT_WITH_PAGE_INDEX(session,
ret = __stat_page(session, next_walk->page, stats));
WT_RET(ret);
}
return (ret == WT_NOTFOUND ? 0 : ret);
}
/*
* __wt_btree_stat_init --
* Initialize the Btree statistics.
*/
int
__wt_btree_stat_init(WT_SESSION_IMPL *session, uint32_t flags)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
WT_DSRC_STATS *stats;
WT_PAGE *page;
btree = S2BT(session);
bm = btree->bm;
stats = &btree->dhandle->stats;
WT_RET(bm->stat(bm, session, stats));
WT_STAT_SET(stats, btree_fixed_len, btree->bitcnt);
WT_STAT_SET(stats, btree_maximum_depth, btree->maximum_depth);
WT_STAT_SET(stats, btree_maxintlitem, btree->maxintlitem);
WT_STAT_SET(stats, btree_maxintlpage, btree->maxintlpage);
WT_STAT_SET(stats, btree_maxleafitem, btree->maxleafitem);
WT_STAT_SET(stats, btree_maxleafpage, btree->maxleafpage);
page = NULL;
if (LF_ISSET(WT_STATISTICS_FAST))
return (0);
while ((ret = __wt_tree_walk(session, &page, 0)) == 0 && page != NULL)
WT_RET(__stat_page(session, page, stats));
return (ret == WT_NOTFOUND ? 0 : ret);
}
/*
* __wt_compact --
* Compact a file.
*/
int
__wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_BM *bm;
WT_CONFIG_ITEM cval;
WT_DECL_RET;
WT_PAGE *page;
int trigger, skip;
bm = S2BT(session)->bm;
WT_DSTAT_INCR(session, session_compact);
WT_RET(__wt_config_gets(session, cfg, "trigger", &cval));
trigger = (int)cval.val;
/* Check if compaction might be useful. */
WT_RET(bm->compact_skip(bm, session, trigger, &skip));
if (skip)
return (0);
/*
* Walk the cache reviewing in-memory pages to see if they need to be
* re-written. This requires looking at page reconciliation results,
* which means the page cannot be reconciled at the same time as it's
* being reviewed for compaction. The underlying functions ensure we
* don't collide with page eviction, but we need to make sure we don't
* collide with checkpoints either, they are the other operation that
* can reconcile a page.
*/
__wt_spin_lock(session, &S2C(session)->metadata_lock);
WT_RET(__wt_bt_cache_op(session, NULL, WT_SYNC_COMPACT));
__wt_spin_unlock(session, &S2C(session)->metadata_lock);
/*
* Walk the tree, reviewing on-disk pages to see if they need to be
* re-written.
*/
for (page = NULL;;) {
WT_RET(__wt_tree_walk(session, &page, WT_TREE_COMPACT));
if (page == NULL)
break;
/*
* The only pages returned by the tree walk function are pages
* we want to re-write; mark the page and tree dirty.
*/
if ((ret = __wt_page_modify_init(session, page)) != 0) {
WT_TRET(__wt_page_release(session, page));
WT_RET(ret);
}
__wt_page_and_tree_modify_set(session, page);
WT_DSTAT_INCR(session, btree_compact_rewrite);
}
return (0);
}
/*
* __wt_btree_stat_init --
* Initialize the Btree statistics.
*/
int
__wt_btree_stat_init(WT_SESSION_IMPL *session, WT_CURSOR_STAT *cst)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
WT_DSRC_STATS **stats;
WT_REF *next_walk;
btree = S2BT(session);
bm = btree->bm;
stats = btree->dhandle->stats;
WT_RET(bm->stat(bm, session, stats[0]));
WT_STAT_SET(session, stats, btree_fixed_len, btree->bitcnt);
WT_STAT_SET(session, stats, btree_maximum_depth, btree->maximum_depth);
WT_STAT_SET(session, stats, btree_maxintlkey, btree->maxintlkey);
WT_STAT_SET(session, stats, btree_maxintlpage, btree->maxintlpage);
WT_STAT_SET(session, stats, btree_maxleafkey, btree->maxleafkey);
WT_STAT_SET(session, stats, btree_maxleafpage, btree->maxleafpage);
WT_STAT_SET(session, stats, btree_maxleafvalue, btree->maxleafvalue);
/* Everything else is really, really expensive. */
if (!F_ISSET(cst, WT_CONN_STAT_ALL))
return (0);
/*
* Clear the statistics we're about to count.
*/
WT_STAT_SET(session, stats, btree_column_deleted, 0);
WT_STAT_SET(session, stats, btree_column_fix, 0);
WT_STAT_SET(session, stats, btree_column_internal, 0);
WT_STAT_SET(session, stats, btree_column_rle, 0);
WT_STAT_SET(session, stats, btree_column_variable, 0);
WT_STAT_SET(session, stats, btree_entries, 0);
WT_STAT_SET(session, stats, btree_overflow, 0);
WT_STAT_SET(session, stats, btree_row_internal, 0);
WT_STAT_SET(session, stats, btree_row_leaf, 0);
next_walk = NULL;
while ((ret = __wt_tree_walk(
session, &next_walk, 0)) == 0 && next_walk != NULL) {
WT_WITH_PAGE_INDEX(session,
ret = __stat_page(session, next_walk->page, stats));
WT_RET(ret);
}
return (ret == WT_NOTFOUND ? 0 : ret);
}
/*
* __sync_file --
* Flush pages for a specific file.
*/
static int
__sync_file(WT_SESSION_IMPL *session, WT_CACHE_OP syncop)
{
struct timespec end, start;
WT_BTREE *btree;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_REF *walk;
WT_TXN *txn;
uint64_t internal_bytes, internal_pages, leaf_bytes, leaf_pages;
uint64_t oldest_id, saved_snap_min;
uint32_t flags;
conn = S2C(session);
btree = S2BT(session);
walk = NULL;
txn = &session->txn;
saved_snap_min = WT_SESSION_TXN_STATE(session)->snap_min;
flags = WT_READ_CACHE | WT_READ_NO_GEN;
internal_bytes = leaf_bytes = 0;
internal_pages = leaf_pages = 0;
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT))
WT_RET(__wt_epoch(session, &start));
switch (syncop) {
case WT_SYNC_WRITE_LEAVES:
/*
* Write all immediately available, dirty in-cache leaf pages.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time.
*/
if (!btree->modified)
return (0);
__wt_spin_lock(session, &btree->flush_lock);
if (!btree->modified) {
__wt_spin_unlock(session, &btree->flush_lock);
return (0);
}
/*
* Save the oldest transaction ID we need to keep around.
* Otherwise, in a busy system, we could be updating pages so
* fast that write leaves never catches up. We deliberately
* have no transaction running at this point that would keep
* the oldest ID from moving forwards as we walk the tree.
*/
oldest_id = __wt_txn_oldest_id(session);
flags |= WT_READ_NO_WAIT | WT_READ_SKIP_INTL;
for (walk = NULL;;) {
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/*
* Write dirty pages if nobody beat us to it. Don't
* try to write hot pages (defined as pages that have
* been updated since the write phase leaves started):
* checkpoint will have to visit them anyway.
*/
page = walk->page;
if (__wt_page_is_modified(page) &&
WT_TXNID_LT(page->modify->update_txn, oldest_id)) {
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
leaf_bytes += page->memory_footprint;
++leaf_pages;
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
}
break;
case WT_SYNC_CHECKPOINT:
/*
* If we are flushing a file at read-committed isolation, which
* is of particular interest for flushing the metadata to make
* schema-changing operation durable, get a transactional
* snapshot now.
*
* All changes committed up to this point should be included.
* We don't update the snapshot in between pages because (a)
* the metadata shouldn't be that big, and (b) if we do ever
*/
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
/*
* We cannot check the tree modified flag in the case of a
* checkpoint, the checkpoint code has already cleared it.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time. We're holding the schema lock, but need the
* lower-level lock as well.
*/
__wt_spin_lock(session, &btree->flush_lock);
/*
* In the final checkpoint pass, child pages cannot be evicted
* from underneath internal pages nor can underlying blocks be
* freed until the checkpoint's block lists are stable. Also,
* we cannot split child pages into parents unless we know the
* final pass will write a consistent view of that namespace.
* Set the checkpointing flag to block such actions and wait for
* any problematic eviction or page splits to complete.
*/
WT_PUBLISH(btree->checkpointing, WT_CKPT_PREPARE);
WT_ERR(__wt_evict_file_exclusive_on(session));
__wt_evict_file_exclusive_off(session);
WT_PUBLISH(btree->checkpointing, WT_CKPT_RUNNING);
/* Write all dirty in-cache pages. */
flags |= WT_READ_NO_EVICT;
for (walk = NULL;;) {
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/* Skip clean pages. */
if (!__wt_page_is_modified(walk->page))
continue;
/*
* Take a local reference to the page modify structure
* now that we know the page is dirty. It needs to be
* done in this order otherwise the page modify
* structure could have been created between taking the
* reference and checking modified.
*/
page = walk->page;
mod = page->modify;
/*
* Write dirty pages, unless we can be sure they only
* became dirty after the checkpoint started.
*
* We can skip dirty pages if:
* (1) they are leaf pages;
* (2) there is a snapshot transaction active (which
* is the case in ordinary application checkpoints
* but not all internal cases); and
* (3) the first dirty update on the page is
* sufficiently recent that the checkpoint
* transaction would skip them.
*
* Mark the tree dirty: the checkpoint marked it clean
* and we can't skip future checkpoints until this page
* is written.
*/
if (!WT_PAGE_IS_INTERNAL(page) &&
F_ISSET(txn, WT_TXN_HAS_SNAPSHOT) &&
WT_TXNID_LT(txn->snap_max, mod->first_dirty_txn)) {
__wt_page_modify_set(session, page);
continue;
}
if (WT_PAGE_IS_INTERNAL(page)) {
internal_bytes += page->memory_footprint;
++internal_pages;
} else {
leaf_bytes += page->memory_footprint;
++leaf_pages;
}
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
break;
case WT_SYNC_CLOSE:
case WT_SYNC_DISCARD:
WT_ILLEGAL_VALUE_ERR(session);
}
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT)) {
WT_ERR(__wt_epoch(session, &end));
WT_ERR(__wt_verbose(session, WT_VERB_CHECKPOINT,
"__sync_file WT_SYNC_%s wrote:\n\t %" PRIu64
" bytes, %" PRIu64 " pages of leaves\n\t %" PRIu64
" bytes, %" PRIu64 " pages of internal\n\t"
"Took: %" PRIu64 "ms",
syncop == WT_SYNC_WRITE_LEAVES ?
"WRITE_LEAVES" : "CHECKPOINT",
leaf_bytes, leaf_pages, internal_bytes, internal_pages,
WT_TIMEDIFF_MS(end, start)));
}
err: /* On error, clear any left-over tree walk. */
if (walk != NULL)
WT_TRET(__wt_page_release(session, walk, flags));
/*
* If we got a snapshot in order to write pages, and there was no
* snapshot active when we started, release it.
*/
if (txn->isolation == WT_ISO_READ_COMMITTED &&
saved_snap_min == WT_TXN_NONE)
__wt_txn_release_snapshot(session);
if (btree->checkpointing != WT_CKPT_OFF) {
/*
* Update the checkpoint generation for this handle so visible
* updates newer than the checkpoint can be evicted.
*
* This has to be published before eviction is enabled again,
* so that eviction knows that the checkpoint has completed.
*/
WT_PUBLISH(btree->checkpoint_gen,
conn->txn_global.checkpoint_gen);
WT_STAT_FAST_DATA_SET(session,
btree_checkpoint_generation, btree->checkpoint_gen);
/*
* Clear the checkpoint flag and push the change; not required,
* but publishing the change means stalled eviction gets moving
* as soon as possible.
*/
btree->checkpointing = WT_CKPT_OFF;
WT_FULL_BARRIER();
/*
* If this tree was being skipped by the eviction server during
* the checkpoint, clear the wait.
*/
btree->evict_walk_period = 0;
/*
* Wake the eviction server, in case application threads have
* stalled while the eviction server decided it couldn't make
* progress. Without this, application threads will be stalled
* until the eviction server next wakes.
*/
WT_TRET(__wt_evict_server_wake(session));
}
__wt_spin_unlock(session, &btree->flush_lock);
/*
* Leaves are written before a checkpoint (or as part of a file close,
* before checkpointing the file). Start a flush to stable storage,
* but don't wait for it.
*/
if (ret == 0 &&
syncop == WT_SYNC_WRITE_LEAVES && F_ISSET(conn, WT_CONN_CKPT_SYNC))
WT_RET(btree->bm->sync(btree->bm, session, true));
return (ret);
}
/*将btree cursor移动到下一个记录*/
int __wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
/*btree 扫描标示*/
flags = WT_READ_SKIP_INTL;
if (truncating)
LF_SET(WT_READ_TRUNCATE);
/*激活一个btree cursor*/
WT_RET(__cursor_func_init(cbt, 0));
/*初始化cursor*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*对btree的扫描*/
for (;;){
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
/*column store append方式，在insert header上做扫描*/
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)){
switch (page->type){
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
/*清除掉column store的标记*/
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
}
else if (page != NULL){
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/*找到对应的记录了，直接返回*/
if (ret != WT_NOTFOUND)
break;
/*假如是column store方式，检查是否要扫描insert header list*/
if (page->type != WT_PAGE_ROW_LEAF && (cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*删除的记录太多，对page进行重组，增大page的填充因子*/
if (page != NULL && (cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD || (newpage && cbt->page_deleted_count > 0))){
__wt_page_evict_soon(page);
}
cbt->page_deleted_count = 0;
/*btree cursor跳转到下一个page上*/
WT_ERR(__wt_tree_walk(session, &cbt->ref, NULL, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err:
if (ret != 0) /*失败了，恢复cursor的状态*/
WT_TRET(__cursor_reset(cbt));
return ret;
}
/*
* __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);
}
/*
* __wt_compact --
* Compact a file.
*/
int
__wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
WT_REF *ref;
bool block_manager_begin, skip;
WT_UNUSED(cfg);
btree = S2BT(session);
bm = btree->bm;
ref = NULL;
block_manager_begin = false;
WT_STAT_FAST_DATA_INCR(session, session_compact);
/*
* Check if compaction might be useful -- the API layer will quit trying
* to compact the data source if we make no progress, set a flag if the
* block layer thinks compaction is possible.
*/
WT_RET(bm->compact_skip(bm, session, &skip));
if (skip)
return (0);
/*
* Reviewing in-memory pages requires looking at page reconciliation
* results, because we care about where the page is stored now, not
* where the page was stored when we first read it into the cache.
* We need to ensure we don't race with page reconciliation as it's
* writing the page modify information.
*
* There are three ways we call reconciliation: checkpoints, threads
* writing leaf pages (usually in preparation for a checkpoint or if
* closing a file), and eviction.
*
* We're holding the schema lock which serializes with checkpoints.
*/
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SCHEMA));
/*
* Get the tree handle's flush lock which blocks threads writing leaf
* pages.
*/
__wt_spin_lock(session, &btree->flush_lock);
/* Start compaction. */
WT_ERR(bm->compact_start(bm, session));
block_manager_begin = true;
/* Walk the tree reviewing pages to see if they should be re-written. */
for (;;) {
/*
* Pages read for compaction aren't "useful"; don't update the
* read generation of pages already in memory, and if a page is
* read, set its generation to a low value so it is evicted
* quickly.
*/
WT_ERR(__wt_tree_walk(session, &ref,
WT_READ_COMPACT | WT_READ_NO_GEN | WT_READ_WONT_NEED));
if (ref == NULL)
break;
WT_ERR(__compact_rewrite(session, ref, &skip));
if (skip)
continue;
session->compact_state = WT_COMPACT_SUCCESS;
/* Rewrite the page: mark the page and tree dirty. */
WT_ERR(__wt_page_modify_init(session, ref->page));
__wt_page_modify_set(session, ref->page);
WT_STAT_FAST_DATA_INCR(session, btree_compact_rewrite);
}
err:
if (ref != NULL)
WT_TRET(__wt_page_release(session, ref, 0));
if (block_manager_begin)
WT_TRET(bm->compact_end(bm, session));
/* Unblock threads writing leaf pages. */
__wt_spin_unlock(session, &btree->flush_lock);
return (ret);
}
/*
* __wt_compact --
* Compact a file.
*/
int
__wt_compact(WT_SESSION_IMPL *session)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
WT_REF *ref;
u_int i;
bool skip;
btree = S2BT(session);
bm = btree->bm;
ref = NULL;
WT_STAT_DATA_INCR(session, session_compact);
/*
* Check if compaction might be useful -- the API layer will quit trying
* to compact the data source if we make no progress, set a flag if the
* block layer thinks compaction is possible.
*/
WT_RET(bm->compact_skip(bm, session, &skip));
if (skip)
return (0);
/*
* Reviewing in-memory pages requires looking at page reconciliation
* results, because we care about where the page is stored now, not
* where the page was stored when we first read it into the cache.
* We need to ensure we don't race with page reconciliation as it's
* writing the page modify information.
*
* There are two ways we call reconciliation: checkpoints and eviction.
* Get the tree's flush lock which blocks threads writing pages for
* checkpoints.
*/
__wt_spin_lock(session, &btree->flush_lock);
/* Walk the tree reviewing pages to see if they should be re-written. */
for (i = 0;;) {
/* Periodically check if we've run out of time. */
if (++i > 100) {
WT_ERR(__wt_session_compact_check_timeout(session));
i = 0;
}
/*
* Pages read for compaction aren't "useful"; don't update the
* read generation of pages already in memory, and if a page is
* read, set its generation to a low value so it is evicted
* quickly.
*/
WT_ERR(__wt_tree_walk(session, &ref,
WT_READ_COMPACT | WT_READ_NO_GEN | WT_READ_WONT_NEED));
if (ref == NULL)
break;
WT_ERR(__compact_rewrite(session, ref, &skip));
if (skip)
continue;
session->compact_state = WT_COMPACT_SUCCESS;
/* Rewrite the page: mark the page and tree dirty. */
WT_ERR(__wt_page_modify_init(session, ref->page));
__wt_page_modify_set(session, ref->page);
WT_STAT_DATA_INCR(session, btree_compact_rewrite);
}
err: if (ref != NULL)
WT_TRET(__wt_page_release(session, ref, 0));
/* Unblock threads writing leaf pages. */
__wt_spin_unlock(session, &btree->flush_lock);
return (ret);
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, bool truncating)
{
WT_CURSOR *cursor;
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
bool newpage;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_CONN_INCR(session, cursor_next);
WT_STAT_DATA_INCR(session, cursor_next);
F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET);
WT_RET(__cursor_func_init(cbt, false));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
flags = WT_READ_SKIP_INTL; /* tree walk flags */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
for (newpage = false;; newpage = true) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* Column-store pages may have appended entries. Handle
* it separately from the usual cursor code, it's in a
* simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*
* If we saw a lot of deleted records on this page, or we went
* all the way through a page and only saw deleted records, try
* to evict the page when we release it. Otherwise repeatedly
* deleting from the beginning of a tree can have quadratic
* performance. Take care not to force eviction of pages that
* are genuinely empty, in new trees.
*/
if (page != NULL &&
(cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD ||
(newpage && cbt->page_deleted_count > 0)))
__wt_page_evict_soon(session, cbt->ref);
cbt->page_deleted_count = 0;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
#ifdef HAVE_DIAGNOSTIC
if (ret == 0)
WT_ERR(__wt_cursor_key_order_check(session, cbt, true));
#endif
if (ret == 0)
F_SET(cursor, WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT);
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_evict_file --
* Discard pages for a specific file.
*/
int
__wt_evict_file(WT_SESSION_IMPL *session, int syncop)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
WT_REF *next_ref, *ref;
int eviction_enabled;
btree = S2BT(session);
eviction_enabled = !F_ISSET(btree, WT_BTREE_NO_EVICTION);
/*
* We need exclusive access to the file -- disable ordinary eviction
* and drain any blocks already queued.
*/
if (eviction_enabled)
WT_RET(__wt_evict_file_exclusive_on(session));
/* Make sure the oldest transaction ID is up-to-date. */
__wt_txn_update_oldest(session);
/* Walk the tree, discarding pages. */
next_ref = NULL;
WT_ERR(__wt_tree_walk(
session, &next_ref, 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, WT_READ_CACHE | WT_READ_NO_EVICT));
switch (syncop) {
case WT_SYNC_CLOSE:
/*
* Evict the page.
* Do not attempt to evict pages expected to be merged
* into their parents, with the exception that the root
* page can't be merged, it must be written.
*/
if (__wt_ref_is_root(ref) ||
page->modify == NULL ||
!F_ISSET(page->modify, WT_PM_REC_EMPTY))
WT_ERR(__wt_evict(session, ref, 1));
break;
case WT_SYNC_DISCARD:
/*
* Ordinary discard of the page, whether clean or dirty.
* If we see a dirty page in an ordinary discard (e.g.,
* from sweep), give up: an update must have happened
* since the file was selected for sweeping.
*/
if (__wt_page_is_modified(page))
WT_ERR(EBUSY);
/*
* If the page contains an update that is too recent to
* evict, stop. This should never happen during
* connection close, but in other paths our caller
* should be prepared to deal with this case.
*/
if (page->modify != NULL &&
!__wt_txn_visible_all(session,
page->modify->rec_max_txn))
WT_ERR(EBUSY);
__wt_evict_page_clean_update(session, ref);
break;
case WT_SYNC_DISCARD_FORCE:
/*
* Forced discard of the page, whether clean or dirty.
* If we see a dirty page in a forced discard, clean
* the page, both to keep statistics correct, and to
* let the page-discard function assert no dirty page
* is ever discarded.
*/
if (__wt_page_is_modified(page)) {
page->modify->write_gen = 0;
__wt_cache_dirty_decr(session, page);
}
F_SET(session, WT_SESSION_DISCARD_FORCE);
__wt_evict_page_clean_update(session, ref);
F_CLR(session, WT_SESSION_DISCARD_FORCE);
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 (eviction_enabled)
__wt_evict_file_exclusive_off(session);
return (ret);
}
/*对文件进行compact操作*/
int __wt_compact(WT_SESSION_IMPL* session, const char* cfg[])
{
WT_BM *bm;
WT_BTREE *btree;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_REF *ref;
int block_manager_begin, evict_reset, skip;
WT_UNUSED(cfg);
conn = S2C(session);
btree = S2BT(session);
bm = btree->bm;
ref = NULL;
block_manager_begin = 0;
WT_STAT_FAST_DATA_INCR(session, session_compact);
/*检查bm对相应的blocks是否可以compact,如果不可以，直接返回*/
WT_RET(bm->compact_skip(bm, session, &skip));
if (skip)
return 0;
/*
* Reviewing in-memory pages requires looking at page reconciliation
* results, because we care about where the page is stored now, not
* where the page was stored when we first read it into the cache.
* We need to ensure we don't race with page reconciliation as it's
* writing the page modify information.
*
* There are three ways we call reconciliation: checkpoints, threads
* writing leaf pages (usually in preparation for a checkpoint or if
* closing a file), and eviction.
*
* We're holding the schema lock which serializes with checkpoints.
*/
WT_ASSERT(session, F_ISSET(session, WT_SESSION_SCHEMA_LOCKED));
/*获得btree flusk_lock,防止在文件空间compact被其他线程flush*/
__wt_spin_lock(session, &btree->flush_lock);
conn->compact_in_memory_pass = 1;
WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
if (evict_reset)
__wt_evict_file_exclusive_off(session);
WT_ERR(bm->compact_start(bm, session));
block_manager_begin = 1;
session->compaction = 1;
for (;;){
WT_ERR(__wt_tree_walk(session, &ref, NULL, WT_READ_COMPACT | WT_READ_NO_GEN | WT_READ_WONT_NEED));
if (ref == NULL)
break;
/*进行compact标记*/
WT_ERR(__compact_rewrite(session, ref, &skip));
if (skip)
continue;
/*如果需要compact的page需要标记为脏page,通过内存驱逐来回写compact结果*/
WT_ERR(__wt_page_modify_init(session, ref->page));
__wt_page_modify_set(session, ref->page);
WT_STAT_FAST_DATA_INCR(session, btree_compact_rewrite);
}
err:
if (ref != NULL)
WT_TRET(__wt_page_release(session, ref, 0));
/*结束compact动作*/
if (block_manager_begin)
WT_TRET(bm->compact_end(bm, session));
/*
* Unlock will be a release barrier, use it to update the compaction
* status for reconciliation.
*/
conn->compact_in_memory_pass = 0;
__wt_spin_unlock(session, &btree->flush_lock);
return ret;
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int skipped, newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
flags = WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__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_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
for (skipped = newpage = 0;; skipped = 0, newpage = 1) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(
cbt, newpage, &skipped);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage, &skipped);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage, &skipped);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* 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 (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*
* If we scanned all the way through a page and only saw
* deleted records, try to evict the page as we release it.
* Otherwise repeatedly deleting from the beginning of a tree
* can have quadratic performance.
*/
if (newpage && skipped)
page->read_gen = WT_READGEN_OLDEST;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __sync_file --
* Flush pages for a specific file.
*/
static int
__sync_file(WT_SESSION_IMPL *session, int syncop)
{
struct timespec end, start;
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_REF *walk;
WT_TXN *txn;
uint64_t internal_bytes, leaf_bytes;
uint64_t internal_pages, leaf_pages;
uint32_t flags;
bool evict_reset;
btree = S2BT(session);
flags = WT_READ_CACHE | WT_READ_NO_GEN;
walk = NULL;
txn = &session->txn;
internal_bytes = leaf_bytes = 0;
internal_pages = leaf_pages = 0;
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT))
WT_RET(__wt_epoch(session, &start));
switch (syncop) {
case WT_SYNC_WRITE_LEAVES:
/*
* Write all immediately available, dirty in-cache leaf pages.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time.
*/
if (!btree->modified)
return (0);
__wt_spin_lock(session, &btree->flush_lock);
if (!btree->modified) {
__wt_spin_unlock(session, &btree->flush_lock);
return (0);
}
flags |= WT_READ_NO_WAIT | WT_READ_SKIP_INTL;
for (walk = NULL;;) {
WT_ERR(__wt_tree_walk(session, &walk, NULL, flags));
if (walk == NULL)
break;
/*
* Write dirty pages if nobody beat us to it. Don't
* try to write the hottest pages: checkpoint will have
* to visit them anyway.
*/
page = walk->page;
if (__wt_page_is_modified(page) &&
__wt_txn_visible_all(
session, page->modify->update_txn)) {
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
leaf_bytes += page->memory_footprint;
++leaf_pages;
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
}
break;
case WT_SYNC_CHECKPOINT:
/*
* We cannot check the tree modified flag in the case of a
* checkpoint, the checkpoint code has already cleared it.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time. We're holding the schema lock, but need the
* lower-level lock as well.
*/
__wt_spin_lock(session, &btree->flush_lock);
/*
* When internal pages are being reconciled by checkpoint their
* child pages cannot disappear from underneath them or be split
* into them, nor can underlying blocks be freed until the block
* lists for the checkpoint are stable. Set the checkpointing
* flag to block eviction of dirty pages until the checkpoint's
* internal page pass is complete, then wait for any existing
* eviction to complete.
*/
btree->checkpointing = 1;
WT_FULL_BARRIER();
WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
if (evict_reset)
__wt_evict_file_exclusive_off(session);
/* Write all dirty in-cache pages. */
flags |= WT_READ_NO_EVICT;
for (walk = NULL;;) {
/*
* If we have a page, and it was ever modified, track
* the highest transaction ID in the tree. We do this
* here because we want the value after reconciling
* dirty pages.
*/
if (walk != NULL && walk->page != NULL &&
(mod = walk->page->modify) != NULL &&
WT_TXNID_LT(btree->rec_max_txn, mod->rec_max_txn))
btree->rec_max_txn = mod->rec_max_txn;
WT_ERR(__wt_tree_walk(session, &walk, NULL, flags));
if (walk == NULL)
break;
page = walk->page;
mod = page->modify;
/* Skip clean pages. */
if (!__wt_page_is_modified(page))
continue;
/*
* Write dirty pages, unless we can be sure they only
* became dirty after the checkpoint started.
*
* We can skip dirty pages if:
* (1) they are leaf pages;
* (2) there is a snapshot transaction active (which
* is the case in ordinary application checkpoints
* but not all internal cases); and
* (3) the first dirty update on the page is
* sufficiently recent that the checkpoint
* transaction would skip them.
*
* Mark the tree dirty: the checkpoint marked it clean
* and we can't skip future checkpoints until this page
* is written.
*/
if (!WT_PAGE_IS_INTERNAL(page) &&
F_ISSET(txn, WT_TXN_HAS_SNAPSHOT) &&
WT_TXNID_LT(txn->snap_max, mod->first_dirty_txn) &&
mod->rec_result != WT_PM_REC_REWRITE) {
__wt_page_modify_set(session, page);
continue;
}
if (WT_PAGE_IS_INTERNAL(page)) {
internal_bytes += page->memory_footprint;
++internal_pages;
} else {
leaf_bytes += page->memory_footprint;
++leaf_pages;
}
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
break;
}
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT)) {
WT_ERR(__wt_epoch(session, &end));
WT_ERR(__wt_verbose(session, WT_VERB_CHECKPOINT,
"__sync_file WT_SYNC_%s wrote:\n\t %" PRIu64
" bytes, %" PRIu64 " pages of leaves\n\t %" PRIu64
" bytes, %" PRIu64 " pages of internal\n\t"
"Took: %" PRIu64 "ms",
syncop == WT_SYNC_WRITE_LEAVES ?
"WRITE_LEAVES" : "CHECKPOINT",
leaf_bytes, leaf_pages, internal_bytes, internal_pages,
WT_TIMEDIFF(end, start) / WT_MILLION));
}
err: /* On error, clear any left-over tree walk. */
if (walk != NULL)
WT_TRET(__wt_page_release(session, walk, flags));
if (txn->isolation == WT_ISO_READ_COMMITTED && session->ncursors == 0)
__wt_txn_release_snapshot(session);
if (btree->checkpointing) {
/*
* Update the checkpoint generation for this handle so visible
* updates newer than the checkpoint can be evicted.
*
* This has to be published before eviction is enabled again,
* so that eviction knows that the checkpoint has completed.
*/
WT_PUBLISH(btree->checkpoint_gen,
S2C(session)->txn_global.checkpoint_gen);
WT_STAT_FAST_DATA_SET(session,
btree_checkpoint_generation, btree->checkpoint_gen);
/*
* Clear the checkpoint flag and push the change; not required,
* but publishing the change means stalled eviction gets moving
* as soon as possible.
*/
btree->checkpointing = 0;
WT_FULL_BARRIER();
/*
* If this tree was being skipped by the eviction server during
* the checkpoint, clear the wait.
*/
btree->evict_walk_period = 0;
/*
* Wake the eviction server, in case application threads have
* stalled while the eviction server decided it couldn't make
* progress. Without this, application threads will be stalled
* until the eviction server next wakes.
*/
WT_TRET(__wt_evict_server_wake(session));
}
__wt_spin_unlock(session, &btree->flush_lock);
/*
* Leaves are written before a checkpoint (or as part of a file close,
* before checkpointing the file). Start a flush to stable storage,
* but don't wait for it.
*/
if (ret == 0 && syncop == WT_SYNC_WRITE_LEAVES)
WT_RET(btree->bm->sync(btree->bm, session, true));
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, int discard)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_DSTAT_INCR(session, cursor_prev);
flags = WT_TREE_SKIP_INTL | WT_TREE_PREV; /* Tree walk flags. */
if (discard)
LF_SET(WT_TREE_DISCARD);
retry: WT_RET(__cursor_func_init(cbt, 0));
__cursor_position_clear(cbt);
/*
* 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);
/*
* If this is a modification, we're about to read information from the
* page, save the write generation.
*/
page = cbt->page;
if (discard && page != NULL) {
WT_ERR(__wt_page_modify_init(session, page));
WT_ORDERED_READ(cbt->write_gen, page->modify->write_gen);
}
/*
* 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 (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 (page != NULL) {
switch (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;
}
cbt->page = NULL;
WT_ERR(__wt_tree_walk(session, &page, flags));
WT_ERR_TEST(page == NULL, WT_NOTFOUND);
WT_ASSERT(session,
page->type != WT_PAGE_COL_INT &&
page->type != WT_PAGE_ROW_INT);
cbt->page = page;
/* Initialize the page's modification information */
if (discard) {
WT_ERR(__wt_page_modify_init(session, page));
WT_ORDERED_READ(
cbt->write_gen, page->modify->write_gen);
}
/*
* 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 (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
err: if (ret == WT_RESTART)
goto retry;
WT_TRET(__cursor_func_resolve(cbt, ret));
return (ret);
}
/*
* __sync_file --
* Flush pages for a specific file.
*/
static int
__sync_file(WT_SESSION_IMPL *session, WT_CACHE_OP syncop)
{
WT_BTREE *btree;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_REF *prev, *walk;
WT_TXN *txn;
uint64_t internal_bytes, internal_pages, leaf_bytes, leaf_pages;
uint64_t oldest_id, saved_pinned_id, time_start, time_stop;
uint32_t flags;
bool timer, tried_eviction;
conn = S2C(session);
btree = S2BT(session);
prev = walk = NULL;
txn = &session->txn;
tried_eviction = false;
time_start = time_stop = 0;
/* Only visit pages in cache and don't bump page read generations. */
flags = WT_READ_CACHE | WT_READ_NO_GEN;
/*
* Skip all deleted pages. For a page to be marked deleted, it must
* have been evicted from cache and marked clean. Checkpoint should
* never instantiate deleted pages: if a truncate is not visible to the
* checkpoint, the on-disk version is correct. If the truncate is
* visible, we skip over the child page when writing its parent. We
* check whether a truncate is visible in the checkpoint as part of
* reconciling internal pages (specifically in __rec_child_modify).
*/
LF_SET(WT_READ_DELETED_SKIP);
internal_bytes = leaf_bytes = 0;
internal_pages = leaf_pages = 0;
saved_pinned_id = WT_SESSION_TXN_STATE(session)->pinned_id;
timer = WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT);
if (timer)
time_start = __wt_clock(session);
switch (syncop) {
case WT_SYNC_WRITE_LEAVES:
/*
* Write all immediately available, dirty in-cache leaf pages.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time.
*/
if (!btree->modified)
return (0);
__wt_spin_lock(session, &btree->flush_lock);
if (!btree->modified) {
__wt_spin_unlock(session, &btree->flush_lock);
return (0);
}
/*
* Save the oldest transaction ID we need to keep around.
* Otherwise, in a busy system, we could be updating pages so
* fast that write leaves never catches up. We deliberately
* have no transaction running at this point that would keep
* the oldest ID from moving forwards as we walk the tree.
*/
oldest_id = __wt_txn_oldest_id(session);
LF_SET(WT_READ_NO_WAIT | WT_READ_SKIP_INTL);
for (;;) {
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/*
* Write dirty pages if nobody beat us to it. Don't
* try to write hot pages (defined as pages that have
* been updated since the write phase leaves started):
* checkpoint will have to visit them anyway.
*/
page = walk->page;
if (__wt_page_is_modified(page) &&
WT_TXNID_LT(page->modify->update_txn, oldest_id)) {
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
leaf_bytes += page->memory_footprint;
++leaf_pages;
WT_ERR(__wt_reconcile(session,
walk, NULL, WT_REC_CHECKPOINT, NULL));
}
}
break;
case WT_SYNC_CHECKPOINT:
/*
* If we are flushing a file at read-committed isolation, which
* is of particular interest for flushing the metadata to make
* a schema-changing operation durable, get a transactional
* snapshot now.
*
* All changes committed up to this point should be included.
* We don't update the snapshot in between pages because the
* metadata shouldn't have many pages. Instead, read-committed
* isolation ensures that all metadata updates completed before
* the checkpoint are included.
*/
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
/*
* We cannot check the tree modified flag in the case of a
* checkpoint, the checkpoint code has already cleared it.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time. We're holding the schema lock, but need the
* lower-level lock as well.
*/
__wt_spin_lock(session, &btree->flush_lock);
/*
* In the final checkpoint pass, child pages cannot be evicted
* from underneath internal pages nor can underlying blocks be
* freed until the checkpoint's block lists are stable. Also,
* we cannot split child pages into parents unless we know the
* final pass will write a consistent view of that namespace.
* Set the checkpointing flag to block such actions and wait for
* any problematic eviction or page splits to complete.
*/
WT_ASSERT(session, btree->syncing == WT_BTREE_SYNC_OFF &&
btree->sync_session == NULL);
btree->sync_session = session;
btree->syncing = WT_BTREE_SYNC_WAIT;
(void)__wt_gen_next_drain(session, WT_GEN_EVICT);
btree->syncing = WT_BTREE_SYNC_RUNNING;
/* Write all dirty in-cache pages. */
LF_SET(WT_READ_NO_EVICT);
/* Read pages with lookaside entries and evict them asap. */
LF_SET(WT_READ_LOOKASIDE | WT_READ_WONT_NEED);
for (;;) {
WT_ERR(__sync_dup_walk(session, walk, flags, &prev));
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/*
* Skip clean pages, but need to make sure maximum
* transaction ID is always updated.
*/
if (!__wt_page_is_modified(walk->page)) {
if (((mod = walk->page->modify) != NULL) &&
mod->rec_max_txn > btree->rec_max_txn)
btree->rec_max_txn = mod->rec_max_txn;
if (mod != NULL &&
btree->rec_max_timestamp <
mod->rec_max_timestamp)
btree->rec_max_timestamp =
mod->rec_max_timestamp;
continue;
}
/*
* Take a local reference to the page modify structure
* now that we know the page is dirty. It needs to be
* done in this order otherwise the page modify
* structure could have been created between taking the
* reference and checking modified.
*/
page = walk->page;
/*
* Write dirty pages, if we can't skip them. If we skip
* a page, mark the tree dirty. The checkpoint marked it
* clean and we can't skip future checkpoints until this
* page is written.
*/
if (__sync_checkpoint_can_skip(session, page)) {
__wt_tree_modify_set(session);
continue;
}
if (WT_PAGE_IS_INTERNAL(page)) {
internal_bytes += page->memory_footprint;
++internal_pages;
} else {
leaf_bytes += page->memory_footprint;
++leaf_pages;
}
/*
* If the page was pulled into cache by our read, try
* to evict it now.
*
* For eviction to have a chance, we first need to move
* the walk point to the next page checkpoint will
* visit. We want to avoid this code being too special
* purpose, so try to reuse the ordinary eviction path.
*
* Regardless of whether eviction succeeds or fails,
* the walk continues from the previous location. We
* remember whether we tried eviction, and don't try
* again. Even if eviction fails (the page may stay in
* cache clean but with history that cannot be
* discarded), that is not wasted effort because
* checkpoint doesn't need to write the page again.
*/
if (!WT_PAGE_IS_INTERNAL(page) &&
page->read_gen == WT_READGEN_WONT_NEED &&
!tried_eviction) {
WT_ERR_BUSY_OK(
__wt_page_release_evict(session, walk));
walk = prev;
prev = NULL;
tried_eviction = true;
continue;
}
tried_eviction = false;
WT_ERR(__wt_reconcile(
session, walk, NULL, WT_REC_CHECKPOINT, NULL));
/*
* Update checkpoint IO tracking data if configured
* to log verbose progress messages.
*/
if (conn->ckpt_timer_start.tv_sec > 0) {
conn->ckpt_write_bytes +=
page->memory_footprint;
++conn->ckpt_write_pages;
/* Periodically log checkpoint progress. */
if (conn->ckpt_write_pages % 5000 == 0)
__wt_checkpoint_progress(
session, false);
}
}
break;
case WT_SYNC_CLOSE:
case WT_SYNC_DISCARD:
WT_ERR(__wt_illegal_value(session, syncop));
break;
}
if (timer) {
time_stop = __wt_clock(session);
__wt_verbose(session, WT_VERB_CHECKPOINT,
"__sync_file WT_SYNC_%s wrote: %" PRIu64
" leaf pages (%" PRIu64 "B), %" PRIu64
" internal pages (%" PRIu64 "B), and took %" PRIu64 "ms",
syncop == WT_SYNC_WRITE_LEAVES ?
"WRITE_LEAVES" : "CHECKPOINT",
leaf_pages, leaf_bytes, internal_pages, internal_bytes,
WT_CLOCKDIFF_MS(time_stop, time_start));
}
err: /* On error, clear any left-over tree walk. */
WT_TRET(__wt_page_release(session, walk, flags));
WT_TRET(__wt_page_release(session, prev, flags));
/*
* If we got a snapshot in order to write pages, and there was no
* snapshot active when we started, release it.
*/
if (txn->isolation == WT_ISO_READ_COMMITTED &&
saved_pinned_id == WT_TXN_NONE)
__wt_txn_release_snapshot(session);
/* Clear the checkpoint flag. */
btree->syncing = WT_BTREE_SYNC_OFF;
btree->sync_session = NULL;
__wt_spin_unlock(session, &btree->flush_lock);
/*
* Leaves are written before a checkpoint (or as part of a file close,
* before checkpointing the file). Start a flush to stable storage,
* but don't wait for it.
*/
if (ret == 0 &&
syncop == WT_SYNC_WRITE_LEAVES && F_ISSET(conn, WT_CONN_CKPT_SYNC))
WT_RET(btree->bm->sync(btree->bm, session, false));
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_prev);
WT_STAT_FAST_DATA_INCR(session, cursor_prev);
flags = WT_READ_PREV | WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__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.
*/
page = cbt->ref == NULL ? NULL : cbt->ref->page;
for (newpage = 0;; newpage = 1) {
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (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 (page != NULL) {
switch (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;
}
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
page = cbt->ref->page;
WT_ASSERT(session,
page->type != WT_PAGE_COL_INT &&
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 (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
err: if (ret != 0)
WT_TRET(__cursor_error_resolve(cbt));
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, bool truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
bool newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_prev);
WT_STAT_FAST_DATA_INCR(session, cursor_prev);
flags = WT_READ_PREV | WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, false));
/*
* 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);
/*
* 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 = false;; newpage = true) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
/*
* 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 (newpage && page != NULL && page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (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 = true;
}
if (page != NULL) {
switch (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;
}
/*
* If we saw a lot of deleted records on this page, or we went
* all the way through a page and only saw deleted records, try
* to evict the page when we release it. Otherwise repeatedly
* deleting from the beginning of a tree can have quadratic
* performance. Take care not to force eviction of pages that
* are genuinely empty, in new trees.
*/
if (page != NULL &&
(cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD ||
(newpage && cbt->page_deleted_count > 0)))
__wt_page_evict_soon(page);
cbt->page_deleted_count = 0;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_compact --
* Compact a file.
*/
int
__wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_BM *bm;
WT_BTREE *btree;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_REF *ref;
int block_manager_begin, evict_reset, skip;
WT_UNUSED(cfg);
conn = S2C(session);
btree = S2BT(session);
bm = btree->bm;
ref = NULL;
block_manager_begin = 0;
WT_STAT_FAST_DATA_INCR(session, session_compact);
/*
* Check if compaction might be useful -- the API layer will quit trying
* to compact the data source if we make no progress, set a flag if the
* block layer thinks compaction is possible.
*/
WT_RET(bm->compact_skip(bm, session, &skip));
if (skip)
return (0);
/*
* Reviewing in-memory pages requires looking at page reconciliation
* results, because we care about where the page is stored now, not
* where the page was stored when we first read it into the cache.
* We need to ensure we don't race with page reconciliation as it's
* writing the page modify information.
*
* There are three ways we call reconciliation: checkpoints, threads
* writing leaf pages (usually in preparation for a checkpoint or if
* closing a file), and eviction.
*
* We're holding the schema lock which serializes with checkpoints.
*/
WT_ASSERT(session, F_ISSET(session, WT_SESSION_SCHEMA_LOCKED));
/*
* Get the tree handle's flush lock which blocks threads writing leaf
* pages.
*/
__wt_spin_lock(session, &btree->flush_lock);
/*
* That leaves eviction, we don't want to block eviction. Set a flag
* so reconciliation knows compaction is running. If reconciliation
* sees the flag it locks the page it's writing, we acquire the same
* lock when reading the page's modify information, serializing access.
* The same page lock blocks work on the page, but compaction is an
* uncommon, heavy-weight operation. If it's ever a problem, there's
* no reason we couldn't use an entirely separate lock than the page
* lock.
*
* We also need to ensure we don't race with an on-going reconciliation.
* After we set the flag, wait for eviction of this file to drain, and
* then let eviction continue;
*/
conn->compact_in_memory_pass = 1;
WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
if (evict_reset)
__wt_evict_file_exclusive_off(session);
/* Start compaction. */
WT_ERR(bm->compact_start(bm, session));
block_manager_begin = 1;
/* Walk the tree reviewing pages to see if they should be re-written. */
session->compaction = 1;
for (;;) {
/*
* Pages read for compaction aren't "useful"; don't update the
* read generation of pages already in memory, and if a page is
* read, set its generation to a low value so it is evicted
* quickly.
*/
WT_ERR(__wt_tree_walk(session, &ref, NULL,
WT_READ_COMPACT | WT_READ_NO_GEN | WT_READ_WONT_NEED));
if (ref == NULL)
break;
WT_ERR(__compact_rewrite(session, ref, &skip));
if (skip)
continue;
/* Rewrite the page: mark the page and tree dirty. */
WT_ERR(__wt_page_modify_init(session, ref->page));
__wt_page_modify_set(session, ref->page);
WT_STAT_FAST_DATA_INCR(session, btree_compact_rewrite);
}
err: if (ref != NULL)
WT_TRET(__wt_page_release(session, ref, 0));
if (block_manager_begin)
WT_TRET(bm->compact_end(bm, session));
/*
* Unlock will be a release barrier, use it to update the compaction
* status for reconciliation.
*/
conn->compact_in_memory_pass = 0;
__wt_spin_unlock(session, &btree->flush_lock);
return (ret);
}
