/*
* __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);
}
/*
* __sync_dup_walk --
* Duplicate a tree walk point.
*/
static inline int
__sync_dup_walk(
WT_SESSION_IMPL *session, WT_REF *walk, uint32_t flags, WT_REF **dupp)
{
WT_REF *old;
bool busy;
if ((old = *dupp) != NULL) {
*dupp = NULL;
WT_RET(__wt_page_release(session, old, flags));
}
/* It is okay to duplicate a walk before it starts. */
if (walk == NULL || __wt_ref_is_root(walk)) {
*dupp = walk;
return (0);
}
/* Get a duplicate hazard pointer. */
for (;;) {
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, walk, &busy, __func__, __LINE__));
#else
WT_RET(__wt_hazard_set(session, walk, &busy));
#endif
/*
* We already have a hazard pointer, we should generally be able
* to get another one. We can get spurious busy errors (e.g., if
* eviction is attempting to lock the page. Keep trying: we have
* one hazard pointer so we should be able to get another one.
*/
if (!busy)
break;
__wt_yield();
}
*dupp = walk;
return (0);
}
/*
* __wt_tree_walk --
* Move to the next/previous page in the tree.
*/
int
__wt_tree_walk(WT_SESSION_IMPL *session,
WT_REF **refp, uint64_t *walkcntp, uint32_t flags)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_INDEX *pindex;
WT_REF *couple, *couple_orig, *ref;
int prev, skip;
uint32_t slot;
btree = S2BT(session);
/*
* Tree walks are special: they look inside page structures that splits
* may want to free. Publish that the tree is active during this
* window.
*/
WT_ENTER_PAGE_INDEX(session);
/*
* !!!
* Fast-truncate currently only works on row-store trees.
*/
if (btree->type != BTREE_ROW)
LF_CLR(WT_READ_TRUNCATE);
prev = LF_ISSET(WT_READ_PREV) ? 1 : 0;
/*
* There are multiple reasons and approaches to walking the in-memory
* tree:
*
* (1) finding pages to evict (the eviction server);
* (2) writing just dirty leaves or internal nodes (checkpoint);
* (3) discarding pages (close);
* (4) truncating pages in a range (fast truncate);
* (5) skipping pages based on outside information (compaction);
* (6) cursor scans (applications).
*
* Except for cursor scans and compaction, the walk is limited to the
* cache, no pages are read. In all cases, hazard pointers protect the
* walked pages from eviction.
*
* Walks use hazard-pointer coupling through the tree and that's OK
* (hazard pointers can't deadlock, so there's none of the usual
* problems found when logically locking up a btree). If the eviction
* thread tries to evict the active page, it fails because of our
* hazard pointer. If eviction tries to evict our parent, that fails
* because the parent has a child page that can't be discarded. We do
* play one game: don't couple up to our parent and then back down to a
* new leaf, couple to the next page to which we're descending, it
* saves a hazard-pointer swap for each cursor page movement.
*
* !!!
* NOTE: we depend on the fact it's OK to release a page we don't hold,
* that is, it's OK to release couple when couple is set to NULL.
*
* Take a copy of any held page and clear the return value. Remember
* the hazard pointer we're currently holding.
*
* We may be passed a pointer to btree->evict_page that we are clearing
* here. We check when discarding pages that we're not discarding that
* page, so this clear must be done before the page is released.
*/
couple = couple_orig = ref = *refp;
*refp = NULL;
/* If no page is active, begin a walk from the start of the tree. */
if (ref == NULL) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
ascend: /*
* If the active page was the root, we've reached the walk's end.
* Release any hazard-pointer we're holding.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(session, couple, flags));
goto done;
}
/* Figure out the current slot in the WT_REF array. */
__wt_page_refp(session, ref, &pindex, &slot);
for (;;) {
/*
* If we're at the last/first slot on the page, return this page
* in post-order traversal. Otherwise we move to the next/prev
* slot and left/right-most element in its subtree.
*/
if ((prev && slot == 0) ||
(!prev && slot == pindex->entries - 1)) {
ref = ref->home->pg_intl_parent_ref;
/* Optionally skip internal pages. */
if (LF_ISSET(WT_READ_SKIP_INTL))
goto ascend;
/*
* We've ascended the tree and are returning an internal
* page. If it's the root, discard our hazard pointer,
* otherwise, swap our hazard pointer for the page we'll
* return.
*/
if (__wt_ref_is_root(ref))
WT_ERR(__wt_page_release(
session, couple, flags));
else {
/*
* Locate the reference to our parent page then
* swap our child hazard pointer for the parent.
* We don't handle restart or not-found returns.
* It would require additional complexity and is
* not a possible return: we're moving to the
* parent of the current child page, our parent
* reference can't have split or been evicted.
*/
__wt_page_refp(session, ref, &pindex, &slot);
if ((ret = __wt_page_swap(
session, couple, ref, flags)) != 0) {
WT_TRET(__wt_page_release(
session, couple, flags));
WT_ERR(ret);
}
}
*refp = ref;
goto done;
}
if (prev)
--slot;
else
++slot;
if (walkcntp != NULL)
++*walkcntp;
for (;;) {
ref = pindex->index[slot];
if (LF_ISSET(WT_READ_CACHE)) {
/*
* Only look at unlocked pages in memory:
* fast-path some common cases.
*/
if (LF_ISSET(WT_READ_NO_WAIT) &&
ref->state != WT_REF_MEM)
break;
} else if (LF_ISSET(WT_READ_TRUNCATE)) {
/*
* Avoid pulling a deleted page back in to try
* to delete it again.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref))
break;
/*
* If deleting a range, try to delete the page
* without instantiating it.
*/
WT_ERR(__wt_delete_page(session, ref, &skip));
if (skip)
break;
} else if (LF_ISSET(WT_READ_COMPACT)) {
/*
* Skip deleted pages, rewriting them doesn't
* seem useful.
*/
if (ref->state == WT_REF_DELETED)
break;
/*
* If the page is in-memory, we want to look at
* it (it may have been modified and written,
* and the current location is the interesting
* one in terms of compaction, not the original
* location). If the page isn't in-memory, test
* if the page will help with compaction, don't
* read it if we don't have to.
*/
if (ref->state == WT_REF_DISK) {
WT_ERR(__wt_compact_page_skip(
session, ref, &skip));
if (skip)
break;
}
} else {
/*
* Try to skip deleted pages visible to us.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref))
break;
}
ret = __wt_page_swap(session, couple, ref, flags);
/*
* Not-found is an expected return when only walking
* in-cache pages.
*/
if (ret == WT_NOTFOUND) {
ret = 0;
break;
}
/*
* The page we're moving to might have split, in which
* case move to the last position we held.
*/
if (ret == WT_RESTART) {
ret = 0;
/*
* If a new walk that never coupled from the
* root to a new saved position in the tree,
* restart the walk.
*/
if (couple == &btree->root) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
/*
* If restarting from some original position,
* repeat the increment or decrement we made at
* that time. Otherwise, couple is an internal
* page we've acquired after moving from that
* starting position and we can treat it as a
* new page. This works because we never acquire
* a hazard pointer on a leaf page we're not
* going to return to our caller, this will quit
* working if that ever changes.
*/
WT_ASSERT(session,
couple == couple_orig ||
WT_PAGE_IS_INTERNAL(couple->page));
ref = couple;
__wt_page_refp(session, ref, &pindex, &slot);
if (couple == couple_orig)
break;
}
WT_ERR(ret);
/*
* A new page: configure for traversal of any internal
* page's children, else return the leaf page.
*/
descend: couple = ref;
page = ref->page;
if (page->type == WT_PAGE_ROW_INT ||
page->type == WT_PAGE_COL_INT) {
WT_INTL_INDEX_GET(session, page, pindex);
slot = prev ? pindex->entries - 1 : 0;
} else {
*refp = ref;
goto done;
}
}
}
done:
err: WT_LEAVE_PAGE_INDEX(session);
return (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);
}
/*
* __verify_tree --
* Verify a tree, recursively descending through it in depth-first fashion.
* The page argument was physically verified (so we know it's correctly formed),
* and the in-memory version built. Our job is to check logical relationships
* in the page and in the tree.
*/
static int
__verify_tree(WT_SESSION_IMPL *session, WT_REF *ref, WT_VSTUFF *vs)
{
WT_BM *bm;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COL *cip;
WT_DECL_RET;
WT_PAGE *page;
WT_REF *child_ref;
uint64_t recno;
uint32_t entry, i;
bool found;
bm = S2BT(session)->bm;
page = ref->page;
unpack = &_unpack;
WT_CLEAR(*unpack); /* -Wuninitialized */
WT_RET(__wt_verbose(session, WT_VERB_VERIFY, "%s %s",
__wt_page_addr_string(session, ref, vs->tmp1),
__wt_page_type_string(page->type)));
/* Optionally dump the address. */
if (vs->dump_address)
WT_RET(__wt_msg(session, "%s %s",
__wt_page_addr_string(session, ref, vs->tmp1),
__wt_page_type_string(page->type)));
/* Track the shape of the tree. */
if (WT_PAGE_IS_INTERNAL(page))
++vs->depth_internal[
WT_MIN(vs->depth, WT_ELEMENTS(vs->depth_internal) - 1)];
else
++vs->depth_leaf[
WT_MIN(vs->depth, WT_ELEMENTS(vs->depth_internal) - 1)];
/*
* The page's physical structure was verified when it was read into
* memory by the read server thread, and then the in-memory version
* of the page was built. Now we make sure the page and tree are
* logically consistent.
*
* !!!
* The problem: (1) the read server has to build the in-memory version
* of the page because the read server is the thread that flags when
* any thread can access the page in the tree; (2) we can't build the
* in-memory version of the page until the physical structure is known
* to be OK, so the read server has to verify at least the physical
* structure of the page; (3) doing complete page verification requires
* reading additional pages (for example, overflow keys imply reading
* overflow pages in order to test the key's order in the page); (4)
* the read server cannot read additional pages because it will hang
* waiting on itself. For this reason, we split page verification
* into a physical verification, which allows the in-memory version
* of the page to be built, and then a subsequent logical verification
* which happens here.
*
* Report progress occasionally.
*/
#define WT_VERIFY_PROGRESS_INTERVAL 100
if (++vs->fcnt % WT_VERIFY_PROGRESS_INTERVAL == 0)
WT_RET(__wt_progress(session, NULL, vs->fcnt));
#ifdef HAVE_DIAGNOSTIC
/* Optionally dump the blocks or page in debugging mode. */
if (vs->dump_blocks)
WT_RET(__wt_debug_disk(session, page->dsk, NULL));
if (vs->dump_pages)
WT_RET(__wt_debug_page(session, page, NULL));
#endif
/*
* Column-store key order checks: check the page's record number and
* then update the total record count.
*/
switch (page->type) {
case WT_PAGE_COL_FIX:
recno = page->pg_fix_recno;
goto recno_chk;
case WT_PAGE_COL_INT:
recno = page->pg_intl_recno;
goto recno_chk;
case WT_PAGE_COL_VAR:
recno = page->pg_var_recno;
recno_chk: if (recno != vs->record_total + 1)
WT_RET_MSG(session, WT_ERROR,
"page at %s has a starting record of %" PRIu64
" when the expected starting record is %" PRIu64,
__wt_page_addr_string(session, ref, vs->tmp1),
recno, vs->record_total + 1);
break;
}
switch (page->type) {
case WT_PAGE_COL_FIX:
vs->record_total += page->pg_fix_entries;
break;
case WT_PAGE_COL_VAR:
recno = 0;
WT_COL_FOREACH(page, cip, i)
if ((cell = WT_COL_PTR(page, cip)) == NULL)
++recno;
else {
__wt_cell_unpack(cell, unpack);
recno += __wt_cell_rle(unpack);
}
vs->record_total += recno;
break;
}
/*
* Row-store leaf page key order check: it's a depth-first traversal,
* the first key on this page should be larger than any key previously
* seen.
*/
switch (page->type) {
case WT_PAGE_ROW_LEAF:
WT_RET(__verify_row_leaf_key_order(session, ref, vs));
break;
}
/* If it's not the root page, unpack the parent cell. */
if (!__wt_ref_is_root(ref)) {
__wt_cell_unpack(ref->addr, unpack);
/* Compare the parent cell against the page type. */
switch (page->type) {
case WT_PAGE_COL_FIX:
if (unpack->raw != WT_CELL_ADDR_LEAF_NO)
goto celltype_err;
break;
case WT_PAGE_COL_VAR:
if (unpack->raw != WT_CELL_ADDR_LEAF &&
unpack->raw != WT_CELL_ADDR_LEAF_NO)
goto celltype_err;
break;
case WT_PAGE_ROW_LEAF:
if (unpack->raw != WT_CELL_ADDR_DEL &&
unpack->raw != WT_CELL_ADDR_LEAF &&
unpack->raw != WT_CELL_ADDR_LEAF_NO)
goto celltype_err;
break;
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
if (unpack->raw != WT_CELL_ADDR_INT)
celltype_err: WT_RET_MSG(session, WT_ERROR,
"page at %s, of type %s, is referenced in "
"its parent by a cell of type %s",
__wt_page_addr_string(
session, ref, vs->tmp1),
__wt_page_type_string(page->type),
__wt_cell_type_string(unpack->raw));
break;
}
}
/*
* Check overflow pages. We check overflow cells separately from other
* tests that walk the page as it's simpler, and I don't care much how
* fast table verify runs.
*/
switch (page->type) {
case WT_PAGE_COL_VAR:
case WT_PAGE_ROW_INT:
case WT_PAGE_ROW_LEAF:
WT_RET(__verify_overflow_cell(session, ref, &found, vs));
if (__wt_ref_is_root(ref) || page->type == WT_PAGE_ROW_INT)
break;
/*
* Object if a leaf-no-overflow address cell references a page
* with overflow keys, but don't object if a leaf address cell
* references a page without overflow keys. Reconciliation
* doesn't guarantee every leaf page without overflow items will
* be a leaf-no-overflow type.
*/
if (found && unpack->raw == WT_CELL_ADDR_LEAF_NO)
WT_RET_MSG(session, WT_ERROR,
"page at %s, of type %s and referenced in its "
"parent by a cell of type %s, contains overflow "
"items",
__wt_page_addr_string(session, ref, vs->tmp1),
__wt_page_type_string(page->type),
__wt_cell_type_string(WT_CELL_ADDR_LEAF_NO));
break;
}
/* Check tree connections and recursively descend the tree. */
switch (page->type) {
case WT_PAGE_COL_INT:
/* For each entry in an internal page, verify the subtree. */
entry = 0;
WT_INTL_FOREACH_BEGIN(session, page, child_ref) {
/*
* It's a depth-first traversal: this entry's starting
* record number should be 1 more than the total records
* reviewed to this point.
*/
++entry;
if (child_ref->key.recno != vs->record_total + 1) {
WT_RET_MSG(session, WT_ERROR,
"the starting record number in entry %"
PRIu32 " of the column internal page at "
"%s is %" PRIu64 " and the expected "
"starting record number is %" PRIu64,
entry,
__wt_page_addr_string(
session, child_ref, vs->tmp1),
child_ref->key.recno,
vs->record_total + 1);
}
/* Verify the subtree. */
++vs->depth;
WT_RET(__wt_page_in(session, child_ref, 0));
ret = __verify_tree(session, child_ref, vs);
WT_TRET(__wt_page_release(session, child_ref, 0));
--vs->depth;
WT_RET(ret);
__wt_cell_unpack(child_ref->addr, unpack);
WT_RET(bm->verify_addr(
bm, session, unpack->data, unpack->size));
} WT_INTL_FOREACH_END;
break;
case WT_PAGE_ROW_INT:
/* For each entry in an internal page, verify the subtree. */
entry = 0;
WT_INTL_FOREACH_BEGIN(session, page, child_ref) {
/*
* It's a depth-first traversal: this entry's starting
* key should be larger than the largest key previously
* reviewed.
*
* The 0th key of any internal page is magic, and we
* can't test against it.
*/
++entry;
if (entry != 1)
WT_RET(__verify_row_int_key_order(
session, page, child_ref, entry, vs));
/* Verify the subtree. */
++vs->depth;
WT_RET(__wt_page_in(session, child_ref, 0));
ret = __verify_tree(session, child_ref, vs);
WT_TRET(__wt_page_release(session, child_ref, 0));
--vs->depth;
WT_RET(ret);
__wt_cell_unpack(child_ref->addr, unpack);
WT_RET(bm->verify_addr(
bm, session, unpack->data, unpack->size));
} WT_INTL_FOREACH_END;
/*
* __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);
}
/*
* __tree_walk_internal --
* Move to the next/previous page in the tree.
*/
static inline int
__tree_walk_internal(WT_SESSION_IMPL *session,
WT_REF **refp, uint64_t *walkcntp, uint64_t *skipleafcntp, uint32_t flags)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE_INDEX *pindex;
WT_REF *couple, *couple_orig, *ref;
bool empty_internal, prev, skip;
uint32_t slot;
btree = S2BT(session);
empty_internal = false;
/*
* Tree walks are special: they look inside page structures that splits
* may want to free. Publish that the tree is active during this
* window.
*/
WT_ENTER_PAGE_INDEX(session);
/* Walk should never instantiate deleted pages. */
LF_SET(WT_READ_NO_EMPTY);
/*
* !!!
* Fast-truncate currently only works on row-store trees.
*/
if (btree->type != BTREE_ROW)
LF_CLR(WT_READ_TRUNCATE);
prev = LF_ISSET(WT_READ_PREV) ? 1 : 0;
/*
* There are multiple reasons and approaches to walking the in-memory
* tree:
*
* (1) finding pages to evict (the eviction server);
* (2) writing just dirty leaves or internal nodes (checkpoint);
* (3) discarding pages (close);
* (4) truncating pages in a range (fast truncate);
* (5) skipping pages based on outside information (compaction);
* (6) cursor scans (applications).
*
* Except for cursor scans and compaction, the walk is limited to the
* cache, no pages are read. In all cases, hazard pointers protect the
* walked pages from eviction.
*
* Walks use hazard-pointer coupling through the tree and that's OK
* (hazard pointers can't deadlock, so there's none of the usual
* problems found when logically locking up a btree). If the eviction
* thread tries to evict the active page, it fails because of our
* hazard pointer. If eviction tries to evict our parent, that fails
* because the parent has a child page that can't be discarded. We do
* play one game: don't couple up to our parent and then back down to a
* new leaf, couple to the next page to which we're descending, it
* saves a hazard-pointer swap for each cursor page movement.
*
* !!!
* NOTE: we depend on the fact it's OK to release a page we don't hold,
* that is, it's OK to release couple when couple is set to NULL.
*
* Take a copy of any held page and clear the return value. Remember
* the hazard pointer we're currently holding.
*
* We may be passed a pointer to btree->evict_page that we are clearing
* here. We check when discarding pages that we're not discarding that
* page, so this clear must be done before the page is released.
*/
couple = couple_orig = ref = *refp;
*refp = NULL;
/* If no page is active, begin a walk from the start of the tree. */
if (ref == NULL) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
/*
* If the active page was the root, we've reached the walk's end.
* Release any hazard-pointer we're holding.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(session, couple, flags));
goto done;
}
/* Figure out the current slot in the WT_REF array. */
__ref_index_slot(session, ref, &pindex, &slot);
for (;;) {
/*
* If we're at the last/first slot on the internal page, return
* it in post-order traversal. Otherwise move to the next/prev
* slot and left/right-most element in that subtree.
*/
while ((prev && slot == 0) ||
(!prev && slot == pindex->entries - 1)) {
/* Ascend to the parent. */
__page_ascend(session, &ref, &pindex, &slot);
/*
* If we got all the way through an internal page and
* all of the child pages were deleted, mark it for
* eviction.
*/
if (empty_internal && pindex->entries > 1) {
__wt_page_evict_soon(ref->page);
empty_internal = false;
}
/*
* If at the root and returning internal pages, return
* the root page, otherwise we're done. Regardless, no
* hazard pointer is required, release the one we hold.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(
session, couple, flags));
if (!LF_ISSET(WT_READ_SKIP_INTL))
*refp = ref;
goto done;
}
/*
* Optionally return internal pages. Swap our previous
* hazard pointer for the page we'll return. We don't
* handle restart or not-found returns, it would require
* additional complexity and is not a possible return:
* we're moving to the parent of the current child page,
* the parent can't have been evicted.
*/
if (!LF_ISSET(WT_READ_SKIP_INTL)) {
WT_ERR(__wt_page_swap(
session, couple, ref, flags));
*refp = ref;
goto done;
}
}
if (prev)
--slot;
else
++slot;
if (walkcntp != NULL)
++*walkcntp;
for (;;) {
/*
* Move to the next slot, and set the reference hint if
* it's wrong (used when we continue the walk). We don't
* update those hints when splitting, so it's common for
* them to be incorrect in some workloads.
*/
ref = pindex->index[slot];
if (ref->pindex_hint != slot)
ref->pindex_hint = slot;
/*
* If we see any child states other than deleted, the
* page isn't empty.
*/
if (ref->state != WT_REF_DELETED &&
!LF_ISSET(WT_READ_TRUNCATE))
empty_internal = false;
if (LF_ISSET(WT_READ_CACHE)) {
/*
* Only look at unlocked pages in memory:
* fast-path some common cases.
*/
if (LF_ISSET(WT_READ_NO_WAIT) &&
ref->state != WT_REF_MEM)
break;
} else if (LF_ISSET(WT_READ_TRUNCATE)) {
/*
* Avoid pulling a deleted page back in to try
* to delete it again.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
/*
* If deleting a range, try to delete the page
* without instantiating it.
*/
WT_ERR(__wt_delete_page(session, ref, &skip));
if (skip)
break;
empty_internal = false;
} else if (LF_ISSET(WT_READ_COMPACT)) {
/*
* Skip deleted pages, rewriting them doesn't
* seem useful.
*/
if (ref->state == WT_REF_DELETED)
break;
/*
* If the page is in-memory, we want to look at
* it (it may have been modified and written,
* and the current location is the interesting
* one in terms of compaction, not the original
* location). If the page isn't in-memory, test
* if the page will help with compaction, don't
* read it if we don't have to.
*/
if (ref->state == WT_REF_DISK) {
WT_ERR(__wt_compact_page_skip(
session, ref, &skip));
if (skip)
break;
}
} else {
/*
* Try to skip deleted pages visible to us.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
}
/*
* Optionally skip leaf pages: skip all leaf pages if
* WT_READ_SKIP_LEAF is set, when the skip-leaf-count
* variable is non-zero, skip some count of leaf pages.
* If this page is disk-based, crack the cell to figure
* out it's a leaf page without reading it.
*
* If skipping some number of leaf pages, decrement the
* count of pages to zero, and then take the next leaf
* page we can. Be cautious around the page decrement,
* if for some reason don't take this particular page,
* we can take the next one, and, there are additional
* tests/decrements when we're about to return a leaf
* page.
*/
if (skipleafcntp != NULL || LF_ISSET(WT_READ_SKIP_LEAF))
if (__ref_is_leaf(ref)) {
if (LF_ISSET(WT_READ_SKIP_LEAF))
break;
if (*skipleafcntp > 0) {
--*skipleafcntp;
break;
}
}
ret = __wt_page_swap(session, couple, ref,
WT_READ_NOTFOUND_OK | WT_READ_RESTART_OK | flags);
/*
* Not-found is an expected return when only walking
* in-cache pages, or if we see a deleted page.
*/
if (ret == WT_NOTFOUND) {
ret = 0;
break;
}
/*
* The page we're moving to might have split, in which
* case move to the last position we held.
*/
if (ret == WT_RESTART) {
ret = 0;
/*
* If a new walk that never coupled from the
* root to a new saved position in the tree,
* restart the walk.
*/
if (couple == &btree->root) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
/*
* If restarting from some original position,
* repeat the increment or decrement we made at
* that time. Otherwise, couple is an internal
* page we've acquired after moving from that
* starting position and we can treat it as a
* new page. This works because we never acquire
* a hazard pointer on a leaf page we're not
* going to return to our caller, this will quit
* working if that ever changes.
*/
WT_ASSERT(session,
couple == couple_orig ||
WT_PAGE_IS_INTERNAL(couple->page));
ref = couple;
__ref_index_slot(session, ref, &pindex, &slot);
if (couple == couple_orig)
break;
}
WT_ERR(ret);
/*
* A new page: configure for traversal of any internal
* page's children, else return the leaf page.
*/
if (WT_PAGE_IS_INTERNAL(ref->page)) {
descend: couple = ref;
empty_internal = true;
__page_descend(
session, ref->page, &pindex, &slot, prev);
} else {
/*
* Optionally skip leaf pages, the second half.
* We didn't have an on-page cell to figure out
* if it was a leaf page, we had to acquire the
* hazard pointer and look at the page.
*/
if (skipleafcntp != NULL ||
LF_ISSET(WT_READ_SKIP_LEAF)) {
couple = ref;
if (LF_ISSET(WT_READ_SKIP_LEAF))
break;
if (*skipleafcntp > 0) {
--*skipleafcntp;
break;
}
}
*refp = ref;
goto done;
}
}
}
done:
err: WT_LEAVE_PAGE_INDEX(session);
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);
}
/*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_page_in --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(
WT_SESSION_IMPL *session, WT_PAGE *parent, WT_REF *ref
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_DECL_RET;
WT_PAGE *page;
WT_TXN *txn;
int busy, oldgen;
txn = &session->txn;
for (oldgen = 0;;) {
switch (ref->state) {
case WT_REF_DISK:
case WT_REF_DELETED:
/*
* The page isn't in memory, attempt to read it.
*
* First make sure there is space in the cache.
*/
WT_RET(__wt_cache_full_check(session, 0));
WT_RET(__wt_cache_read(session, parent, ref));
oldgen = F_ISSET(session, WT_SESSION_NO_CACHE) ? 1 : 0;
continue;
case WT_REF_LOCKED:
case WT_REF_READING:
/*
* The page is being read or considered for eviction --
* wait for that to be resolved.
*/
break;
case WT_REF_EVICT_WALK:
case WT_REF_MEM:
/*
* The page is in memory: get a hazard pointer, update
* the page's LRU and return. The expected reason we
* can't get a hazard pointer is because the page is
* being evicted; yield and try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy)
break;
page = ref->page;
WT_ASSERT(session,
page != NULL && !WT_PAGE_IS_ROOT(page));
/*
* Make sure the page isn't too big. Only do this
* check once per transaction: it is not a common case,
* and we don't want to get stuck if it isn't possible
* to evict the page.
*/
if (!F_ISSET(txn, TXN_FORCE_EVICT) &&
__wt_eviction_page_force(session, page)) {
F_SET(txn, TXN_FORCE_EVICT);
page->read_gen = WT_READ_GEN_OLDEST;
WT_RET(__wt_page_release(session, page));
break;
}
/* Check if we need an autocommit transaction. */
if ((ret = __wt_txn_autocommit_check(session)) != 0) {
WT_TRET(__wt_hazard_clear(session, page));
return (ret);
}
/*
* If this page has ever been considered for eviction,
* and its generation is aging, update it.
*/
if (page->read_gen != WT_READ_GEN_NOTSET &&
page->read_gen < __wt_cache_read_gen(session))
page->read_gen =
__wt_cache_read_gen_set(session);
/*
* If we read the page and we are configured to not
* trash the cache, set the oldest read generation so
* the page is forcibly evicted as soon as possible.
*/
if (oldgen && page->read_gen == WT_READ_GEN_NOTSET)
page->read_gen = WT_READ_GEN_OLDEST;
return (0);
WT_ILLEGAL_VALUE(session);
}
/* We failed to get the page -- yield before retrying. */
__wt_yield();
}
}
/*
* __wt_page_in --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(
WT_SESSION_IMPL *session, WT_PAGE *parent, WT_REF *ref
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_DECL_RET;
WT_PAGE *page;
int busy, oldgen;
for (oldgen = 0;;) {
switch (ref->state) {
case WT_REF_DISK:
case WT_REF_DELETED:
/*
* The page isn't in memory, attempt to read it.
*
* First make sure there is space in the cache.
*/
WT_RET(__wt_cache_full_check(session, 0));
WT_RET(__wt_cache_read(session, parent, ref));
oldgen = F_ISSET(session, WT_SESSION_NO_CACHE) ? 1 : 0;
continue;
case WT_REF_EVICT_FORCE:
case WT_REF_LOCKED:
case WT_REF_READING:
/*
* The page is being read or considered for eviction --
* wait for that to be resolved.
*/
break;
case WT_REF_EVICT_WALK:
case WT_REF_MEM:
/*
* The page is in memory: get a hazard pointer, update
* the page's LRU and return. The expected reason we
* can't get a hazard pointer is because the page is
* being evicted; yield and try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy)
break;
page = ref->page;
WT_ASSERT(session,
page != NULL && !WT_PAGE_IS_ROOT(page));
/*
* Ensure the page doesn't have ancient updates on it.
* If it did, reading the page could ignore committed
* updates. This should be extremely unlikely in real
* applications, wait for eviction of the page to avoid
* the issue.
*/
if (page->modify != NULL &&
__wt_txn_ancient(session, page->modify->first_id)) {
page->read_gen = WT_READ_GEN_OLDEST;
WT_RET(__wt_page_release(session, page));
break;
}
/* Check if we need an autocommit transaction. */
if ((ret = __wt_txn_autocommit_check(session)) != 0) {
WT_TRET(__wt_hazard_clear(session, page));
return (ret);
}
/*
* If this page has ever been considered for eviction,
* and its generation is aging, update it.
*/
if (page->read_gen != WT_READ_GEN_NOTSET &&
page->read_gen < __wt_cache_read_gen(session))
page->read_gen =
__wt_cache_read_gen_set(session);
/*
* If we read the page and we are configured to not
* trash the cache, set the oldest read generation so
* the page is forcibly evicted as soon as possible.
*/
if (oldgen && page->read_gen == WT_READ_GEN_NOTSET)
page->read_gen = WT_READ_GEN_OLDEST;
return (0);
WT_ILLEGAL_VALUE(session);
}
/* We failed to get the page -- yield before retrying. */
__wt_yield();
}
}
/*
* __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);
}
/*
* __verify_tree --
* Verify a tree, recursively descending through it in depth-first fashion.
* The page argument was physically verified (so we know it's correctly formed),
* and the in-memory version built. Our job is to check logical relationships
* in the page and in the tree.
*/
static int
__verify_tree(WT_SESSION_IMPL *session, WT_PAGE *page, WT_VSTUFF *vs)
{
WT_BM *bm;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COL *cip;
WT_DECL_RET;
WT_REF *ref;
uint64_t recno;
uint32_t entry, i;
int found, lno;
bm = S2BT(session)->bm;
unpack = &_unpack;
WT_VERBOSE_RET(session, verify, "%s %s",
__wt_page_addr_string(session, vs->tmp1, page),
__wt_page_type_string(page->type));
#ifdef HAVE_DIAGNOSTIC
if (vs->dump_address)
WT_RET(__wt_msg(session, "%s %s",
__wt_page_addr_string(session, vs->tmp1, page),
__wt_page_type_string(page->type)));
#endif
/*
* The page's physical structure was verified when it was read into
* memory by the read server thread, and then the in-memory version
* of the page was built. Now we make sure the page and tree are
* logically consistent.
*
* !!!
* The problem: (1) the read server has to build the in-memory version
* of the page because the read server is the thread that flags when
* any thread can access the page in the tree; (2) we can't build the
* in-memory version of the page until the physical structure is known
* to be OK, so the read server has to verify at least the physical
* structure of the page; (3) doing complete page verification requires
* reading additional pages (for example, overflow keys imply reading
* overflow pages in order to test the key's order in the page); (4)
* the read server cannot read additional pages because it will hang
* waiting on itself. For this reason, we split page verification
* into a physical verification, which allows the in-memory version
* of the page to be built, and then a subsequent logical verification
* which happens here.
*
* Report progress every 10 pages.
*/
if (++vs->fcnt % 10 == 0)
WT_RET(__wt_progress(session, NULL, vs->fcnt));
#ifdef HAVE_DIAGNOSTIC
/* Optionally dump the page in debugging mode. */
if (vs->dump_blocks && page->dsk != NULL)
WT_RET(__wt_debug_disk(session, page->dsk, NULL));
if (vs->dump_pages)
WT_RET(__wt_debug_page(session, page, NULL));
#endif
/*
* Column-store key order checks: check the page's record number and
* then update the total record count.
*/
switch (page->type) {
case WT_PAGE_COL_FIX:
recno = page->u.col_fix.recno;
goto recno_chk;
case WT_PAGE_COL_INT:
recno = page->u.intl.recno;
goto recno_chk;
case WT_PAGE_COL_VAR:
recno = page->u.col_var.recno;
recno_chk: if (recno != vs->record_total + 1)
WT_RET_MSG(session, WT_ERROR,
"page at %s has a starting record of %" PRIu64
" when the expected starting record is %" PRIu64,
__wt_page_addr_string(session, vs->tmp1, page),
recno, vs->record_total + 1);
break;
}
switch (page->type) {
case WT_PAGE_COL_FIX:
vs->record_total += page->entries;
break;
case WT_PAGE_COL_VAR:
recno = 0;
WT_COL_FOREACH(page, cip, i)
if ((cell = WT_COL_PTR(page, cip)) == NULL)
++recno;
else {
__wt_cell_unpack(cell, unpack);
recno += __wt_cell_rle(unpack);
}
vs->record_total += recno;
break;
}
/*
* Row-store leaf page key order check: it's a depth-first traversal,
* the first key on this page should be larger than any key previously
* seen.
*/
switch (page->type) {
case WT_PAGE_ROW_LEAF:
WT_RET(__verify_row_leaf_key_order(session, page, vs));
break;
}
/*
* Check overflow pages. We check overflow cells separately from other
* tests that walk the page as it's simpler, and I don't care much how
* fast table verify runs.
*
* Object if a leaf-no-overflow address cell references a page that has
* overflow keys, but don't object if a standard address cell references
* a page without overflow keys. The leaf-no-overflow address cell is
* an optimization for trees without few, if any, overflow items, and
* may not be set by reconciliation in all possible cases.
*/
if (WT_PAGE_IS_ROOT(page))
lno = 0;
else {
__wt_cell_unpack(page->ref->addr, unpack);
lno = unpack->raw == WT_CELL_ADDR_LNO ? 1 : 0;
}
switch (page->type) {
case WT_PAGE_COL_FIX:
break;
case WT_PAGE_COL_VAR:
case WT_PAGE_ROW_INT:
case WT_PAGE_ROW_LEAF:
WT_RET(__verify_overflow_cell(session, page, &found, vs));
if (found && lno)
WT_RET_MSG(session, WT_ERROR,
"page at %s referenced in its parent by a cell of "
"type %s illegally contains overflow items",
__wt_page_addr_string(session, vs->tmp1, page),
__wt_cell_type_string(WT_CELL_ADDR_LNO));
break;
default:
if (lno)
WT_RET_MSG(session, WT_ERROR,
"page at %s is of type %s and is illegally "
"referenced in its parent by a cell of type %s",
__wt_page_addr_string(session, vs->tmp1, page),
__wt_page_type_string(page->type),
__wt_cell_type_string(WT_CELL_ADDR_LNO));
break;
}
/* Check tree connections and recursively descend the tree. */
switch (page->type) {
case WT_PAGE_COL_INT:
/* For each entry in an internal page, verify the subtree. */
entry = 0;
WT_REF_FOREACH(page, ref, i) {
/*
* It's a depth-first traversal: this entry's starting
* record number should be 1 more than the total records
* reviewed to this point.
*/
++entry;
if (ref->u.recno != vs->record_total + 1) {
__wt_cell_unpack(ref->addr, unpack);
WT_RET_MSG(session, WT_ERROR,
"the starting record number in entry %"
PRIu32 " of the column internal page at "
"%s is %" PRIu64 " and the expected "
"starting record number is %" PRIu64,
entry,
__wt_page_addr_string(
session, vs->tmp1, page),
ref->u.recno,
vs->record_total + 1);
}
/* Verify the subtree. */
WT_RET(__wt_page_in(session, page, ref));
ret = __verify_tree(session, ref->page, vs);
WT_TRET(__wt_page_release(session, ref->page));
WT_RET(ret);
__wt_cell_unpack(ref->addr, unpack);
WT_RET(bm->verify_addr(
bm, session, unpack->data, unpack->size));
}
break;
case WT_PAGE_ROW_INT:
/* For each entry in an internal page, verify the subtree. */
entry = 0;
WT_REF_FOREACH(page, ref, i) {
/*
* It's a depth-first traversal: this entry's starting
* key should be larger than the largest key previously
* reviewed.
*
* The 0th key of any internal page is magic, and we
* can't test against it.
*/
++entry;
if (entry != 1)
WT_RET(__verify_row_int_key_order(
session, page, ref, entry, vs));
/* Verify the subtree. */
WT_RET(__wt_page_in(session, page, ref));
ret = __verify_tree(session, ref->page, vs);
WT_TRET(__wt_page_release(session, ref->page));
WT_RET(ret);
__wt_cell_unpack(ref->addr, unpack);
WT_RET(bm->verify_addr(
bm, session, unpack->data, unpack->size));
}
/*
* __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);
}
/*
* __tree_walk_internal --
* Move to the next/previous page in the tree.
*/
static inline int
__tree_walk_internal(WT_SESSION_IMPL *session,
WT_REF **refp, uint64_t *walkcntp,
int (*skip_func)(WT_SESSION_IMPL *, WT_REF *, void *, bool *),
void *func_cookie, uint32_t flags)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE_INDEX *pindex;
WT_REF *couple, *couple_orig, *ref;
uint32_t slot;
bool empty_internal, initial_descent, prev, skip;
btree = S2BT(session);
pindex = NULL;
empty_internal = initial_descent = false;
/*
* Tree walks are special: they look inside page structures that splits
* may want to free. Publish that the tree is active during this
* window.
*/
WT_ENTER_PAGE_INDEX(session);
/* Walk should never instantiate deleted pages. */
LF_SET(WT_READ_NO_EMPTY);
/*
* !!!
* Fast-truncate currently only works on row-store trees.
*/
if (btree->type != BTREE_ROW)
LF_CLR(WT_READ_TRUNCATE);
prev = LF_ISSET(WT_READ_PREV) ? 1 : 0;
/*
* There are multiple reasons and approaches to walking the in-memory
* tree:
*
* (1) finding pages to evict (the eviction server);
* (2) writing just dirty leaves or internal nodes (checkpoint);
* (3) discarding pages (close);
* (4) truncating pages in a range (fast truncate);
* (5) skipping pages based on outside information (compaction);
* (6) cursor scans (applications).
*
* Except for cursor scans and compaction, the walk is limited to the
* cache, no pages are read. In all cases, hazard pointers protect the
* walked pages from eviction.
*
* Walks use hazard-pointer coupling through the tree and that's OK
* (hazard pointers can't deadlock, so there's none of the usual
* problems found when logically locking up a btree). If the eviction
* thread tries to evict the active page, it fails because of our
* hazard pointer. If eviction tries to evict our parent, that fails
* because the parent has a child page that can't be discarded. We do
* play one game: don't couple up to our parent and then back down to a
* new leaf, couple to the next page to which we're descending, it
* saves a hazard-pointer swap for each cursor page movement.
*
* !!!
* NOTE: we depend on the fact it's OK to release a page we don't hold,
* that is, it's OK to release couple when couple is set to NULL.
*
* Take a copy of any held page and clear the return value. Remember
* the hazard pointer we're currently holding.
*
* Clear the returned value, it makes future error handling easier.
*/
couple = couple_orig = ref = *refp;
*refp = NULL;
/* If no page is active, begin a walk from the start/end of the tree. */
if (ref == NULL) {
restart: /*
* We can be here with a NULL or root WT_REF; the page release
* function handles them internally, don't complicate this code
* by calling them out.
*/
WT_ERR(__wt_page_release(session, couple, flags));
/*
* We're not supposed to walk trees without root pages. As this
* has not always been the case, assert to debug that change.
*/
WT_ASSERT(session, btree->root.page != NULL);
couple = couple_orig = ref = &btree->root;
initial_descent = true;
goto descend;
}
/*
* If the active page was the root, we've reached the walk's end; we
* only get here if we've returned the root to our caller, so we're
* holding no hazard pointers.
*/
if (__wt_ref_is_root(ref))
goto done;
/* Figure out the current slot in the WT_REF array. */
__ref_index_slot(session, ref, &pindex, &slot);
for (;;) {
/*
* If we're at the last/first slot on the internal page, return
* it in post-order traversal. Otherwise move to the next/prev
* slot and left/right-most element in that subtree.
*/
while ((prev && slot == 0) ||
(!prev && slot == pindex->entries - 1)) {
/* Ascend to the parent. */
__ref_ascend(session, &ref, &pindex, &slot);
/*
* If at the root and returning internal pages, return
* the root page, otherwise we're done. Regardless, no
* hazard pointer is required, release the one we hold.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(
session, couple, flags));
if (!LF_ISSET(WT_READ_SKIP_INTL))
*refp = ref;
goto done;
}
/*
* If we got all the way through an internal page and
* all of the child pages were deleted, mark it for
* eviction.
*/
if (empty_internal && pindex->entries > 1) {
__wt_page_evict_soon(session, ref);
empty_internal = false;
}
/*
* Optionally return internal pages. Swap our previous
* hazard pointer for the page we'll return. We don't
* handle restart or not-found returns, it would require
* additional complexity and is not a possible return:
* we're moving to the parent of the current child page,
* the parent can't have been evicted.
*/
if (!LF_ISSET(WT_READ_SKIP_INTL)) {
WT_ERR(__wt_page_swap(
session, couple, ref, flags));
*refp = ref;
goto done;
}
}
if (prev)
--slot;
else
++slot;
if (walkcntp != NULL)
++*walkcntp;
for (;;) {
/*
* Move to the next slot, and set the reference hint if
* it's wrong (used when we continue the walk). We don't
* always update the hints when splitting, it's expected
* for them to be incorrect in some workloads.
*/
ref = pindex->index[slot];
if (ref->pindex_hint != slot)
ref->pindex_hint = slot;
/*
* If we see any child states other than deleted, the
* page isn't empty.
*/
if (ref->state != WT_REF_DELETED &&
!LF_ISSET(WT_READ_TRUNCATE))
empty_internal = false;
if (LF_ISSET(WT_READ_CACHE)) {
/*
* Only look at unlocked pages in memory:
* fast-path some common cases.
*/
if (LF_ISSET(WT_READ_NO_WAIT) &&
ref->state != WT_REF_MEM)
break;
/* Skip lookaside pages if not requested. */
if (ref->state == WT_REF_LOOKASIDE &&
!LF_ISSET(WT_READ_LOOKASIDE))
break;
} else if (LF_ISSET(WT_READ_TRUNCATE)) {
/*
* Avoid pulling a deleted page back in to try
* to delete it again.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
/*
* If deleting a range, try to delete the page
* without instantiating it.
*/
WT_ERR(__wt_delete_page(session, ref, &skip));
if (skip)
break;
empty_internal = false;
} else if (skip_func != NULL) {
WT_ERR(skip_func(session,
ref, func_cookie, &skip));
if (skip)
break;
} else {
/*
* Try to skip deleted pages visible to us.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
}
ret = __wt_page_swap(session, couple, ref,
WT_READ_NOTFOUND_OK | WT_READ_RESTART_OK | flags);
/*
* Not-found is an expected return when only walking
* in-cache pages, or if we see a deleted page.
*/
if (ret == WT_NOTFOUND) {
ret = 0;
break;
}
/*
* The page we're moving to might have split, in which
* case move to the last position we held.
*/
if (ret == WT_RESTART) {
ret = 0;
/*
* If a cursor is setting up at the end of the
* tree, we can't use our parent page's index,
* because it may have already split; restart
* the walk.
*/
if (prev && initial_descent)
goto restart;
/*
* If a new walk that never coupled from the
* root to a new saved position in the tree,
* restart the walk.
*/
if (couple == &btree->root)
goto restart;
/*
* If restarting from some original position,
* repeat the increment or decrement we made at
* that time. Otherwise, couple is an internal
* page we've acquired after moving from that
* starting position and we can treat it as a
* new page. This works because we never acquire
* a hazard pointer on a leaf page we're not
* going to return to our caller, this will quit
* working if that ever changes.
*/
WT_ASSERT(session,
couple == couple_orig ||
WT_PAGE_IS_INTERNAL(couple->page));
ref = couple;
__ref_index_slot(session, ref, &pindex, &slot);
if (couple == couple_orig)
break;
}
WT_ERR(ret);
couple = ref;
/*
* A new page: configure for traversal of any internal
* page's children, else return the leaf page.
*/
if (WT_PAGE_IS_INTERNAL(ref->page)) {
descend: empty_internal = true;
/*
* There's a split race when a cursor is setting
* up at the end of the tree or moving backwards
* through the tree and descending a level. When
* splitting an internal page into its parent,
* we move the WT_REF structures and update the
* parent's page index before updating the split
* page's page index, and it's not an atomic
* update. A thread can read the parent page's
* replacement page index, then read the split
* page's original index, or the parent page's
* original and the split page's replacement.
*
* This isn't a problem for a cursor setting up
* at the start of the tree or moving forwards
* through the tree because we do right-hand
* splits on internal pages and the initial part
* of the split page's namespace won't change as
* part of a split. A thread reading the parent
* page's and split page's indexes will move to
* the same slot no matter what order of indexes
* are read.
*
* Handle a cursor setting up at the end of the
* tree or moving backwards through the tree.
*/
if (!prev) {
WT_INTL_INDEX_GET(
session, ref->page, pindex);
slot = 0;
} else if (initial_descent) {
if (!__ref_initial_descent_prev(
session, ref, &pindex))
goto restart;
slot = pindex->entries - 1;
} else {
__ref_descend_prev(
session, ref, &pindex);
slot = pindex->entries - 1;
}
continue;
}
/*
* The tree-walk restart code knows we return any leaf
* page we acquire (never hazard-pointer coupling on
* after acquiring a leaf page), and asserts no restart
* happens while holding a leaf page. This page must be
* returned to our caller.
*/
*refp = ref;
goto done;
}
}
done:
err: WT_LEAVE_PAGE_INDEX(session);
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);
}
/*
* __wt_page_in_func --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_DECL_RET;
WT_PAGE *page;
int busy, force_attempts, oldgen;
for (force_attempts = oldgen = 0;;) {
switch (ref->state) {
case WT_REF_DISK:
case WT_REF_DELETED:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
/*
* The page isn't in memory, attempt to read it.
* Make sure there is space in the cache.
*/
WT_RET(__wt_cache_full_check(session));
WT_RET(__wt_cache_read(session, ref));
oldgen = LF_ISSET(WT_READ_WONT_NEED) ||
F_ISSET(session, WT_SESSION_NO_CACHE);
continue;
case WT_REF_READING:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
/* FALLTHROUGH */
case WT_REF_LOCKED:
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
/* The page is busy -- wait. */
break;
case WT_REF_SPLIT:
return (WT_RESTART);
case WT_REF_MEM:
/*
* The page is in memory: get a hazard pointer, update
* the page's LRU and return. The expected reason we
* can't get a hazard pointer is because the page is
* being evicted; yield and try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy)
break;
page = ref->page;
WT_ASSERT(session, page != NULL);
/* Forcibly evict pages that are too big. */
if (!LF_ISSET(WT_READ_NO_EVICT) &&
force_attempts < 10 &&
__evict_force_check(session, page)) {
++force_attempts;
WT_RET(__wt_page_release(session, ref, flags));
break;
}
/* Check if we need an autocommit transaction. */
if ((ret = __wt_txn_autocommit_check(session)) != 0) {
WT_TRET(__wt_hazard_clear(session, page));
return (ret);
}
/*
* If we read the page and we are configured to not
* trash the cache, set the oldest read generation so
* the page is forcibly evicted as soon as possible.
*
* Otherwise, update the page's read generation.
*/
if (oldgen && page->read_gen == WT_READGEN_NOTSET)
__wt_page_evict_soon(page);
else if (!LF_ISSET(WT_READ_NO_GEN) &&
page->read_gen < __wt_cache_read_gen(session))
page->read_gen =
__wt_cache_read_gen_set(session);
return (0);
WT_ILLEGAL_VALUE(session);
}
/* We failed to get the page -- yield before retrying. */
__wt_yield();
}
}
/*
* __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);
}
