/*
* __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_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);
}
/*
* __evict_stat_walk --
* Walk all the pages in cache for a dhandle gathering stats information
*/
static void
__evict_stat_walk(WT_SESSION_IMPL *session)
{
WT_BTREE *btree;
WT_CACHE *cache;
WT_PAGE *page;
WT_REF *next_walk;
uint64_t dsk_size, gen_gap, gen_gap_max, gen_gap_sum, max_pagesize;
uint64_t min_written_size, num_memory, num_not_queueable, num_queued;
uint64_t num_smaller_allocsz, pages_clean, pages_dirty, pages_internal;
uint64_t pages_leaf, seen_count, size, visited_count;
uint64_t visited_age_gap_sum, unvisited_count, unvisited_age_gap_sum;
uint64_t walk_count, written_size_cnt, written_size_sum;
btree = S2BT(session);
cache = S2C(session)->cache;
next_walk = NULL;
gen_gap_max = gen_gap_sum = max_pagesize = 0;
num_memory = num_not_queueable = num_queued = 0;
num_smaller_allocsz = pages_clean = pages_dirty = pages_internal = 0;
pages_leaf = seen_count = size = visited_count = 0;
visited_age_gap_sum = unvisited_count = unvisited_age_gap_sum = 0;
walk_count = written_size_cnt = written_size_sum = 0;
min_written_size = UINT64_MAX;
while (__wt_tree_walk_count(session, &next_walk, &walk_count,
WT_READ_CACHE | WT_READ_NO_EVICT |
WT_READ_NO_GEN | WT_READ_NO_WAIT) == 0 &&
next_walk != NULL) {
++seen_count;
page = next_walk->page;
size = page->memory_footprint;
if (__wt_page_is_modified(page))
++pages_dirty;
else
++pages_clean;
if (!__wt_ref_is_root(next_walk) &&
!__wt_page_can_evict(session, next_walk, NULL))
++num_not_queueable;
if (F_ISSET_ATOMIC(page, WT_PAGE_EVICT_LRU))
++num_queued;
if (size > max_pagesize)
max_pagesize = size;
dsk_size = page->dsk != NULL ? page->dsk->mem_size : 0;
if (dsk_size != 0) {
if (dsk_size < btree->allocsize)
++num_smaller_allocsz;
if (dsk_size < min_written_size)
min_written_size = dsk_size;
++written_size_cnt;
written_size_sum += dsk_size;
} else
++num_memory;
if (WT_PAGE_IS_INTERNAL(page))
++pages_internal;
else
++pages_leaf;
/* Skip root pages since they are never considered */
if (__wt_ref_is_root(next_walk))
continue;
if (page->evict_pass_gen == 0) {
unvisited_age_gap_sum +=
(cache->evict_pass_gen - page->cache_create_gen);
++unvisited_count;
} else {
visited_age_gap_sum +=
(cache->evict_pass_gen - page->cache_create_gen);
gen_gap = cache->evict_pass_gen - page->evict_pass_gen;
if (gen_gap > gen_gap_max)
gen_gap_max = gen_gap;
gen_gap_sum += gen_gap;
++visited_count;
}
}
WT_STAT_DATA_SET(session, cache_state_gen_avg_gap,
visited_count == 0 ? 0 : gen_gap_sum / visited_count);
WT_STAT_DATA_SET(session, cache_state_avg_unvisited_age,
unvisited_count == 0 ? 0 : unvisited_age_gap_sum / unvisited_count);
WT_STAT_DATA_SET(session, cache_state_avg_visited_age,
visited_count == 0 ? 0 : visited_age_gap_sum / visited_count);
WT_STAT_DATA_SET(session, cache_state_avg_written_size,
written_size_cnt == 0 ? 0 : written_size_sum / written_size_cnt);
WT_STAT_DATA_SET(session, cache_state_gen_max_gap, gen_gap_max);
WT_STAT_DATA_SET(session, cache_state_max_pagesize, max_pagesize);
WT_STAT_DATA_SET(session,
cache_state_min_written_size, min_written_size);
WT_STAT_DATA_SET(session, cache_state_memory, num_memory);
WT_STAT_DATA_SET(session, cache_state_queued, num_queued);
WT_STAT_DATA_SET(session, cache_state_not_queueable, num_not_queueable);
WT_STAT_DATA_SET(session, cache_state_pages, walk_count);
WT_STAT_DATA_SET(session, cache_state_pages_clean, pages_clean);
WT_STAT_DATA_SET(session, cache_state_pages_dirty, pages_dirty);
WT_STAT_DATA_SET(session, cache_state_pages_internal, pages_internal);
WT_STAT_DATA_SET(session, cache_state_pages_leaf, pages_leaf);
WT_STAT_DATA_SET(session,
cache_state_refs_skipped, walk_count - seen_count);
WT_STAT_DATA_SET(session,
cache_state_smaller_alloc_size, num_smaller_allocsz);
WT_STAT_DATA_SET(session,
cache_state_unvisited_count, unvisited_count);
}
/*
* __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.
*/
for (newpage = 0;; newpage = 1) {
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 = 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;
}
/*
* 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);
}
/*
* __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_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 saved_snap_min;
uint32_t flags;
bool evict_reset;
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);
}
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 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:
/*
* 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);
/*
* 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.
*/
WT_PUBLISH(btree->checkpointing, WT_CKPT_PREPARE);
WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
if (evict_reset)
__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,
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 = 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)
WT_RET(btree->bm->sync(btree->bm, session, true));
return (ret);
}
/*
* __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;
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;
}
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. */
__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;
/*
* 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;
}
/* 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.
*/
__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 (;;) {
/*
* 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;
}
ret = __wt_page_swap(session, couple, ref, 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;
__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 (WT_PAGE_IS_INTERNAL(page)) {
WT_INTL_INDEX_GET(session, page, pindex);
slot = prev ? pindex->entries - 1 : 0;
empty_internal = true;
} 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, 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;
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, flags));
if (walk == NULL)
break;
/* Write dirty pages if nobody beat us to it. */
page = walk->page;
if (__wt_page_is_modified(page)) {
if (txn->isolation == TXN_ISO_READ_COMMITTED)
__wt_txn_refresh(session, 1);
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;
if (!F_ISSET(btree, WT_BTREE_NO_EVICTION)) {
WT_ERR(__wt_evict_file_exclusive_on(session));
__wt_evict_file_exclusive_off(session);
}
/* 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;
/*
* 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.
*/
page = walk->page;
mod = page->modify;
if (__wt_page_is_modified(page) &&
(WT_PAGE_IS_INTERNAL(page) ||
!F_ISSET(txn, TXN_HAS_SNAPSHOT) ||
TXNID_LE(mod->first_dirty_txn, txn->snap_max))) {
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;
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(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 == TXN_ISO_READ_COMMITTED && session->ncursors == 0)
__wt_txn_release_snapshot(session);
if (btree->checkpointing) {
/*
* 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();
/*
* 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, 1));
return (ret);
}