/*
* __wt_compact --
* Compact a file.
*/
int
__wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_BM *bm;
WT_CONFIG_ITEM cval;
WT_DECL_RET;
WT_PAGE *page;
int trigger, skip;
bm = S2BT(session)->bm;
WT_DSTAT_INCR(session, session_compact);
WT_RET(__wt_config_gets(session, cfg, "trigger", &cval));
trigger = (int)cval.val;
/* Check if compaction might be useful. */
WT_RET(bm->compact_skip(bm, session, trigger, &skip));
if (skip)
return (0);
/*
* Walk the cache reviewing in-memory pages to see if they need to be
* re-written. This requires looking at page reconciliation results,
* which means the page cannot be reconciled at the same time as it's
* being reviewed for compaction. The underlying functions ensure we
* don't collide with page eviction, but we need to make sure we don't
* collide with checkpoints either, they are the other operation that
* can reconcile a page.
*/
__wt_spin_lock(session, &S2C(session)->metadata_lock);
WT_RET(__wt_bt_cache_op(session, NULL, WT_SYNC_COMPACT));
__wt_spin_unlock(session, &S2C(session)->metadata_lock);
/*
* Walk the tree, reviewing on-disk pages to see if they need to be
* re-written.
*/
for (page = NULL;;) {
WT_RET(__wt_tree_walk(session, &page, WT_TREE_COMPACT));
if (page == NULL)
break;
/*
* The only pages returned by the tree walk function are pages
* we want to re-write; mark the page and tree dirty.
*/
if ((ret = __wt_page_modify_init(session, page)) != 0) {
WT_TRET(__wt_page_release(session, page));
WT_RET(ret);
}
__wt_page_and_tree_modify_set(session, page);
WT_DSTAT_INCR(session, btree_compact_rewrite);
}
return (0);
}
/*
* __wt_bt_cache_force_write --
* Dirty the root page of the tree so it gets written.
*/
int
__wt_bt_cache_force_write(WT_SESSION_IMPL *session)
{
WT_BTREE *btree;
WT_PAGE *page;
btree = session->btree;
page = btree->root_page;
/* Dirty the root page to ensure a write. */
WT_RET(__wt_page_modify_init(session, page));
__wt_page_modify_set(session, page);
return (0);
}
/*
* __merge_new_page --
* Create a new in-memory internal page.
*/
static int
__merge_new_page(WT_SESSION_IMPL *session,
uint8_t type, uint32_t entries, int merge, WT_PAGE **pagep)
{
WT_DECL_RET;
WT_PAGE *newpage;
/* Allocate a new internal page and fill it in. */
WT_RET(__wt_page_alloc(session, type, entries, &newpage));
newpage->read_gen = WT_READ_GEN_NOTSET;
newpage->entries = entries;
WT_ERR(__wt_page_modify_init(session, newpage));
if (merge)
F_SET(newpage->modify, WT_PM_REC_SPLIT_MERGE);
else
__wt_page_modify_set(session, newpage);
*pagep = newpage;
return (0);
err: __wt_page_out(session, &newpage);
return (ret);
}
/*
* __wt_delete_page_instantiate --
* Instantiate an entirely deleted row-store leaf page.
*/
int
__wt_delete_page_instantiate(WT_SESSION_IMPL *session, WT_REF *ref)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_DELETED *page_del;
WT_UPDATE **upd_array, *upd;
size_t size;
uint32_t i;
btree = S2BT(session);
page = ref->page;
page_del = ref->page_del;
/*
* Give the page a modify structure.
*
* If the tree is already dirty and so will be written, mark the page
* dirty. (We'd like to free the deleted pages, but if the handle is
* read-only or if the application never modifies the tree, we're not
* able to do so.)
*/
if (btree->modified) {
WT_RET(__wt_page_modify_init(session, page));
__wt_page_modify_set(session, page);
}
/*
* An operation is accessing a "deleted" page, and we're building an
* in-memory version of the page (making it look like all entries in
* the page were individually updated by a remove operation). There
* are two cases where we end up here:
*
* First, a running transaction used a truncate call to delete the page
* without reading it, in which case the page reference includes a
* structure with a transaction ID; the page we're building might split
* in the future, so we update that structure to include references to
* all of the update structures we create, so the transaction can abort.
*
* Second, a truncate call deleted a page and the truncate committed,
* but an older transaction in the system forced us to keep the old
* version of the page around, then we crashed and recovered, and now
* we're being forced to read that page.
*
* In the first case, we have a page reference structure, in the second
* second, we don't.
*
* Allocate the per-reference update array; in the case of instantiating
* a page, deleted by a running transaction that might eventually abort,
* we need a list of the update structures so we can do that abort. The
* hard case is if a page splits: the update structures might be moved
* to different pages, and we still have to find them all for an abort.
*/
if (page_del != NULL)
WT_RET(__wt_calloc_def(
session, page->pg_row_entries + 1, &page_del->update_list));
/* Allocate the per-page update array. */
WT_ERR(__wt_calloc_def(session, page->pg_row_entries, &upd_array));
page->pg_row_upd = upd_array;
/*
* Fill in the per-reference update array with references to update
* structures, fill in the per-page update array with references to
* deleted items.
*/
for (i = 0, size = 0; i < page->pg_row_entries; ++i) {
WT_ERR(__wt_calloc_one(session, &upd));
WT_UPDATE_DELETED_SET(upd);
if (page_del == NULL)
upd->txnid = WT_TXN_NONE; /* Globally visible */
else {
upd->txnid = page_del->txnid;
page_del->update_list[i] = upd;
}
upd->next = upd_array[i];
upd_array[i] = upd;
size += sizeof(WT_UPDATE *) + WT_UPDATE_MEMSIZE(upd);
}
__wt_cache_page_inmem_incr(session, page, size);
return (0);
err: /*
* There's no need to free the page update structures on error, our
* caller will discard the page and do that work for us. We could
* similarly leave the per-reference update array alone because it
* won't ever be used by any page that's not in-memory, but cleaning
* it up makes sense, especially if we come back in to this function
* attempting to instantiate this page again.
*/
if (page_del != NULL)
__wt_free(session, page_del->update_list);
return (ret);
}
/*
* __wt_row_modify --
* Row-store insert, update and delete.
*/
int
__wt_row_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int is_remove)
{
WT_DECL_RET;
WT_INSERT *ins;
WT_INSERT_HEAD *ins_head, **ins_headp;
WT_ITEM *key, *value;
WT_PAGE *page;
WT_UPDATE *old_upd, *upd, **upd_entry;
size_t ins_size, upd_size;
uint32_t ins_slot;
u_int i, skipdepth;
int logged;
key = &cbt->iface.key;
value = is_remove ? NULL : &cbt->iface.value;
page = cbt->page;
/* If we don't yet have a modify structure, we'll need one. */
WT_RET(__wt_page_modify_init(session, page));
ins = NULL;
upd = NULL;
logged = 0;
/*
* Modify: allocate an update array as necessary, build a WT_UPDATE
* structure, and call a serialized function to insert the WT_UPDATE
* structure.
*
* Insert: allocate an insert array as necessary, build a WT_INSERT
* and WT_UPDATE structure pair, and call a serialized function to
* insert the WT_INSERT structure.
*/
if (cbt->compare == 0) {
if (cbt->ins == NULL) {
/* Allocate an update array as necessary. */
WT_PAGE_ALLOC_AND_SWAP(session, page,
page->u.row.upd, upd_entry, page->entries);
/* Set the WT_UPDATE array reference. */
upd_entry = &page->u.row.upd[cbt->slot];
} else
upd_entry = &cbt->ins->upd;
/* Make sure the update can proceed. */
WT_ERR(__wt_txn_update_check(session, old_upd = *upd_entry));
/* Allocate the WT_UPDATE structure and transaction ID. */
WT_ERR(__wt_update_alloc(session, value, &upd, &upd_size));
WT_ERR(__wt_txn_modify(session, cbt, upd));
logged = 1;
/*
* Point the new WT_UPDATE item to the next element in the list.
* If we get it right, the serialization function lock acts as
* our memory barrier to flush this write.
*/
upd->next = old_upd;
/* Serialize the update. */
WT_ERR(__wt_update_serial(
session, page, upd_entry, &upd, upd_size));
} else {
/*
* Allocate the insert array as necessary.
*
* We allocate an additional insert array slot for insert keys
* sorting less than any key on the page. The test to select
* that slot is baroque: if the search returned the first page
* slot, we didn't end up processing an insert list, and the
* comparison value indicates the search key was smaller than
* the returned slot, then we're using the smallest-key insert
* slot. That's hard, so we set a flag.
*/
WT_PAGE_ALLOC_AND_SWAP(session, page,
page->u.row.ins, ins_headp, page->entries + 1);
ins_slot = F_ISSET(cbt, WT_CBT_SEARCH_SMALLEST) ?
page->entries : cbt->slot;
ins_headp = &page->u.row.ins[ins_slot];
/* Allocate the WT_INSERT_HEAD structure as necessary. */
WT_PAGE_ALLOC_AND_SWAP(session, page, *ins_headp, ins_head, 1);
ins_head = *ins_headp;
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth();
/*
* Allocate a WT_INSERT/WT_UPDATE pair and transaction ID, and
* update the cursor to reference it.
*/
WT_ERR(__wt_row_insert_alloc(
session, key, skipdepth, &ins, &ins_size));
WT_ERR(__wt_update_alloc(session, value, &upd, &upd_size));
ins->upd = upd;
ins_size += upd_size;
/*
* Update the cursor: the WT_INSERT_HEAD might be allocated,
* the WT_INSERT was allocated.
*/
cbt->ins_head = ins_head;
cbt->ins = ins;
WT_ERR(__wt_txn_modify(session, cbt, upd));
logged = 1;
/*
* If there was no insert list during the search, the cursor's
* information cannot be correct, search couldn't have
* initialized it.
*
* Otherwise, point the new WT_INSERT item's skiplist to the
* next elements in the insert list (which we will check are
* still valid inside the serialization function).
*
* The serial mutex acts as our memory barrier to flush these
* writes before inserting them into the list.
*/
if (WT_SKIP_FIRST(ins_head) == NULL)
for (i = 0; i < skipdepth; i++) {
cbt->ins_stack[i] = &ins_head->head[i];
ins->next[i] = cbt->next_stack[i] = NULL;
}
else
for (i = 0; i < skipdepth; i++)
ins->next[i] = cbt->next_stack[i];
/* Insert the WT_INSERT structure. */
WT_ERR(__wt_insert_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, skipdepth));
}
if (0) {
err: /*
* Remove the update from the current transaction, so we don't
* try to modify it on rollback.
*/
if (logged)
__wt_txn_unmodify(session);
__wt_free(session, ins);
cbt->ins = NULL;
__wt_free(session, upd);
}
return (ret);
}
/*
* __wt_compact --
* Compact a file.
*/
int
__wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
WT_REF *ref;
bool block_manager_begin, skip;
WT_UNUSED(cfg);
btree = S2BT(session);
bm = btree->bm;
ref = NULL;
block_manager_begin = false;
WT_STAT_FAST_DATA_INCR(session, session_compact);
/*
* Check if compaction might be useful -- the API layer will quit trying
* to compact the data source if we make no progress, set a flag if the
* block layer thinks compaction is possible.
*/
WT_RET(bm->compact_skip(bm, session, &skip));
if (skip)
return (0);
/*
* Reviewing in-memory pages requires looking at page reconciliation
* results, because we care about where the page is stored now, not
* where the page was stored when we first read it into the cache.
* We need to ensure we don't race with page reconciliation as it's
* writing the page modify information.
*
* There are three ways we call reconciliation: checkpoints, threads
* writing leaf pages (usually in preparation for a checkpoint or if
* closing a file), and eviction.
*
* We're holding the schema lock which serializes with checkpoints.
*/
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SCHEMA));
/*
* Get the tree handle's flush lock which blocks threads writing leaf
* pages.
*/
__wt_spin_lock(session, &btree->flush_lock);
/* Start compaction. */
WT_ERR(bm->compact_start(bm, session));
block_manager_begin = true;
/* Walk the tree reviewing pages to see if they should be re-written. */
for (;;) {
/*
* Pages read for compaction aren't "useful"; don't update the
* read generation of pages already in memory, and if a page is
* read, set its generation to a low value so it is evicted
* quickly.
*/
WT_ERR(__wt_tree_walk(session, &ref,
WT_READ_COMPACT | WT_READ_NO_GEN | WT_READ_WONT_NEED));
if (ref == NULL)
break;
WT_ERR(__compact_rewrite(session, ref, &skip));
if (skip)
continue;
session->compact_state = WT_COMPACT_SUCCESS;
/* Rewrite the page: mark the page and tree dirty. */
WT_ERR(__wt_page_modify_init(session, ref->page));
__wt_page_modify_set(session, ref->page);
WT_STAT_FAST_DATA_INCR(session, btree_compact_rewrite);
}
err:
if (ref != NULL)
WT_TRET(__wt_page_release(session, ref, 0));
if (block_manager_begin)
WT_TRET(bm->compact_end(bm, session));
/* Unblock threads writing leaf pages. */
__wt_spin_unlock(session, &btree->flush_lock);
return (ret);
}
/*
* __wt_col_modify --
* Column-store delete, insert, and update.
*/
int
__wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
uint64_t recno, WT_ITEM *value, WT_UPDATE *upd, int is_remove)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_INSERT *ins;
WT_INSERT_HEAD *ins_head, **ins_headp;
WT_ITEM _value;
WT_PAGE *page;
WT_UPDATE *old_upd;
size_t ins_size, upd_size;
u_int i, skipdepth;
int append, logged;
btree = cbt->btree;
ins = NULL;
page = cbt->ref->page;
append = logged = 0;
/* This code expects a remove to have a NULL value. */
if (is_remove) {
if (btree->type == BTREE_COL_FIX) {
value = &_value;
value->data = "";
value->size = 1;
} else
value = NULL;
} else {
/*
* There's some chance the application specified a record past
* the last record on the page. If that's the case, and we're
* inserting a new WT_INSERT/WT_UPDATE pair, it goes on the
* append list, not the update list. In addition, a recno of 0
* implies an append operation, we're allocating a new row.
*/
if (recno == 0 ||
recno > (btree->type == BTREE_COL_VAR ?
__col_var_last_recno(page) : __col_fix_last_recno(page)))
append = 1;
}
/* If we don't yet have a modify structure, we'll need one. */
WT_RET(__wt_page_modify_init(session, page));
/*
* Delete, insert or update a column-store entry.
*
* If modifying a previously modified record, create a new WT_UPDATE
* entry and have a serialized function link it into an existing
* WT_INSERT entry's WT_UPDATE list.
*
* Else, allocate an insert array as necessary, build a WT_INSERT and
* WT_UPDATE structure pair, and call a serialized function to insert
* the WT_INSERT structure.
*/
if (cbt->compare == 0 && cbt->ins != NULL) {
/*
* If we are restoring updates that couldn't be evicted, the
* key must not exist on the new page.
*/
WT_ASSERT(session, upd == NULL);
/* Make sure the update can proceed. */
WT_ERR(__wt_txn_update_check(
session, old_upd = cbt->ins->upd));
/* Allocate a WT_UPDATE structure and transaction ID. */
WT_ERR(__wt_update_alloc(session, value, &upd, &upd_size));
WT_ERR(__wt_txn_modify(session, upd));
logged = 1;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
/*
* Point the new WT_UPDATE item to the next element in the list.
* If we get it right, the serialization function lock acts as
* our memory barrier to flush this write.
*/
upd->next = old_upd;
/* Serialize the update. */
WT_ERR(__wt_update_serial(
session, page, &cbt->ins->upd, &upd, upd_size));
} else {
/* Allocate the append/update list reference as necessary. */
if (append) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, page->modify->mod_append, ins_headp, 1);
ins_headp = &page->modify->mod_append[0];
} else if (page->type == WT_PAGE_COL_FIX) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, page->modify->mod_update, ins_headp, 1);
ins_headp = &page->modify->mod_update[0];
} else {
WT_PAGE_ALLOC_AND_SWAP(session,
page, page->modify->mod_update, ins_headp,
page->pg_var_entries);
ins_headp = &page->modify->mod_update[cbt->slot];
}
/* Allocate the WT_INSERT_HEAD structure as necessary. */
WT_PAGE_ALLOC_AND_SWAP(session, page, *ins_headp, ins_head, 1);
ins_head = *ins_headp;
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth(session);
/*
* Allocate a WT_INSERT/WT_UPDATE pair and transaction ID, and
* update the cursor to reference it (the WT_INSERT_HEAD might
* be allocated, the WT_INSERT was allocated).
*/
WT_ERR(__col_insert_alloc(
session, recno, skipdepth, &ins, &ins_size));
cbt->ins_head = ins_head;
cbt->ins = ins;
if (upd == NULL) {
WT_ERR(
__wt_update_alloc(session, value, &upd, &upd_size));
WT_ERR(__wt_txn_modify(session, upd));
logged = 1;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
} else
upd_size = __wt_update_list_memsize(upd);
ins->upd = upd;
ins_size += upd_size;
/*
* If there was no insert list during the search, or there was
* no search because the record number has not been allocated
* yet, the cursor's information cannot be correct, search
* couldn't have initialized it.
*
* Otherwise, point the new WT_INSERT item's skiplist to the
* next elements in the insert list (which we will check are
* still valid inside the serialization function).
*
* The serial mutex acts as our memory barrier to flush these
* writes before inserting them into the list.
*/
if (WT_SKIP_FIRST(ins_head) == NULL || recno == 0)
for (i = 0; i < skipdepth; i++) {
cbt->ins_stack[i] = &ins_head->head[i];
ins->next[i] = cbt->next_stack[i] = NULL;
}
else
for (i = 0; i < skipdepth; i++)
ins->next[i] = cbt->next_stack[i];
/* Append or insert the WT_INSERT structure. */
if (append)
WT_ERR(__wt_col_append_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, &cbt->recno, skipdepth));
else
WT_ERR(__wt_insert_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, skipdepth));
}
/* If the update was successful, add it to the in-memory log. */
if (logged)
WT_ERR(__wt_txn_log_op(session, cbt));
if (0) {
err: /*
* Remove the update from the current transaction, so we don't
* try to modify it on rollback.
*/
if (logged)
__wt_txn_unmodify(session);
__wt_free(session, ins);
__wt_free(session, upd);
}
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, int discard)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_DSTAT_INCR(session, cursor_prev);
flags = WT_TREE_SKIP_INTL | WT_TREE_PREV; /* Tree walk flags. */
if (discard)
LF_SET(WT_TREE_DISCARD);
retry: WT_RET(__cursor_func_init(cbt, 0));
__cursor_position_clear(cbt);
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_PREV))
__wt_btcur_iterate_setup(cbt, 0);
/*
* If this is a modification, we're about to read information from the
* page, save the write generation.
*/
page = cbt->page;
if (discard && page != NULL) {
WT_ERR(__wt_page_modify_init(session, page));
WT_ORDERED_READ(cbt->write_gen, page->modify->write_gen);
}
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the previous page, until we reach the start
* of the file.
*/
for (newpage = 0;; newpage = 1) {
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
newpage = 1;
}
if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_prev(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
}
cbt->page = NULL;
WT_ERR(__wt_tree_walk(session, &page, flags));
WT_ERR_TEST(page == NULL, WT_NOTFOUND);
WT_ASSERT(session,
page->type != WT_PAGE_COL_INT &&
page->type != WT_PAGE_ROW_INT);
cbt->page = page;
/* Initialize the page's modification information */
if (discard) {
WT_ERR(__wt_page_modify_init(session, page));
WT_ORDERED_READ(
cbt->write_gen, page->modify->write_gen);
}
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
err: if (ret == WT_RESTART)
goto retry;
WT_TRET(__cursor_func_resolve(cbt, ret));
return (ret);
}
/*
* __wt_col_modify --
* Column-store delete, insert, and update.
*/
int
__wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int op)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_INSERT *ins, *ins_copy;
WT_INSERT_HEAD **inshead, *new_inshead, **new_inslist;
WT_ITEM *value, _value;
WT_PAGE *page;
WT_UPDATE *old_upd, *upd, *upd_obsolete;
size_t ins_size, new_inshead_size, new_inslist_size, upd_size;
uint64_t recno;
u_int skipdepth;
int i, logged;
btree = cbt->btree;
page = cbt->page;
recno = cbt->iface.recno;
logged = 0;
WT_ASSERT(session, op != 1);
switch (op) {
case 2: /* Remove */
if (btree->type == BTREE_COL_FIX) {
value = &_value;
value->data = "";
value->size = 1;
} else
value = NULL;
break;
case 3: /* Insert/Update */
default:
value = &cbt->iface.value;
/*
* There's some chance the application specified a record past
* the last record on the page. If that's the case, and we're
* inserting a new WT_INSERT/WT_UPDATE pair, it goes on the
* append list, not the update list.
*/
if (recno == 0 || recno > __col_last_recno(page))
op = 1;
break;
}
/* If we don't yet have a modify structure, we'll need one. */
WT_RET(__wt_page_modify_init(session, page));
ins = NULL;
new_inshead = NULL;
new_inslist = NULL;
upd = NULL;
/*
* Delete, insert or update a column-store entry.
*
* If modifying a previously modified record, create a new WT_UPDATE
* entry and have a serialized function link it into an existing
* WT_INSERT entry's WT_UPDATE list.
*
* Else, allocate an insert array as necessary, build a WT_INSERT and
* WT_UPDATE structure pair, and call a serialized function to insert
* the WT_INSERT structure.
*/
if (cbt->compare == 0 && cbt->ins != NULL) {
/* Make sure the update can proceed. */
WT_ERR(
__wt_update_check(session, page, old_upd = cbt->ins->upd));
/* Allocate the WT_UPDATE structure and transaction ID. */
WT_ERR(__wt_update_alloc(session, value, &upd, &upd_size));
WT_ERR(__wt_txn_modify(session, &upd->txnid));
logged = 1;
/* Serialize the update. */
WT_ERR(__wt_update_serial(session, page,
cbt->write_gen, &cbt->ins->upd, old_upd,
NULL, 0, &upd, upd_size, &upd_obsolete));
/* Discard any obsolete WT_UPDATE structures. */
if (upd_obsolete != NULL)
__wt_update_obsolete_free(session, page, upd_obsolete);
} else {
/* Make sure the update can proceed. */
WT_ERR(__wt_update_check(session, page, NULL));
/* There may be no insert list, allocate as necessary. */
new_inshead_size = new_inslist_size = 0;
if (op == 1) {
if (page->modify->append == NULL) {
new_inslist_size = 1 * sizeof(WT_INSERT_HEAD *);
WT_ERR(
__wt_calloc_def(session, 1, &new_inslist));
inshead = &new_inslist[0];
} else
inshead = &page->modify->append[0];
cbt->ins_head = *inshead;
} else if (page->type == WT_PAGE_COL_FIX) {
if (page->modify->update == NULL) {
new_inslist_size = 1 * sizeof(WT_INSERT_HEAD *);
WT_ERR(
__wt_calloc_def(session, 1, &new_inslist));
inshead = &new_inslist[0];
} else
inshead = &page->modify->update[0];
} else {
if (page->modify->update == NULL) {
new_inslist_size =
page->entries * sizeof(WT_INSERT_HEAD *);
WT_ERR(__wt_calloc_def(
session, page->entries, &new_inslist));
inshead = &new_inslist[cbt->slot];
} else
inshead = &page->modify->update[cbt->slot];
}
/* There may be no WT_INSERT list, allocate as necessary. */
if (*inshead == NULL) {
new_inshead_size = sizeof(WT_INSERT_HEAD);
WT_ERR(__wt_calloc_def(session, 1, &new_inshead));
for (i = 0; i < WT_SKIP_MAXDEPTH; i++) {
cbt->ins_stack[i] = &new_inshead->head[i];
cbt->next_stack[i] = NULL;
}
cbt->ins_head = new_inshead;
}
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth();
/*
* Allocate a WT_INSERT/WT_UPDATE pair and transaction ID, and
* update the cursor to reference it.
*/
WT_ERR(__col_insert_alloc(
session, recno, skipdepth, &ins, &ins_size));
WT_ERR(__wt_update_alloc(session, value, &upd, &upd_size));
WT_ERR(__wt_txn_modify(session, &upd->txnid));
logged = 1;
ins->upd = upd;
ins_size += upd_size;
cbt->ins = ins;
/* Insert or append the WT_INSERT structure. */
if (op == 1) {
/*
* The serialized function clears ins: take a copy of
* the pointer so we can look up the record number.
*/
ins_copy = ins;
WT_ERR(__wt_col_append_serial(session,
page, cbt->write_gen, inshead,
cbt->ins_stack, cbt->next_stack,
&new_inslist, new_inslist_size,
&new_inshead, new_inshead_size,
&ins, ins_size, skipdepth));
/* Put the new recno into the cursor. */
cbt->recno = WT_INSERT_RECNO(ins_copy);
} else
WT_ERR(__wt_insert_serial(session,
page, cbt->write_gen, inshead,
cbt->ins_stack, cbt->next_stack,
&new_inslist, new_inslist_size,
&new_inshead, new_inshead_size,
&ins, ins_size, skipdepth));
}
if (0) {
err: /*
* Remove the update from the current transaction, so we don't
* try to modify it on rollback.
*/
if (logged)
__wt_txn_unmodify(session);
__wt_free(session, ins);
__wt_free(session, upd);
}
__wt_free(session, new_inslist);
__wt_free(session, new_inshead);
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_col_modify --
* Column-store delete, insert, and update.
*/
int
__wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
uint64_t recno, const WT_ITEM *value,
WT_UPDATE *upd_arg, u_int modify_type, bool exclusive)
{
static const WT_ITEM col_fix_remove = { "", 1, NULL, 0, 0 };
WT_BTREE *btree;
WT_DECL_RET;
WT_INSERT *ins;
WT_INSERT_HEAD *ins_head, **ins_headp;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_UPDATE *old_upd, *upd;
size_t ins_size, upd_size;
u_int i, skipdepth;
bool append, logged;
btree = cbt->btree;
ins = NULL;
page = cbt->ref->page;
upd = upd_arg;
append = logged = false;
if (upd_arg == NULL) {
if (modify_type == WT_UPDATE_RESERVE ||
modify_type == WT_UPDATE_TOMBSTONE) {
/*
* Fixed-size column-store doesn't have on-page deleted
* values, it's a nul byte.
*/
if (modify_type == WT_UPDATE_TOMBSTONE &&
btree->type == BTREE_COL_FIX) {
modify_type = WT_UPDATE_STANDARD;
value = &col_fix_remove;
}
}
/*
* There's a chance the application specified a record past the
* last record on the page. If that's the case and we're
* inserting a new WT_INSERT/WT_UPDATE pair, it goes on the
* append list, not the update list. Also, an out-of-band recno
* implies an append operation, we're allocating a new row.
* Ignore any information obtained from the search.
*/
WT_ASSERT(session, recno != WT_RECNO_OOB || cbt->compare != 0);
if (cbt->compare != 0 &&
(recno == WT_RECNO_OOB ||
recno > (btree->type == BTREE_COL_VAR ?
__col_var_last_recno(cbt->ref) :
__col_fix_last_recno(cbt->ref)))) {
append = true;
cbt->ins = NULL;
cbt->ins_head = NULL;
}
}
/* We're going to modify the page, we should have loaded history. */
WT_ASSERT(session, cbt->ref->state != WT_REF_LIMBO);
/* If we don't yet have a modify structure, we'll need one. */
WT_RET(__wt_page_modify_init(session, page));
mod = page->modify;
/*
* If modifying a record not previously modified, but which is in the
* same update slot as a previously modified record, cursor.ins will
* not be set because there's no list of update records for this recno,
* but cursor.ins_head will be set to point to the correct update slot.
* Acquire the necessary insert information, then create a new update
* entry and link it into the existing list. We get here if a page has
* a single cell representing multiple records (the records have the
* same value), and then a record in the cell is updated or removed,
* creating the update list for the cell, and then a cursor iterates
* into that same cell to update/remove a different record. We find the
* correct slot in the update array, but we don't find an update list
* (because it doesn't exist), and don't have the information we need
* to do the insert. Normally, we wouldn't care (we could fail and do
* a search for the record which would configure everything for the
* insert), but range truncation does this pattern for every record in
* the cell, and the performance is terrible. For that reason, catch it
* here.
*/
if (cbt->ins == NULL && cbt->ins_head != NULL) {
cbt->ins = __col_insert_search(
cbt->ins_head, cbt->ins_stack, cbt->next_stack, recno);
if (cbt->ins != NULL) {
if (WT_INSERT_RECNO(cbt->ins) == recno)
cbt->compare = 0;
else {
/*
* The test below is for cursor.compare set to 0
* and cursor.ins set: cursor.compare wasn't set
* by the search we just did, and has an unknown
* value. Clear cursor.ins to avoid the test.
*/
cbt->ins = NULL;
}
}
}
/*
* Delete, insert or update a column-store entry.
*
* If modifying a previously modified record, cursor.ins will be set to
* point to the correct update list. Create a new update entry and link
* it into the existing list.
*
* Else, allocate an insert array as necessary, build an insert/update
* structure pair, and link it into place.
*/
if (cbt->compare == 0 && cbt->ins != NULL) {
/*
* If we are restoring updates that couldn't be evicted, the
* key must not exist on the new page.
*/
WT_ASSERT(session, upd_arg == NULL);
/* Make sure the update can proceed. */
WT_ERR(__wt_txn_update_check(session, old_upd = cbt->ins->upd));
/* Allocate a WT_UPDATE structure and transaction ID. */
WT_ERR(__wt_update_alloc(session,
value, &upd, &upd_size, modify_type));
WT_ERR(__wt_txn_modify(session, upd));
logged = true;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
/*
* Point the new WT_UPDATE item to the next element in the list.
* If we get it right, the serialization function lock acts as
* our memory barrier to flush this write.
*/
upd->next = old_upd;
/* Serialize the update. */
WT_ERR(__wt_update_serial(
session, page, &cbt->ins->upd, &upd, upd_size, false));
} else {
/* Allocate the append/update list reference as necessary. */
if (append) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, mod->mod_col_append, ins_headp, 1);
ins_headp = &mod->mod_col_append[0];
} else if (page->type == WT_PAGE_COL_FIX) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, mod->mod_col_update, ins_headp, 1);
ins_headp = &mod->mod_col_update[0];
} else {
WT_PAGE_ALLOC_AND_SWAP(session, page,
mod->mod_col_update, ins_headp, page->entries);
ins_headp = &mod->mod_col_update[cbt->slot];
}
/* Allocate the WT_INSERT_HEAD structure as necessary. */
WT_PAGE_ALLOC_AND_SWAP(session, page, *ins_headp, ins_head, 1);
ins_head = *ins_headp;
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth(session);
/*
* Allocate a WT_INSERT/WT_UPDATE pair and transaction ID, and
* update the cursor to reference it (the WT_INSERT_HEAD might
* be allocated, the WT_INSERT was allocated).
*/
WT_ERR(__col_insert_alloc(
session, recno, skipdepth, &ins, &ins_size));
cbt->ins_head = ins_head;
cbt->ins = ins;
/*
* Check for insert split and checkpoint races in column-store:
* it's easy (as opposed to in row-store) and a difficult bug to
* otherwise diagnose.
*/
WT_ASSERT(session, mod->mod_col_split_recno == WT_RECNO_OOB ||
(recno != WT_RECNO_OOB &&
mod->mod_col_split_recno > recno));
if (upd_arg == NULL) {
WT_ERR(__wt_update_alloc(session,
value, &upd, &upd_size, modify_type));
WT_ERR(__wt_txn_modify(session, upd));
logged = true;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
} else
upd_size = __wt_update_list_memsize(upd);
ins->upd = upd;
ins_size += upd_size;
/*
* If there was no insert list during the search, or there was
* no search because the record number has not been allocated
* yet, the cursor's information cannot be correct, search
* couldn't have initialized it.
*
* Otherwise, point the new WT_INSERT item's skiplist to the
* next elements in the insert list (which we will check are
* still valid inside the serialization function).
*
* The serial mutex acts as our memory barrier to flush these
* writes before inserting them into the list.
*/
if (cbt->ins_stack[0] == NULL || recno == WT_RECNO_OOB)
for (i = 0; i < skipdepth; i++) {
cbt->ins_stack[i] = &ins_head->head[i];
ins->next[i] = cbt->next_stack[i] = NULL;
}
else
for (i = 0; i < skipdepth; i++)
ins->next[i] = cbt->next_stack[i];
/* Append or insert the WT_INSERT structure. */
if (append)
WT_ERR(__wt_col_append_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, &cbt->recno, skipdepth, exclusive));
else
WT_ERR(__wt_insert_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, skipdepth, exclusive));
}
/* If the update was successful, add it to the in-memory log. */
if (logged && modify_type != WT_UPDATE_RESERVE) {
WT_ERR(__wt_txn_log_op(session, cbt));
/*
* In case of append, the recno (key) for the value is assigned
* now. Set the recno in the transaction operation to be used
* incase this transaction is prepared to retrieve the update
* corresponding to this operation.
*/
__wt_txn_op_set_recno(session, cbt->recno);
}
if (0) {
err: /*
* Remove the update from the current transaction, so we don't
* try to modify it on rollback.
*/
if (logged)
__wt_txn_unmodify(session);
__wt_free(session, ins);
if (upd_arg == NULL)
__wt_free(session, upd);
}
return (ret);
}
/*
* __wt_compact_evict --
* Helper routine to decide if a file's size would benefit from re-writing
* this page.
*/
int
__wt_compact_evict(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_BM *bm;
WT_PAGE_MODIFY *mod;
int skip;
uint32_t addr_size;
const uint8_t *addr;
bm = S2BT(session)->bm;
mod = page->modify;
/*
* We have to review page reconciliation information as an in-memory
* page's original disk addresses might have been fine for compaction
* but its replacement addresses might be a problem. To review page
* reconciliation information, we have to lock out both eviction and
* checkpoints, as those are the other two operations that can write
* a page.
*
* Ignore the root: it may not have a replacement address, and besides,
* if anything else gets written, so will it.
*/
if (WT_PAGE_IS_ROOT(page))
return (0);
/*
* If the page is already dirty, skip some work, it will be written in
* any case.
*/
if (__wt_page_is_modified(page))
return (0);
/*
* If the page is clean, test the original addresses.
* If the page is a 1-to-1 replacement, test the replacement addresses.
* If the page is a split, ignore it, it will be merged into the parent.
*/
if (mod == NULL)
goto disk;
switch (F_ISSET(mod, WT_PM_REC_MASK)) {
case 0:
disk: __wt_get_addr(page->parent, page->ref, &addr, &addr_size);
if (addr == NULL)
return (0);
WT_RET(
bm->compact_page_skip(bm, session, addr, addr_size, &skip));
if (skip)
return (0);
break;
case WT_PM_REC_EMPTY:
return (0);
case WT_PM_REC_REPLACE:
WT_RET(bm->compact_page_skip(bm,
session, mod->u.replace.addr, mod->u.replace.size, &skip));
if (skip)
return (0);
break;
case WT_PM_REC_SPLIT:
case WT_PM_REC_SPLIT_MERGE:
return (0);
}
/* Mark the page and tree dirty, we want to write this page. */
WT_RET(__wt_page_modify_init(session, page));
__wt_page_and_tree_modify_set(session, page);
WT_DSTAT_INCR(session, btree_compact_rewrite);
return (0);
}
/*
* __wt_compact --
* Compact a file.
*/
int
__wt_compact(WT_SESSION_IMPL *session)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
WT_REF *ref;
u_int i;
bool skip;
btree = S2BT(session);
bm = btree->bm;
ref = NULL;
WT_STAT_DATA_INCR(session, session_compact);
/*
* Check if compaction might be useful -- the API layer will quit trying
* to compact the data source if we make no progress, set a flag if the
* block layer thinks compaction is possible.
*/
WT_RET(bm->compact_skip(bm, session, &skip));
if (skip)
return (0);
/*
* Reviewing in-memory pages requires looking at page reconciliation
* results, because we care about where the page is stored now, not
* where the page was stored when we first read it into the cache.
* We need to ensure we don't race with page reconciliation as it's
* writing the page modify information.
*
* There are two ways we call reconciliation: checkpoints and eviction.
* Get the tree's flush lock which blocks threads writing pages for
* checkpoints.
*/
__wt_spin_lock(session, &btree->flush_lock);
/* Walk the tree reviewing pages to see if they should be re-written. */
for (i = 0;;) {
/* Periodically check if we've run out of time. */
if (++i > 100) {
WT_ERR(__wt_session_compact_check_timeout(session));
i = 0;
}
/*
* Pages read for compaction aren't "useful"; don't update the
* read generation of pages already in memory, and if a page is
* read, set its generation to a low value so it is evicted
* quickly.
*/
WT_ERR(__wt_tree_walk(session, &ref,
WT_READ_COMPACT | WT_READ_NO_GEN | WT_READ_WONT_NEED));
if (ref == NULL)
break;
WT_ERR(__compact_rewrite(session, ref, &skip));
if (skip)
continue;
session->compact_state = WT_COMPACT_SUCCESS;
/* Rewrite the page: mark the page and tree dirty. */
WT_ERR(__wt_page_modify_init(session, ref->page));
__wt_page_modify_set(session, ref->page);
WT_STAT_DATA_INCR(session, btree_compact_rewrite);
}
err: if (ref != NULL)
WT_TRET(__wt_page_release(session, ref, 0));
/* Unblock threads writing leaf pages. */
__wt_spin_unlock(session, &btree->flush_lock);
return (ret);
}
/*
* __wt_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);
}
