/*
* __wt_page_modify_alloc --
* Allocate a page's modification structure.
*/
int
__wt_page_modify_alloc(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_CONNECTION_IMPL *conn;
WT_PAGE_MODIFY *modify;
conn = S2C(session);
WT_RET(__wt_calloc_one(session, &modify));
/*
* Select a spinlock for the page; let the barrier immediately below
* keep things from racing too badly.
*/
modify->page_lock = ++conn->page_lock_cnt % WT_PAGE_LOCKS;
/*
* Multiple threads of control may be searching and deciding to modify
* a page. If our modify structure is used, update the page's memory
* footprint, else discard the modify structure, another thread did the
* work.
*/
if (__wt_atomic_cas_ptr(&page->modify, NULL, modify))
__wt_cache_page_inmem_incr(session, page, sizeof(*modify));
else
__wt_free(session, modify);
return (0);
}
/*
* __wt_ovfl_txnc_add --
* Add a new entry to the page's list of transaction-cached overflow
* records.
*/
int
__wt_ovfl_txnc_add(WT_SESSION_IMPL *session, WT_PAGE *page,
const uint8_t *addr, size_t addr_size,
const void *value, size_t value_size)
{
WT_OVFL_TXNC **head, **stack[WT_SKIP_MAXDEPTH], *txnc;
size_t size;
u_int i, skipdepth;
uint8_t *p;
if (page->modify->ovfl_track == NULL)
WT_RET(__ovfl_track_init(session, page));
head = page->modify->ovfl_track->ovfl_txnc;
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth(session);
/*
* Allocate the WT_OVFL_TXNC structure, next pointers for the skip
* list, room for the address and value, then copy everything into
* place.
*
* To minimize the WT_OVFL_TXNC structure size, the address offset
* and size are single bytes: that's safe because the address follows
* the structure (which can't be more than about 100B), and address
* cookies are limited to 255B.
*/
size = sizeof(WT_OVFL_TXNC) +
skipdepth * sizeof(WT_OVFL_TXNC *) + addr_size + value_size;
WT_RET(__wt_calloc(session, 1, size, &txnc));
p = (uint8_t *)txnc +
sizeof(WT_OVFL_TXNC) + skipdepth * sizeof(WT_OVFL_TXNC *);
txnc->addr_offset = (uint8_t)WT_PTRDIFF(p, txnc);
txnc->addr_size = (uint8_t)addr_size;
memcpy(p, addr, addr_size);
p += addr_size;
txnc->value_offset = WT_PTRDIFF32(p, txnc);
txnc->value_size = WT_STORE_SIZE(value_size);
memcpy(p, value, value_size);
txnc->current = __wt_txn_new_id(session);
__wt_cache_page_inmem_incr(
session, page, WT_OVFL_SIZE(txnc, WT_OVFL_TXNC));
/* Insert the new entry into the skiplist. */
__ovfl_txnc_skip_search_stack(head, stack, addr, addr_size);
for (i = 0; i < skipdepth; ++i) {
txnc->next[i] = *stack[i];
*stack[i] = txnc;
}
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(__ovfl_txnc_verbose(session, page, txnc, "add"));
return (0);
}
/*
* __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);
}
/*
* __las_page_instantiate --
* Instantiate lookaside update records in a recently read page.
*/
static int
__las_page_instantiate(WT_SESSION_IMPL *session,
WT_REF *ref, uint32_t read_id, const uint8_t *addr, size_t addr_size)
{
WT_CURSOR *cursor;
WT_CURSOR_BTREE cbt;
WT_DECL_ITEM(current_key);
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_ITEM(las_value);
WT_DECL_RET;
WT_PAGE *page;
WT_UPDATE *first_upd, *last_upd, *upd;
size_t incr, total_incr;
uint64_t current_recno, las_counter, las_txnid, recno, upd_txnid;
uint32_t las_id, upd_size, session_flags;
int exact;
const uint8_t *p;
cursor = NULL;
page = ref->page;
first_upd = last_upd = upd = NULL;
total_incr = 0;
current_recno = recno = WT_RECNO_OOB;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
__wt_btcur_init(session, &cbt);
__wt_btcur_open(&cbt);
WT_ERR(__wt_scr_alloc(session, 0, ¤t_key));
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
WT_ERR(__wt_scr_alloc(session, 0, &las_value));
/* Open a lookaside table cursor. */
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/*
* The lookaside records are in key and update order, that is, there
* will be a set of in-order updates for a key, then another set of
* in-order updates for a subsequent key. We process all of the updates
* for a key and then insert those updates into the page, then all the
* updates for the next key, and so on.
*
* Search for the block's unique prefix, stepping through any matching
* records.
*/
las_addr->data = addr;
las_addr->size = addr_size;
las_key->size = 0;
cursor->set_key(
cursor, read_id, las_addr, (uint64_t)0, (uint32_t)0, las_key);
if ((ret = cursor->search_near(cursor, &exact)) == 0 && exact < 0)
ret = cursor->next(cursor);
for (; ret == 0; ret = cursor->next(cursor)) {
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* Confirm the search using the unique prefix; if not a match,
* we're done searching for records for this page.
*/
if (las_id != read_id ||
las_addr->size != addr_size ||
memcmp(las_addr->data, addr, addr_size) != 0)
break;
/*
* If the on-page value has become globally visible, this record
* is no longer needed.
*/
if (__wt_txn_visible_all(session, las_txnid))
continue;
/* Allocate the WT_UPDATE structure. */
WT_ERR(cursor->get_value(
cursor, &upd_txnid, &upd_size, las_value));
WT_ERR(__wt_update_alloc(session,
(upd_size == WT_UPDATE_DELETED_VALUE) ? NULL : las_value,
&upd, &incr));
total_incr += incr;
upd->txnid = upd_txnid;
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
p = las_key->data;
WT_ERR(__wt_vunpack_uint(&p, 0, &recno));
if (current_recno == recno)
break;
WT_ASSERT(session, current_recno < recno);
if (first_upd != NULL) {
WT_ERR(__col_instantiate(session,
current_recno, ref, &cbt, first_upd));
first_upd = NULL;
}
current_recno = recno;
break;
case WT_PAGE_ROW_LEAF:
if (current_key->size == las_key->size &&
memcmp(current_key->data,
las_key->data, las_key->size) == 0)
break;
if (first_upd != NULL) {
WT_ERR(__row_instantiate(session,
current_key, ref, &cbt, first_upd));
first_upd = NULL;
}
WT_ERR(__wt_buf_set(session,
current_key, las_key->data, las_key->size));
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Append the latest update to the list. */
if (first_upd == NULL)
first_upd = last_upd = upd;
else {
last_upd->next = upd;
last_upd = upd;
}
upd = NULL;
}
WT_ERR_NOTFOUND_OK(ret);
/* Insert the last set of updates, if any. */
if (first_upd != NULL)
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
WT_ERR(__col_instantiate(session,
current_recno, ref, &cbt, first_upd));
first_upd = NULL;
break;
case WT_PAGE_ROW_LEAF:
WT_ERR(__row_instantiate(session,
current_key, ref, &cbt, first_upd));
first_upd = NULL;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Discard the cursor. */
WT_ERR(__wt_las_cursor_close(session, &cursor, session_flags));
if (total_incr != 0) {
__wt_cache_page_inmem_incr(session, page, total_incr);
/*
* We've modified/dirtied the page, but that's not necessary and
* if we keep the page clean, it's easier to evict. We leave the
* lookaside table updates in place, so if we evict this page
* without dirtying it, any future instantiation of it will find
* the records it needs. If the page is dirtied before eviction,
* then we'll write any needed lookaside table records for the
* new location of the page.
*/
__wt_page_modify_clear(session, page);
}
err: WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
WT_TRET(__wt_btcur_close(&cbt, 1));
/*
* On error, upd points to a single unlinked WT_UPDATE structure,
* first_upd points to a list.
*/
if (upd != NULL)
__wt_free(session, upd);
if (first_upd != NULL)
__wt_free_update_list(session, first_upd);
__wt_scr_free(session, ¤t_key);
__wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
__wt_scr_free(session, &las_value);
return (ret);
}
/*
* __wt_page_inmem --
* Build in-memory page information.
*/
int
__wt_page_inmem(WT_SESSION_IMPL *session,
WT_REF *ref, const void *image, uint32_t flags, WT_PAGE **pagep)
{
WT_DECL_RET;
WT_PAGE *page;
const WT_PAGE_HEADER *dsk;
uint32_t alloc_entries;
size_t size;
*pagep = NULL;
dsk = image;
alloc_entries = 0;
/*
* Figure out how many underlying objects the page references so we can
* allocate them along with the page.
*/
switch (dsk->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_INT:
case WT_PAGE_COL_VAR:
/*
* Column-store leaf page entries map one-to-one to the number
* of physical entries on the page (each physical entry is a
* value item).
*
* Column-store internal page entries map one-to-one to the
* number of physical entries on the page (each entry is a
* location cookie).
*/
alloc_entries = dsk->u.entries;
break;
case WT_PAGE_ROW_INT:
/*
* Row-store internal page entries map one-to-two to the number
* of physical entries on the page (each entry is a key and
* location cookie pair).
*/
alloc_entries = dsk->u.entries / 2;
break;
case WT_PAGE_ROW_LEAF:
/*
* If the "no empty values" flag is set, row-store leaf page
* entries map one-to-one to the number of physical entries
* on the page (each physical entry is a key or value item).
* If that flag is not set, there are more keys than values,
* we have to walk the page to figure it out.
*/
if (F_ISSET(dsk, WT_PAGE_EMPTY_V_ALL))
alloc_entries = dsk->u.entries;
else if (F_ISSET(dsk, WT_PAGE_EMPTY_V_NONE))
alloc_entries = dsk->u.entries / 2;
else
WT_RET(__inmem_row_leaf_entries(
session, dsk, &alloc_entries));
break;
WT_ILLEGAL_VALUE(session);
}
/* Allocate and initialize a new WT_PAGE. */
WT_RET(__wt_page_alloc(
session, dsk->type, dsk->recno, alloc_entries, 1, &page));
page->dsk = dsk;
F_SET_ATOMIC(page, flags);
/*
* Track the memory allocated to build this page so we can update the
* cache statistics in a single call.
*/
size = LF_ISSET(WT_PAGE_DISK_ALLOC) ? dsk->mem_size : 0;
switch (page->type) {
case WT_PAGE_COL_FIX:
__inmem_col_fix(session, page);
break;
case WT_PAGE_COL_INT:
__inmem_col_int(session, page);
break;
case WT_PAGE_COL_VAR:
WT_ERR(__inmem_col_var(session, page, &size));
break;
case WT_PAGE_ROW_INT:
WT_ERR(__inmem_row_int(session, page, &size));
break;
case WT_PAGE_ROW_LEAF:
WT_ERR(__inmem_row_leaf(session, page));
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Update the page's in-memory size and the cache statistics. */
__wt_cache_page_inmem_incr(session, page, size);
/* Link the new internal page to the parent. */
if (ref != NULL) {
switch (page->type) {
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
page->pg_intl_parent_ref = ref;
break;
}
ref->page = page;
}
*pagep = page;
return (0);
err: __wt_page_out(session, &page);
return (ret);
}
/*
* __wt_page_alloc --
* Create or read a page into the cache.
*/
int
__wt_page_alloc(WT_SESSION_IMPL *session, uint8_t type,
uint64_t recno, uint32_t alloc_entries, int alloc_refs, WT_PAGE **pagep)
{
WT_CACHE *cache;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_INDEX *pindex;
size_t size;
uint32_t i;
void *p;
*pagep = NULL;
cache = S2C(session)->cache;
page = NULL;
size = sizeof(WT_PAGE);
switch (type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
break;
case WT_PAGE_COL_VAR:
/*
* Variable-length column-store leaf page: allocate memory to
* describe the page's contents with the initial allocation.
*/
size += alloc_entries * sizeof(WT_COL);
break;
case WT_PAGE_ROW_LEAF:
/*
* Row-store leaf page: allocate memory to describe the page's
* contents with the initial allocation.
*/
size += alloc_entries * sizeof(WT_ROW);
break;
WT_ILLEGAL_VALUE(session);
}
WT_RET(__wt_calloc(session, 1, size, &page));
page->type = type;
page->read_gen = WT_READGEN_NOTSET;
switch (type) {
case WT_PAGE_COL_FIX:
page->pg_fix_recno = recno;
page->pg_fix_entries = alloc_entries;
break;
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
page->pg_intl_recno = recno;
/*
* Internal pages have an array of references to objects so they
* can split. Allocate the array of references and optionally,
* the objects to which they point.
*/
WT_ERR(__wt_calloc(session, 1,
sizeof(WT_PAGE_INDEX) + alloc_entries * sizeof(WT_REF *),
&p));
size +=
sizeof(WT_PAGE_INDEX) + alloc_entries * sizeof(WT_REF *);
pindex = p;
pindex->index = (WT_REF **)((WT_PAGE_INDEX *)p + 1);
pindex->entries = alloc_entries;
WT_INTL_INDEX_SET(page, pindex);
if (alloc_refs)
for (i = 0; i < pindex->entries; ++i) {
WT_ERR(__wt_calloc_def(
session, 1, &pindex->index[i]));
size += sizeof(WT_REF);
}
if (0) {
err: if ((pindex = WT_INTL_INDEX_COPY(page)) != NULL) {
for (i = 0; i < pindex->entries; ++i)
__wt_free(session, pindex->index[i]);
__wt_free(session, pindex);
}
__wt_free(session, page);
return (ret);
}
break;
case WT_PAGE_COL_VAR:
page->pg_var_recno = recno;
page->pg_var_d = (WT_COL *)((uint8_t *)page + sizeof(WT_PAGE));
page->pg_var_entries = alloc_entries;
break;
case WT_PAGE_ROW_LEAF:
page->pg_row_d = (WT_ROW *)((uint8_t *)page + sizeof(WT_PAGE));
page->pg_row_entries = alloc_entries;
break;
WT_ILLEGAL_VALUE(session);
}
/* Increment the cache statistics. */
__wt_cache_page_inmem_incr(session, page, size);
(void)WT_ATOMIC_ADD8(cache->pages_inmem, 1);
*pagep = page;
return (0);
}
/*
* __wt_page_inmem --
* Build in-memory page information.
*/
int
__wt_page_inmem(
WT_SESSION_IMPL *session, WT_PAGE *parent, WT_REF *parent_ref,
WT_PAGE_HEADER *dsk, int disk_not_alloc, WT_PAGE **pagep)
{
WT_DECL_RET;
WT_PAGE *page;
uint32_t alloc_entries;
size_t size;
alloc_entries = 0;
*pagep = NULL;
/*
* Figure out how many underlying objects the page references so
* we can allocate them along with the page.
*/
switch (dsk->type) {
case WT_PAGE_COL_FIX:
break;
case WT_PAGE_COL_INT:
/*
* Column-store internal page entries map one-to-one to the
* number of physical entries on the page (each physical entry
* is an offset object).
*/
alloc_entries = dsk->u.entries;
break;
case WT_PAGE_COL_VAR:
/*
* Column-store leaf page entries map one-to-one to the number
* of physical entries on the page (each physical entry is a
* data item).
*/
alloc_entries = dsk->u.entries;
break;
case WT_PAGE_ROW_INT:
/*
* Row-store internal page entries map one-to-two to the number
* of physical entries on the page (each in-memory entry is a
* key item and location cookie).
*/
alloc_entries = dsk->u.entries / 2;
break;
case WT_PAGE_ROW_LEAF:
/*
* Row-store leaf page entries map in an indeterminate way to
* the physical entries on the page, we have to walk the page
* to figure it out.
*/
WT_RET(__inmem_row_leaf_entries(session, dsk, &alloc_entries));
break;
WT_ILLEGAL_VALUE(session);
}
/* Allocate and initialize a new WT_PAGE. */
WT_RET(__wt_page_alloc(session, dsk->type, alloc_entries, &page));
page->dsk = dsk;
page->read_gen = WT_READ_GEN_NOTSET;
if (disk_not_alloc)
F_SET_ATOMIC(page, WT_PAGE_DISK_NOT_ALLOC);
/*
* Track the memory allocated to build this page so we can update the
* cache statistics in a single call.
*/
size = disk_not_alloc ? 0 : dsk->mem_size;
switch (page->type) {
case WT_PAGE_COL_FIX:
page->entries = dsk->u.entries;
page->u.col_fix.recno = dsk->recno;
__inmem_col_fix(session, page);
break;
case WT_PAGE_COL_INT:
page->entries = dsk->u.entries;
page->u.intl.recno = dsk->recno;
__inmem_col_int(session, page);
break;
case WT_PAGE_COL_VAR:
page->entries = dsk->u.entries;
page->u.col_var.recno = dsk->recno;
WT_ERR(__inmem_col_var(session, page, &size));
break;
case WT_PAGE_ROW_INT:
page->entries = dsk->u.entries / 2;
WT_ERR(__inmem_row_int(session, page, &size));
break;
case WT_PAGE_ROW_LEAF:
page->entries = alloc_entries;
WT_ERR(__inmem_row_leaf(session, page));
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Update the page's in-memory size and the cache statistics. */
__wt_cache_page_inmem_incr(session, page, size);
/* Link the new page into the parent. */
if (parent_ref != NULL)
WT_LINK_PAGE(parent, parent_ref, page);
*pagep = page;
return (0);
err: __wt_page_out(session, &page);
return (ret);
}
/*
* __wt_page_alloc --
* Create or read a page into the cache.
*/
int
__wt_page_alloc(WT_SESSION_IMPL *session,
uint8_t type, uint32_t alloc_entries, WT_PAGE **pagep)
{
WT_CACHE *cache;
WT_PAGE *page;
size_t size;
void *p;
*pagep = NULL;
cache = S2C(session)->cache;
/*
* Allocate a page, and for most page types, the additional information
* it needs to describe the disk image.
*/
size = sizeof(WT_PAGE);
switch (type) {
case WT_PAGE_COL_FIX:
break;
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
size += alloc_entries * sizeof(WT_REF);
break;
case WT_PAGE_COL_VAR:
size += alloc_entries * sizeof(WT_COL);
break;
case WT_PAGE_ROW_LEAF:
size += alloc_entries * sizeof(WT_ROW);
break;
WT_ILLEGAL_VALUE(session);
}
WT_RET(__wt_calloc(session, 1, size, &page));
p = (uint8_t *)page + sizeof(WT_PAGE);
switch (type) {
case WT_PAGE_COL_FIX:
break;
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
page->u.intl.t = p;
break;
case WT_PAGE_COL_VAR:
page->u.col_var.d = p;
break;
case WT_PAGE_ROW_LEAF:
page->u.row.d = p;
break;
WT_ILLEGAL_VALUE(session);
}
/* Increment the cache statistics. */
__wt_cache_page_inmem_incr(session, page, size);
(void)WT_ATOMIC_ADD(cache->pages_inmem, 1);
/* The one page field we set is the type. */
page->type = type;
*pagep = page;
return (0);
}
/*
* __ovfl_reuse_wrapup --
* Resolve the page's overflow reuse list after a page is written.
*/
static int
__ovfl_reuse_wrapup(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_BM *bm;
WT_OVFL_REUSE **e, **head, *reuse;
size_t incr, decr;
int i;
bm = S2BT(session)->bm;
head = page->modify->ovfl_track->ovfl_reuse;
/*
* Discard any overflow records that aren't in-use, freeing underlying
* blocks.
*
* First, walk the overflow reuse lists (except for the lowest one),
* fixing up skiplist links.
*/
for (i = WT_SKIP_MAXDEPTH - 1; i > 0; --i)
for (e = &head[i]; *e != NULL;) {
if (F_ISSET(*e, WT_OVFL_REUSE_INUSE)) {
e = &(*e)->next[i];
continue;
}
*e = (*e)->next[i];
}
/*
* Second, discard any overflow record without an in-use flag, clear
* the flags for the next run.
*
* As part of the pass through the lowest level, figure out how much
* space we added/subtracted from the page, and update its footprint.
* We don't get it exactly correct because we don't know the depth of
* the skiplist here, but it's close enough, and figuring out the
* memory footprint change in the reconciliation wrapup code means
* fewer atomic updates and less code overall.
*/
incr = decr = 0;
for (e = &head[0]; (reuse = *e) != NULL;) {
if (F_ISSET(reuse, WT_OVFL_REUSE_INUSE)) {
if (F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED))
incr += WT_OVFL_SIZE(WT_OVFL_REUSE) +
reuse->addr_size + reuse->value_size;
F_CLR(reuse,
WT_OVFL_REUSE_INUSE | WT_OVFL_REUSE_JUST_ADDED);
e = &(*e)->next[0];
continue;
}
*e = (*e)->next[0];
WT_ASSERT(session, !F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED));
decr += WT_OVFL_SIZE(WT_OVFL_REUSE) +
reuse->addr_size + reuse->value_size;
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_reuse_verbose(session, page, reuse, "free"));
WT_RET(bm->free(
bm, session, WT_OVFL_REUSE_ADDR(reuse), reuse->addr_size));
__wt_free(session, reuse);
}
if (incr > decr)
__wt_cache_page_inmem_incr(session, page, incr - decr);
if (decr > incr)
__wt_cache_page_inmem_decr(session, page, decr - incr);
return (0);
}
/*
* __wt_rec_track --
* Add an object to the page's list of tracked objects.
*/
int
__wt_rec_track(WT_SESSION_IMPL *session, WT_PAGE *page,
const uint8_t *addr, uint32_t addr_size,
const void *data, uint32_t data_size, uint32_t flags)
{
WT_PAGE_MODIFY *mod;
WT_PAGE_TRACK *empty, *track;
uint8_t *p;
uint32_t i;
mod = page->modify;
/* Find an empty slot. */
empty = NULL;
for (track = mod->track, i = 0; i < mod->track_entries; ++track, ++i)
if (!F_ISSET(track, WT_TRK_OBJECT)) {
empty = track;
break;
}
/* Reallocate space as necessary. */
if (empty == NULL) {
WT_RET(__rec_track_extend(session, page));
empty = &mod->track[mod->track_entries - 1];
}
track = empty;
/*
* Minor optimization: allocate a single chunk of space instead of two
* separate ones: be careful when it's freed.
*/
WT_RET(__wt_calloc_def(session, addr_size + data_size, &p));
/*
* Set the just-added flag so we clean up should reconciliation fail,
* except for cached overflow values, which don't get discarded, even
* if reconciliation fails.
*/
track->flags = (uint8_t)flags | WT_TRK_OBJECT;
if (!F_ISSET(track, WT_TRK_OVFL_VALUE))
F_SET(track, WT_TRK_JUST_ADDED);
track->addr.addr = p;
track->addr.size = addr_size;
memcpy(track->addr.addr, addr, addr_size);
if (data_size) {
p += addr_size;
track->data = p;
track->size = data_size;
memcpy(track->data, data, data_size);
}
/*
* Overflow items are potentially large and on-page items remain in the
* tracking list until the page is evicted. If we're tracking a lot of
* them, their memory might matter: increment the page and cache memory
* totals. This is unlikely to matter, but it's inexpensive (unless
* there are lots of them, in with case I guess the memory matters).
*
* If this reconciliation were to fail, we would reasonably perform the
* inverse operation in __wt_rec_track_wrapup_err. I'm not bothering
* with that because we'd have to crack the structure itself to figure
* out how much to decrement and I don't think it's worth the effort.
* The potential problem is repeatedly failing reconciliation of a page
* with a large number of overflow items, which causes the page's memory
* memory footprint to become incorrectly high, causing us to push the
* page out of cache unnecessarily. Like I said, not worth the effort.
*/
if (LF_ISSET(WT_TRK_ONPAGE))
__wt_cache_page_inmem_incr(
session, page, addr_size + data_size);
if (WT_VERBOSE_ISSET(session, reconcile))
WT_RET(__track_msg(session, page, "add", track));
return (0);
}
