/*
* __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);
}
/*
* __ovfl_txnc_wrapup --
* Resolve the page's transaction-cache list.
*/
static int
__ovfl_txnc_wrapup(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_OVFL_TXNC **e, **head, *txnc;
uint64_t oldest_txn;
size_t decr;
int i;
head = page->modify->ovfl_track->ovfl_txnc;
/*
* Take a snapshot of the oldest transaction ID we need to keep alive.
* Since we do two passes through entries in the structure, the normal
* visibility check could give different results as the global ID moves
* forward.
*/
oldest_txn = __wt_txn_oldest_id(session);
/*
* Discard any transaction-cache records with transaction IDs earlier
* than any in the system.
*
* First, walk the overflow transaction-cache skip lists (except for
* the lowest level), fixing up links.
*/
for (i = WT_SKIP_MAXDEPTH - 1; i > 0; --i)
for (e = &head[i]; (txnc = *e) != NULL;) {
if (WT_TXNID_LE(oldest_txn, txnc->current)) {
e = &txnc->next[i];
continue;
}
*e = txnc->next[i];
}
/* Second, discard any no longer needed transaction-cache records. */
decr = 0;
for (e = &head[0]; (txnc = *e) != NULL;) {
if (WT_TXNID_LE(oldest_txn, txnc->current)) {
e = &txnc->next[0];
continue;
}
*e = txnc->next[0];
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_txnc_verbose(session, page, txnc, "free"));
decr += WT_OVFL_SIZE(txnc, WT_OVFL_TXNC);
__wt_free(session, txnc);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
return (0);
}
/*
* __ovfl_reuse_wrapup_err --
* Resolve the page's overflow reuse list after an error occurs.
*/
static int
__ovfl_reuse_wrapup_err(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_BM *bm;
WT_DECL_RET;
WT_OVFL_REUSE **e, **head, *reuse;
size_t decr;
int i;
bm = S2BT(session)->bm;
head = page->modify->ovfl_track->ovfl_reuse;
/*
* Discard any overflow records that were just added, 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]; (reuse = *e) != NULL;) {
if (!F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED)) {
e = &reuse->next[i];
continue;
}
*e = reuse->next[i];
}
/*
* Second, discard any overflow record with a just-added flag, clear the
* flags for the next run.
*/
decr = 0;
for (e = &head[0]; (reuse = *e) != NULL;) {
if (!F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED)) {
F_CLR(reuse, WT_OVFL_REUSE_INUSE);
e = &reuse->next[0];
continue;
}
*e = reuse->next[0];
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_reuse_verbose(session, page, reuse, "free"));
WT_TRET(bm->free(
bm, session, WT_OVFL_REUSE_ADDR(reuse), reuse->addr_size));
decr += WT_OVFL_SIZE(reuse, WT_OVFL_REUSE);
__wt_free(session, reuse);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
return (0);
}
/*
* __ovfl_txnc_wrapup --
* Resolve the page's transaction-cache list.
*/
static int
__ovfl_txnc_wrapup(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_OVFL_TXNC **e, **head, *txnc;
size_t decr;
int i;
head = page->modify->ovfl_track->ovfl_txnc;
/*
* Discard any transaction-cache records with transaction IDs earlier
* than any in the system.
*
* First, walk the overflow transaction-cache skip lists (except for
* the lowest level), fixing up links.
*/
for (i = WT_SKIP_MAXDEPTH - 1; i > 0; --i)
for (e = &head[i]; *e != NULL;) {
if (!__wt_txn_visible_all(session, (*e)->current)) {
e = &(*e)->next[i];
continue;
}
*e = (*e)->next[i];
}
/* Second, discard any no longer needed transaction-cache records. */
decr = 0;
for (e = &head[0]; (txnc = *e) != NULL;) {
if (!__wt_txn_visible_all(session, txnc->current)) {
e = &(*e)->next[0];
continue;
}
*e = (*e)->next[0];
decr += WT_OVFL_SIZE(WT_OVFL_TXNC) +
txnc->addr_size + txnc->value_size;
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_txnc_verbose(session, page, txnc, "free"));
__wt_free(session, txnc);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
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 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]; (reuse = *e) != NULL;) {
if (F_ISSET(reuse, WT_OVFL_REUSE_INUSE)) {
e = &reuse->next[i];
continue;
}
*e = reuse->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.
*/
decr = 0;
for (e = &head[0]; (reuse = *e) != NULL;) {
if (F_ISSET(reuse, WT_OVFL_REUSE_INUSE)) {
F_CLR(reuse,
WT_OVFL_REUSE_INUSE | WT_OVFL_REUSE_JUST_ADDED);
e = &reuse->next[0];
continue;
}
*e = reuse->next[0];
WT_ASSERT(session, !F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED));
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));
decr += WT_OVFL_SIZE(reuse, WT_OVFL_REUSE);
__wt_free(session, reuse);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
return (0);
}
