/*
* __wt_txn_get_snapshot --
* Set up a snapshot in the current transaction, without allocating an ID.
*/
void
__wt_txn_get_snapshot(
WT_SESSION_IMPL *session, wt_txnid_t my_id, wt_txnid_t max_id, int force)
{
WT_CONNECTION_IMPL *conn;
WT_TXN *txn;
WT_TXN_GLOBAL *txn_global;
WT_TXN_STATE *s, *txn_state;
wt_txnid_t current_id, id, oldest_snap_min;
uint32_t i, n, session_cnt;
conn = S2C(session);
txn = &session->txn;
txn_global = &conn->txn_global;
txn_state = &txn_global->states[session->id];
/* If nothing has changed since last time, we're done. */
if (!force && txn->last_id == txn_global->current &&
txn->last_gen == txn_global->gen) {
txn_state->snap_min = txn->snap_min;
return;
}
do {
/* Take a copy of the current session ID. */
txn->last_gen = txn->last_oldest_gen = txn_global->gen;
txn->last_id = oldest_snap_min = current_id =
txn_global->current;
/* Copy the array of concurrent transactions. */
WT_ORDERED_READ(session_cnt, conn->session_cnt);
for (i = n = 0, s = txn_global->states;
i < session_cnt;
i++, s++) {
/* Ignore the session's own transaction. */
if (i == session->id)
continue;
if ((id = s->snap_min) != WT_TXN_NONE)
if (TXNID_LT(id, oldest_snap_min))
oldest_snap_min = id;
if ((id = s->id) == WT_TXN_NONE)
continue;
else if (max_id == WT_TXN_NONE || TXNID_LT(id, max_id))
txn->snapshot[n++] = id;
}
/*
* Ensure the snapshot reads are scheduled before re-checking
* the global current ID.
*/
WT_READ_BARRIER();
} while (current_id != txn_global->current);
__txn_sort_snapshot(session, n,
(max_id != WT_TXN_NONE) ? max_id : current_id,
oldest_snap_min);
txn_state->snap_min =
(my_id == WT_TXN_NONE || TXNID_LT(txn->snap_min, my_id)) ?
txn->snap_min : my_id;
}
/*
* __wt_txn_get_oldest --
* Update the current transaction's cached copy of the oldest snap_min
* value.
*/
void
__wt_txn_get_oldest(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_TXN *txn;
WT_TXN_GLOBAL *txn_global;
WT_TXN_STATE *s;
wt_txnid_t id, oldest_snap_min;
uint32_t i, session_cnt;
conn = S2C(session);
txn = &session->txn;
txn_global = &conn->txn_global;
/* If nothing has changed since last time, we're done. */
if (txn->last_oldest_gen == txn_global->gen)
return;
txn->last_oldest_gen = txn_global->gen;
oldest_snap_min =
(txn->id != WT_TXN_NONE) ? txn->id : txn_global->current;
WT_ORDERED_READ(session_cnt, conn->session_cnt);
for (i = 0, s = txn_global->states;
i < session_cnt;
i++, s++) {
if ((id = s->snap_min) != WT_TXN_NONE &&
TXNID_LT(id, oldest_snap_min))
oldest_snap_min = id;
}
txn->oldest_snap_min = oldest_snap_min;
}
/*
* __wt_txnid_cmp --
* Compare transaction IDs for sorting / searching.
*/
int
__wt_txnid_cmp(const void *v1, const void *v2)
{
uint64_t id1, id2;
id1 = *(uint64_t *)v1;
id2 = *(uint64_t *)v2;
return ((id1 == id2) ? 0 : TXNID_LT(id1, id2) ? -1 : 1);
}
/*
* __txn_sort_snapshot --
* Sort a snapshot for faster searching and set the min/max bounds.
*/
static void
__txn_sort_snapshot(WT_SESSION_IMPL *session,
uint32_t n, wt_txnid_t id, wt_txnid_t oldest_snap_min)
{
WT_TXN *txn;
txn = &session->txn;
if (n > 1)
qsort(txn->snapshot, n, sizeof(wt_txnid_t), __wt_txnid_cmp);
txn->snapshot_count = n;
txn->snap_min = (n == 0) ? id : txn->snapshot[0];
txn->snap_max = id;
WT_ASSERT(session, txn->snap_min != WT_TXN_NONE);
txn->oldest_snap_min = TXNID_LT(oldest_snap_min, txn->snap_min) ?
oldest_snap_min : txn->snap_min;
}
/*
* __wt_txn_refresh --
* Allocate a transaction ID and/or a snapshot.
*/
void
__wt_txn_refresh(WT_SESSION_IMPL *session, uint64_t max_id, int get_snapshot)
{
WT_CONNECTION_IMPL *conn;
WT_TXN *txn;
WT_TXN_GLOBAL *txn_global;
WT_TXN_STATE *s, *txn_state;
uint64_t current_id, id, snap_min, oldest_id, prev_oldest_id;
uint32_t i, n, session_cnt;
int32_t count;
conn = S2C(session);
txn = &session->txn;
txn_global = &conn->txn_global;
txn_state = &txn_global->states[session->id];
prev_oldest_id = txn_global->oldest_id;
current_id = snap_min = txn_global->current;
/* For pure read-only workloads, avoid updates to shared state. */
if (!get_snapshot) {
/*
* If we are trying to update the oldest ID and it is already
* equal to the current ID, there is no point scanning.
*/
if (prev_oldest_id == current_id)
return;
} else if (txn->id == max_id &&
txn->snapshot_count == 0 &&
txn->snap_min == snap_min &&
TXNID_LE(prev_oldest_id, snap_min)) {
txn_state->snap_min = txn->snap_min;
/* If nothing has changed in the meantime, we're done. */
if (txn_global->scan_count == 0 &&
txn_global->oldest_id == prev_oldest_id)
return;
}
/*
* We're going to scan. Increment the count of scanners to prevent the
* oldest ID from moving forwards. Spin if the count is negative,
* which indicates that some thread is moving the oldest ID forwards.
*/
do {
if ((count = txn_global->scan_count) < 0)
WT_PAUSE();
} while (count < 0 ||
!WT_ATOMIC_CAS(txn_global->scan_count, count, count + 1));
/* The oldest ID cannot change until the scan count goes to zero. */
prev_oldest_id = txn_global->oldest_id;
current_id = snap_min = txn_global->current;
/* If the maximum ID is constrained, so is the oldest. */
oldest_id = (max_id != WT_TXN_NONE) ? max_id : snap_min;
/* Walk the array of concurrent transactions. */
WT_ORDERED_READ(session_cnt, conn->session_cnt);
for (i = n = 0, s = txn_global->states; i < session_cnt; i++, s++) {
/*
* Ignore the ID if we are committing (indicated by max_id
* being set): it is about to be released.
*
* Also ignore the ID if it is older than the oldest ID we saw.
* This can happen if we race with a thread that is allocating
* an ID -- the ID will not be used because the thread will
* keep spinning until it gets a valid one.
*/
if ((id = s->id) != WT_TXN_NONE && id + 1 != max_id &&
TXNID_LE(prev_oldest_id, id)) {
if (get_snapshot)
txn->snapshot[n++] = id;
if (TXNID_LT(id, snap_min))
snap_min = id;
}
/*
* Ignore the session's own snap_min if we are in the process
* of updating it.
*/
if (get_snapshot && s == txn_state)
continue;
/*
* !!!
* Note: Don't ignore snap_min values older than the previous
* oldest ID. Read-uncommitted operations publish snap_min
* values without incrementing scan_count to protect the global
* table. See the comment in __wt_txn_cursor_op for
* more details.
*/
if ((id = s->snap_min) != WT_TXN_NONE &&
TXNID_LT(id, oldest_id))
oldest_id = id;
}
if (TXNID_LT(snap_min, oldest_id))
oldest_id = snap_min;
if (get_snapshot) {
WT_ASSERT(session, TXNID_LE(prev_oldest_id, snap_min));
WT_ASSERT(session, prev_oldest_id == txn_global->oldest_id);
txn_state->snap_min = snap_min;
}
/*
* Update the last running ID if we have a much newer value or we are
* forcing an update.
*/
if (!get_snapshot || snap_min > txn_global->last_running + 100)
txn_global->last_running = snap_min;
/*
* Update the oldest ID if we have a newer ID and we can get exclusive
* access. During normal snapshot refresh, only do this if we have a
* much newer value. Once we get exclusive access, do another pass to
* make sure nobody else is using an earlier ID.
*/
if (max_id == WT_TXN_NONE &&
TXNID_LT(prev_oldest_id, oldest_id) &&
(!get_snapshot || oldest_id - prev_oldest_id > 100) &&
WT_ATOMIC_CAS(txn_global->scan_count, 1, -1)) {
WT_ORDERED_READ(session_cnt, conn->session_cnt);
for (i = 0, s = txn_global->states; i < session_cnt; i++, s++) {
if ((id = s->id) != WT_TXN_NONE &&
TXNID_LT(id, oldest_id))
oldest_id = id;
if ((id = s->snap_min) != WT_TXN_NONE &&
TXNID_LT(id, oldest_id))
oldest_id = id;
}
if (TXNID_LT(txn_global->oldest_id, oldest_id))
txn_global->oldest_id = oldest_id;
txn_global->scan_count = 0;
} else {
WT_ASSERT(session, txn_global->scan_count > 0);
(void)WT_ATOMIC_SUB(txn_global->scan_count, 1);
}
if (get_snapshot)
__txn_sort_snapshot(session, n, current_id);
}
/*
* __clsm_enter --
* Start an operation on an LSM cursor, update if the tree has changed.
*/
static inline int
__clsm_enter(WT_CURSOR_LSM *clsm, int reset, int update)
{
WT_CURSOR *c;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_SESSION_IMPL *session;
uint64_t *txnid_maxp;
uint64_t id, myid, snap_min;
session = (WT_SESSION_IMPL *)clsm->iface.session;
/* Merge cursors never update. */
if (F_ISSET(clsm, WT_CLSM_MERGE))
return (0);
if (reset) {
c = &clsm->iface;
/* Copy out data before resetting chunk cursors. */
if (F_ISSET(c, WT_CURSTD_KEY_INT) &&
!WT_DATA_IN_ITEM(&c->key))
WT_RET(__wt_buf_set(
session, &c->key, c->key.data, c->key.size));
if (F_ISSET(c, WT_CURSTD_VALUE_INT) &&
!WT_DATA_IN_ITEM(&c->value))
WT_RET(__wt_buf_set(
session, &c->value, c->value.data, c->value.size));
WT_RET(__clsm_reset_cursors(clsm, NULL));
}
for (;;) {
/*
* If the cursor looks up-to-date, check if the cache is full.
* In case this call blocks, the check will be repeated before
* proceeding.
*/
if (clsm->dsk_gen != clsm->lsm_tree->dsk_gen)
goto open;
WT_RET(__wt_cache_full_check(session));
if (clsm->dsk_gen != clsm->lsm_tree->dsk_gen)
goto open;
/* Update the maximum transaction ID in the primary chunk. */
if (update && (chunk = clsm->primary_chunk) != NULL) {
WT_RET(__wt_txn_autocommit_check(session));
for (id = chunk->txnid_max, myid = session->txn.id;
!TXNID_LE(myid, id);
id = chunk->txnid_max) {
WT_ASSERT(session, myid != WT_TXN_NONE);
(void)WT_ATOMIC_CAS(
chunk->txnid_max, id, myid);
}
}
/*
* Figure out how many updates are required for snapshot
* isolation.
*
* This is not a normal visibility check on the maximum
* transaction ID in each chunk: any transaction ID that
* overlaps with our snapshot is a potential conflict.
*/
clsm->nupdates = 1;
if (session->txn.isolation == TXN_ISO_SNAPSHOT &&
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
snap_min = session->txn.snap_min;
for (txnid_maxp = &clsm->txnid_max[clsm->nchunks - 2];
clsm->nupdates < clsm->nchunks;
clsm->nupdates++, txnid_maxp--)
if (TXNID_LT(*txnid_maxp, snap_min))
break;
}
/*
* Stop when we are up-to-date, as long as this is:
* - a snapshot isolation update and the cursor is set up for
* that;
* - an update operation with a primary chunk, or
* - a read operation and the cursor is open for reading.
*/
if ((!update ||
session->txn.isolation != TXN_ISO_SNAPSHOT ||
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) &&
((update && clsm->primary_chunk != NULL) ||
(!update && F_ISSET(clsm, WT_CLSM_OPEN_READ))))
break;
open: WT_WITH_SCHEMA_LOCK(session,
ret = __clsm_open_cursors(clsm, update, 0, 0));
WT_RET(ret);
}
if (!F_ISSET(clsm, WT_CLSM_ACTIVE)) {
WT_RET(__cursor_enter(session));
F_SET(clsm, WT_CLSM_ACTIVE);
}
return (0);
}
/*
* __wt_txn_begin --
* Begin a transaction.
*/
int
__wt_txn_begin(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_TXN *txn;
WT_TXN_GLOBAL *txn_global;
WT_TXN_STATE *s, *txn_state;
wt_txnid_t id, oldest_snap_min;
uint32_t i, n, session_cnt;
conn = S2C(session);
txn = &session->txn;
txn_global = &conn->txn_global;
txn_state = &txn_global->states[session->id];
WT_ASSERT(session, txn_state->id == WT_TXN_NONE);
WT_RET(__wt_config_gets_defno(session, cfg, "isolation", &cval));
if (cval.len == 0)
txn->isolation = session->isolation;
else
txn->isolation =
WT_STRING_MATCH("snapshot", cval.str, cval.len) ?
TXN_ISO_SNAPSHOT :
WT_STRING_MATCH("read-committed", cval.str, cval.len) ?
TXN_ISO_READ_COMMITTED : TXN_ISO_READ_UNCOMMITTED;
F_SET(txn, TXN_RUNNING);
do {
/*
* Allocate a transaction ID.
*
* We use an atomic increment to ensure that we get a unique
* ID, then publish that to the global state table.
*
* If two threads race to allocate an ID, only the latest ID
* will proceed. The winning thread can be sure its snapshot
* contains all of the earlier active IDs. Threads that race
* and get an earlier ID may not appear in the snapshot,
* but they will loop and allocate a new ID before proceeding
* to make any updates.
*
* This potentially wastes transaction IDs when threads race to
* begin transactions, but that is the price we pay to keep
* this path latch free.
*/
do {
txn->id = WT_ATOMIC_ADD(txn_global->current, 1);
} while (txn->id == WT_TXN_NONE || txn->id == WT_TXN_ABORTED);
WT_PUBLISH(txn_state->id, txn->id);
/*
* If we are starting a snapshot isolation transaction, get
* a snapshot of the running transactions.
*
* If we already have a snapshot (e.g., for an auto-commit
* operation), update it so that the newly-allocated ID is
* visible.
*/
if (txn->isolation == TXN_ISO_SNAPSHOT) {
txn->last_gen = txn->last_oldest_gen = txn_global->gen;
oldest_snap_min = txn->id;
/* Copy the array of concurrent transactions. */
WT_ORDERED_READ(session_cnt, conn->session_cnt);
for (i = n = 0, s = txn_global->states;
i < session_cnt;
i++, s++) {
if ((id = s->snap_min) != WT_TXN_NONE)
if (TXNID_LT(id, oldest_snap_min))
oldest_snap_min = id;
if ((id = s->id) == WT_TXN_NONE)
continue;
else
txn->snapshot[n++] = id;
}
__txn_sort_snapshot(
session, n, txn->id, oldest_snap_min);
txn_state->snap_min = txn->snap_min;
}
/*
* Ensure the snapshot reads are complete before re-checking
* the global current ID.
*/
WT_READ_BARRIER();
} while (txn->id != txn_global->current);
return (0);
}