/*
* __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 *txn_state;
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_def(session, cfg, "isolation", 0, &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;
/*
* Allocate a transaction ID.
*
* We use an atomic compare and swap to ensure that we get a
* unique ID that is published before the global counter is
* updated.
*
* 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: that is the price we pay to keep this
* path latch free.
*/
do {
txn_state->id = txn->id = txn_global->current;
} while (!WT_ATOMIC_CAS(txn_global->current, txn->id, txn->id + 1));
/*
* If we have used 64-bits of transaction IDs, there is nothing
* more we can do.
*/
if (txn->id == WT_TXN_ABORTED)
WT_RET_MSG(session, ENOMEM, "Out of transaction IDs");
F_SET(txn, TXN_RUNNING);
if (txn->isolation == TXN_ISO_SNAPSHOT)
__wt_txn_refresh(session, WT_TXN_NONE, 1);
return (0);
}
/*
* __wt_txn_commit --
* Commit the current transaction.
*/
int
__wt_txn_commit(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_DECL_RET;
WT_TXN *txn;
WT_TXN_OP *op;
u_int i;
WT_UNUSED(cfg);
txn = &session->txn;
WT_ASSERT(session, !F_ISSET(txn, TXN_ERROR));
if (!F_ISSET(txn, TXN_RUNNING))
WT_RET_MSG(session, EINVAL, "No transaction is active");
/* Commit notification. */
if (txn->notify != NULL)
WT_TRET(txn->notify->notify(txn->notify,
(WT_SESSION *)session, txn->id, 1));
/* If we are logging, write a commit log record. */
if (ret == 0 &&
txn->mod_count > 0 && S2C(session)->logging &&
!F_ISSET(session, WT_SESSION_LOGGING_DISABLED))
ret = __wt_txn_log_commit(session, cfg);
/*
* If anything went wrong, roll back.
*
* !!!
* Nothing can fail after this point.
*/
if (ret != 0) {
WT_TRET(__wt_txn_rollback(session, cfg));
return (ret);
}
/* Free memory associated with updates. */
for (i = 0, op = txn->mod; i < txn->mod_count; i++, op++)
__wt_txn_op_free(session, op);
/*
* Auto-commit transactions need a new transaction snapshot so that the
* committed changes are visible to subsequent reads. However, cursor
* keys and values will point to the data that was just modified, so
* the snapshot cannot be so new that updates could be freed underneath
* the cursor. Get the new snapshot before releasing the ID for the
* commit.
*/
if (session->ncursors > 0 && txn->isolation != TXN_ISO_READ_UNCOMMITTED)
__wt_txn_refresh(session, txn->id + 1, 1);
__wt_txn_release(session);
return (0);
}
/*
* __wt_txn_update_oldest --
* Sweep the running transactions to update the oldest ID required.
*/
void
__wt_txn_update_oldest(WT_SESSION_IMPL *session)
{
/*
* !!!
* If a data-source is calling the WT_EXTENSION_API.transaction_oldest
* method (for the oldest transaction ID not yet visible to a running
* transaction), and then comparing that oldest ID against committed
* transactions to see if updates for a committed transaction are still
* visible to running transactions, the oldest transaction ID may be
* the same as the last committed transaction ID, if the transaction
* state wasn't refreshed after the last transaction committed. Push
* past the last committed transaction.
*/
__wt_txn_refresh(session, WT_TXN_NONE, 0);
}
/*
* __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;
int evict_reset;
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, NULL, 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 == 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;
WT_FULL_BARRIER();
WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
if (evict_reset)
__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, NULL, 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;
}
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();
/*
* 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, 1));
return (ret);
}