Merge pull request #2331 from wiredtiger/WT-2237

WT-2237 Have threads publish unique transaction IDs so that updates always become visible immediately on commit.

(cherry picked from commit 0a52a80a39fc47145fb755d792792ae820b266ed)

6 files changed, 68 insertions, 65 deletions

diff --git a/src/btree/bt_split.c b/src/btree/bt_split.c
index 6f31ff89aa7..9d8e463feb0 100644
--- a/src/btree/bt_split.c
+++ b/src/btree/bt_split.c
@@ -502,7 +502,7 @@ __split_deepen(WT_SESSION_IMPL *session, WT_PAGE *parent)
 		 * array, a thread might see a freed WT_REF.  Set the eviction
 		 * transaction requirement for the newly created internal pages.
 		 */
-		child->modify->mod_split_txn = __wt_txn_new_id(session);
+		child->modify->mod_split_txn = __wt_txn_id_alloc(session, false);
 
 		/*
 		 * The newly allocated child's page index references the same
@@ -1383,7 +1383,7 @@ __split_insert(WT_SESSION_IMPL *session, WT_REF *ref)
 	 * threads will not try to forcibly evict the page again until
 	 * all concurrent transactions commit.
 	 */
-	page->modify->inmem_split_txn = __wt_txn_new_id(session);
+	page->modify->inmem_split_txn = __wt_txn_id_alloc(session, false);
 
 	/*
 	 * Update the page accounting.
diff --git a/src/include/txn.h b/src/include/txn.h
index 0d5f337fff0..634363b9216 100644
--- a/src/include/txn.h
+++ b/src/include/txn.h
@@ -34,6 +34,7 @@ struct WT_COMPILER_TYPE_ALIGN(WT_CACHE_LINE_ALIGNMENT) __wt_txn_state {
 };
 
 struct __wt_txn_global {
+	uint64_t alloc;			/* Transaction ID to allocate. */
 	volatile uint64_t current;	/* Current transaction ID. */
 
 	/* The oldest running transaction ID (may race). */
diff --git a/src/include/txn.i b/src/include/txn.i
index 73d7f1f0518..3701d81522a 100644
--- a/src/include/txn.i
+++ b/src/include/txn.i
@@ -280,23 +280,6 @@ __wt_txn_autocommit_check(WT_SESSION_IMPL *session)
 }
 
 /*
- * __wt_txn_new_id --
- *	Allocate a new transaction ID.
- */
-static inline uint64_t
-__wt_txn_new_id(WT_SESSION_IMPL *session)
-{
-	/*
-	 * We want the global value to lead the allocated values, so that any
-	 * allocated transaction ID eventually becomes globally visible.  When
-	 * there are no transactions running, the oldest_id will reach the
-	 * global current ID, so we want post-increment semantics.  Our atomic
-	 * add primitive does pre-increment, so adjust the result here.
-	 */
-	return (__wt_atomic_addv64(&S2C(session)->txn_global.current, 1) - 1);
-}
-
-/*
  * __wt_txn_idle_cache_check --
  *	If there is no transaction active in this thread and we haven't checked
  *	if the cache is full, do it now.  If we have to block for eviction,
@@ -323,6 +306,55 @@ __wt_txn_idle_cache_check(WT_SESSION_IMPL *session)
 }
 
 /*
+ * __wt_txn_id_alloc --
+ *	Allocate a new transaction ID.
+ */
+static inline uint64_t
+__wt_txn_id_alloc(WT_SESSION_IMPL *session, bool publish)
+{
+	WT_TXN_GLOBAL *txn_global;
+	uint64_t id;
+
+	txn_global = &S2C(session)->txn_global;
+
+	/*
+	 * Allocating transaction IDs involves several steps.
+	 *
+	 * Firstly, we do an atomic increment to allocate a unique ID.  The
+	 * field we increment is not used anywhere else.
+	 *
+	 * Then we optionally publish the allocated ID into the global
+	 * transaction table.  It is critical that this becomes visible before
+	 * the global current value moves past our ID, or some concurrent
+	 * reader could get a snapshot that makes our changes visible before we
+	 * commit.
+	 *
+	 * Lastly, we spin to update the current ID.  This is the only place
+	 * that the current ID is updated, and it is in the same cache line as
+	 * the field we allocate from, so we should usually succeed on the
+	 * first try.
+	 *
+	 * We want the global value to lead the allocated values, so that any
+	 * allocated transaction ID eventually becomes globally visible.  When
+	 * there are no transactions running, the oldest_id will reach the
+	 * global current ID, so we want post-increment semantics.  Our atomic
+	 * add primitive does pre-increment, so adjust the result here.
+	 */
+	id = __wt_atomic_addv64(&S2C(session)->txn_global.alloc, 1) - 1;
+
+	if (publish) {
+		session->txn.id = id;
+		WT_PUBLISH(WT_SESSION_TXN_STATE(session)->id, id);
+	}
+
+	while (txn_global->current != id ||
+	    !__wt_atomic_casv64(&txn_global->current, id, id + 1))
+		__wt_yield();
+
+	 return (id);
+}
+
+/*
  * __wt_txn_id_check --
  *	A transaction is going to do an update, start an auto commit
  *	transaction if required and allocate a transaction ID.
@@ -330,57 +362,27 @@ __wt_txn_idle_cache_check(WT_SESSION_IMPL *session)
 static inline int
 __wt_txn_id_check(WT_SESSION_IMPL *session)
 {
-	WT_CONNECTION_IMPL *conn;
 	WT_TXN *txn;
-	WT_TXN_GLOBAL *txn_global;
-	WT_TXN_STATE *txn_state;
 
 	txn = &session->txn;
 
 	WT_ASSERT(session, F_ISSET(txn, WT_TXN_RUNNING));
 
+	if (F_ISSET(txn, WT_TXN_HAS_ID))
+		return (0);
+
 	/* If the transaction is idle, check that the cache isn't full. */
 	WT_RET(__wt_txn_idle_cache_check(session));
 
-	if (!F_ISSET(txn, WT_TXN_HAS_ID)) {
-		conn = S2C(session);
-		txn_global = &conn->txn_global;
-		txn_state = WT_SESSION_TXN_STATE(session);
+	(void)__wt_txn_id_alloc(session, true);
 
-		WT_ASSERT(session, txn_state->id == WT_TXN_NONE);
-
-		/*
-		 * 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_casv64(
-		    &txn_global->current, txn->id, txn->id + 1) ||
-		    WT_TXNID_LT(txn->id, txn_global->last_running));
-
-		/*
-		 * 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, WT_TXN_HAS_ID);
-	}
+	/*
+	 * 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, WT_TXN_HAS_ID);
 
 	return (0);
 }
diff --git a/src/lsm/lsm_tree.c b/src/lsm/lsm_tree.c
index 0680d98bdde..d3979da0da1 100644
--- a/src/lsm/lsm_tree.c
+++ b/src/lsm/lsm_tree.c
@@ -858,7 +858,7 @@ __wt_lsm_tree_switch(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
 	 */
 	if (last_chunk != NULL && last_chunk->switch_txn == WT_TXN_NONE &&
 	    !F_ISSET(last_chunk, WT_LSM_CHUNK_ONDISK))
-		last_chunk->switch_txn = __wt_txn_new_id(session);
+		last_chunk->switch_txn = __wt_txn_id_alloc(session, false);
 
 	/*
 	 * If a maximum number of chunks are configured, drop the any chunks
@@ -1257,7 +1257,7 @@ __wt_lsm_compact(WT_SESSION_IMPL *session, const char *name, bool *skipp)
 	if (lsm_tree->nchunks > 0 &&
 	    (chunk = lsm_tree->chunk[lsm_tree->nchunks - 1]) != NULL) {
 		if (chunk->switch_txn == WT_TXN_NONE)
-			chunk->switch_txn = __wt_txn_new_id(session);
+			chunk->switch_txn = __wt_txn_id_alloc(session, false);
 		/*
 		 * If we have a chunk, we want to look for it to be on-disk.
 		 * So we need to add a reference to keep it available.
diff --git a/src/reconcile/rec_track.c b/src/reconcile/rec_track.c
index 36e85713421..17ad1c5fdc4 100644
--- a/src/reconcile/rec_track.c
+++ b/src/reconcile/rec_track.c
@@ -820,7 +820,7 @@ __wt_ovfl_txnc_add(WT_SESSION_IMPL *session, WT_PAGE *page,
 	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);
+	txnc->current = __wt_txn_id_alloc(session, false);
 
 	__wt_cache_page_inmem_incr(
 	    session, page, WT_OVFL_SIZE(txnc, WT_OVFL_TXNC));
diff --git a/src/txn/txn.c b/src/txn/txn.c
index eb2b621f315..f89a8ae1b53 100644
--- a/src/txn/txn.c
+++ b/src/txn/txn.c
@@ -594,8 +594,8 @@ __wt_txn_global_init(WT_SESSION_IMPL *session, const char *cfg[])
 	conn = S2C(session);
 
 	txn_global = &conn->txn_global;
-	txn_global->current = txn_global->last_running =
-	    txn_global->oldest_id = WT_TXN_FIRST;
+	txn_global->alloc = txn_global->current =
+	    txn_global->last_running = txn_global->oldest_id = WT_TXN_FIRST;
 
 	WT_RET(__wt_calloc_def(
 	    session, conn->session_size, &txn_global->states));