Import wiredtiger: bb8086c4ed645854f19554198f391ea8c821f603 from branch mongodb-4.2

ref: 67066a329d..bb8086c4ed
for: 4.1.10

WT-4614       Trigger sweep immediately to reclaim space after a table drop
WT-4631       Always clear the read timestamp for transactions
WT-4636       Fix strace in syscall test
WT-4650       Coverity complaint: Constant expression result (CONSTANT_EXPRESSION_RESULT)
WT-4668       lint

1 files changed, 55 insertions, 31 deletions

diff --git a/src/third_party/wiredtiger/src/txn/txn_timestamp.c b/src/third_party/wiredtiger/src/txn/txn_timestamp.c
index 594be12e6b1..f4e340535ef 100644
--- a/src/third_party/wiredtiger/src/txn/txn_timestamp.c
+++ b/src/third_party/wiredtiger/src/txn/txn_timestamp.c
@@ -150,6 +150,20 @@ __wt_txn_parse_timestamp(WT_SESSION_IMPL *session, const char *name,
 }
 
 /*
+ * __txn_get_read_timestamp --
+ *	Get the read timestamp from the transaction. Additionally
+ *	return bool to specify whether the transaction has set
+ *	clear read queue flag.
+ */
+static bool
+__txn_get_read_timestamp(
+    WT_TXN *txn, wt_timestamp_t *read_timestampp)
+{
+	WT_ORDERED_READ(*read_timestampp, txn->read_timestamp);
+	return (!txn->clear_read_q);
+}
+
+/*
  * __txn_get_pinned_timestamp --
  *	Calculate the current pinned timestamp.
  */
@@ -160,7 +174,7 @@ __txn_get_pinned_timestamp(
 	WT_CONNECTION_IMPL *conn;
 	WT_TXN *txn;
 	WT_TXN_GLOBAL *txn_global;
-	wt_timestamp_t tmp_ts;
+	wt_timestamp_t tmp_read_ts, tmp_ts;
 	bool include_oldest, txn_has_write_lock;
 
 	conn = S2C(session);
@@ -189,15 +203,18 @@ __txn_get_pinned_timestamp(
 	TAILQ_FOREACH(txn, &txn_global->read_timestamph, read_timestampq) {
 		/*
 		 * Skip any transactions on the queue that are not active.
+		 * Copy out value of read timestamp to prevent possible
+		 * race where a transaction resets its read timestamp while
+		 * we traverse the queue.
 		 */
-		if (txn->clear_read_q)
+		if (!__txn_get_read_timestamp(txn, &tmp_read_ts))
 			continue;
 		/*
 		 * A zero timestamp is possible here only when the oldest
 		 * timestamp is not accounted for.
 		 */
-		if (tmp_ts == 0 || txn->read_timestamp < tmp_ts)
-			tmp_ts = txn->read_timestamp;
+		if (tmp_ts == 0 || tmp_read_ts < tmp_ts)
+			tmp_ts = tmp_read_ts;
 		/*
 		 * We break on the first active txn on the list.
 		 */
@@ -638,22 +655,18 @@ __wt_txn_commit_timestamp_validate(WT_SESSION_IMPL *session, const char *name,
 	 * Compare against the prepare timestamp of the current transaction.
 	 * Return an error if the given timestamp is older than the prepare
 	 * timestamp.
+	 *
+	 * If roundup timestamps is configured, the commit timestamp will be
+	 * rounded up to the prepare timestamp.
 	 */
-	if (F_ISSET(txn, WT_TXN_PREPARE) && ts < txn->prepare_timestamp) {
-		/*
-		 * If roundup timestamps is configured, the commit timestamp
-		 * will be rounded up to the prepare timestamp.
-		 */
-		if (F_ISSET(txn, WT_TXN_TS_ROUND_PREPARED))
-			ts = txn->prepare_timestamp;
-		else {
-			__wt_timestamp_to_string(
-			    txn->prepare_timestamp, ts_string[0]);
-			WT_RET_MSG(session, EINVAL,
-			    "%s timestamp %.*s older than the prepare "
-			    "timestamp %s for this transaction",
-			    name, (int)cval->len, cval->str, ts_string[0]);
-		}
+	if (F_ISSET(txn, WT_TXN_PREPARE) &&
+	    !F_ISSET(txn, WT_TXN_TS_ROUND_PREPARED) &&
+	    ts < txn->prepare_timestamp) {
+		__wt_timestamp_to_string(txn->prepare_timestamp, ts_string[0]);
+		WT_RET_MSG(session, EINVAL,
+		    "%s timestamp %.*s older than the prepare "
+		    "timestamp %s for this transaction",
+		    name, (int)cval->len, cval->str, ts_string[0]);
 	}
 
 	if (F_ISSET(txn, WT_TXN_HAS_TS_DURABLE) &&
@@ -754,7 +767,7 @@ __wt_txn_parse_prepare_timestamp(
 	WT_CONFIG_ITEM cval;
 	WT_TXN *prev, *txn;
 	WT_TXN_GLOBAL *txn_global;
-	wt_timestamp_t oldest_ts, timestamp;
+	wt_timestamp_t oldest_ts, timestamp, tmp_timestamp;
 	char ts_string[2][WT_TS_INT_STRING_SIZE];
 
 	txn = &session->txn;
@@ -779,18 +792,23 @@ __wt_txn_parse_prepare_timestamp(
 	prev = TAILQ_LAST(
 	    &txn_global->read_timestamph, __wt_txn_rts_qh);
 	while (prev != NULL) {
-		/* Skip self and non-active transactions. */
-		if (prev->clear_read_q || prev == txn) {
+		/*
+		 * Skip self and non-active transactions. Copy out value of
+		 * read timestamp to prevent possible race where a transaction
+		 * resets its read timestamp while we traverse the queue.
+		 */
+		if (!__txn_get_read_timestamp(prev, &tmp_timestamp) ||
+		    prev == txn) {
 			prev = TAILQ_PREV(
 			    prev, __wt_txn_rts_qh, read_timestampq);
 			continue;
 		}
 
-		if (prev->read_timestamp >= timestamp) {
+		if (tmp_timestamp >= timestamp) {
 			__wt_readunlock(session,
 			    &txn_global->read_timestamp_rwlock);
 			__wt_timestamp_to_string(
-			    prev->read_timestamp, ts_string[0]);
+			    tmp_timestamp, ts_string[0]);
 			WT_RET_MSG(session, EINVAL,
 			    "prepare timestamp %.*s must be greater than the "
 			    "latest active read timestamp %s ",
@@ -1076,6 +1094,7 @@ __wt_txn_set_read_timestamp(WT_SESSION_IMPL *session)
 {
 	WT_TXN *qtxn, *txn, *txn_tmp;
 	WT_TXN_GLOBAL *txn_global;
+	wt_timestamp_t tmp_timestamp;
 	uint64_t walked;
 
 	txn = &session->txn;
@@ -1127,11 +1146,14 @@ __wt_txn_set_read_timestamp(WT_SESSION_IMPL *session)
 		 */
 		qtxn = TAILQ_LAST(
 		     &txn_global->read_timestamph, __wt_txn_rts_qh);
-		while (qtxn != NULL &&
-		    qtxn->read_timestamp > txn->read_timestamp) {
-			++walked;
-			qtxn = TAILQ_PREV(
-			    qtxn, __wt_txn_rts_qh, read_timestampq);
+		while (qtxn != NULL) {
+			if (!__txn_get_read_timestamp(qtxn, &tmp_timestamp) ||
+			    tmp_timestamp > txn->read_timestamp) {
+				++walked;
+				qtxn = TAILQ_PREV(qtxn,
+				    __wt_txn_rts_qh, read_timestampq);
+			} else
+				break;
 		}
 		if (qtxn == NULL) {
 			TAILQ_INSERT_HEAD(&txn_global->read_timestamph,
@@ -1165,9 +1187,10 @@ __wt_txn_clear_read_timestamp(WT_SESSION_IMPL *session)
 
 	txn = &session->txn;
 
-	if (!F_ISSET(txn, WT_TXN_PUBLIC_TS_READ))
+	if (!F_ISSET(txn, WT_TXN_PUBLIC_TS_READ)) {
+		txn->read_timestamp = WT_TS_NONE;
 		return;
-
+	}
 #ifdef HAVE_DIAGNOSTIC
 	{
 	WT_TXN_GLOBAL *txn_global;
@@ -1188,6 +1211,7 @@ __wt_txn_clear_read_timestamp(WT_SESSION_IMPL *session)
 	 */
 	WT_PUBLISH(txn->clear_read_q, true);
 	WT_PUBLISH(txn->flags, flags);
+	txn->read_timestamp = WT_TS_NONE;
 }
 
 /*