Import wiredtiger: 5a74e438ea34cc6737438f6c99ce2e5e25519a0e from branch mongodb-4.4

ref: 5e6daf7d42..5a74e438ea
for: 4.4.0-rc11

WT-6412       Fix extended stalls being seen during MongoDB performance testing

7 files changed, 87 insertions, 32 deletions

diff --git a/src/third_party/wiredtiger/import.data b/src/third_party/wiredtiger/import.data
index 6ba5f527c47..93efc061937 100644
--- a/src/third_party/wiredtiger/import.data
+++ b/src/third_party/wiredtiger/import.data
@@ -2,5 +2,5 @@
     "vendor": "wiredtiger",
     "github": "wiredtiger/wiredtiger.git",
     "branch": "mongodb-4.4",
-    "commit": "5e6daf7d42727e3d86b2603c20852d1426dee55f"
+    "commit": "5a74e438ea34cc6737438f6c99ce2e5e25519a0e"
 }
diff --git a/src/third_party/wiredtiger/src/history/hs.c b/src/third_party/wiredtiger/src/history/hs.c
index e94270eb37c..3b7c9a69de7 100644
--- a/src/third_party/wiredtiger/src/history/hs.c
+++ b/src/third_party/wiredtiger/src/history/hs.c
@@ -1105,14 +1105,14 @@ err:
 }
 
 /*
- * __wt_find_hs_upd --
+ * __wt_hs_find_upd --
  *     Scan the history store for a record the btree cursor wants to position on. Create an update
  *     for the record and return to the caller. The caller may choose to optionally allow prepared
  *     updates to be returned regardless of whether prepare is being ignored globally. Otherwise, a
  *     prepare conflict will be returned upon reading a prepared update.
  */
 int
-__wt_find_hs_upd(WT_SESSION_IMPL *session, WT_ITEM *key, const char *value_format, uint64_t recno,
+__wt_hs_find_upd(WT_SESSION_IMPL *session, WT_ITEM *key, const char *value_format, uint64_t recno,
   WT_UPDATE_VALUE *upd_value, bool allow_prepare, WT_ITEM *on_disk_buf)
 {
     WT_CURSOR *hs_cursor;
@@ -1202,6 +1202,13 @@ __wt_find_hs_upd(WT_SESSION_IMPL *session, WT_ITEM *key, const char *value_forma
             goto done;
 
         /*
+         * If the stop time pair on the tombstone in the history store is already globally visible
+         * we can skip it.
+         */
+        if (__wt_txn_visible_all(
+              session, hs_cbt->upd_value->tw.stop_txn, hs_cbt->upd_value->tw.durable_stop_ts))
+            continue;
+        /*
          * If the stop time point of a record is visible to us, we won't be able to see anything for
          * this entire key. Just jump straight to the end.
          */
@@ -1521,10 +1528,17 @@ __hs_fixup_out_of_order_from_pos(WT_SESSION_IMPL *session, WT_CURSOR *hs_cursor,
         WT_ERR(hs_cursor->get_key(hs_cursor, &hs_btree_id, &hs_key, &hs_ts, &hs_counter));
         if (hs_btree_id != btree->id)
             break;
+
         WT_ERR(__wt_compare(session, NULL, &hs_key, key, &cmp));
         if (cmp != 0)
             break;
-
+        /*
+         * If the stop time pair on the tombstone in the history store is already globally visible
+         * we can skip it.
+         */
+        if (__wt_txn_visible_all(
+              session, hs_cbt->upd_value->tw.stop_txn, hs_cbt->upd_value->tw.durable_stop_ts))
+            continue;
         /*
          * If we got here, we've got out-of-order updates in the history store.
          *
@@ -1617,6 +1631,14 @@ __hs_delete_key_from_pos(
         WT_RET(__wt_compare(session, NULL, &hs_key, key, &cmp));
         if (cmp != 0)
             break;
+
+        /*
+         * If the stop time pair on the tombstone in the history store is already globally visible
+         * we can skip it.
+         */
+        if (__wt_txn_visible_all(
+              session, hs_cbt->upd_value->tw.stop_txn, hs_cbt->upd_value->tw.durable_stop_ts))
+            continue;
         /*
          * Since we're using internal functions to modify the row structure, we need to manually set
          * the comparison to an exact match.
diff --git a/src/third_party/wiredtiger/src/include/cursor.h b/src/third_party/wiredtiger/src/include/cursor.h
index 95e5c8de1fb..dfcc4f7888f 100644
--- a/src/third_party/wiredtiger/src/include/cursor.h
+++ b/src/third_party/wiredtiger/src/include/cursor.h
@@ -503,9 +503,5 @@ struct __wt_cursor_table {
 
 #define WT_CURSOR_RECNO(cursor) WT_STREQ((cursor)->key_format, "r")
 
-#define WT_CURSOR_IS_DUMP(cursor) \
-    F_ISSET(cursor,               \
-      (WT_CURSTD_DUMP_HEX | WT_CURSTD_DUMP_JSON | WT_CURSTD_DUMP_PRETTY | WT_CURSTD_DUMP_PRINT))
-
 #define WT_CURSOR_RAW_OK \
     (WT_CURSTD_DUMP_HEX | WT_CURSTD_DUMP_PRETTY | WT_CURSTD_DUMP_PRINT | WT_CURSTD_RAW)
diff --git a/src/third_party/wiredtiger/src/include/extern.h b/src/third_party/wiredtiger/src/include/extern.h
index 31e24f65edd..bfb6647c88f 100644
--- a/src/third_party/wiredtiger/src/include/extern.h
+++ b/src/third_party/wiredtiger/src/include/extern.h
@@ -724,9 +724,6 @@ extern int __wt_filename(WT_SESSION_IMPL *session, const char *name, char **path
 extern int __wt_filename_construct(WT_SESSION_IMPL *session, const char *path,
   const char *file_prefix, uintmax_t id_1, uint32_t id_2, WT_ITEM *buf)
   WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
-extern int __wt_find_hs_upd(WT_SESSION_IMPL *session, WT_ITEM *key, const char *value_format,
-  uint64_t recno, WT_UPDATE_VALUE *upd_value, bool allow_prepare, WT_ITEM *on_disk_buf)
-  WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern int __wt_fopen(WT_SESSION_IMPL *session, const char *name, uint32_t open_flags,
   uint32_t flags, WT_FSTREAM **fstrp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern int __wt_fsync_background(WT_SESSION_IMPL *session)
@@ -767,6 +764,9 @@ extern int __wt_hs_cursor_position(WT_SESSION_IMPL *session, WT_CURSOR *cursor,
   WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern int __wt_hs_delete_key_from_ts(WT_SESSION_IMPL *session, uint32_t btree_id,
   const WT_ITEM *key, wt_timestamp_t ts) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_hs_find_upd(WT_SESSION_IMPL *session, WT_ITEM *key, const char *value_format,
+  uint64_t recno, WT_UPDATE_VALUE *upd_value, bool allow_prepare, WT_ITEM *on_disk_buf)
+  WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern int __wt_hs_get_btree(WT_SESSION_IMPL *session, WT_BTREE **hs_btreep)
   WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern int __wt_hs_insert_updates(WT_SESSION_IMPL *session, WT_PAGE *page, WT_MULTI *multi)
diff --git a/src/third_party/wiredtiger/src/include/txn.i b/src/third_party/wiredtiger/src/include/txn.i
index b99a54532b0..530b40ef9ed 100644
--- a/src/third_party/wiredtiger/src/include/txn.i
+++ b/src/third_party/wiredtiger/src/include/txn.i
@@ -854,10 +854,12 @@ __wt_txn_read_upd_list(
         /*
          * If the cursor is configured to ignore tombstones, copy the timestamps from the tombstones
          * to the stop time window of the update value being returned to the caller. Caller can
-         * process the stop time window to decide if there was a tombstone on the update chain.
+         * process the stop time window to decide if there was a tombstone on the update chain. If
+         * the time window already has a stop time set then we must've seen a tombstone prior to
+         * ours in the update list, and therefore don't need to do this again.
          */
         if (type == WT_UPDATE_TOMBSTONE && F_ISSET(&cbt->iface, WT_CURSTD_IGNORE_TOMBSTONE) &&
-          !__wt_txn_upd_visible_all(session, upd)) {
+          !WT_TIME_WINDOW_HAS_STOP(&cbt->upd_value->tw)) {
             cbt->upd_value->tw.durable_stop_ts = upd->durable_ts;
             cbt->upd_value->tw.stop_ts = upd->start_ts;
             cbt->upd_value->tw.stop_txn = upd->txnid;
@@ -918,7 +920,7 @@ __wt_txn_read(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_ITEM *key, uint
 {
     WT_TIME_WINDOW tw;
     WT_UPDATE *prepare_upd;
-
+    bool have_stop_tw;
     prepare_upd = NULL;
 
     WT_RET(__wt_txn_read_upd_list(session, cbt, upd, &prepare_upd));
@@ -933,6 +935,12 @@ __wt_txn_read(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_ITEM *key, uint
         return (0);
     }
 
+    /*
+     * When we inspected the update list we may have seen a tombstone leaving us with a valid stop
+     * time window, we don't want to overwrite this stop time window.
+     */
+    have_stop_tw = WT_TIME_WINDOW_HAS_STOP(&cbt->upd_value->tw);
+
     /* Check the ondisk value. */
     if (vpack == NULL) {
         WT_TIME_WINDOW_INIT(&tw);
@@ -949,9 +957,8 @@ __wt_txn_read(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_ITEM *key, uint
      * return "not found", except scanning the history store during rollback to stable and when we
      * are told to ignore non-globally visible tombstones.
      */
-    if (__wt_txn_tw_stop_visible(session, &tw) &&
-      (!F_ISSET(&cbt->iface, WT_CURSTD_IGNORE_TOMBSTONE) ||
-          (__wt_txn_tw_stop_visible_all(session, &tw) && !WT_CURSOR_IS_DUMP(&cbt->iface)))) {
+    if (!have_stop_tw && __wt_txn_tw_stop_visible(session, &tw) &&
+      !F_ISSET(&cbt->iface, WT_CURSTD_IGNORE_TOMBSTONE)) {
         cbt->upd_value->buf.data = NULL;
         cbt->upd_value->buf.size = 0;
         cbt->upd_value->tw.durable_stop_ts = tw.durable_stop_ts;
@@ -963,7 +970,7 @@ __wt_txn_read(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_ITEM *key, uint
     }
 
     /* Store the stop time pair of the history store record that is returning. */
-    if (WT_TIME_WINDOW_HAS_STOP(&tw) && WT_IS_HS(S2BT(session))) {
+    if (!have_stop_tw && WT_TIME_WINDOW_HAS_STOP(&tw) && WT_IS_HS(S2BT(session))) {
         cbt->upd_value->tw.durable_stop_ts = tw.durable_stop_ts;
         cbt->upd_value->tw.stop_ts = tw.stop_ts;
         cbt->upd_value->tw.stop_txn = tw.stop_txn;
@@ -986,7 +993,7 @@ __wt_txn_read(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_ITEM *key, uint
 
     /* If there's no visible update in the update chain or ondisk, check the history store file. */
     if (F_ISSET(S2C(session), WT_CONN_HS_OPEN) && !F_ISSET(S2BT(session), WT_BTREE_HS))
-        WT_RET_NOTFOUND_OK(__wt_find_hs_upd(session, key, cbt->iface.value_format, recno,
+        WT_RET_NOTFOUND_OK(__wt_hs_find_upd(session, key, cbt->iface.value_format, recno,
           cbt->upd_value, false, &cbt->upd_value->buf));
 
     /*
diff --git a/src/third_party/wiredtiger/src/txn/txn.c b/src/third_party/wiredtiger/src/txn/txn.c
index b63dadbf9b5..571ec8fd6bc 100644
--- a/src/third_party/wiredtiger/src/txn/txn.c
+++ b/src/third_party/wiredtiger/src/txn/txn.c
@@ -654,23 +654,37 @@ __txn_append_hs_record(WT_SESSION_IMPL *session, WT_CURSOR *hs_cursor, WT_ITEM *
     WT_ERR(__wt_scr_alloc(session, 0, &hs_key));
     WT_ERR(__wt_scr_alloc(session, 0, &hs_value));
 
-    WT_ERR(hs_cursor->get_key(hs_cursor, &hs_btree_id, hs_key, &hs_start_ts, &hs_counter));
+    for (; ret == 0; ret = hs_cursor->prev(hs_cursor)) {
+        WT_ERR(hs_cursor->get_key(hs_cursor, &hs_btree_id, hs_key, &hs_start_ts, &hs_counter));
 
-    /* Not found if we cross the tree boundary. */
-    if (hs_btree_id != S2BT(session)->id) {
-        ret = WT_NOTFOUND;
-        goto done;
+        /* Stop before crossing over to the next btree */
+        if (hs_btree_id != S2BT(session)->id) {
+            ret = WT_NOTFOUND;
+            goto done;
+        }
+
+        /*
+         * Keys are sorted in an order, skip the ones before the desired key, and bail out if we
+         * have crossed over the desired key and not found the record we are looking for.
+         */
+        WT_ERR(__wt_compare(session, NULL, hs_key, key, &cmp));
+        if (cmp != 0) {
+            ret = WT_NOTFOUND;
+            goto done;
+        }
+
+        /*
+         * If the stop time pair on the tombstone in the history store is already globally visible
+         * we can skip it.
+         */
+        if (!__wt_txn_visible_all(
+              session, hs_cbt->upd_value->tw.stop_txn, hs_cbt->upd_value->tw.durable_stop_ts))
+            break;
     }
 
-    /*
-     * Keys are sorted in an order, skip the ones before the desired key, and bail out if we have
-     * crossed over the desired key and not found the record we are looking for.
-     */
-    WT_ERR(__wt_compare(session, NULL, hs_key, key, &cmp));
-    if (cmp != 0) {
-        ret = WT_NOTFOUND;
+    /* We walked off the top of the history store. */
+    if (ret == WT_NOTFOUND)
         goto done;
-    }
 
     /*
      * As part of the history store search, we never get an exact match based on our search criteria
diff --git a/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c b/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c
index afc3508e577..bb4a96431f0 100644
--- a/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c
+++ b/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c
@@ -228,6 +228,14 @@ __rollback_row_ondisk_fixup_key(WT_SESSION_IMPL *session, WT_PAGE *page, WT_ROW
             break;
 
         /*
+         * If the stop time pair on the tombstone in the history store is already globally visible
+         * we can skip it.
+         */
+        if (__wt_txn_visible_all(
+              session, cbt->upd_value->tw.stop_txn, cbt->upd_value->tw.durable_stop_ts))
+            continue;
+
+        /*
          * As part of the history store search, we never get an exact match based on our search
          * criteria as we always search for a maximum record for that key. Make sure that we set the
          * comparison result as an exact match to remove this key as part of rollback to stable. In
@@ -1074,6 +1082,14 @@ __rollback_to_stable_btree_hs_truncate(WT_SESSION_IMPL *session, uint32_t btree_
         if (btree_id != hs_btree_id)
             break;
 
+        /*
+         * If the stop time pair on the tombstone in the history store is already globally visible
+         * we can skip it.
+         */
+        if (__wt_txn_visible_all(
+              session, cbt->upd_value->tw.stop_txn, cbt->upd_value->tw.durable_stop_ts))
+            continue;
+
         /* Set this comparison as exact match of the search for later use. */
         cbt->compare = 0;
         __wt_verbose(session, WT_VERB_RTS,