WT-7757 Skip internal and leaf pages with content obsolete to the reader

dist/s_void

@@ -60,6 +60,7 @@ func_ok()
 	    -e '/int __wt_block_manager_size$/d' \
 	    -e '/int __wt_block_tiered_load$/d' \
 	    -e '/int __wt_block_write_size$/d' \
+	    -e '/int __wt_btcur_skip_page$/d' \
 	    -e '/int __wt_buf_catfmt$/d' \
 	    -e '/int __wt_buf_fmt$/d' \
 	    -e '/int __wt_count_birthmarks$/d' \

dist/stat_data.py

@@ -811,13 +811,11 @@ def __init__(self, name, desc, flags=''):
     CursorStat('cursor_next_hs_tombstone', 'cursor next calls that skip due to a globally visible history store tombstone'),
     CursorStat('cursor_next_skip_ge_100', 'cursor next calls that skip greater than or equal to 100 entries'),
     CursorStat('cursor_next_skip_lt_100', 'cursor next calls that skip less than 100 entries'),
-    CursorStat('cursor_next_skip_page_count', 'Total number of pages skipped without reading by cursor next calls'),
     CursorStat('cursor_next_skip_total', 'Total number of entries skipped by cursor next calls'),
     CursorStat('cursor_open_count', 'open cursor count', 'no_clear,no_scale'),
     CursorStat('cursor_prev_hs_tombstone', 'cursor prev calls that skip due to a globally visible history store tombstone'),
     CursorStat('cursor_prev_skip_ge_100', 'cursor prev calls that skip greater than or equal to 100 entries'),
     CursorStat('cursor_prev_skip_lt_100', 'cursor prev calls that skip less than 100 entries'),
-    CursorStat('cursor_prev_skip_page_count', 'Total number of pages skipped without reading by cursor prev calls'),
     CursorStat('cursor_prev_skip_total', 'Total number of entries skipped by cursor prev calls'),
     CursorStat('cursor_search_near_prefix_fast_paths', 'Total number of times a search near has exited due to prefix config'),
     CursorStat('cursor_skip_hs_cur_position', 'Total number of entries skipped to position the history store cursor'),

src/btree/bt_curnext.c

@@ -644,13 +644,12 @@ __wt_btcur_next_prefix(WT_CURSOR_BTREE *cbt, WT_ITEM *prefix, bool truncating)
     WT_DECL_RET;
     WT_PAGE *page;
     WT_SESSION_IMPL *session;
-    size_t pages_skipped_count, total_skipped, skipped;
+    size_t total_skipped, skipped;
     uint32_t flags;
     bool newpage, restart;
 
     cursor = &cbt->iface;
     session = CUR2S(cbt);
-    pages_skipped_count = 0;
     total_skipped = 0;
 
     WT_STAT_CONN_DATA_INCR(session, cursor_next);
@@ -670,24 +669,14 @@ __wt_btcur_next_prefix(WT_CURSOR_BTREE *cbt, WT_ITEM *prefix, bool truncating)
         __wt_btcur_iterate_setup(cbt);
 
     /*
-     * Walk any page we're holding until the underlying call returns not- found. Then, move to the
+     * Walk any page we're holding until the underlying call returns not-found. Then, move to the
      * next page, until we reach the end of the file.
      */
     restart = F_ISSET(cbt, WT_CBT_ITERATE_RETRY_NEXT);
     F_CLR(cbt, WT_CBT_ITERATE_RETRY_NEXT);
     for (newpage = false;; newpage = true, restart = false) {
         page = cbt->ref == NULL ? NULL : cbt->ref->page;
 
-        /*
-         * Determine if all records on the page have been deleted and all the tombstones are visible
-         * to our transaction. If so, we can avoid reading the records on the page and move to the
-         * next page.
-         */
-        if (__wt_btcur_skip_page(cbt)) {
-            pages_skipped_count++;
-            goto skip_page;
-        }
-
         if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
             /* The page cannot be NULL if the above flag is set. */
             WT_ASSERT(session, page != NULL);
@@ -759,16 +748,26 @@ __wt_btcur_next_prefix(WT_CURSOR_BTREE *cbt, WT_ITEM *prefix, bool truncating)
             WT_STAT_CONN_INCR(session, cache_eviction_force_delete);
         }
         cbt->page_deleted_count = 0;
-skip_page:
+
         if (F_ISSET(cbt, WT_CBT_READ_ONCE))
             LF_SET(WT_READ_WONT_NEED);
-        WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
+
+        /*
+         * If we are running with snapshot isolation, and not interested in returning tombstones, we
+         * could potentially skip pages. The skip function looks at the aggregated timestamp
+         * information to determine if something is visible on the page. If nothing is, the page is
+         * skipped.
+         */
+        if (session->txn->isolation == WT_ISO_SNAPSHOT &&
+          !F_ISSET(&cbt->iface, WT_CURSTD_IGNORE_TOMBSTONE))
+            WT_ERR(
+              __wt_tree_walk_custom_skip(session, &cbt->ref, __wt_btcur_skip_page, NULL, flags));
+        else
+            WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
         WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND, false);
     }
 
 err:
-    WT_STAT_CONN_DATA_INCRV(session, cursor_next_skip_page_count, pages_skipped_count);
-
     if (total_skipped < 100)
         WT_STAT_CONN_DATA_INCR(session, cursor_next_skip_lt_100);
     else

src/btree/bt_curprev.c

@@ -590,13 +590,12 @@ __wt_btcur_prev_prefix(WT_CURSOR_BTREE *cbt, WT_ITEM *prefix, bool truncating)
     WT_DECL_RET;
     WT_PAGE *page;
     WT_SESSION_IMPL *session;
-    size_t pages_skipped_count, total_skipped, skipped;
+    size_t total_skipped, skipped;
     uint32_t flags;
     bool newpage, restart;
 
     cursor = &cbt->iface;
     session = CUR2S(cbt);
-    pages_skipped_count = 0;
     total_skipped = 0;
 
     WT_STAT_CONN_DATA_INCR(session, cursor_prev);
@@ -617,24 +616,14 @@ __wt_btcur_prev_prefix(WT_CURSOR_BTREE *cbt, WT_ITEM *prefix, bool truncating)
         __wt_btcur_iterate_setup(cbt);
 
     /*
-     * Walk any page we're holding until the underlying call returns not- found. Then, move to the
+     * Walk any page we're holding until the underlying call returns not-found. Then, move to the
      * previous page, until we reach the start of the file.
      */
     restart = F_ISSET(cbt, WT_CBT_ITERATE_RETRY_PREV);
     F_CLR(cbt, WT_CBT_ITERATE_RETRY_PREV);
     for (newpage = false;; newpage = true, restart = false) {
         page = cbt->ref == NULL ? NULL : cbt->ref->page;
 
-        /*
-         * Determine if all records on the page have been deleted and all the tombstones are visible
-         * to our transaction. If so, we can avoid reading the records on the page and move to the
-         * next page.
-         */
-        if (__wt_btcur_skip_page(cbt)) {
-            pages_skipped_count++;
-            goto skip_page;
-        }
-
         /*
          * Column-store pages may have appended entries. Handle it separately from the usual cursor
          * code, it's in a simple format.
@@ -707,16 +696,26 @@ __wt_btcur_prev_prefix(WT_CURSOR_BTREE *cbt, WT_ITEM *prefix, bool truncating)
             WT_STAT_CONN_INCR(session, cache_eviction_force_delete);
         }
         cbt->page_deleted_count = 0;
-skip_page:
+
         if (F_ISSET(cbt, WT_CBT_READ_ONCE))
             LF_SET(WT_READ_WONT_NEED);
-        WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
+
+        /*
+         * If we are running with snapshot isolation, and not interested in returning tombstones, we
+         * could potentially skip pages. The skip function looks at the aggregated timestamp
+         * information to determine if something is visible on the page. If nothing is, the page is
+         * skipped.
+         */
+        if (session->txn->isolation == WT_ISO_SNAPSHOT &&
+          !F_ISSET(&cbt->iface, WT_CURSTD_IGNORE_TOMBSTONE))
+            WT_ERR(
+              __wt_tree_walk_custom_skip(session, &cbt->ref, __wt_btcur_skip_page, NULL, flags));
+        else
+            WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
         WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND, false);
     }
 
 err:
-    WT_STAT_CONN_DATA_INCRV(session, cursor_prev_skip_page_count, pages_skipped_count);
-
     if (total_skipped < 100)
         WT_STAT_CONN_DATA_INCR(session, cursor_prev_skip_lt_100);
     else

src/include/btree_inline.h

@@ -2024,50 +2024,36 @@ __wt_bt_col_var_cursor_walk_txn_read(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *
  *     Return if the cursor is pointing to a page with deleted records and can be skipped for cursor
  *     traversal.
  */
-static inline bool
-__wt_btcur_skip_page(WT_CURSOR_BTREE *cbt)
+static inline int
+__wt_btcur_skip_page(WT_SESSION_IMPL *session, WT_REF *ref, void *context, bool *skipp)
 {
     WT_ADDR_COPY addr;
-    WT_PAGE *page;
-    WT_REF *ref;
-    WT_SESSION_IMPL *session;
     uint8_t previous_state;
-    bool can_skip;
 
-    session = CUR2S(cbt);
-    ref = cbt->ref;
-    page = cbt->ref == NULL ? NULL : cbt->ref->page;
+    WT_UNUSED(context);
 
-    if (page == NULL)
-        return false;
-
-    previous_state = ref->state;
-    can_skip = false;
+    *skipp = false; /* Default to reading */
 
     /*
      * Determine if all records on the page have been deleted and all the tombstones are visible to
      * our transaction. If so, we can avoid reading the records on the page and move to the next
      * page. We base this decision on the aggregate stop point added to the page during the last
-     * reconciliation. We can skip this test if the page has been modified since it was reconciled
-     * or the underlying cursor is configured to ignore tombstones.
+     * reconciliation. We can skip this test if the page has been modified since it was reconciled.
      *
      * We are making these decisions while holding a lock for the page as checkpoint or eviction can
-     * make changes to the data structures (i.e., aggregate timestamps) we are reading.
+     * make changes to the data structures (i.e., aggregate timestamps) we are reading. It is okay
+     * if the page is not in memory, or gets evicted before we lock it. In such a case, we can forgo
+     * checking if the page has been modified. So, only do a page modified check if the page was in
+     * memory before locking.
      */
-    if (session->txn->isolation == WT_ISO_SNAPSHOT && !__wt_page_is_modified(page) &&
-      !F_ISSET(&cbt->iface, WT_CURSTD_IGNORE_TOMBSTONE) && previous_state == WT_REF_MEM) {
-
-        /* We only try to lock the page once. */
-        if (!WT_REF_CAS_STATE(session, ref, previous_state, WT_REF_LOCKED))
-            return false;
+    WT_REF_LOCK(session, ref, &previous_state);
+    if ((previous_state == WT_REF_DISK || previous_state == WT_REF_DELETED ||
+          (previous_state == WT_REF_MEM && !__wt_page_is_modified(ref->page))) &&
+      __wt_ref_addr_copy(session, ref, &addr) &&
+      __wt_txn_visible(session, addr.ta.newest_stop_txn, addr.ta.newest_stop_ts) &&
+      __wt_txn_visible(session, addr.ta.newest_stop_txn, addr.ta.newest_stop_durable_ts))
+        *skipp = true;
+    WT_REF_UNLOCK(ref, previous_state);
 
-        if (__wt_ref_addr_copy(session, ref, &addr) &&
-          __wt_txn_visible(session, addr.ta.newest_stop_txn, addr.ta.newest_stop_ts) &&
-          __wt_txn_visible(session, addr.ta.newest_stop_txn, addr.ta.newest_stop_durable_ts))
-            can_skip = true;
-
-        WT_REF_SET_STATE(ref, previous_state);
-    }
-
-    return (can_skip);
+    return (0);
 }

src/include/extern.h

@@ -1880,8 +1880,6 @@ static inline WT_IKEY *__wt_ref_key_instantiated(WT_REF *ref)
   WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 static inline WT_VISIBLE_TYPE __wt_txn_upd_visible_type(WT_SESSION_IMPL *session, WT_UPDATE *upd)
   WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
-static inline bool __wt_btcur_skip_page(WT_CURSOR_BTREE *cbt)
-  WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 static inline bool __wt_btree_can_evict_dirty(WT_SESSION_IMPL *session)
   WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 static inline bool __wt_btree_dominating_cache(WT_SESSION_IMPL *session, WT_BTREE *btree)
@@ -1968,6 +1966,8 @@ static inline double __wt_eviction_dirty_target(WT_CACHE *cache)
 static inline int __wt_bt_col_var_cursor_walk_txn_read(WT_SESSION_IMPL *session,
   WT_CURSOR_BTREE *cbt, WT_PAGE *page, WT_CELL_UNPACK_KV *unpack, WT_COL *cip)
   WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+static inline int __wt_btcur_skip_page(WT_SESSION_IMPL *session, WT_REF *ref, void *context,
+  bool *skipp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 static inline int __wt_btree_block_free(WT_SESSION_IMPL *session, const uint8_t *addr,
   size_t addr_size) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 static inline int __wt_buf_extend(WT_SESSION_IMPL *session, WT_ITEM *buf, size_t size)

src/include/stat.h

@@ -503,8 +503,6 @@ struct __wt_connection_stats {
     int64_t cursor_next_skip_total;
     int64_t cursor_prev_skip_total;
     int64_t cursor_skip_hs_cur_position;
-    int64_t cursor_next_skip_page_count;
-    int64_t cursor_prev_skip_page_count;
     int64_t cursor_search_near_prefix_fast_paths;
     int64_t cursor_cached_count;
     int64_t cursor_insert_bulk;
@@ -932,8 +930,6 @@ struct __wt_dsrc_stats {
     int64_t cursor_next_skip_total;
     int64_t cursor_prev_skip_total;
     int64_t cursor_skip_hs_cur_position;
-    int64_t cursor_next_skip_page_count;
-    int64_t cursor_prev_skip_page_count;
     int64_t cursor_search_near_prefix_fast_paths;
     int64_t cursor_insert_bulk;
     int64_t cursor_reopen;