closes #5804, merge 51139, work to read in entire internal nodes into memory,...

closes #5804, merge 51139, work to read in entire internal nodes into memory, compressed, as opposed to using two I/O's, to main.

git-svn-id: file:///svn/toku/tokudb@51185 c7de825b-a66e-492c-adef-691d508d4ae1

ft/ft-internal.h

@@ -87,10 +87,13 @@ struct ftnode_fetch_extra {
     // and the user is doing a dictionary wide scan, then
     // even though a query may only want one basement node,
     // we fetch all basement nodes in a leaf node.
-    bool disable_prefetching; 
+    bool disable_prefetching;
     // this value will be set during the fetch_callback call by toku_ftnode_fetch_callback or toku_ftnode_pf_req_callback
     // thi callbacks need to evaluate this anyway, so we cache it here so the search code does not reevaluate it
     int child_to_read;
+    // when we read internal nodes, we want to read all the data off disk in one I/O
+    // then we'll treat it as normal and only decompress the needed partitions etc.
+    bool read_all_partitions;
     // Accounting: How many bytes were fetched, and how much time did it take?
     tokutime_t bytes_read;
     uint64_t read_time;
@@ -724,6 +727,7 @@ static inline void fill_bfe_for_full_read(struct ftnode_fetch_extra *bfe, FT h)
     bfe->right_is_pos_infty = false;
     bfe->child_to_read = -1;
     bfe->disable_prefetching = false;
+    bfe->read_all_partitions = false;
     bfe->bytes_read = 0;
     bfe->read_time = 0;
 }
@@ -742,7 +746,8 @@ static inline void fill_bfe_for_subset_read(
     DBT *right,
     bool left_is_neg_infty,
     bool right_is_pos_infty,
-    bool disable_prefetching
+    bool disable_prefetching,
+    bool read_all_partitions
     )
 {
     paranoid_invariant(h->h->type == FT_CURRENT);
@@ -755,6 +760,7 @@ static inline void fill_bfe_for_subset_read(
     bfe->right_is_pos_infty = right_is_pos_infty;
     bfe->child_to_read = -1;
     bfe->disable_prefetching = disable_prefetching;
+    bfe->read_all_partitions = read_all_partitions;
     bfe->bytes_read = 0;
     bfe->read_time = 0;
 }
@@ -776,6 +782,7 @@ static inline void fill_bfe_for_min_read(struct ftnode_fetch_extra *bfe, FT h) {
     bfe->right_is_pos_infty = false;
     bfe->child_to_read = -1;
     bfe->disable_prefetching = false;
+    bfe->read_all_partitions = false;
     bfe->bytes_read = 0;
     bfe->read_time = 0;
 }
@@ -824,6 +831,7 @@ static inline void fill_bfe_for_prefetch(struct ftnode_fetch_extra *bfe,
     bfe->right_is_pos_infty = c->right_is_pos_infty;
     bfe->child_to_read = -1;
     bfe->disable_prefetching = c->disable_prefetching;
+    bfe->read_all_partitions = false;
     bfe->bytes_read = 0;
     bfe->read_time = 0;
 }

ft/ft-ops.cc

@@ -4860,6 +4860,9 @@ ft_search_child(FT_HANDLE brt, FTNODE node, int childnum, ft_search_t *search, F
     uint32_t fullhash = compute_child_fullhash(brt->ft->cf, node, childnum);
     FTNODE childnode;
 
+    // If the current node's height is greater than 1, then its child is an internal node.
+    // Therefore, to warm the cache better (#5798), we want to read all the partitions off disk in one shot.
+    bool read_all_partitions = node->height > 1;
     struct ftnode_fetch_extra bfe;
     fill_bfe_for_subset_read(
         &bfe,
@@ -4869,7 +4872,8 @@ ft_search_child(FT_HANDLE brt, FTNODE node, int childnum, ft_search_t *search, F
         &ftcursor->range_lock_right_key,
         ftcursor->left_is_neg_infty,
         ftcursor->right_is_pos_infty,
-        ftcursor->disable_prefetching
+        ftcursor->disable_prefetching,
+        read_all_partitions
         );
     bool msgs_applied = false;
     {
@@ -5195,7 +5199,8 @@ try_again:
         &ftcursor->range_lock_right_key,
         ftcursor->left_is_neg_infty,
         ftcursor->right_is_pos_infty,
-        ftcursor->disable_prefetching
+        ftcursor->disable_prefetching,
+        true // We may as well always read the whole root into memory, if it's a leaf node it's a tiny tree anyway.
         );
     FTNODE node = NULL;
     {

ft/ft_node-serialize.cc

@@ -2445,9 +2445,15 @@ toku_deserialize_ftnode_from (int fd,
     toku_trace("deserial start");
     int r = 0;
     struct rbuf rb = RBUF_INITIALIZER;
-    read_ftnode_header_from_fd_into_rbuf_if_small_enough(fd, blocknum, bfe->h, &rb, bfe);
 
-    r = deserialize_ftnode_header_from_rbuf_if_small_enough(ftnode, ndd, blocknum, fullhash, bfe, &rb, fd);
+    if (!bfe->read_all_partitions) {
+        read_ftnode_header_from_fd_into_rbuf_if_small_enough(fd, blocknum, bfe->h, &rb, bfe);
+
+        r = deserialize_ftnode_header_from_rbuf_if_small_enough(ftnode, ndd, blocknum, fullhash, bfe, &rb, fd);
+    } else {
+        // force us to do it the old way
+        r = -1;
+    }
     if (r != 0) {
         // Something went wrong, go back to doing it the old way.
         r = deserialize_ftnode_from_fd(fd, blocknum, fullhash, ftnode, ndd, bfe, NULL);

ft/tests/ft-bfe-query.cc

@@ -192,6 +192,7 @@ test_subset_read(int fd, FT_HANDLE UU(brt), FT brt_h) {
         &right,
         false,
         false,
+        false,
         false
         );
     

ft/tests/ft-clock-test.cc

@@ -164,6 +164,7 @@ test2(int fd, FT brt_h, FTNODE *dn) {
         &right,
         true,
         true,
+        false,
         false
         );
     FTNODE_DISK_DATA ndd = NULL;