fixes #156 Add a parameter for the nonleaf partial eviction strategy

buildheader/make_tdb.cc

@@ -828,6 +828,7 @@ int main (int argc, char *const argv[] __attribute__((__unused__))) {
     printf("int toku_set_trace_file (const char *fname) %s;\n", VISIBLE);
     printf("int toku_close_trace_file (void) %s;\n", VISIBLE);
     printf("void db_env_set_direct_io (bool direct_io_on) %s;\n", VISIBLE);
+    printf("void db_env_set_compress_buffers_before_eviction (bool compress_buffers) %s;\n", VISIBLE);
     printf("void db_env_set_func_fsync (int (*)(int)) %s;\n", VISIBLE);
     printf("void db_env_set_func_free (void (*)(void*)) %s;\n", VISIBLE);
     printf("void db_env_set_func_malloc (void *(*)(size_t)) %s;\n", VISIBLE);

ft/ft-ops.cc

@@ -979,6 +979,12 @@ int toku_ftnode_fetch_callback (CACHEFILE UU(cachefile), PAIR p, int fd, BLOCKNU
     return r;
 }
 
+static bool ft_compress_buffers_before_eviction = true;
+
+void toku_ft_set_compress_buffers_before_eviction(bool compress_buffers) {
+    ft_compress_buffers_before_eviction = compress_buffers;
+}
+
 void toku_ftnode_pe_est_callback(
     void* ftnode_pv,
     void* disk_data,
@@ -1010,16 +1016,20 @@ void toku_ftnode_pe_est_callback(
             // we compress this node and add it to
             // bytes_to_free
 
-            // first get an estimate for how much space will be taken
+            if (ft_compress_buffers_before_eviction) {
-            // after compression, it is simply the size of compressed
+                // first get an estimate for how much space will be taken
-            // data on disk plus the size of the struct that holds it
+                // after compression, it is simply the size of compressed
-            FTNODE_DISK_DATA ndd = (FTNODE_DISK_DATA) disk_data;
+                // data on disk plus the size of the struct that holds it
-            uint32_t compressed_data_size = BP_SIZE(ndd, i);
+                FTNODE_DISK_DATA ndd = (FTNODE_DISK_DATA) disk_data;
-            compressed_data_size += sizeof(struct sub_block);
+                uint32_t compressed_data_size = BP_SIZE(ndd, i);
+                compressed_data_size += sizeof(struct sub_block);
 
-            // now get the space taken now
+                // now get the space taken now
-            uint32_t decompressed_data_size = get_avail_internal_node_partition_size(node,i);
+                uint32_t decompressed_data_size = get_avail_internal_node_partition_size(node,i);
-            bytes_to_free += (decompressed_data_size - compressed_data_size);
+                bytes_to_free += (decompressed_data_size - compressed_data_size);
+            } else {
+                bytes_to_free += get_avail_internal_node_partition_size(node, i);
+            }
         }
     }
 
@@ -1088,9 +1098,20 @@ int toku_ftnode_pe_callback (void *ftnode_pv, PAIR_ATTR UU(old_attr), PAIR_ATTR*
                     if (num_partial_evictions++ == 0) {
                         size_before = ftnode_memory_size(node);
                     }
-                    compress_internal_node_partition(node, i, 
+                    if (ft_compress_buffers_before_eviction) {
-                                                     // When partially evicting, always compress with quicklz,
+                        // When partially evicting, always compress with quicklz
-                                                     TOKU_QUICKLZ_METHOD);
+                        compress_internal_node_partition(
+                            node,
+                            i,
+                            TOKU_QUICKLZ_METHOD
+                            );
+                    } else {
+                        // We're not compressing buffers before eviction. Simply
+                        // detach the buffer and set the child's state to on-disk.
+                        destroy_nonleaf_childinfo(BNC(node, i));
+                        set_BNULL(node, i);
+                        BP_STATE(node, i) = PT_ON_DISK;
+                    }
                 }
                 else {
                     BP_SWEEP_CLOCK(node,i);

ft/ft-ops.h

@@ -351,5 +351,8 @@ int toku_ft_strerror_r(int error, char *buf, size_t buflen);
 
 extern bool garbage_collection_debug;
 
+// This is a poor place to put global options like these.
 void toku_ft_set_direct_io(bool direct_io_on);
+void toku_ft_set_compress_buffers_before_eviction(bool compress_buffers);
+
 #endif

src/export.map

@@ -5,6 +5,7 @@
    db_strerror;
    db_version;
    db_env_set_direct_io;
+   db_env_set_compress_buffers_before_eviction;
    db_env_set_func_fsync;
    db_env_set_func_malloc;
    db_env_set_func_realloc;

src/ydb_env_func.cc

@@ -117,6 +117,10 @@ void db_env_set_direct_io (bool direct_io_on) {
     toku_ft_set_direct_io(direct_io_on);
 }
 
+void db_env_set_compress_buffers_before_eviction (bool compress_buffers) {
+    toku_ft_set_compress_buffers_before_eviction(compress_buffers);
+}
+
 void db_env_set_func_fsync (int (*fsync_function)(int)) {
     toku_set_func_fsync(fsync_function);
 }