[t:4541], merge new fix to main

git-svn-id: file:///svn/toku/tokudb@41583 c7de825b-a66e-492c-adef-691d508d4ae1

buildheader/make_tdb.c

@@ -203,6 +203,7 @@ static void print_defines (void) {
     //printf("#define DB_PRELOCKED_FILE_READ 0x00200000\n"); // private tokudb. No longer supported in #4472
     printf("#define DB_IS_HOT_INDEX 0x00100000\n"); // private tokudb
     printf("#define DBC_DISABLE_PREFETCHING 0x20000000\n"); // private tokudb
+    printf("#define DB_UPDATE_CMP_DESCRIPTOR 0x40000000\n"); // private tokudb
 
     {
         //dbt flags
@@ -424,7 +425,6 @@ static void print_db_struct (void) {
 			 "DESCRIPTOR descriptor /* saved row/dictionary descriptor for aiding in comparisons */",
 			 "DESCRIPTOR cmp_descriptor /* saved row/dictionary descriptor for aiding in comparisons */",
 			 "int (*change_descriptor) (DB*, DB_TXN*, const DBT* descriptor, u_int32_t) /* change row/dictionary descriptor for a db.  Available only while db is open */",
-			 "int (*update_cmp_descriptor) (DB*) /* Update cmp descriptor.  Available only while db is open */",
 			 "int (*getf_set)(DB*, DB_TXN*, u_int32_t, DBT*, YDB_CALLBACK_FUNCTION, void*) /* same as DBC->c_getf_set without a persistent cursor) */",
 			 "int (*optimize)(DB*) /* Run garbage collecion and promote all transactions older than oldest. Amortized (happens during flattening) */",
 			 "int (*hot_optimize)(DB*, int (*progress_callback)(void *progress_extra, float progress), void *progress_extra)",

include/db.h

@@ -156,6 +156,7 @@ typedef enum {
 #define DB_PRELOCKED_WRITE 0x00400000
 #define DB_IS_HOT_INDEX 0x00100000
 #define DBC_DISABLE_PREFETCHING 0x20000000
+#define DB_UPDATE_CMP_DESCRIPTOR 0x40000000
 #define DB_DBT_APPMALLOC 1
 #define DB_DBT_DUPOK 2
 #define DB_DBT_MALLOC 8
@@ -305,7 +306,6 @@ struct __toku_db {
   DESCRIPTOR descriptor /* saved row/dictionary descriptor for aiding in comparisons */;
   DESCRIPTOR cmp_descriptor /* saved row/dictionary descriptor for aiding in comparisons */;
   int (*change_descriptor) (DB*, DB_TXN*, const DBT* descriptor, u_int32_t) /* change row/dictionary descriptor for a db.  Available only while db is open */;
-  int (*update_cmp_descriptor) (DB*) /* Update cmp descriptor.  Available only while db is open */;
   int (*getf_set)(DB*, DB_TXN*, u_int32_t, DBT*, YDB_CALLBACK_FUNCTION, void*) /* same as DBC->c_getf_set without a persistent cursor) */;
   int (*optimize)(DB*) /* Run garbage collecion and promote all transactions older than oldest. Amortized (happens during flattening) */;
   int (*hot_optimize)(DB*, int (*progress_callback)(void *progress_extra, float progress), void *progress_extra);

newbrt/brt.c

@@ -3504,8 +3504,8 @@ cleanup:
     return r;
 }
 
-void 
-toku_brt_update_cmp_descriptor(BRT t) {
+static void 
+brt_update_cmp_descriptor(BRT t) {
     t->h->cmp_descriptor.dbt.size = t->h->descriptor.dbt.size;
     t->h->cmp_descriptor.dbt.data = toku_xmemdup(
         t->h->descriptor.dbt.data, 
@@ -3520,7 +3520,8 @@ toku_brt_change_descriptor(
     const DBT* old_descriptor, 
     const DBT* new_descriptor, 
     BOOL do_log, 
-    TOKUTXN txn
+    TOKUTXN txn,
+    BOOL update_cmp_descriptor
     ) 
 {
     int r = 0;
@@ -3543,27 +3544,33 @@ toku_brt_change_descriptor(
     if (r != 0) { goto cleanup; }
 
     if (do_log) {
-	TOKULOGGER logger = toku_txn_logger(txn);
-	TXNID xid = toku_txn_get_txnid(txn);
-	r = toku_log_change_fdescriptor(
-	    logger, NULL, 0, 
-	    toku_cachefile_filenum(t->cf),
-	    xid,
-	    old_desc_bs,
-	    new_desc_bs
-	    );
-	if (r != 0) { goto cleanup; }
+        TOKULOGGER logger = toku_txn_logger(txn);
+        TXNID xid = toku_txn_get_txnid(txn);
+        r = toku_log_change_fdescriptor(
+            logger, NULL, 0, 
+            toku_cachefile_filenum(t->cf),
+            xid,
+            old_desc_bs,
+            new_desc_bs,
+            update_cmp_descriptor
+            );
+        if (r != 0) { goto cleanup; }
     }
 
     // write new_descriptor to header
     new_d.dbt = *new_descriptor;
     fd = toku_cachefile_get_and_pin_fd (t->cf);
     r = toku_update_descriptor(t->h, &new_d, fd);
-    if (r == 0)	 // very infrequent operation, worth precise threadsafe count
-	STATUS_VALUE(BRT_DESCRIPTOR_SET)++;
+    // very infrequent operation, worth precise threadsafe count
+    if (r == 0) {
+        STATUS_VALUE(BRT_DESCRIPTOR_SET)++;
+    }
     toku_cachefile_unpin_fd(t->cf);
     if (r!=0) goto cleanup;
 
+    if (update_cmp_descriptor) {
+        brt_update_cmp_descriptor(t);
+    }
 cleanup:
     return r;
 }

newbrt/brt.h

@@ -32,8 +32,7 @@ C_BEGIN
 typedef int(*BRT_GET_CALLBACK_FUNCTION)(ITEMLEN keylen, bytevec key, ITEMLEN vallen, bytevec val, void *extra, bool lock_only);
 
 int toku_open_brt (const char *fname, int is_create, BRT *, int nodesize, int basementnodesize, CACHETABLE, TOKUTXN, int(*)(DB *,const DBT*,const DBT*), DB*) __attribute__ ((warn_unused_result));
-void toku_brt_update_cmp_descriptor(BRT t);
-int toku_brt_change_descriptor(BRT t, const DBT* old_descriptor, const DBT* new_descriptor, BOOL do_log, TOKUTXN txn);
+int toku_brt_change_descriptor(BRT t, const DBT* old_descriptor, const DBT* new_descriptor, BOOL do_log, TOKUTXN txn, BOOL update_cmp_descriptor);
 int toku_update_descriptor(struct brt_header * h, DESCRIPTOR d, int fd);
 // Note: See the locking discussion in brt.c for toku_brt_change_descriptor and toku_update_descriptor.
 

newbrt/logformat.c

@@ -210,6 +210,7 @@ const struct logtype logtypes[] = {
                             {"TXNID",      "xid", 0},
                             {"BYTESTRING", "old_descriptor", 0},
                             {"BYTESTRING", "new_descriptor", 0},
+                            {"BOOL",       "update_cmp_descriptor", 0},
                             NULLFIELD}},
     {0,0,FA{NULLFIELD}}
 };

newbrt/recover.c

@@ -869,7 +869,8 @@ static int toku_recover_change_fdescriptor (struct logtype_change_fdescriptor *l
             &old_descriptor, 
             &new_descriptor, 
             FALSE, 
-            txn
+            txn,
+            l->update_cmp_descriptor
             );
         assert(r==0);
     }    

src/tests/recover-compare-db-descriptor.c

@@ -30,7 +30,7 @@ change_descriptor(DB_ENV* env, DB* db) {
     DBT descriptor;
     dbt_init(&descriptor, descriptor_contents, sizeof(descriptor_contents));
     IN_TXN_COMMIT(env, NULL, txn_desc, 0, {
-        CHK(db->change_descriptor(db, txn_desc, &descriptor, 0));
+        CHK(db->change_descriptor(db, txn_desc, &descriptor, DB_UPDATE_CMP_DESCRIPTOR));
       });
 #endif
 }
@@ -52,12 +52,10 @@ do_x1_shutdown (BOOL do_commit, BOOL do_abort) {
     r = db_create(&dba, env, 0);                                                        CKERR(r);
     r = dba->open(dba, NULL, namea, NULL, DB_BTREE, DB_AUTO_COMMIT|DB_CREATE, 0666);    CKERR(r);
     change_descriptor(env, dba);
-    dba->update_cmp_descriptor(dba);
     r = db_create(&dbb, env, 0);                                                        CKERR(r);
     r = dbb->open(dbb, NULL, nameb, NULL, DB_BTREE, DB_AUTO_COMMIT|DB_CREATE, 0666);    CKERR(r);
     DB_TXN *txn;
     change_descriptor(env, dbb);
-    dbb->update_cmp_descriptor(dbb);
     r = env->txn_begin(env, NULL, &txn, 0);                                             CKERR(r);
     {
 	DBT a={.data="a", .size=2};

src/tests/test_4015.c

@@ -116,8 +116,7 @@ int test_main(int argc, char * const argv[]) {
     DBT desc;
     dbt_init(&desc, "foo", sizeof("foo"));
     IN_TXN_COMMIT(env, NULL, txn, 0,
-		  CHK(db->change_descriptor(db, txn, &desc, 0)));
-    CHK(db->update_cmp_descriptor(db));
+		  CHK(db->change_descriptor(db, txn, &desc, DB_UPDATE_CMP_DESCRIPTOR)));
     pthread_t thd;
     CHK(toku_pthread_create(&thd, NULL, startA, NULL));
 

src/tests/test_cmp_descriptor.c

@@ -157,10 +157,8 @@ static void run_test(void) {
             CHK(db->open(db, txn_create, "foo.db", NULL, DB_BTREE, DB_CREATE, 0666));
             assert(db->descriptor->dbt.size == 0);
             assert(db->cmp_descriptor->dbt.size == 0);
-            CHK(db->change_descriptor(db, txn_create, &orig_desc, 0));
+            CHK(db->change_descriptor(db, txn_create, &orig_desc, DB_UPDATE_CMP_DESCRIPTOR));
             assert_desc_four(db);
-            assert(db->cmp_descriptor->dbt.size == 0);
-            CHK(db->update_cmp_descriptor(db));
             assert_cmp_desc_valid(db);
             r = env->create_loader(env, txn_create, &loader, db, 1, &db, NULL, NULL, 0); 
             CKERR(r);

src/ydb_db.c

@@ -499,12 +499,6 @@ toku_db_rename(DB * db, const char *fname, const char *dbname, const char *newna
     return r;
 }
 
-static int
-toku_db_update_cmp_descriptor(DB *db) {
-    toku_brt_update_cmp_descriptor(db->i->brt);
-    return 0;
-}
-
 //
 // This function is the only way to set a descriptor of a DB.
 //
@@ -516,6 +510,7 @@ toku_db_change_descriptor(DB *db, DB_TXN* txn, const DBT* descriptor, u_int32_t
     TOKUTXN ttxn = txn ? db_txn_struct_i(txn)->tokutxn : NULL;
     DBT old_descriptor;
     BOOL is_db_hot_index  = ((flags & DB_IS_HOT_INDEX) != 0);
+    BOOL update_cmp_descriptor = ((flags & DB_UPDATE_CMP_DESCRIPTOR) != 0);
 
     toku_init_dbt(&old_descriptor);
     if (!db_opened(db) || !txn || !descriptor || (descriptor->size>0 && !descriptor->data)){
@@ -533,7 +528,14 @@ toku_db_change_descriptor(DB *db, DB_TXN* txn, const DBT* descriptor, u_int32_t
     
     old_descriptor.size = db->descriptor->dbt.size;
     old_descriptor.data = toku_memdup(db->descriptor->dbt.data, db->descriptor->dbt.size);
-    r = toku_brt_change_descriptor(db->i->brt, &old_descriptor, descriptor, TRUE, ttxn);
+    r = toku_brt_change_descriptor(
+        db->i->brt, 
+        &old_descriptor, 
+        descriptor, 
+        TRUE, 
+        ttxn, 
+        update_cmp_descriptor
+        );
 cleanup:
     if (old_descriptor.data) toku_free(old_descriptor.data);
     return r;
@@ -1077,7 +1079,6 @@ toku_db_create(DB ** db, DB_ENV * env, u_int32_t flags) {
 #undef SDB
     // methods that take the ydb lock in some capacity,
     // but not from beginning to end
-    result->update_cmp_descriptor = toku_db_update_cmp_descriptor;
     result->del = autotxn_db_del;
     result->put = autotxn_db_put;
     result->update = autotxn_db_update;