merge tokudb.1844 to main close[t:1844]

git-svn-id: file:///svn/toku/tokudb@14180 c7de825b-a66e-492c-adef-691d508d4ae1

buildheader/db.h_4_6

@@ -12,9 +12,9 @@ extern "C" {
 #define TOKUDB_NATIVE_H 0
 #define DB_VERSION_MAJOR 4
 #define DB_VERSION_MINOR 6
-#define DB_VERSION_PATCH 19
+#define DB_VERSION_PATCH 21
 #ifndef _TOKUDB_WRAP_H
-#define DB_VERSION_STRING "Tokutek: TokuDB 4.6.19"
+#define DB_VERSION_STRING "Tokutek: TokuDB 4.6.21"
 #else
 #define DB_VERSION_STRING_ydb "Tokutek: TokuDB (wrapped bdb)"
 #endif
@@ -63,6 +63,7 @@ typedef enum {
 #define DB_PRIVATE 8388608
 #define DB_RDONLY 32
 #define DB_RECOVER 64
+#define DB_RUNRECOVERY -30975
 #define DB_THREAD 128
 #define DB_TXN_NOSYNC 512
 #define DB_LOCK_DEFAULT 1

buildheader/make_db_h.c

@@ -60,6 +60,7 @@ void print_defines (void) {
     dodefine(DB_PRIVATE);
     dodefine(DB_RDONLY);
     dodefine(DB_RECOVER);
+    dodefine(DB_RUNRECOVERY);
     dodefine(DB_THREAD);
     dodefine(DB_TXN_NOSYNC);
 

newbrt/brt.c

@@ -213,6 +213,9 @@ brtnode_put_cmd (BRT t, BRTNODE node, BRT_CMD cmd, enum reactivity *re, BOOL *di
 static int
 flush_this_child (BRT t, BRTNODE node, int childnum, enum reactivity *child_re, BOOL *did_io);
 
+static void brt_verify_flags(BRT brt, BRTNODE node) {
+    assert(brt->flags == node->flags);
+}
 
 int toku_brt_debug_mode = 0;
 
@@ -2502,10 +2505,9 @@ int toku_brt_root_put_cmd(BRT brt, BRT_CMD cmd, TOKULOGGER logger)
     node=node_v;
 
     VERIFY_NODE(brt, node);
-
     assert(node->fullhash==fullhash);
+    brt_verify_flags(brt, node);
 
-    VERIFY_NODE(brt, node);
     verify_local_fingerprint_nonleaf(node);
     if ((r = push_something_at_root(brt, &node, rootp, cmd, logger))) {
 	toku_unpin_brtnode(brt, node); // ignore any error code on the unpin.

newbrt/recover.c

@@ -13,6 +13,8 @@
 #include "log_header.h"
 #include "varray.h"
 
+static int toku_recover_trace = 0;
+
 //#define DO_VERIFY_COUNTS
 #ifdef DO_VERIFY_COUNTS
 #define VERIFY_COUNTS(n) toku_verify_counts(n)
@@ -41,6 +43,8 @@ int toku_recover_init (void) {
     assert(r == 0);
     toku_logger_write_log_files(recover_logger, FALSE);
     toku_logger_set_cachetable(recover_logger, recover_ct);
+    if (toku_recover_trace)
+        printf("%s:%d\n", __FUNCTION__, __LINE__);
     return r;
 }
 
@@ -71,6 +75,9 @@ void toku_recover_cleanup (void) {
     
     r = toku_cachetable_close(&recover_ct);
     assert(r == 0);
+
+    if (toku_recover_trace)
+        printf("%s:%d\n", __FUNCTION__, __LINE__);
 }
 
 // Null function supplied to transaction commit and abort
@@ -492,11 +499,10 @@ static int toku_delete_rolltmp_files (const char *log_dir) {
     return result;
 }
 
-// Does this log environment need recovery?
-// Effects: If there are no log files, or if there is a "null" checkpoint at the end of the log,
+// Effects: If there are no log files, or if there is a "clean" checkpoint at the end of the log,
 // then we don't need recovery to run.  We skip the optional shutdown log entry.
 // Returns: TRUE if we need recovery, FALSE if we do not need recovery.
-static int tokudb_needs_recovery(const char *log_dir) {
+int tokudb_needs_recovery(const char *log_dir) {
     int needs_recovery;
     int r;
     TOKULOGCURSOR logcursor = NULL;
@@ -569,8 +575,6 @@ static void abort_live_txn(void *v, void *UU(extra)) {
     toku_txn_close_txn(txn);
 }
 
-static const int toku_recover_trace = 0;
-
 static int really_do_recovery(const char *data_dir, const char *log_dir) {
     int r;
 
@@ -602,12 +606,14 @@ static int really_do_recovery(const char *data_dir, const char *log_dir) {
     while (1) {
         le = NULL;
         r = toku_logcursor_prev(logcursor, &le);
-        if (toku_recover_trace) printf("%s:%d r=%d cmd=%c\n", __FUNCTION__, __LINE__, r, le ? le->cmd : '?');
+        if (toku_recover_trace) 
+            printf("%s:%d r=%d cmd=%c\n", __FUNCTION__, __LINE__, r, le ? le->cmd : '?');
         if (r != 0)
             break;
         logtype_dispatch_assign(le, toku_recover_backward_, r, &bs);
         if (r != 0) {
-            if (toku_recover_trace) printf("%s:%d r=%d cmd=%c\n", __FUNCTION__, __LINE__, r, le ? le->cmd : '?');
+            if (toku_recover_trace) 
+                printf("%s:%d r=%d cmd=%c\n", __FUNCTION__, __LINE__, r, le ? le->cmd : '?');
             logtype_dispatch_args(le, toku_recover_);
             break;
         }
@@ -617,7 +623,8 @@ static int really_do_recovery(const char *data_dir, const char *log_dir) {
     while (1) {
         le = NULL;
         r = toku_logcursor_next(logcursor, &le);
-        if (toku_recover_trace) printf("%s:%d r=%d cmd=%c\n", __FUNCTION__, __LINE__, r, le ? le->cmd : '?');
+        if (toku_recover_trace) 
+            printf("%s:%d r=%d cmd=%c\n", __FUNCTION__, __LINE__, r, le ? le->cmd : '?');
         if (r != 0)
             break;
         logtype_dispatch_args(le, toku_recover_);

newbrt/recover.h

@@ -15,7 +15,15 @@
 #include "x1764.h"
 
 int toku_recover_init(void);
+
 void toku_recover_cleanup(void);
+
 int tokudb_recover(const char *datadir, const char *logdir);
 
+// Effect: Check the tokudb logs to determine whether or not we need to run recovery.
+// If the log is empty or if there is a clean shutdown at the end of the log, then we
+// dont need to run recovery.
+// Returns: TRUE if we need recovery, otherwise FALSE.
+int tokudb_needs_recovery(const char *logdir);
+
 #endif // TOKURECOVER_H

src/tests/Makefile

@@ -143,6 +143,9 @@ TDB_TESTS_THAT_SHOULD_FAIL= \
 	test_groupcommit_count \
 	test944 \
 	test_truncate_txn_abort \
+	test_log8 \
+	test_log9 \
+	test_log10 \
 #\ ends prev line
 ifneq ($(OS_CHOICE),windows)
     TDB_TESTS_THAT_SHOULD_FAIL+= \

src/tests/checkpoint_1.c

@@ -23,7 +23,7 @@ checkpoint_test_1(u_int32_t flags, u_int32_t n, int snap_all) {
         fflush(stdout); 
     }
     dir_create();
-    env_startup(0);
+    env_startup(0, FALSE);
     int run;
     int r;
     DICTIONARY_S db_control;
@@ -61,7 +61,7 @@ checkpoint_test_2(u_int32_t flags, u_int32_t n) {
         fflush(stdout); 
     }
     dir_create();
-    env_startup(0);
+    env_startup(0, FALSE);
     int run;
     int r;
     DICTIONARY_S db_control;

src/tests/checkpoint_stress.c

@@ -39,6 +39,7 @@ Each iteration does:
 #define NUM_DICTIONARIES 4       // any more than 3 is overkill to exercise linked list logic
 
 static int oper_per_iter = 5001;  // not-very-nice odd number (not a multiple of a power of two)
+static int do_log_recover = 0;
 
 static toku_pthread_t thread;
 
@@ -219,7 +220,7 @@ run_test (int iter, int die) {
     if (verbose)
 	printf("checkpoint_stress: iter = %d, cachesize (bytes) = 0x%08"PRIx64"\n", iter, cachebytes);
     
-    env_startup(cachebytes);
+    env_startup(cachebytes, do_log_recover);
 
     // create array of dictionaries
     // for each dictionary verify previous iterations and perform new inserts
@@ -265,8 +266,9 @@ run_test (int iter, int die) {
 
 static void
 usage(char *progname) {
-    fprintf(stderr, "Usage:\n%s [-i n] [-q|-v]\n"
-                    "      \n%s [-h]\n", progname, progname);
+    fprintf(stderr, "Usage:\n%s [-c] [-C] [-i N] [-n N] [-l] [-q|-v]\n"
+                    "      \n%s [-h]\n", progname, 
+                                         progname);
 }
 
 
@@ -279,7 +281,7 @@ test_main (int argc, char *argv[]) {
 
     int c;
     int crash = 0;
-    while ((c = getopt(argc, argv, "cChi:qvn:")) != -1) {
+    while ((c = getopt(argc, argv, "cChi:qvn:l:")) != -1) {
 	switch(c) {
         case 'c':
             crash = 1;
@@ -293,6 +295,9 @@ test_main (int argc, char *argv[]) {
 	case 'n':
 	    oper_per_iter = atoi(optarg);
 	    break;
+        case 'l':
+            do_log_recover = 1;
+            break;
 	case 'v':
 	    verbose++;
             break;

src/tests/checkpoint_test.h

@@ -96,7 +96,7 @@ dir_create(void) {
 
 // pass in zeroes for default cachesize
 static void  UU()
-env_startup(int64_t bytes) {
+    env_startup(int64_t bytes, BOOL do_log_recover) {
     int r;
     r = db_env_create(&env, 0);
         CKERR(r);
@@ -108,7 +108,10 @@ env_startup(int64_t bytes) {
 	r = env->set_cachesize(env, bytes >> 30, bytes % (1<<30), 1);
         CKERR(r);
     }
-    r = env->open(env, ENVDIR, DB_INIT_LOCK|DB_INIT_LOG|DB_INIT_MPOOL|DB_INIT_TXN|DB_CREATE|DB_PRIVATE, S_IRWXU+S_IRWXG+S_IRWXO);
+    int envflags = DB_INIT_LOCK | DB_INIT_MPOOL | DB_INIT_TXN | DB_CREATE | DB_PRIVATE;
+    if (do_log_recover)
+        envflags += DB_INIT_LOG | DB_RECOVER;
+    r = env->open(env, ENVDIR, envflags, S_IRWXU+S_IRWXG+S_IRWXO);
         CKERR(r);
     env->set_errfile(env, stderr);
     r = env->checkpointing_set_period(env, 0); //Disable auto-checkpointing.

src/tests/checkpoint_truncate_1.c

@@ -42,7 +42,7 @@ checkpoint_truncate_test(u_int32_t flags, u_int32_t n) {
         fflush(stdout); 
     }
     dir_create();
-    env_startup(0);
+    env_startup(0, FALSE);
 
     DICTIONARY_S db_control;
     init_dictionary(&db_control, flags, "control");

src/tests/recover-fassociate-dupsort.c

+#include <sys/stat.h>
+#include <sys/wait.h>
+#include "test.h"
+
+const int envflags = DB_INIT_MPOOL|DB_CREATE|DB_THREAD |DB_INIT_LOCK|DB_INIT_LOG|DB_INIT_TXN;
+
+char *namea="a.db";
+char *nameb="b.db";
+
+static void
+run_test (void) {
+    int r;
+    system("rm -rf " ENVDIR);
+    toku_os_mkdir(ENVDIR, S_IRWXU+S_IRWXG+S_IRWXO);
+    DB_ENV *env;
+    DB *dba, *dbb;
+    r = db_env_create(&env, 0);                                                         CKERR(r);
+    r = env->open(env, ENVDIR, envflags, S_IRWXU+S_IRWXG+S_IRWXO);                      CKERR(r);
+
+    r = db_create(&dba, env, 0);                                                        CKERR(r);
+    r = dba->set_flags(dba, DB_DUPSORT);                                                CKERR(r);
+    r = dba->open(dba, NULL, namea, NULL, DB_BTREE, DB_AUTO_COMMIT|DB_CREATE, 0666);    CKERR(r);
+
+    r = db_create(&dbb, env, 0);                                                        CKERR(r);
+    r = dbb->set_flags(dbb, DB_DUPSORT);                                                CKERR(r);
+    r = dbb->open(dbb, NULL, nameb, NULL, DB_BTREE, DB_AUTO_COMMIT|DB_CREATE, 0666);    CKERR(r);
+
+    DB_TXN *txn;
+    r = env->txn_begin(env, NULL, &txn, 0);                                             CKERR(r);
+    {
+	DBT a={.data="a", .size=2};
+	DBT b={.data="b", .size=2};
+	r = dba->put(dba, txn, &a, &b, DB_YESOVERWRITE);                                CKERR(r);
+	r = env->txn_checkpoint(env, 0, 0, 0);                                          CKERR(r);
+	r = dbb->put(dbb, txn, &b, &a, DB_YESOVERWRITE);                                CKERR(r);
+    }
+
+    r = txn->commit(txn, 0);                                                            CKERR(r);
+
+    abort();
+}
+
+static void
+run_recover (void) {
+    DB_ENV *env;
+    int r;
+
+    r = db_env_create(&env, 0);                                                             CKERR(r);
+    r = env->open(env, ENVDIR, envflags + DB_RECOVER, S_IRWXU+S_IRWXG+S_IRWXO);             CKERR(r);
+    r = env->close(env, 0);                                                                 CKERR(r);
+    exit(0);
+}
+
+static void
+run_no_recover (void) {
+    DB_ENV *env;
+    int r;
+
+    r = db_env_create(&env, 0);                                                             CKERR(r);
+    r = env->open(env, ENVDIR, envflags & ~DB_RECOVER, S_IRWXU+S_IRWXG+S_IRWXO);            CKERR(r);
+    r = env->close(env, 0);                                                                 CKERR(r);
+    exit(0);
+}
+
+const char *cmd;
+
+BOOL do_commit=FALSE, do_abort=FALSE, do_explicit_abort=FALSE, do_recover_committed=FALSE,  do_recover_aborted=FALSE, do_recover_only=FALSE, do_no_recover = FALSE;
+
+static void test_parse_args (int argc, char *argv[]) {
+    int resultcode;
+    cmd = argv[0];
+    argc--; argv++;
+    while (argc>0) {
+	if (strcmp(argv[0], "-v") == 0) {
+	    verbose++;
+	} else if (strcmp(argv[0],"-q")==0) {
+	    verbose--;
+	    if (verbose<0) verbose=0;
+	} else if (strcmp(argv[0], "--commit")==0) {
+	    do_commit=TRUE;
+	} else if (strcmp(argv[0], "--abort")==0) {
+	    do_abort=TRUE;
+	} else if (strcmp(argv[0], "--explicit-abort")==0) {
+	    do_explicit_abort=TRUE;
+	} else if (strcmp(argv[0], "--recover-committed")==0) {
+	    do_recover_committed=TRUE;
+	} else if (strcmp(argv[0], "--recover-aborted")==0) {
+	    do_recover_aborted=TRUE;
+        } else if (strcmp(argv[0], "--recover-only") == 0) {
+            do_recover_only=TRUE;
+        } else if (strcmp(argv[0], "--no-recover") == 0) {
+            do_no_recover=TRUE;
+	} else if (strcmp(argv[0], "-h")==0) {
+	    resultcode=0;
+	do_usage:
+	    fprintf(stderr, "Usage:\n%s [-v|-q]* [-h] {--commit | --abort | --explicit-abort | --recover-committed | --recover-aborted } \n", cmd);
+	    exit(resultcode);
+	} else {
+	    fprintf(stderr, "Unknown arg: %s\n", argv[0]);
+	    resultcode=1;
+	    goto do_usage;
+	}
+	argc--;
+	argv++;
+    }
+    {
+	int n_specified=0;
+	if (do_commit)            n_specified++;
+	if (do_abort)             n_specified++;
+	if (do_explicit_abort)    n_specified++;
+	if (do_recover_committed) n_specified++;
+	if (do_recover_aborted)   n_specified++;
+	if (do_recover_only)      n_specified++;
+	if (do_no_recover)        n_specified++;
+	if (n_specified>1) {
+	    printf("Specify only one of --commit or --abort or --recover-committed or --recover-aborted\n");
+	    resultcode=1;
+	    goto do_usage;
+	}
+    }
+}
+
+int
+test_main (int argc, char *argv[]) {
+    test_parse_args(argc, argv);
+    if (do_commit) {
+	run_test();
+    } else if (do_recover_only) {
+        run_recover();
+    } else if (do_no_recover) {
+        run_no_recover();
+    } 
+    return 0;
+}

src/tests/recover-fassociate-nodup.c

+#include <sys/stat.h>
+#include <sys/wait.h>
+#include "test.h"
+
+const int envflags = DB_INIT_MPOOL|DB_CREATE|DB_THREAD |DB_INIT_LOCK|DB_INIT_LOG|DB_INIT_TXN;
+
+char *namea="a.db";
+char *nameb="b.db";
+
+static void
+run_test (void) {
+    int r;
+    system("rm -rf " ENVDIR);
+    toku_os_mkdir(ENVDIR, S_IRWXU+S_IRWXG+S_IRWXO);
+    DB_ENV *env;
+    DB *dba, *dbb;
+    r = db_env_create(&env, 0);                                                         CKERR(r);
+    r = env->open(env, ENVDIR, envflags, S_IRWXU+S_IRWXG+S_IRWXO);                      CKERR(r);
+
+    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);
+
+    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;
+    r = env->txn_begin(env, NULL, &txn, 0);                                             CKERR(r);
+    {
+	DBT a={.data="a", .size=2};
+	DBT b={.data="b", .size=2};
+	r = dba->put(dba, txn, &a, &b, DB_YESOVERWRITE);                                CKERR(r);
+	r = env->txn_checkpoint(env, 0, 0, 0);                                          CKERR(r);
+	r = dbb->put(dbb, txn, &b, &a, DB_YESOVERWRITE);                                CKERR(r);
+    }
+
+    r = txn->commit(txn, 0);                                                            CKERR(r);
+
+    abort();
+}
+
+static void
+run_recover (void) {
+    DB_ENV *env;
+    int r;
+
+    r = db_env_create(&env, 0);                                                             CKERR(r);
+    r = env->open(env, ENVDIR, envflags + DB_RECOVER, S_IRWXU+S_IRWXG+S_IRWXO);             CKERR(r);
+    r = env->close(env, 0);                                                                 CKERR(r);
+    exit(0);
+}
+
+static void
+run_no_recover (void) {
+    DB_ENV *env;
+    int r;
+
+    r = db_env_create(&env, 0);                                                             CKERR(r);
+    r = env->open(env, ENVDIR, envflags & ~DB_RECOVER, S_IRWXU+S_IRWXG+S_IRWXO);            CKERR(r);
+    r = env->close(env, 0);                                                                 CKERR(r);
+    exit(0);
+}
+
+const char *cmd;
+
+BOOL do_commit=FALSE, do_abort=FALSE, do_explicit_abort=FALSE, do_recover_committed=FALSE,  do_recover_aborted=FALSE, do_recover_only=FALSE, do_no_recover = FALSE;
+
+static void test_parse_args (int argc, char *argv[]) {
+    int resultcode;
+    cmd = argv[0];
+    argc--; argv++;
+    while (argc>0) {
+	if (strcmp(argv[0], "-v") == 0) {
+	    verbose++;
+	} else if (strcmp(argv[0],"-q")==0) {
+	    verbose--;
+	    if (verbose<0) verbose=0;
+	} else if (strcmp(argv[0], "--commit")==0) {
+	    do_commit=TRUE;
+	} else if (strcmp(argv[0], "--abort")==0) {
+	    do_abort=TRUE;
+	} else if (strcmp(argv[0], "--explicit-abort")==0) {
+	    do_explicit_abort=TRUE;
+	} else if (strcmp(argv[0], "--recover-committed")==0) {
+	    do_recover_committed=TRUE;
+	} else if (strcmp(argv[0], "--recover-aborted")==0) {
+	    do_recover_aborted=TRUE;
+        } else if (strcmp(argv[0], "--recover-only") == 0) {
+            do_recover_only=TRUE;
+        } else if (strcmp(argv[0], "--no-recover") == 0) {
+            do_no_recover=TRUE;
+	} else if (strcmp(argv[0], "-h")==0) {
+	    resultcode=0;
+	do_usage:
+	    fprintf(stderr, "Usage:\n%s [-v|-q]* [-h] {--commit | --abort | --explicit-abort | --recover-committed | --recover-aborted } \n", cmd);
+	    exit(resultcode);
+	} else {
+	    fprintf(stderr, "Unknown arg: %s\n", argv[0]);
+	    resultcode=1;
+	    goto do_usage;
+	}
+	argc--;
+	argv++;
+    }
+    {
+	int n_specified=0;
+	if (do_commit)            n_specified++;
+	if (do_abort)             n_specified++;
+	if (do_explicit_abort)    n_specified++;
+	if (do_recover_committed) n_specified++;
+	if (do_recover_aborted)   n_specified++;
+	if (do_recover_only)      n_specified++;
+	if (do_no_recover)        n_specified++;
+	if (n_specified>1) {
+	    printf("Specify only one of --commit or --abort or --recover-committed or --recover-aborted\n");
+	    resultcode=1;
+	    goto do_usage;
+	}
+    }
+}
+
+int
+test_main (int argc, char *argv[]) {
+    test_parse_args(argc, argv);
+    if (do_commit) {
+	run_test();
+    } else if (do_recover_only) {
+        run_recover();
+    } else if (do_no_recover) {
+        run_no_recover();
+    } 
+    return 0;
+}

src/tests/x1.c

@@ -120,6 +120,17 @@ do_x1_recover_only (void) {
     exit(0);
 }
 
+static void
+do_x1_no_recover (void) {
+    DB_ENV *env;
+    int r;
+
+    r = db_env_create(&env, 0);                                                             CKERR(r);
+    r = env->open(env, ENVDIR, envflags & ~DB_RECOVER, S_IRWXU+S_IRWXG+S_IRWXO);            CKERR(r);
+    r = env->close(env, 0);                                                                 CKERR(r);
+    exit(0);
+}
+
 const char *cmd;
 
 static void
@@ -162,7 +173,7 @@ do_test (void) {
     do_test_internal(FALSE);
 }
 
-BOOL do_commit=FALSE, do_abort=FALSE, do_explicit_abort=FALSE, do_recover_committed=FALSE,  do_recover_aborted=FALSE, do_recover_only=FALSE;
+BOOL do_commit=FALSE, do_abort=FALSE, do_explicit_abort=FALSE, do_recover_committed=FALSE,  do_recover_aborted=FALSE, do_recover_only=FALSE, do_no_recover = FALSE;
 
 static void
 x1_parse_args (int argc, char *argv[]) {
@@ -176,17 +187,19 @@ x1_parse_args (int argc, char *argv[]) {
 	    verbose--;
 	    if (verbose<0) verbose=0;
 	} else if (strcmp(argv[0], "--commit")==0) {
-	    do_commit=1;
+	    do_commit=TRUE;
 	} else if (strcmp(argv[0], "--abort")==0) {
-	    do_abort=1;
+	    do_abort=TRUE;
 	} else if (strcmp(argv[0], "--explicit-abort")==0) {
-	    do_explicit_abort=1;
+	    do_explicit_abort=TRUE;
 	} else if (strcmp(argv[0], "--recover-committed")==0) {
-	    do_recover_committed=1;
+	    do_recover_committed=TRUE;
 	} else if (strcmp(argv[0], "--recover-aborted")==0) {
-	    do_recover_aborted=1;
+	    do_recover_aborted=TRUE;
         } else if (strcmp(argv[0], "--recover-only") == 0) {
-            do_recover_only=1;
+            do_recover_only=TRUE;
+        } else if (strcmp(argv[0], "--no-recover") == 0) {
+            do_no_recover=TRUE;
 	} else if (strcmp(argv[0], "-h")==0) {
 	    resultcode=0;
 	do_usage:
@@ -208,6 +221,7 @@ x1_parse_args (int argc, char *argv[]) {
 	if (do_recover_committed) n_specified++;
 	if (do_recover_aborted)   n_specified++;
 	if (do_recover_only)      n_specified++;
+	if (do_no_recover)        n_specified++;
 	if (n_specified>1) {
 	    printf("Specify only one of --commit or --abort or --recover-committed or --recover-aborted\n");
 	    resultcode=1;
@@ -232,6 +246,8 @@ test_main (int argc, char *argv[])
 	do_x1_recover(FALSE);
     } else if (do_recover_only) {
         do_x1_recover_only();
+    } else if (do_no_recover) {
+        do_x1_no_recover();
     } else {
 	do_test();
     }

src/ydb.c

@@ -302,29 +302,21 @@ static int do_recovery (DB_ENV *env) {
     } else {
 	logdir = toku_strdup(env->i->dir);
     }
-    
-#if 0
-    // want to do recovery in its own process
-    pid_t pid;
-    if ((pid=fork())==0) {
-	int r=tokudb_recover(datadir, logdir);
-	assert(r==0);
-	toku_free(logdir); // the child must also free.
-	exit(0);
-    }
-    int status;
-    waitpid(pid, &status, 0);
-    if (!WIFEXITED(status) || WEXITSTATUS(status)!=0)  {
-	toku_free(logdir);
-	return toku_ydb_do_error(env, -1, "Recovery failed\n");
-    }
-    toku_free(logdir);
-    return 0;
-#else
     int r = tokudb_recover(datadir, logdir);
     toku_free(logdir);
     return r;
-#endif
+}
+
+static int needs_recovery (DB_ENV *env) {
+    char *logdir;
+    if (env->i->lg_dir) {
+	logdir = construct_full_name(env->i->dir, env->i->lg_dir);
+    } else {
+	logdir = toku_strdup(env->i->dir);
+    }
+    BOOL recovery_needed = tokudb_needs_recovery(logdir);
+    toku_free(logdir);
+    return recovery_needed ? DB_RUNRECOVERY : 0;
 }
 
 static int toku_env_open(DB_ENV * env, const char *home, u_int32_t flags, int mode) {
@@ -396,8 +388,11 @@ static int toku_env_open(DB_ENV * env, const char *home, u_int32_t flags, int mo
     unused_flags &= ~DB_INIT_TXN & ~DB_INIT_LOG; 
 
     if (flags&DB_RECOVER) {
-	r=do_recovery(env);
-	if (r!=0) return r;
+	r = do_recovery(env);
+	if (r != 0) return r;
+    } else {
+        r = needs_recovery(env);
+        if (r != 0) return r;
     }
 
     if (flags & (DB_INIT_TXN | DB_INIT_LOG)) {