merge -c 14118 tokudb.1960 to main [t:1960]

git-svn-id: file:///svn/toku/tokudb@14119 c7de825b-a66e-492c-adef-691d508d4ae1

newbrt/recover.c

@@ -410,12 +410,13 @@ static int toku_recover_backward_xstillopen (struct logtype_xstillopen *l, struc
 static int toku_recover_xbegin (LSN lsn, TXNID parent_xid) {
     int r;
 
-    // TODO do something with the parent
-    assert(parent_xid == 0);
+    // lookup the parent
     TOKUTXN parent = NULL;
+    r = toku_txnid2txn(recover_logger, parent_xid, &parent);
+    assert(r == 0);
 
     // create a transaction and bind it to the transaction id
-    TOKUTXN txn;
+    TOKUTXN txn = NULL;
     r = toku_txn_begin_with_xid(parent, &txn, recover_logger, lsn.lsn);
     assert(r == 0);
     return 0;
@@ -538,6 +539,13 @@ static int tokudb_needs_recovery(const char *log_dir) {
     return needs_recovery;
 }
 
+// Sort the transactions in reverse order of transaction id
+static int compare_txn(const void *a, const void *b) {
+    TOKUTXN atxn = (TOKUTXN) * (void **) a;
+    TOKUTXN btxn = (TOKUTXN) * (void **) b;
+    return - (atxn->txnid64 - btxn->txnid64);
+}
+
 // Append a transaction to the set of live transactions
 static int append_live_txn(OMTVALUE v, u_int32_t UU(i), void *extra) {
     TOKUTXN txn = (TOKUTXN) v;
@@ -618,11 +626,12 @@ static int really_do_recovery(const char *data_dir, const char *log_dir) {
     // restart logging
     toku_logger_restart(recover_logger, lastlsn);
 
-    // abort all of the remaining live transactions
+    // abort all of the remaining live transactions in reverse transaction id order
     struct varray *live_txns = NULL;
     r = varray_create(&live_txns, 1);
     assert(r == 0);
     toku_omt_iterate(recover_logger->live_txns, append_live_txn, live_txns);
+    varray_sort(live_txns, compare_txn);
     varray_iterate(live_txns, abort_live_txn, NULL);
     varray_destroy(&live_txns);
 

newbrt/varray.c

@@ -50,9 +50,12 @@ int varray_append(struct varray *va, void *p) {
     va->a[va->c++] = p;
     return 0;
 }
+
+void varray_sort(struct varray *va, int (*compare)(const void *a, const void *b)) {
+    qsort(va->a, va->c, sizeof (void *), compare);
+}
                   
 void varray_iterate(struct varray *va, void (*f)(void *p, void *extra), void *extra) {
-    int i;
-    for (i=0; i<va->c; i++) 
+    for (int i = 0; i < va->c; i++) 
         f(va->a[i], extra);
 }

newbrt/varray.h

@@ -22,6 +22,9 @@ void varray_destroy(struct varray **vap);
 // element is the new element
 // Returns: 0 if success
 int varray_append(struct varray *va, void *p);
+
+// Sort the array elements
+void varray_sort(struct varray *va, int (*)(const void *, const void *));
                   
 // Apply a function to all of the elements in the array
 // Effect: The function f is called for every element in the array, with

src/tests/x1-nested2.c

+/* Transaction consistency:  
+ *  fork a process:
+ *   Open two tables, T1 and T2
+ *   begin transaction
+ *   store A in T1
+ *   checkpoint
+ *   store B in T2
+ *   commit (or abort)
+ *   signal to end the process abruptly
+ *  wait for the process to finish
+ *   open the environment doing recovery
+ *   check to see if both A and B are present (or absent)
+ */
+#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 |DB_RECOVER;
+char *namea="a.db";
+char *nameb="b.db";
+
+
+static void
+do_x1_shutdown (BOOL do_commit, BOOL do_abort) {
+    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 *txn0;
+    r = env->txn_begin(env, NULL, &txn0, 0);                                            CKERR(r);
+    DB_TXN *txn;
+    r = env->txn_begin(env, txn0, &txn, 0);                                             CKERR(r);
+    {
+	DBT a={.data="a", .size=2};
+	DBT b={.data="b", .size=2};
+	r = dba->put(dba, txn, &a, &b, 0);                                              CKERR(r);
+	r = env->txn_checkpoint(env, 0, 0, 0);                                          CKERR(r);
+	r = dbb->put(dbb, txn, &b, &a, 0);                                              CKERR(r);
+    }
+    //printf("opened\n");
+    if (do_commit) {
+	r = txn->commit(txn, 0);                                                        CKERR(r);
+        r = txn->commit(txn0, 0);                                                       CKERR(r);
+    } else if (do_abort) {
+        r = txn->abort(txn);                                                            CKERR(r);
+        r = txn->abort(txn0);                                                           CKERR(r);
+        
+        // force an fsync of the log
+        r = env->txn_begin(env, NULL, &txn, 0);                                         CKERR(r);
+        r = txn->commit(txn, 0);                                                        CKERR(r);
+    }
+    //printf("shutdown\n");
+    abort();
+}
+
+static void
+do_x1_recover (BOOL did_commit) {
+    DB_ENV *env;
+    DB *dba, *dbb;
+    int r;
+    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 = dba->open(dbb, NULL, nameb, NULL, DB_BTREE, DB_AUTO_COMMIT|DB_CREATE, 0666);        CKERR(r);
+    DBT aa={.size=0}, ab={.size=0};
+    DBT ba={.size=0}, bb={.size=0};
+    DB_TXN *txn;
+    DBC *ca,*cb;
+    r = env->txn_begin(env, NULL, &txn, 0);                                                 CKERR(r);
+    r = dba->cursor(dba, txn, &ca, 0);                                                      CKERR(r);
+    r = dbb->cursor(dbb, txn, &cb, 0);                                                      CKERR(r);
+    int ra = ca->c_get(ca, &aa, &ab, DB_FIRST);                                             CKERR(r);
+    int rb = cb->c_get(cb, &ba, &bb, DB_FIRST);                                             CKERR(r);
+    if (did_commit) {
+	assert(ra==0);
+	assert(rb==0);
+	// verify key-value pairs
+	assert(aa.size==2);
+	assert(ab.size==2);
+	assert(ba.size==2);
+	assert(bb.size==2);
+	const char a[2] = "a";
+	const char b[2] = "b";
+        assert(memcmp(aa.data, &a, 2)==0);
+        assert(memcmp(ab.data, &b, 2)==0);
+        assert(memcmp(ab.data, &b, 2)==0);
+        assert(memcmp(bb.data, &a, 2)==0);
+	// make sure no other entries in DB
+	assert(ca->c_get(ca, &aa, &ab, DB_NEXT) == DB_NOTFOUND);
+	assert(cb->c_get(cb, &ba, &bb, DB_NEXT) == DB_NOTFOUND);
+	fprintf(stderr, "Both verified. Yay!\n");
+    } else {
+	// It wasn't committed (it also wasn't aborted), but a checkpoint happened.
+	assert(ra==DB_NOTFOUND);
+	assert(rb==DB_NOTFOUND);
+	fprintf(stderr, "Neither present. Yay!\n");
+    }
+    r = ca->c_close(ca);                                                                    CKERR(r);
+    r = cb->c_close(cb);                                                                    CKERR(r);
+    r = txn->commit(txn, 0);                                                                CKERR(r);
+    r = dba->close(dba, 0);                                                                 CKERR(r);
+    r = dbb->close(dbb, 0);                                                                 CKERR(r);
+    r = env->close(env, 0);                                                                 CKERR(r);
+    exit(0);
+}
+
+static void
+do_x1_recover_only (void) {
+    DB_ENV *env;
+    int r;
+
+    r = db_env_create(&env, 0);                                                             CKERR(r);
+    r = env->open(env, ENVDIR, envflags, S_IRWXU+S_IRWXG+S_IRWXO);                          CKERR(r);
+    r = env->close(env, 0);                                                                 CKERR(r);
+    exit(0);
+}
+
+const char *cmd;
+
+static void
+do_test_internal (BOOL commit) {
+    pid_t pid;
+    if (0 == (pid=fork())) {
+	int r=execl(cmd, verbose ? "-v" : "-q", commit ? "--commit" : "--abort", NULL);
+	assert(r==-1);
+	printf("execl failed: %d (%s)\n", errno, strerror(errno));
+	assert(0);
+    }
+    {
+	int r;
+	int status;
+	r = waitpid(pid, &status, 0);
+	//printf("signaled=%d sig=%d\n", WIFSIGNALED(status), WTERMSIG(status));
+	assert(WIFSIGNALED(status) && WTERMSIG(status)==SIGABRT);
+    }
+    // Now find out what happend
+    
+    if (0 == (pid = fork())) {
+	int r=execl(cmd, verbose ? "-v" : "-q", commit ? "--recover-committed" : "--recover-aborted", NULL);
+	assert(r==-1);
+	printf("execl failed: %d (%s)\n", errno, strerror(errno));
+	assert(0);
+    }
+    {
+	int r;
+	int status;
+	r = waitpid(pid, &status, 0);
+	//printf("recovery exited=%d\n", WIFEXITED(status));
+	assert(WIFEXITED(status) && WEXITSTATUS(status)==0);
+    }
+}
+
+static void
+do_test (void) {
+    do_test_internal(TRUE);
+    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;
+
+static void
+x1_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=1;
+	} else if (strcmp(argv[0], "--abort")==0) {
+	    do_abort=1;
+	} else if (strcmp(argv[0], "--explicit-abort")==0) {
+	    do_explicit_abort=1;
+	} else if (strcmp(argv[0], "--recover-committed")==0) {
+	    do_recover_committed=1;
+	} else if (strcmp(argv[0], "--recover-aborted")==0) {
+	    do_recover_aborted=1;
+        } else if (strcmp(argv[0], "--recover-only") == 0) {
+            do_recover_only=1;
+	} 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 (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[]) {
+    x1_parse_args(argc, argv);
+    if (do_commit) {
+	do_x1_shutdown (TRUE, FALSE);
+    } else if (do_abort) {
+	do_x1_shutdown (FALSE, FALSE);
+    } else if (do_explicit_abort) {
+        do_x1_shutdown(FALSE, TRUE);
+    } else if (do_recover_committed) {
+	do_x1_recover(TRUE);
+    } else if (do_recover_aborted) {
+	do_x1_recover(FALSE);
+    } else if (do_recover_only) {
+        do_x1_recover_only();
+    } else {
+	do_test();
+    }
+    return 0;
+}