Contrary to r18222, the new group commit code was not on the main truck.

{{{
svn merge -r17893:18056 https://svn.tokutek.com/tokudb/toku/tokudb.2370c
}}}

Refs #2370, #2385.  [t:2370] [t:2385].


git-svn-id: file:///svn/toku/tokudb@18259 c7de825b-a66e-492c-adef-691d508d4ae1

buildheader/make_db_h.c

@@ -427,7 +427,6 @@ int main (int argc __attribute__((__unused__)), char *const argv[] __attribute__
     printf("  u_int64_t        time_ydb_lock_held_unavailable;  /* number of times a thread migrated and theld is unavailable */ \n");
     printf("  u_int64_t        max_time_ydb_lock_held;  /* max time a client thread held the ydb lock  */ \n");
     printf("  u_int64_t        total_time_ydb_lock_held;/* total time client threads held the ydb lock  */ \n");
-    printf("  u_int64_t        logger_lock_ctr;         /* how many times has logger lock been taken/released */ \n");
     printf("  u_int32_t        checkpoint_period;       /* delay between automatic checkpoints  */ \n");
     printf("  u_int32_t        checkpoint_footprint;    /* state of checkpoint procedure        */ \n");
     printf("  char             checkpoint_time_begin[26]; /* time of last checkpoint begin      */ \n");

linux/file.c

@@ -5,6 +5,7 @@
 #include <errno.h>
 #include <toku_assert.h>
 #include <stdio.h>
+#include <string.h>
 
 //Print any necessary errors
 //Return whether we should try the write again.
@@ -130,19 +131,29 @@ toku_set_func_fsync(int (*fsync_function)(int)) {
     return 0;
 }
 
-// keep trying if fsync fails because of EINTR
 int
-toku_file_fsync(int fd) {
+toku_file_fsync_without_accounting (int fd) {
     int r = -1;
-    uint64_t tstart = get_tnow();
     while (r != 0) {
 	if (t_fsync)
 	    r = t_fsync(fd);
 	else 
 	    r = fsync(fd);
-	if (r) 
-	    assert(errno==EINTR);
+	if (r) {
+	    int rr = errno;
+	    if (rr!=EINTR) printf("rr=%d (%s)\n", rr, strerror(rr));
+	    assert(rr==EINTR);
 	}
+    }
+    return r;
+}
+
+
+// keep trying if fsync fails because of EINTR
+int
+toku_file_fsync(int fd) {
+    uint64_t tstart = get_tnow();
+    int r = toku_file_fsync_without_accounting(fd);
     toku_sync_fetch_and_increment_uint64(&toku_fsync_count);
     toku_sync_fetch_and_add_uint64(&toku_fsync_time, get_tnow() - tstart);
     return r;

newbrt/brt.c

@@ -5459,8 +5459,6 @@ toku_brt_lock_init(void) {
     int r = 0;
     if (r==0)
         r = toku_pwrite_lock_init();
-    if (r==0)
-        r = toku_logger_lock_init();
     return r;
 }
 
@@ -5469,8 +5467,6 @@ toku_brt_lock_destroy(void) {
     int r = 0;
     if (r==0) 
         r = toku_pwrite_lock_destroy();
-    if (r==0) 
-        r = toku_logger_lock_destroy();
     return r;
 }
 

newbrt/log-internal.h

@@ -59,7 +59,12 @@ struct logbuf {
 
 struct tokulogger {
     enum typ_tag tag; // must be first
-    struct mylock  input_lock, output_lock; // acquired in that order
+    struct mylock  input_lock;
+
+    toku_pthread_mutex_t output_condition_lock; // if you need both this lock and input_lock, acquire the output_lock first, then input_lock. More typical is to get the output_is_available condition to be false, and then acquire the input_lock.
+    toku_pthread_cond_t  output_condition;      //
+    BOOL output_is_available;                  // this is part of the predicate for the output condition.  It's true if no thread is modifying the output (either doing an fsync or otherwise fiddling with the output).
+
     BOOL is_open;
     BOOL is_panicked;
     BOOL write_log_files;
@@ -75,14 +80,15 @@ struct tokulogger {
     OMT live_txns; // a sorted tree.  Old comment said should be a hashtable.  Do we still want that?
     struct logbuf inbuf; // data being accumulated for the write
 
-    // To access these, you must have the output lock
+    // To access these, you must have the output condition lock.
     LSN written_lsn; // the last lsn written
-    LSN fsynced_lsn; // What is the LSN of the highest fsynced log entry
+    LSN fsynced_lsn; // What is the LSN of the highest fsynced log entry  (accessed only while holding the output lock, and updated only when the output lock and output permission are held)
     LSN checkpoint_lsn;     // What is the LSN of the most recent completed checkpoint.
     long long next_log_file_number;
     struct logbuf outbuf; // data being written to the file
     int  n_in_file; // The amount of data in the current file
 
+    // To access the logfilemgr you must have the output condition lock.
     TOKULOGFILEMGR logfilemgr;
 
     u_int32_t write_block_size;       // How big should the blocks be written to various logs?

newbrt/logger.h

@@ -12,8 +12,6 @@ int toku_logger_open (const char *directory, TOKULOGGER logger);
 int toku_logger_shutdown(TOKULOGGER logger);
 int toku_logger_close(TOKULOGGER *loggerp);
 
-u_int32_t toku_logger_get_lock_ctr(void);
-
 int toku_logger_fsync (TOKULOGGER logger);
 void toku_logger_panic (TOKULOGGER logger, int err);
 int toku_logger_panicked(TOKULOGGER logger);

newbrt/recover.c

@@ -931,7 +931,10 @@ static int toku_recover_commit (struct logtype_commit *l, RECOVER_ENV renv) {
     }
 
     // commit the transaction
-    r = toku_txn_commit_with_lsn(txn, TRUE, recover_yield, NULL, l->lsn, NULL, NULL);
+    r = toku_txn_commit_with_lsn(txn, TRUE, recover_yield, NULL, l->lsn,
+				 NULL, NULL,
+				 // No need to release locks during recovery.
+				 NULL, NULL, NULL);
     assert(r == 0);
 
     // close the transaction

newbrt/txn.c

@@ -78,12 +78,16 @@ died:
 
 // Doesn't close the txn, just performs the commit operations.
 int toku_txn_commit_txn(TOKUTXN txn, int nosync, YIELDF yield, void *yieldv,
-                        TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra) {
-    return toku_txn_commit_with_lsn(txn, nosync, yield, yieldv, ZERO_LSN, poll, poll_extra);
+                        TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
+			void (*release_locks)(void*), void(*reacquire_locks)(void*), void *locks_thunk) {
+    return toku_txn_commit_with_lsn(txn, nosync, yield, yieldv, ZERO_LSN,
+				    poll, poll_extra,
+				    release_locks, reacquire_locks, locks_thunk);
 }
 
 int toku_txn_commit_with_lsn(TOKUTXN txn, int nosync, YIELDF yield, void *yieldv, LSN oplsn,
-                             TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra) {
+                             TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
+			     void (*release_locks)(void*), void(*reacquire_locks)(void*), void *locks_thunk) {
     int r;
     // panic handled in log_commit
 
@@ -93,7 +97,9 @@ int toku_txn_commit_with_lsn(TOKUTXN txn, int nosync, YIELDF yield, void *yieldv
     txn->progress_poll_fun = poll;
     txn->progress_poll_fun_extra = poll_extra;
 
+    if (release_locks) release_locks(locks_thunk);
     r = toku_log_commit(txn->logger, (LSN*)0, do_fsync, txn->txnid64); // exits holding neither of the tokulogger locks.
+    if (reacquire_locks) reacquire_locks(locks_thunk);
     if (r!=0)
         return r;
     r = toku_rollback_commit(txn, yield, yieldv, oplsn);

newbrt/txn.h

@@ -9,9 +9,11 @@ int toku_txn_begin_txn (TOKUTXN parent_tokutxn, TOKUTXN *tokutxn, TOKULOGGER log
 int toku_txn_begin_with_xid (TOKUTXN parent_tokutxn, TOKUTXN *tokutxn, TOKULOGGER logger, TXNID xid);
 
 int toku_txn_commit_txn (TOKUTXN txn, int nosync, YIELDF yield, void *yieldv,
-                         TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra);
+			 TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
+			 void (*release_locks)(void*), void(*reacquire_locks)(void*), void *locks_thunk);
 int toku_txn_commit_with_lsn(TOKUTXN txn, int nosync, YIELDF yield, void *yieldv, LSN oplsn,
-                             TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra);
+			     TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
+			     void (*release_locks)(void*), void(*reacquire_locks)(void*), void *locks_thunk);
 
 int toku_txn_abort_txn(TOKUTXN txn, YIELDF yield, void *yieldv,
                        TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra);

src/tests/Makefile

@@ -174,7 +174,6 @@ BDB_BINS = $(patsubst %.c,%.bdb$(BINSUF),$(filter-out $(patsubst %,%.c,$(BDB_DON
 endif
 
 TDB_TESTS_THAT_SHOULD_FAIL= \
-	test_groupcommit_count \
 	test944 \
 	test_truncate_txn_abort \
 	test_db_no_env
@@ -215,6 +214,8 @@ TLRECOVER = 2 3 4 5 6 7 8 9 10
 
 EXTRA_TDB_TESTS = \
 	$(patsubst %,test_log%.recover,$(TLRECOVER)) \
+	test_groupcommit_count_hgrind.tdbrun \
+	test_groupcommit_count_vgrind.tdbrun \
 #\ ends prev line
 
 ifeq ($(OS_CHOICE),windows)
@@ -602,6 +603,12 @@ helgrind2.bdbrun: BDBVGRIND=$(HGRIND)
 helgrind3.tdbrun: TDBVGRIND=$(HGRIND)
 helgrind3.bdbrun: BDBVGRIND=$(HGRIND)
 
+test_groupcommit_count_hgrind.tdbrun: HGRIND+=--suppressions=helgrind.suppressions
+test_groupcommit_count_hgrind.tdbrun: test_groupcommit_count.tdb$(BINSUF)
+	$(HGRIND) ./test_groupcommit_count.tdb$(BINSUF) $(VERBVERBOSE) -n 1 -p hgrind $(SUMMARIZE_CMD)
+test_groupcommit_count_vgrind.tdbrun: test_groupcommit_count.tdb$(BINSUF)
+	$(VGRIND) ./test_groupcommit_count.tdb$(BINSUF) $(VERBVERBOSE) -n 1 -p vgrind $(SUMMARIZE_CMD)
+
 # we deliberately don't close the env, so recovery runs
 # lets avoid all of the valgrind errors
 test-recover1.tdbrun: VGRIND=

src/tests/helgrind.suppressions

+{
+   helgrind_3.5.0_false_positive_against_pthread_create
+   Helgrind:Race
+   fun:mythread_wrapper
+}

src/tests/test_groupcommit_count.c

@@ -12,6 +12,7 @@
 
 DB_ENV *env;
 DB *db;
+int do_sync=1;
 
 #define NITER 100
 
@@ -28,18 +29,20 @@ static void *start_a_thread (void *i_p) {
 		  dbt_init(&key, keystr, 1+strlen(keystr)),
 		  dbt_init(&data, keystr, 1+strlen(keystr)),
 		  0);
-	r=tid->commit(tid, 0); CKERR(r);
+	r=tid->commit(tid, do_sync ? 0 : DB_TXN_NOSYNC); CKERR(r);
     }
     return 0;
 }
 
+char *env_path;
+
 static void
 test_groupcommit (int nthreads) {
     int r;
     DB_TXN *tid;
 
     r=db_env_create(&env, 0); assert(r==0);
-    r=env->open(env, ENVDIR, DB_INIT_LOCK|DB_INIT_LOG|DB_INIT_MPOOL|DB_INIT_TXN|DB_CREATE|DB_PRIVATE|DB_THREAD, S_IRWXU+S_IRWXG+S_IRWXO); CKERR(r);
+    r=env->open(env, env_path, DB_INIT_LOCK|DB_INIT_LOG|DB_INIT_MPOOL|DB_INIT_TXN|DB_CREATE|DB_PRIVATE|DB_THREAD, S_IRWXU+S_IRWXG+S_IRWXO); CKERR(r);
     r=db_create(&db, env, 0); CKERR(r);
     r=env->txn_begin(env, 0, &tid, 0); assert(r==0);
     r=db->open(db, tid, "foo.db", 0, DB_BTREE, DB_CREATE, S_IRWXU+S_IRWXG+S_IRWXO); CKERR(r);
@@ -59,6 +62,7 @@ test_groupcommit (int nthreads) {
     r=db->close(db, 0); assert(r==0);
     r=env->close(env, 0); assert(r==0);
 
+    //if (verbose) printf(" That's a total of %d commits\n", nthreads*NITER);
 }
 
 // helgrind doesn't understand that pthread_join removes a race condition.   I'm not impressed... -Bradley
@@ -88,42 +92,102 @@ static struct timeval prevtime;
 static int prev_count;
 
 static void
-printtdiff (char *str) {
+printtdiff (int N) {
     struct timeval thistime;
     gettimeofday(&thistime, 0);
     double diff = toku_tdiff(&thistime, &prevtime);
     int fcount=get_fsync_count();
-    if (verbose) printf("%s: %10.6fs %d fsyncs for %s\n", progname, diff, fcount-prev_count, str);
+    if (verbose) printf("%s: %10.6fs %4d fsyncs for %4d threads %s %8.1f tps, %8.1f tps/thread\n", progname, diff, fcount-prev_count,
+			N,
+			do_sync ? "with sync         " : "with DB_TXN_NOSYNC",
+			NITER*(N/diff), NITER/diff);
     prevtime=thistime;
     prev_count=fcount;
 }
 
+static void
+do_test (int N) {
+    for (do_sync = 0; do_sync<2; do_sync++) {
+	int count_before = get_fsync_count();
+	test_groupcommit(N);
+	printtdiff(N);
+	if (get_fsync_count()-count_before>= N*NITER) {
+	    if (verbose) printf("It looks like too many fsyncs.  Group commit doesn't appear to be occuring.\n");
+	    exit(1);
+	}
+    }
+}
+
+int log_max_n_threads_over_10 = 3;
+
+static void
+my_parse_args (int argc, char *argv[]) {
+    verbose=1; // use -q to turn off the talking.
+    env_path = toku_strdup(ENVDIR);
+    const char *argv0=argv[0];
+    while (argc>1) {
+	int resultcode=0;
+	if (strcmp(argv[1], "-v")==0) {
+	    verbose++;
+	} else if (strcmp(argv[1],"-q")==0) {
+	    verbose--;
+	    if (verbose<0) verbose=0;
+	} else if (strcmp(argv[1],"-n")==0) {
+	    argc--;
+	    argv++;
+	    if (argc<=1) { resultcode=1; goto do_usage; }
+	    errno = 0;
+	    char *end;
+	    log_max_n_threads_over_10 = strtol(argv[1], &end, 10);
+	    if (errno!=0 || *end) {
+		resultcode=1;
+		goto do_usage;
+	    }
+	} else if (strcmp(argv[1],"-p")==0) {
+	    argc--;
+	    argv++;
+	    if (argc<=1) { resultcode=1; goto do_usage; }
+	    int size = strlen(ENVDIR) + 10 + strlen(argv[1]);
+	    REALLOC_N(size, env_path);
+	    assert(env_path);
+	    snprintf(env_path, size, "%s.%s", ENVDIR, argv[1]);
+	} else if (strcmp(argv[1], "-h")==0) {
+	do_usage:
+	    fprintf(stderr, "Usage:\n%s [-v|-q] [-n LOG(MAX_N_THREADS/10)] [-h]\n", argv0);
+	    exit(resultcode);
+	} else {
+	    resultcode=1;
+	    goto do_usage;
+	}
+	argc--;
+	argv++;
+    }
+}
+
+
 int
 test_main (int argc, char *const argv[]) {
     progname=argv[0];
-    parse_args(argc, argv);
+    my_parse_args(argc, argv);
 
     gettimeofday(&prevtime, 0);
     prev_count=0;
 
     { int r = db_env_set_func_fsync(do_fsync); CKERR(r); }
 
-    system("rm -rf " ENVDIR);
-    { int r=toku_os_mkdir(ENVDIR, S_IRWXU+S_IRWXG+S_IRWXO);       assert(r==0); }
-
-    test_groupcommit(1);  printtdiff("1 thread");
-    test_groupcommit(2);  printtdiff("2 threads");
-    int count_before_10 = get_fsync_count();
-    test_groupcommit(10); printtdiff("10 threads");
-    if (get_fsync_count()-count_before_10 >= 10*NITER) {
-	if (verbose) printf("It looks like too many fsyncs.  Group commit doesn't appear to be occuring.\n");
-	exit(1);
+    {
+	int size=20+strlen(env_path);
+	char command[size];
+	snprintf(command, size, "rm -rf %s", env_path);
+	system(command);
     }
-    int count_before_20 = get_fsync_count();
-    test_groupcommit(20); printtdiff("20 threads");
-    if (get_fsync_count()-count_before_20 >= 20*NITER) {
-	if (verbose) printf("It looks like too many fsyncs.  Group commit doesn't appear to be occuring.\n");
-	exit(1);
+    { int r=toku_os_mkdir(env_path, S_IRWXU+S_IRWXG+S_IRWXO);       assert(r==0); }
+
+    test_groupcommit(1);  printtdiff(1);
+    test_groupcommit(2);  printtdiff(2);
+    for (int i=0; i<log_max_n_threads_over_10; i++) {
+	do_test(10 << i);
     }
+    toku_free(env_path);
     return 0;
 }

src/tests/test_groupcommit_perf.c

@@ -91,6 +91,7 @@ printtdiff (char *str) {
     struct timeval thistime;
     gettimeofday(&thistime, 0);
     if (verbose) printf("%10.6f %s\n", toku_tdiff(&thistime, &prevtime), str);
+    prevtime = thistime;
 }
 
 int

src/ydb.c

@@ -1323,8 +1323,6 @@ env_get_engine_status(DB_ENV * env, ENGINE_STATUS * engstat) {
 	time_t now = time(NULL);
         format_time(&now, engstat->now);
 
-	engstat->logger_lock_ctr = toku_logger_get_lock_ctr();
-
 	{
 	    SCHEDULE_STATUS_S schedstat;
 	    toku_ydb_lock_get_status(&schedstat);
@@ -1418,7 +1416,6 @@ env_get_engine_status_text(DB_ENV * env, char * buff, int bufsiz) {
     n += snprintf(buff + n, bufsiz - n, "time_ydb_lock_held_unavailable   %"PRIu64"\n", engstat.time_ydb_lock_held_unavailable);
     n += snprintf(buff + n, bufsiz - n, "max_time_ydb_lock_held           %"PRIu64"\n", engstat.max_time_ydb_lock_held);
     n += snprintf(buff + n, bufsiz - n, "total_time_ydb_lock_held         %"PRIu64"\n", engstat.total_time_ydb_lock_held);
-    n += snprintf(buff + n, bufsiz - n, "logger_lock_ctr                  %"PRIu64"\n", engstat.logger_lock_ctr);
     n += snprintf(buff + n, bufsiz - n, "checkpoint_period                %d \n", engstat.checkpoint_period);
     n += snprintf(buff + n, bufsiz - n, "checkpoint_footprint             %d \n", engstat.checkpoint_footprint);
     n += snprintf(buff + n, bufsiz - n, "checkpoint_time_begin            %s \n", engstat.checkpoint_time_begin);
@@ -1608,6 +1605,15 @@ static void ydb_yield (voidfp f, void *UU(v)) {
     toku_ydb_lock();
 }
 
+static void release_ydb_lock_callback (void *ignore __attribute__((__unused__))) {
+    //printf("%8.6fs Thread %ld release\n", get_tdiff(), pthread_self());
+    toku_ydb_unlock(); 
+}
+static void reacquire_ydb_lock_callback (void *ignore __attribute__((__unused__))) {
+    //printf("%8.6fs Thread %ld reacquire\n", get_tdiff(), pthread_self());
+    toku_ydb_lock(); 
+}
+
 static int toku_txn_commit(DB_TXN * txn, u_int32_t flags,
 			   TXN_PROGRESS_POLL_FUNCTION poll, void* poll_extra) {
     if (!txn) return EINVAL;
@@ -1648,7 +1654,9 @@ static int toku_txn_commit(DB_TXN * txn, u_int32_t flags,
 	// frees the tokutxn
 	// Calls ydb_yield(NULL) occasionally
         //r = toku_logger_commit(db_txn_struct_i(txn)->tokutxn, nosync, ydb_yield, NULL);
-        r = toku_txn_commit_txn(db_txn_struct_i(txn)->tokutxn, nosync, ydb_yield, NULL, poll, poll_extra);
+        r = toku_txn_commit_txn(db_txn_struct_i(txn)->tokutxn, nosync, ydb_yield, NULL,
+				poll, poll_extra,
+				release_ydb_lock_callback, reacquire_ydb_lock_callback, NULL);
 
     if (r!=0 && !toku_env_is_panicked(txn->mgrp)) {
         txn->mgrp->i->is_panicked = r;
@@ -5292,8 +5300,10 @@ include_toku_pthread_yield (void) {
 // For test purposes only, translate dname to iname
 static int 
 env_get_iname(DB_ENV* env, DBT* dname_dbt, DBT* iname_dbt) {
+    toku_ydb_lock();
     DB *directory = env->i->directory;
     int r = autotxn_db_get(directory, NULL, dname_dbt, iname_dbt, DB_PRELOCKED); // allocates memory for iname
+    toku_ydb_unlock();
     return r;
 }
 

toku_include/toku_portability.h

@@ -139,6 +139,7 @@ void toku_os_full_write (int fd, const void *buf, size_t len) __attribute__((__v
 int toku_os_write (int fd, const void *buf, size_t len) __attribute__((__visibility__("default")));
 
 // wrapper around fsync
+int toku_file_fsync_without_accounting(int fd);
 int toku_file_fsync(int fd);
 
 // get the number of fsync calls and the fsync times

windows/file.c

@@ -249,9 +249,8 @@ static uint64_t get_tnow(void) {
 
 // keep trying if fsync fails because of EINTR
 int
-toku_file_fsync(int fd) {
+toku_file_fsync_without_accounting (int fd) {
     int r = -1;
-    uint64_t tstart = get_tnow();
     while (r != 0) {
 	if (t_fsync)
 	    r = t_fsync(fd);
@@ -260,6 +259,13 @@ toku_file_fsync(int fd) {
 	if (r) 
 	    assert(errno==EINTR);
     }
+    return r;
+}
+
+int
+toku_file_fsync(int fd) {
+    uint64_t tstart = get_tnow();
+    int r = toku_file_fsync_without_accounting(fd);
 #if TOKU_WINDOWS_HAS_ATOMIC_64 
     toku_sync_fetch_and_increment_uint64(&toku_fsync_count);
     toku_sync_fetch_and_add_uint64(&toku_fsync_time, get_tnow() - tstart);