closes #5149 Skips taking MO lock for readonly transactions on

commit, abort, and prepare

git-svn-id: file:///svn/toku/tokudb@44989 c7de825b-a66e-492c-adef-691d508d4ae1

ft/txn.c

@@ -287,18 +287,12 @@ int toku_txn_commit_with_lsn(TOKUTXN txn, int nosync, LSN oplsn,
     txn->progress_poll_fun = poll;
     txn->progress_poll_fun_extra = poll_extra;
 
-    if (txn->begin_was_logged) {
+    if (!toku_txn_is_read_only(txn)) {
         r = toku_log_xcommit(txn->logger, &txn->do_fsync_lsn, 0, txn, txn->txnid64);
         if (r != 0) {
             goto cleanup;
         }
     }
-    else {
-        // Did no work.
-        invariant(txn->roll_info.num_rollentries == 0);
-        // Was not prepared.
-        invariant(txn->do_fsync_lsn.lsn == ZERO_LSN.lsn);
-    }
     // If !txn->begin_was_logged, we could skip toku_rollback_commit
     // but it's cheap (only a number of function calls that return immediately)
     // since there were no writes.  Skipping it would mean we would need to be careful
@@ -328,18 +322,12 @@ int toku_txn_abort_with_lsn(TOKUTXN txn, LSN oplsn,
     int r;
     txn->do_fsync = FALSE;
 
-    if (txn->begin_was_logged) {
+    if (!toku_txn_is_read_only(txn)) {
         r = toku_log_xabort(txn->logger, &txn->do_fsync_lsn, 0, txn, txn->txnid64);
         if (r != 0) {
             goto cleanup;
         }
     }
-    else {
-        // Did no work.
-        invariant(txn->roll_info.num_rollentries == 0);
-        // Was not prepared.
-        invariant(txn->do_fsync_lsn.lsn == ZERO_LSN.lsn);
-    }
     // If !txn->begin_was_logged, we could skip toku_rollback_abort
     // but it's cheap (only a number of function calls that return immediately)
     // since there were no writes.  Skipping it would mean we would need to be careful
@@ -360,7 +348,7 @@ static void copy_xid (TOKU_XA_XID *dest, TOKU_XA_XID *source) {
 
 int toku_txn_prepare_txn (TOKUTXN txn, TOKU_XA_XID *xa_xid) {
     int r = 0;
-    if (txn->parent || !txn->begin_was_logged) {
+    if (txn->parent || toku_txn_is_read_only(txn)) {
         // nothing to do if there's a parent, or if it's read-only
         goto cleanup;
     }
@@ -526,6 +514,19 @@ toku_maybe_log_begin_txn_for_write_operation(TOKUTXN txn) {
     toku_txn_unlock(txn);
 }
 
+bool
+toku_txn_is_read_only(TOKUTXN txn) {
+    if (!txn->begin_was_logged) {
+        // Did no work.
+        invariant(txn->roll_info.num_rollentries == 0);
+        // Was not prepared.
+        invariant(txn->do_fsync_lsn.lsn == ZERO_LSN.lsn);
+        invariant(toku_omt_size(txn->open_fts) == 0);
+        return true;
+    }
+    return false;
+}
+
 #include <valgrind/helgrind.h>
 void __attribute__((__constructor__)) toku_txn_status_helgrind_ignore(void);
 void

ft/txn.h

@@ -114,6 +114,7 @@ struct tokulogger_preplist {
 int toku_logger_recover_txn (TOKULOGGER logger, struct tokulogger_preplist preplist[/*count*/], long count, /*out*/ long *retp, u_int32_t flags);
 
 void toku_maybe_log_begin_txn_for_write_operation(TOKUTXN txn);
+bool toku_txn_is_read_only(TOKUTXN txn);
 
 #if defined(__cplusplus) || defined(__cilkplusplus)
 }

src/ydb_txn.c

@@ -13,9 +13,10 @@
 #include <valgrind/helgrind.h>
 #include "ft/txn_manager.h"
 
-static int toku_txn_commit(DB_TXN * txn, u_int32_t flags, TXN_PROGRESS_POLL_FUNCTION poll, 
-        void *poll_extra, bool release_multi_operation_client_lock);
-static int toku_txn_abort(DB_TXN * txn, TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra);
+static int toku_txn_commit(DB_TXN * txn, u_int32_t flags, TXN_PROGRESS_POLL_FUNCTION poll,
+        void *poll_extra, bool release_mo_client_lock_if_held, bool *holds_mo_lock);
+static int toku_txn_abort(DB_TXN * txn, TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
+        bool *holds_mo_lock);
 
 static int 
 toku_txn_release_locks(DB_TXN* txn) {
@@ -56,15 +57,23 @@ toku_txn_destroy(DB_TXN *txn) {
     toku_free(txn);
 }
 
-static int 
-toku_txn_commit_only(DB_TXN * txn, u_int32_t flags,
-                     TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
-                     bool release_multi_operation_client_lock) {
+static int
+toku_txn_commit(DB_TXN * txn, u_int32_t flags,
+                TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
+                bool release_mo_lock_if_held,
+                bool *holds_mo_lock) {
     HANDLE_PANICKED_ENV(txn->mgrp);
+    // Don't take multi-operation lock until we see a non-readonly txn.
+    if (!*holds_mo_lock &&
+        !toku_txn_is_read_only(db_txn_struct_i(txn)->tokutxn)) {
+        toku_multi_operation_client_lock();
+        *holds_mo_lock = true;
+    }
     //Recursively kill off children
     if (db_txn_struct_i(txn)->child) {
         //commit of child sets the child pointer to NULL
-        int r_child = toku_txn_commit(db_txn_struct_i(txn)->child, flags, NULL, NULL, false);
+        int r_child = toku_txn_commit(db_txn_struct_i(txn)->child,
+                flags, NULL, NULL, false, holds_mo_lock);
         if (r_child !=0 && !toku_env_is_panicked(txn->mgrp)) {
             env_panic(txn->mgrp, r_child, "Recursive child commit failed during parent commit.\n");
         }
@@ -116,19 +125,15 @@ toku_txn_commit_only(DB_TXN * txn, u_int32_t flags,
     // this lock must be held until the references to the open FTs is released
     // begin checkpoint logs these associations, so we must be protect
     // the changing of these associations with checkpointing
-    if (release_multi_operation_client_lock) {
+    if (release_mo_lock_if_held && *holds_mo_lock) {
         toku_multi_operation_client_unlock();
     }
     toku_txn_maybe_fsync_log(logger, do_fsync_lsn, do_fsync);
-    if (flags!=0) return EINVAL;
-    return r;
-}
-
-static int 
-toku_txn_commit(DB_TXN * txn, u_int32_t flags,
-                TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
-                bool release_multi_operation_client_lock) {
-    int r = toku_txn_commit_only(txn, flags, poll, poll_extra, release_multi_operation_client_lock);
+    if (flags!=0) {
+        r = EINVAL;
+        goto cleanup;
+    }
+cleanup:
     toku_txn_destroy(txn);
     return r;
 }
@@ -148,13 +153,21 @@ toku_txn_id64(DB_TXN * txn) {
 }
 
 static int 
-toku_txn_abort_only(DB_TXN * txn,
-                    TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra) {
+toku_txn_abort(DB_TXN * txn,
+               TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra,
+               bool *holds_mo_lock) {
     HANDLE_PANICKED_ENV(txn->mgrp);
+    // Don't take multi-operation lock until we see a non-readonly txn.
+    if (!*holds_mo_lock &&
+        !toku_txn_is_read_only(db_txn_struct_i(txn)->tokutxn)) {
+        toku_multi_operation_client_lock();
+        *holds_mo_lock = true;
+    }
     //Recursively kill off children (abort or commit are both correct, commit is cheaper)
     if (db_txn_struct_i(txn)->child) {
         //commit of child sets the child pointer to NULL
-        int r_child = toku_txn_commit(db_txn_struct_i(txn)->child, DB_TXN_NOSYNC, NULL, NULL, false);
+        int r_child = toku_txn_commit(db_txn_struct_i(txn)->child,
+                DB_TXN_NOSYNC, NULL, NULL, false, holds_mo_lock);
         if (r_child !=0 && !toku_env_is_panicked(txn->mgrp)) {
             env_panic(txn->mgrp, r_child, "Recursive child commit failed during parent abort.\n");
         }
@@ -176,29 +189,34 @@ toku_txn_abort_only(DB_TXN * txn,
     assert_zero(r);
     toku_txn_complete_txn(db_txn_struct_i(txn)->tokutxn);
     r = toku_txn_release_locks(txn);
+    toku_txn_destroy(txn);
     return r;
 }
 
-// requires: must hold the multi operation lock. it is
-//           released here before the fsync.
 static int
 toku_txn_xa_prepare (DB_TXN *txn, TOKU_XA_XID *xid) {
     int r = 0;
     if (!txn) {
-        toku_multi_operation_client_unlock();
         r = EINVAL;
         goto exit;
     }
     if (txn->parent) {
-        toku_multi_operation_client_unlock();
         r = 0; // make this a NO-OP, MySQL calls this
         goto exit;
     }
     HANDLE_PANICKED_ENV(txn->mgrp);
+    // Take the mo lock as soon as a non-readonly txn is found
+    bool holds_mo_lock = false;
+    if (!toku_txn_is_read_only(db_txn_struct_i(txn)->tokutxn)) {
+        toku_multi_operation_client_lock();
+        holds_mo_lock = true;
+    }
     //Recursively commit any children.
     if (db_txn_struct_i(txn)->child) {
         //commit of child sets the child pointer to NULL
-        int r_child = toku_txn_commit(db_txn_struct_i(txn)->child, 0, NULL, NULL, false);
+
+        // toku_txn_commit will take the mo_lock if not held and a non-readonly txn is found.
+        int r_child = toku_txn_commit(db_txn_struct_i(txn)->child, 0, NULL, NULL, false, &holds_mo_lock);
         if (r_child !=0 && !toku_env_is_panicked(txn->mgrp)) {
             env_panic(txn->mgrp, r_child, "Recursive child commit failed during parent commit.\n");
         }
@@ -213,9 +231,11 @@ toku_txn_xa_prepare (DB_TXN *txn, TOKU_XA_XID *xid) {
     bool do_fsync;
     toku_txn_get_fsync_info(ttxn, &do_fsync, &do_fsync_lsn);
     // release the multi operation lock before fsyncing the log
-    toku_multi_operation_client_unlock();
+    if (holds_mo_lock) {
+        toku_multi_operation_client_unlock();
+    }
     toku_txn_maybe_fsync_log(logger, do_fsync_lsn, do_fsync);
-exit:    
+exit:
     return r;
 }
 
@@ -232,14 +252,6 @@ toku_txn_prepare (DB_TXN *txn, u_int8_t gid[DB_GID_SIZE]) {
     return toku_txn_xa_prepare(txn, &xid);
 }
 
-static int 
-toku_txn_abort(DB_TXN * txn,
-               TXN_PROGRESS_POLL_FUNCTION poll, void *poll_extra) {
-    int r = toku_txn_abort_only(txn, poll, poll_extra);
-    toku_txn_destroy(txn);
-    return r;
-}   
-
 static u_int32_t 
 locked_txn_id(DB_TXN *txn) {
     u_int32_t r = toku_txn_id(txn); 
@@ -265,11 +277,13 @@ locked_txn_commit_with_progress(DB_TXN *txn, u_int32_t flags,
     if (toku_txn_requires_checkpoint(ttxn)) {
         toku_checkpoint(txn->mgrp->i->cachetable, txn->mgrp->i->logger, NULL, NULL, NULL, NULL, TXN_COMMIT_CHECKPOINT);
     }
-    // cannot begin a checkpoint. the multi operation lock is taken here, 
-    // but released in toku_txn_commit_only. this way, we don't hold it 
-    // while we fsync the log.
-    toku_multi_operation_client_lock();
-    int r = toku_txn_commit(txn, flags, poll, poll_extra, true);
+    bool holds_mo_lock = false;
+    // cannot begin a checkpoint.
+    // the multi operation lock is taken the first time we
+    // see a non-readonly txn in the recursive commit.
+    // But released in the first-level toku_txn_commit (if taken),
+    // this way, we don't hold it while we fsync the log.
+    int r = toku_txn_commit(txn, flags, poll, poll_extra, true, &holds_mo_lock);
     return r;
 }
 
@@ -277,9 +291,14 @@ static int
 locked_txn_abort_with_progress(DB_TXN *txn,
                                TXN_PROGRESS_POLL_FUNCTION poll, void* poll_extra) {
     // cannot begin a checkpoint
-    toku_multi_operation_client_lock();
-    int r = toku_txn_abort(txn, poll, poll_extra);
-    toku_multi_operation_client_unlock();
+    // the multi operation lock is taken the first time we
+    // see a non-readonly txn in the abort (or recursive commit).
+    // But released here so we don't have to hold additional state.
+    bool holds_mo_lock = false;
+    int r = toku_txn_abort(txn, poll, poll_extra, &holds_mo_lock);
+    if (holds_mo_lock) {
+        toku_multi_operation_client_unlock();
+    }
     return r;
 }
 
@@ -295,22 +314,21 @@ locked_txn_abort(DB_TXN *txn) {
     return r;
 }
 
-static int
-locked_txn_prepare (DB_TXN *txn, u_int8_t gid[DB_GID_SIZE]) {
-    // toku_txn_prepare eventually releases the multi operation lock
-    // before fsyncing the log
-    toku_multi_operation_client_lock();
-    int r = toku_txn_prepare(txn, gid); 
-    return r;
-}
-
-static int
-locked_txn_xa_prepare (DB_TXN *txn, TOKU_XA_XID *xid) {
-    // toku_txn_xa_prepare eventually releases the multi operation lock
-    // before fsyncing the log
-    toku_multi_operation_client_lock();
-    int r = toku_txn_xa_prepare(txn, xid); 
-    return r;
+static inline void
+txn_func_init(DB_TXN *txn) {
+#define STXN(name) txn->name = locked_txn_ ## name
+    STXN(abort);
+    STXN(commit);
+    STXN(abort_with_progress);
+    STXN(commit_with_progress);
+    STXN(id);
+    STXN(txn_stat);
+#undef STXN
+#define SUTXN(name) txn->name = toku_txn_ ## name
+    SUTXN(prepare);
+    SUTXN(xa_prepare);
+#undef SUTXN
+    txn->id64 = toku_txn_id64;
 }
 
 int 
@@ -398,17 +416,7 @@ toku_txn_begin(DB_ENV *env, DB_TXN * stxn, DB_TXN ** txn, u_int32_t flags) {
 
     //toku_ydb_notef("parent=%p flags=0x%x\n", stxn, flags);
     result->mgrp = env;
-#define STXN(name) result->name = locked_txn_ ## name
-    STXN(abort);
-    STXN(commit);
-    STXN(abort_with_progress);
-    STXN(commit_with_progress);
-    STXN(id);
-    STXN(prepare);
-    STXN(xa_prepare);
-    STXN(txn_stat);
-#undef STXN
-    result->id64 = toku_txn_id64;
+    txn_func_init(result);
 
     result->parent = stxn;
 #if !TOKUDB_NATIVE_H
@@ -477,16 +485,7 @@ void toku_keep_prepared_txn_callback (DB_ENV *env, TOKUTXN tokutxn) {
     memset(eresult, 0, sizeof(*eresult));
     DB_TXN *result = &eresult->external_part;
     result->mgrp = env;
-#define STXN(name) result->name = locked_txn_ ## name
-    STXN(abort);
-    STXN(commit);
-    STXN(abort_with_progress);
-    STXN(commit_with_progress);
-    STXN(id);
-    STXN(prepare);
-    STXN(txn_stat);
-#undef STXN
-    result->id64 = toku_txn_id64;
+    txn_func_init(result);
     
     result->parent = NULL;