WT-3356 Use atomic reads of rwlocks. (#3454)

* WT-3356 Use atomic reads of rwlocks.
  
  Previously we had some conditions that checked several fields within a rwlock by indirecting to the live structure.  Switch to always doing a read of the full 64-bit value, then using local reads from the copy.
  
  Otherwise, we're relying on the compiler and the memory model to order the structure accesses in "code execution order".  That could explain assertion failures and/or incorrect behavior with the new rwlock implementation.

* Change all waits to 10ms.

  Previously when stalling waiting to get into the lock we would wait for 1ms, but once queued we waited forever.  The former is probably too aggressive (burns too much CPU when we should be able to wait for a notification), and the latter is dangerous if a notification is ever lost (a thread with a ticket may never wake up).

bench/wtperf/runners/checkpoint_schema_race.wtperf

@@ -0,0 +1,20 @@
+# Check create and drop behavior concurrent with checkpoints (WT-2798).
+# Setup a multiple tables and a cache size large enough that checkpoints can
+# take a long time.
+conn_config="cache_size=8GB,log=(enabled=false),checkpoint=(wait=30)"
+table_config="leaf_page_max=4k,internal_page_max=16k,type=file"
+icount=10000000
+table_count=100
+table_count_idle=100
+# Turn on create/drop of idle tables, but don't worry if individual operations
+# take a long time.
+idle_table_cycle=120
+populate_threads=5
+checkpoint_threads=0
+report_interval=5
+# 100 million
+random_range=10000000
+run_time=300
+# Setup a workload that dirties a lot of the cache
+threads=((count=2,reads=1),(count=2,inserts=1),(count=2,updates=1))
+value_sz=500

dist/s_prototypes

@@ -42,9 +42,6 @@ proto()
 	    -e '# Add the warn_unused_result attribute to any external' \
 	    -e '# functions that return an int.' \
 	    -e '/^extern int /s/$/ WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result))/' \
-	    -e '# Add the hidden attribute to any external functions without' \
-	    -e '# an explicit visibility.' \
-	    -e '/visibility/!s/$/ WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")))/' \
 	    -e 's/$/;/' \
 	    -e p < $1
 }

src/async/async_api.c

@@ -440,7 +440,7 @@ __wt_async_destroy(WT_SESSION_IMPL *session)
 			    session, async->worker_tids[i]));
 			async->worker_tids[i] = 0;
 		}
-	WT_TRET(__wt_cond_destroy(session, &async->flush_cond));
+	__wt_cond_destroy(session, &async->flush_cond);
 
 	/* Close the server threads' sessions. */
 	for (i = 0; i < conn->async_workers; i++)

src/btree/bt_compact.c

@@ -60,7 +60,7 @@ __compact_rewrite(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
 	 */
 	if (mod->rec_result == WT_PM_REC_REPLACE ||
 	    mod->rec_result == WT_PM_REC_MULTIBLOCK)
-		__wt_writelock(session, &page->page_lock);
+		WT_PAGE_LOCK(session, page);
 
 	if (mod->rec_result == WT_PM_REC_REPLACE)
 		ret = bm->compact_page_skip(bm, session,
@@ -80,7 +80,7 @@ __compact_rewrite(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
 
 	if (mod->rec_result == WT_PM_REC_REPLACE ||
 	    mod->rec_result == WT_PM_REC_MULTIBLOCK)
-		__wt_writeunlock(session, &page->page_lock);
+		WT_PAGE_UNLOCK(session, page);
 
 	return (ret);
 }

src/btree/bt_cursor.c

@@ -369,7 +369,7 @@ __cursor_col_modify(
     WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, bool is_remove)
 {
 	return (__wt_col_modify(session,
-	    cbt, cbt->iface.recno, &cbt->iface.value, NULL, is_remove));
+	    cbt, cbt->iface.recno, &cbt->iface.value, NULL, is_remove, false));
 }
 
 /*
@@ -381,7 +381,7 @@ __cursor_row_modify(
     WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, bool is_remove)
 {
 	return (__wt_row_modify(session,
-	    cbt, &cbt->iface.key, &cbt->iface.value, NULL, is_remove));
+	    cbt, &cbt->iface.key, &cbt->iface.value, NULL, is_remove, false));
 }
 
 /*

src/btree/bt_debug.c

@@ -689,8 +689,6 @@ __debug_page_metadata(WT_DBG *ds, WT_REF *ref)
 	WT_RET(ds->f(ds, ", entries %" PRIu32, entries));
 	WT_RET(ds->f(ds,
 	    ", %s", __wt_page_is_modified(page) ? "dirty" : "clean"));
-	WT_RET(ds->f(ds, ", %s", __wt_rwlock_islocked(
-	    session, &page->page_lock) ? "locked" : "unlocked"));
 
 	if (F_ISSET_ATOMIC(page, WT_PAGE_BUILD_KEYS))
 		WT_RET(ds->f(ds, ", keys-built"));

src/btree/bt_discard.c

@@ -98,7 +98,6 @@ __page_out_int(WT_SESSION_IMPL *session, WT_PAGE **pagep, bool rewrite)
 	 */
 	WT_ASSERT(session, !__wt_page_is_modified(page));
 	WT_ASSERT(session, !F_ISSET_ATOMIC(page, WT_PAGE_EVICT_LRU));
-	WT_ASSERT(session, !__wt_rwlock_islocked(session, &page->page_lock));
 
 	/*
 	 * If a root page split, there may be one or more pages linked from the
@@ -254,6 +253,7 @@ __free_page_modify(WT_SESSION_IMPL *session, WT_PAGE *page)
 	__wt_ovfl_discard_free(session, page);
 
 	__wt_free(session, page->modify->ovfl_track);
+	__wt_spin_destroy(session, &page->modify->page_lock);
 
 	__wt_free(session, page->modify);
 }

src/btree/bt_handle.c

@@ -444,7 +444,7 @@ __btree_conf(WT_SESSION_IMPL *session, WT_CKPT *ckpt)
 	}
 
 	/* Initialize locks. */
-	__wt_rwlock_init(session, &btree->ovfl_lock);
+	WT_RET(__wt_rwlock_init(session, &btree->ovfl_lock));
 	WT_RET(__wt_spin_init(session, &btree->flush_lock, "btree flush"));
 
 	btree->checkpointing = WT_CKPT_OFF;		/* Not checkpointing */

src/btree/bt_read.c

@@ -90,7 +90,7 @@ __col_instantiate(WT_SESSION_IMPL *session,
 {
 	/* Search the page and add updates. */
 	WT_RET(__wt_col_search(session, recno, ref, cbt));
-	WT_RET(__wt_col_modify(session, cbt, recno, NULL, upd, false));
+	WT_RET(__wt_col_modify(session, cbt, recno, NULL, upd, false, false));
 	return (0);
 }
 
@@ -104,7 +104,7 @@ __row_instantiate(WT_SESSION_IMPL *session,
 {
 	/* Search the page and add updates. */
 	WT_RET(__wt_row_search(session, key, ref, cbt, true));
-	WT_RET(__wt_row_modify(session, cbt, key, NULL, upd, false));
+	WT_RET(__wt_row_modify(session, cbt, key, NULL, upd, false, false));
 	return (0);
 }
 

src/btree/bt_split.c

@@ -1300,13 +1300,19 @@ __split_internal_lock(WT_SESSION_IMPL *session, WT_REF *ref, bool trylock,
 	for (;;) {
 		parent = ref->home;
 
+		/*
+		 * The page will be marked dirty, and we can only lock a page
+		 * with a modify structure.
+		 */
+		WT_RET(__wt_page_modify_init(session, parent));
+
 		if (trylock)
-			WT_RET(__wt_try_writelock(session, &parent->page_lock));
+			WT_RET(WT_PAGE_TRYLOCK(session, parent));
 		else
-			__wt_writelock(session, &parent->page_lock);
+			WT_PAGE_LOCK(session, parent);
 		if (parent == ref->home)
 			break;
-		__wt_writeunlock(session, &parent->page_lock);
+		WT_PAGE_UNLOCK(session, parent);
 	}
 
 	/*
@@ -1329,7 +1335,7 @@ __split_internal_lock(WT_SESSION_IMPL *session, WT_REF *ref, bool trylock,
 	*parentp = parent;
 	return (0);
 
-err:	__wt_writeunlock(session, &parent->page_lock);
+err:	WT_PAGE_UNLOCK(session, parent);
 	return (ret);
 }
 
@@ -1345,7 +1351,7 @@ __split_internal_unlock(WT_SESSION_IMPL *session, WT_PAGE *parent, bool hazard)
 	if (hazard)
 		ret = __wt_hazard_clear(session, parent->pg_intl_parent_ref);
 
-	__wt_writeunlock(session, &parent->page_lock);
+	WT_PAGE_UNLOCK(session, parent);
 	return (ret);
 }
 
@@ -1559,7 +1565,7 @@ __split_multi_inmem(
 
 			/* Apply the modification. */
 			WT_ERR(__wt_col_modify(
-			    session, &cbt, recno, NULL, upd, false));
+			    session, &cbt, recno, NULL, upd, false, true));
 			break;
 		case WT_PAGE_ROW_LEAF:
 			/* Build a key. */
@@ -1581,7 +1587,7 @@ __split_multi_inmem(
 
 			/* Apply the modification. */
 			WT_ERR(__wt_row_modify(
-			    session, &cbt, key, NULL, upd, false));
+			    session, &cbt, key, NULL, upd, false, true));
 			break;
 		WT_ILLEGAL_VALUE_ERR(session);
 		}

src/btree/col_modify.c

@@ -17,7 +17,8 @@ static int __col_insert_alloc(
  */
 int
 __wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
-    uint64_t recno, WT_ITEM *value, WT_UPDATE *upd_arg, bool is_remove)
+    uint64_t recno, WT_ITEM *value,
+    WT_UPDATE *upd_arg, bool is_remove, bool exclusive)
 {
 	WT_BTREE *btree;
 	WT_DECL_RET;
@@ -103,7 +104,7 @@ __wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
 
 		/* Serialize the update. */
 		WT_ERR(__wt_update_serial(
-		    session, page, &cbt->ins->upd, &upd, upd_size));
+		    session, page, &cbt->ins->upd, &upd, upd_size, false));
 	} else {
 		/* Allocate the append/update list reference as necessary. */
 		if (append) {
@@ -185,11 +186,11 @@ __wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
 		if (append)
 			WT_ERR(__wt_col_append_serial(
 			    session, page, cbt->ins_head, cbt->ins_stack,
-			    &ins, ins_size, &cbt->recno, skipdepth));
+			    &ins, ins_size, &cbt->recno, skipdepth, exclusive));
 		else
 			WT_ERR(__wt_insert_serial(
 			    session, page, cbt->ins_head, cbt->ins_stack,
-			    &ins, ins_size, skipdepth));
+			    &ins, ins_size, skipdepth, exclusive));
 	}
 
 	/* If the update was successful, add it to the in-memory log. */

src/btree/row_modify.c

@@ -15,18 +15,13 @@
 int
 __wt_page_modify_alloc(WT_SESSION_IMPL *session, WT_PAGE *page)
 {
-	WT_CONNECTION_IMPL *conn;
+	WT_DECL_RET;
 	WT_PAGE_MODIFY *modify;
 
-	conn = S2C(session);
-
 	WT_RET(__wt_calloc_one(session, &modify));
 
-	/*
-	 * Select a spinlock for the page; let the barrier immediately below
-	 * keep things from racing too badly.
-	 */
-	modify->page_lock = ++conn->page_lock_cnt % WT_PAGE_LOCKS;
+	/* Initialize the spinlock for the page. */
+	WT_ERR(__wt_spin_init(session, &modify->page_lock, "btree page"));
 
 	/*
 	 * Multiple threads of control may be searching and deciding to modify
@@ -37,8 +32,8 @@ __wt_page_modify_alloc(WT_SESSION_IMPL *session, WT_PAGE *page)
 	if (__wt_atomic_cas_ptr(&page->modify, NULL, modify))
 		__wt_cache_page_inmem_incr(session, page, sizeof(*modify));
 	else
-		__wt_free(session, modify);
-	return (0);
+err:		__wt_free(session, modify);
+	return (ret);
 }
 
 /*
@@ -47,7 +42,8 @@ __wt_page_modify_alloc(WT_SESSION_IMPL *session, WT_PAGE *page)
  */
 int
 __wt_row_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
-    WT_ITEM *key, WT_ITEM *value, WT_UPDATE *upd_arg, bool is_remove)
+    WT_ITEM *key, WT_ITEM *value,
+    WT_UPDATE *upd_arg, bool is_remove, bool exclusive)
 {
 	WT_DECL_RET;
 	WT_INSERT *ins;
@@ -132,7 +128,7 @@ __wt_row_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
 
 		/* Serialize the update. */
 		WT_ERR(__wt_update_serial(
-		    session, page, upd_entry, &upd, upd_size));
+		    session, page, upd_entry, &upd, upd_size, exclusive));
 	} else {
 		/*
 		 * Allocate the insert array as necessary.
@@ -207,7 +203,7 @@ __wt_row_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
 		/* Insert the WT_INSERT structure. */
 		WT_ERR(__wt_insert_serial(
 		    session, page, cbt->ins_head, cbt->ins_stack,
-		    &ins, ins_size, skipdepth));
+		    &ins, ins_size, skipdepth, exclusive));
 	}
 
 	if (logged)

src/conn/conn_cache.c

@@ -312,7 +312,7 @@ __wt_cache_destroy(WT_SESSION_IMPL *session)
 		    cache->bytes_dirty_intl + cache->bytes_dirty_leaf,
 		    cache->pages_dirty_intl + cache->pages_dirty_leaf);
 
-	WT_TRET(__wt_cond_destroy(session, &cache->evict_cond));
+	__wt_cond_destroy(session, &cache->evict_cond);
 	__wt_spin_destroy(session, &cache->evict_pass_lock);
 	__wt_spin_destroy(session, &cache->evict_queue_lock);
 	__wt_spin_destroy(session, &cache->evict_walk_lock);

src/conn/conn_cache_pool.c

@@ -225,7 +225,7 @@ err:	__wt_spin_unlock(session, &__wt_process.spinlock);
 		__wt_free(session, pool_name);
 	if (ret != 0 && created) {
 		__wt_free(session, cp->name);
-		WT_TRET(__wt_cond_destroy(session, &cp->cache_pool_cond));
+		__wt_cond_destroy(session, &cp->cache_pool_cond);
 		__wt_free(session, cp);
 	}
 	return (ret);
@@ -391,7 +391,7 @@ __wt_conn_cache_pool_destroy(WT_SESSION_IMPL *session)
 		__wt_free(session, cp->name);
 
 		__wt_spin_destroy(session, &cp->cache_pool_lock);
-		WT_TRET(__wt_cond_destroy(session, &cp->cache_pool_cond));
+		__wt_cond_destroy(session, &cp->cache_pool_cond);
 		__wt_free(session, cp);
 	}
 

src/conn/conn_ckpt.c

@@ -22,6 +22,8 @@ __ckpt_server_config(WT_SESSION_IMPL *session, const char **cfg, bool *startp)
 
 	*startp = false;
 
+	*startp = false;
+
 	conn = S2C(session);
 
 	WT_RET(__wt_config_gets(session, cfg, "checkpoint.wait", &cval));
@@ -231,7 +233,7 @@ __wt_checkpoint_server_destroy(WT_SESSION_IMPL *session)
 		WT_TRET(__wt_thread_join(session, conn->ckpt_tid));
 		conn->ckpt_tid_set = false;
 	}
-	WT_TRET(__wt_cond_destroy(session, &conn->ckpt_cond));
+	__wt_cond_destroy(session, &conn->ckpt_cond);
 
 	/* Close the server thread's session. */
 	if (conn->ckpt_session != NULL) {

src/conn/conn_dhandle.c

@@ -52,7 +52,7 @@ __wt_conn_dhandle_alloc(
 
 	WT_RET(__wt_calloc_one(session, &dhandle));
 
-	__wt_rwlock_init(session, &dhandle->rwlock);
+	WT_ERR(__wt_rwlock_init(session, &dhandle->rwlock));
 	dhandle->name_hash = __wt_hash_city64(uri, strlen(uri));
 	WT_ERR(__wt_strdup(session, uri, &dhandle->name));
 	WT_ERR(__wt_strdup(session, checkpoint, &dhandle->checkpoint));

src/conn/conn_handle.c

@@ -62,14 +62,9 @@ __wt_connection_init(WT_CONNECTION_IMPL *conn)
 	WT_RET(__wt_spin_init(session, &conn->turtle_lock, "turtle file"));
 
 	/* Read-write locks */
-	__wt_rwlock_init(session, &conn->dhandle_lock);
-	__wt_rwlock_init(session, &conn->hot_backup_lock);
-	__wt_rwlock_init(session, &conn->table_lock);
-
-	WT_RET(__wt_calloc_def(session, WT_PAGE_LOCKS, &conn->page_lock));
-	for (i = 0; i < WT_PAGE_LOCKS; ++i)
-		WT_RET(
-		    __wt_spin_init(session, &conn->page_lock[i], "btree page"));
+	WT_RET(__wt_rwlock_init(session, &conn->dhandle_lock));
+	WT_RET(__wt_rwlock_init(session, &conn->hot_backup_lock));
+	WT_RET(__wt_rwlock_init(session, &conn->table_lock));
 
 	/* Setup the spin locks for the LSM manager queues. */
 	WT_RET(__wt_spin_init(session,
@@ -113,7 +108,6 @@ void
 __wt_connection_destroy(WT_CONNECTION_IMPL *conn)
 {
 	WT_SESSION_IMPL *session;
-	u_int i;
 
 	/* Check there's something to destroy. */
 	if (conn == NULL)
@@ -144,9 +138,6 @@ __wt_connection_destroy(WT_CONNECTION_IMPL *conn)
 	__wt_spin_destroy(session, &conn->schema_lock);
 	__wt_rwlock_destroy(session, &conn->table_lock);
 	__wt_spin_destroy(session, &conn->turtle_lock);
-	for (i = 0; i < WT_PAGE_LOCKS; ++i)
-		__wt_spin_destroy(session, &conn->page_lock[i]);
-	__wt_free(session, conn->page_lock);
 
 	/* Free allocated memory. */
 	__wt_free(session, conn->cfg);

src/conn/conn_log.c

@@ -880,7 +880,7 @@ __wt_logmgr_create(WT_SESSION_IMPL *session, const char *cfg[])
 	WT_RET(__wt_spin_init(session, &log->log_sync_lock, "log sync"));
 	WT_RET(__wt_spin_init(session, &log->log_writelsn_lock,
 	    "log write LSN"));
-	__wt_rwlock_init(session, &log->log_archive_lock);
+	WT_RET(__wt_rwlock_init(session, &log->log_archive_lock));
 	if (FLD_ISSET(conn->direct_io, WT_DIRECT_IO_LOG))
 		log->allocsize = (uint32_t)
 		    WT_MAX(conn->buffer_alignment, WT_LOG_ALIGN);
@@ -1043,12 +1043,12 @@ __wt_logmgr_destroy(WT_SESSION_IMPL *session)
 	}
 
 	/* Destroy the condition variables now that all threads are stopped */
-	WT_TRET(__wt_cond_destroy(session, &conn->log_cond));
-	WT_TRET(__wt_cond_destroy(session, &conn->log_file_cond));
-	WT_TRET(__wt_cond_destroy(session, &conn->log_wrlsn_cond));
+	__wt_cond_destroy(session, &conn->log_cond);
+	__wt_cond_destroy(session, &conn->log_file_cond);
+	__wt_cond_destroy(session, &conn->log_wrlsn_cond);
 
-	WT_TRET(__wt_cond_destroy(session, &conn->log->log_sync_cond));
-	WT_TRET(__wt_cond_destroy(session, &conn->log->log_write_cond));
+	__wt_cond_destroy(session, &conn->log->log_sync_cond);
+	__wt_cond_destroy(session, &conn->log->log_write_cond);
 	__wt_rwlock_destroy(session, &conn->log->log_archive_lock);
 	__wt_spin_destroy(session, &conn->log->log_lock);
 	__wt_spin_destroy(session, &conn->log->log_slot_lock);

src/conn/conn_stat.c

@@ -648,7 +648,7 @@ __wt_statlog_destroy(WT_SESSION_IMPL *session, bool is_close)
 		WT_TRET(__wt_thread_join(session, conn->stat_tid));
 		conn->stat_tid_set = false;
 	}
-	WT_TRET(__wt_cond_destroy(session, &conn->stat_cond));
+	__wt_cond_destroy(session, &conn->stat_cond);
 
 	/* Log a set of statistics on shutdown if configured. */
 	if (is_close)

src/conn/conn_sweep.c

@@ -432,7 +432,7 @@ __wt_sweep_destroy(WT_SESSION_IMPL *session)
 		WT_TRET(__wt_thread_join(session, conn->sweep_tid));
 		conn->sweep_tid_set = 0;
 	}
-	WT_TRET(__wt_cond_destroy(session, &conn->sweep_cond));
+	__wt_cond_destroy(session, &conn->sweep_cond);
 
 	if (conn->sweep_session != NULL) {
 		wt_session = &conn->sweep_session->iface;

src/cursor/cur_file.c

@@ -486,7 +486,14 @@ __curfile_create(WT_SESSION_IMPL *session,
 	WT_STAT_DATA_INCR(session, cursor_create);
 
 	if (0) {
-err:		WT_TRET(__curfile_close(cursor));
+err:		/*
+		 * Our caller expects to release the data handle if we fail.
+		 * Disconnect it from the cursor before closing.
+		 */
+		if (session->dhandle != NULL)
+			__wt_cursor_dhandle_decr_use(session);
+		cbt->btree = NULL;
+		WT_TRET(__curfile_close(cursor));
 		*cursorp = NULL;
 	}
 

src/include/btmem.h

@@ -414,18 +414,20 @@ struct __wt_page_modify {
 		size_t	  discard_allocated;
 	} *ovfl_track;
 
+#define	WT_PAGE_LOCK(s, p) 						\
+	__wt_spin_lock((s), &(p)->modify->page_lock)
+#define	WT_PAGE_TRYLOCK(s, p)						\
+	__wt_spin_trylock((s), &(p)->modify->page_lock)
+#define	WT_PAGE_UNLOCK(s, p) 						\
+	__wt_spin_unlock((s), &(p)->modify->page_lock)
+	WT_SPINLOCK page_lock;		/* Page's spinlock */
+
 	/*
 	 * The write generation is incremented when a page is modified, a page
 	 * is clean if the write generation is 0.
 	 */
 	uint32_t write_gen;
 
-#define	WT_PAGE_LOCK(s, p)						\
-	__wt_spin_lock((s), &S2C(s)->page_lock[(p)->modify->page_lock])
-#define	WT_PAGE_UNLOCK(s, p)						\
-	__wt_spin_unlock((s), &S2C(s)->page_lock[(p)->modify->page_lock])
-	uint8_t page_lock;		/* Page's spinlock */
-
 #define	WT_PM_REC_EMPTY		1	/* Reconciliation: no replacement */
 #define	WT_PM_REC_MULTIBLOCK	2	/* Reconciliation: multiple blocks */
 #define	WT_PM_REC_REPLACE	3	/* Reconciliation: single block */
@@ -602,13 +604,6 @@ struct __wt_page {
 
 	uint8_t unused[2];		/* Unused padding */
 
-	/*
-	 * Used to protect and co-ordinate splits for internal pages and
-	 * reconciliation for all pages. Only used to co-ordinate among the
-	 * uncommon cases that require exclusive access to a page.
-	 */
-	WT_RWLOCK page_lock;
-
 	/*
 	 * The page's read generation acts as an LRU value for each page in the
 	 * tree; it is used by the eviction server thread to select pages to be
@@ -635,8 +630,6 @@ struct __wt_page {
 #define	WT_READGEN_STEP		100
 	uint64_t read_gen;
 
-	uint64_t evict_pass_gen;	/* Eviction pass generation */
-
 	size_t memory_footprint;	/* Memory attached to the page */
 
 	/* Page's on-disk representation: NULL for pages created in memory. */
@@ -644,6 +637,10 @@ struct __wt_page {
 
 	/* If/when the page is modified, we need lots more information. */
 	WT_PAGE_MODIFY *modify;
+
+	/* This is the 64 byte boundary, try to keep hot fields above here. */
+
+	uint64_t evict_pass_gen;	/* Eviction pass generation */
 };
 
 /*

src/include/connection.h

@@ -175,21 +175,6 @@ struct __wt_connection_impl {
 	WT_SPINLOCK turtle_lock;	/* Turtle file spinlock */
 	WT_RWLOCK dhandle_lock;		/* Data handle list lock */
 
-	/*
-	 * We distribute the btree page locks across a set of spin locks. Don't
-	 * use too many: they are only held for very short operations, each one
-	 * is 64 bytes, so 256 will fill the L1 cache on most CPUs.
-	 *
-	 * Use a prime number of buckets rather than assuming a good hash
-	 * (Reference Sedgewick, Algorithms in C, "Hash Functions").
-	 *
-	 * Note: this can't be an array, we impose cache-line alignment and gcc
-	 * doesn't support that for arrays smaller than the alignment.
-	 */
-#define	WT_PAGE_LOCKS		17
-	WT_SPINLOCK *page_lock;	        /* Btree page spinlocks */
-	u_int	     page_lock_cnt;	/* Next spinlock to use */
-
 					/* Connection queue */
 	TAILQ_ENTRY(__wt_connection_impl) q;
 					/* Cache pool queue */

src/include/extern_posix.h

@@ -1,32 +1,32 @@
 /* DO NOT EDIT: automatically built by dist/s_prototypes. */
 
-extern int __wt_posix_directory_list(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, const char *directory, const char *prefix, char ***dirlistp, uint32_t *countp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_posix_directory_list_free(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, char **dirlist, uint32_t count) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_dlopen(WT_SESSION_IMPL *session, const char *path, WT_DLH **dlhp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_dlsym(WT_SESSION_IMPL *session, WT_DLH *dlh, const char *name, bool fail, void *sym_ret) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_dlclose(WT_SESSION_IMPL *session, WT_DLH *dlh) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_posix_file_extend( WT_FILE_HANDLE *file_handle, WT_SESSION *wt_session, wt_off_t offset) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_os_posix(WT_SESSION_IMPL *session) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
+extern int __wt_posix_directory_list(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, const char *directory, const char *prefix, char ***dirlistp, uint32_t *countp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_posix_directory_list_free(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, char **dirlist, uint32_t count) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_dlopen(WT_SESSION_IMPL *session, const char *path, WT_DLH **dlhp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_dlsym(WT_SESSION_IMPL *session, WT_DLH *dlh, const char *name, bool fail, void *sym_ret) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_dlclose(WT_SESSION_IMPL *session, WT_DLH *dlh) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_posix_file_extend( WT_FILE_HANDLE *file_handle, WT_SESSION *wt_session, wt_off_t offset) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_os_posix(WT_SESSION_IMPL *session) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern int __wt_getenv(WT_SESSION_IMPL *session, const char *variable, const char **envp) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default"))) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
-extern int __wt_posix_map(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, void *mapped_regionp, size_t *lenp, void *mapped_cookiep) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_posix_map_preload(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, const void *map, size_t length, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_posix_map_discard(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, void *map, size_t length, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_posix_unmap(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, void *mapped_region, size_t len, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_cond_alloc(WT_SESSION_IMPL *session, const char *name, WT_CONDVAR **condp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_cond_wait_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond, uint64_t usecs, bool (*run_func)(WT_SESSION_IMPL *), bool *signalled) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_cond_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_cond_destroy(WT_SESSION_IMPL *session, WT_CONDVAR **condp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_once(void (*init_routine)(void)) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_get_vm_pagesize(void) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern bool __wt_absolute_path(const char *path) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern const char *__wt_path_separator(void) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern bool __wt_has_priv(void) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
+extern int __wt_posix_map(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, void *mapped_regionp, size_t *lenp, void *mapped_cookiep) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_posix_map_preload(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, const void *map, size_t length, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_posix_map_discard(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, void *map, size_t length, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_posix_unmap(WT_FILE_HANDLE *fh, WT_SESSION *wt_session, void *mapped_region, size_t len, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_cond_alloc(WT_SESSION_IMPL *session, const char *name, WT_CONDVAR **condp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern void __wt_cond_wait_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond, uint64_t usecs, bool (*run_func)(WT_SESSION_IMPL *), bool *signalled);
+extern void __wt_cond_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond);
+extern void __wt_cond_destroy(WT_SESSION_IMPL *session, WT_CONDVAR **condp);
+extern int __wt_once(void (*init_routine)(void)) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_get_vm_pagesize(void) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern bool __wt_absolute_path(const char *path);
+extern const char *__wt_path_separator(void);
+extern bool __wt_has_priv(void);
 extern void __wt_stream_set_line_buffer(FILE *fp) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default")));
 extern void __wt_stream_set_no_buffer(FILE *fp) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default")));
 extern void __wt_sleep(uint64_t seconds, uint64_t micro_seconds) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default")));
 extern int __wt_vsnprintf_len_incr( char *buf, size_t size, size_t *retsizep, const char *fmt, va_list ap) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default"))) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
-extern int __wt_thread_create(WT_SESSION_IMPL *session, wt_thread_t *tidret, WT_THREAD_CALLBACK(*func)(void *), void *arg) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_thread_join(WT_SESSION_IMPL *session, wt_thread_t tid) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
+extern int __wt_thread_create(WT_SESSION_IMPL *session, wt_thread_t *tidret, WT_THREAD_CALLBACK(*func)(void *), void *arg) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_thread_join(WT_SESSION_IMPL *session, wt_thread_t tid) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern int __wt_thread_id(char *buf, size_t buflen) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default"))) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
 extern void __wt_epoch(WT_SESSION_IMPL *session, struct timespec *tsp) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default")));
 extern void __wt_yield(void) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("default")));

src/include/extern_win.h

@@ -1,35 +1,35 @@
 /* DO NOT EDIT: automatically built by dist/s_prototypes. */
 
-extern int __wt_win_directory_list(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, const char *directory, const char *prefix, char ***dirlistp, uint32_t *countp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_win_directory_list_free(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, char **dirlist, uint32_t count) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_dlopen(WT_SESSION_IMPL *session, const char *path, WT_DLH **dlhp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_dlsym(WT_SESSION_IMPL *session, WT_DLH *dlh, const char *name, bool fail, void *sym_ret) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_dlclose(WT_SESSION_IMPL *session, WT_DLH *dlh) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_win_fs_size(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, const char *name, wt_off_t *sizep) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_os_win(WT_SESSION_IMPL *session) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_getenv(WT_SESSION_IMPL *session, const char *variable, const char **envp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_win_map(WT_FILE_HANDLE *file_handle, WT_SESSION *wt_session, void *mapped_regionp, size_t *lenp, void *mapped_cookiep) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_win_unmap(WT_FILE_HANDLE *file_handle, WT_SESSION *wt_session, void *mapped_region, size_t length, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_cond_alloc(WT_SESSION_IMPL *session, const char *name, WT_CONDVAR **condp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_cond_wait_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond, uint64_t usecs, bool (*run_func)(WT_SESSION_IMPL *), bool *signalled) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_cond_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_cond_destroy(WT_SESSION_IMPL *session, WT_CONDVAR **condp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_once(void (*init_routine)(void)) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_get_vm_pagesize(void) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern bool __wt_absolute_path(const char *path) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern const char *__wt_path_separator(void) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern bool __wt_has_priv(void) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_stream_set_line_buffer(FILE *fp) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_stream_set_no_buffer(FILE *fp) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_sleep(uint64_t seconds, uint64_t micro_seconds) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_vsnprintf_len_incr( char *buf, size_t size, size_t *retsizep, const char *fmt, va_list ap) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_thread_create(WT_SESSION_IMPL *session, wt_thread_t *tidret, WT_THREAD_CALLBACK(*func)(void *), void *arg) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_thread_join(WT_SESSION_IMPL *session, wt_thread_t tid) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_thread_id(char *buf, size_t buflen) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_epoch(WT_SESSION_IMPL *session, struct timespec *tsp) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_to_utf16_string( WT_SESSION_IMPL *session, const char*utf8, WT_ITEM **outbuf) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_to_utf8_string( WT_SESSION_IMPL *session, const wchar_t*wide, WT_ITEM **outbuf) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern DWORD __wt_getlasterror(void) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern int __wt_map_windows_error(DWORD windows_error) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result)) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern const char *__wt_formatmessage(WT_SESSION_IMPL *session, DWORD windows_error) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
-extern void __wt_yield(void) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
+extern int __wt_win_directory_list(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, const char *directory, const char *prefix, char ***dirlistp, uint32_t *countp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_win_directory_list_free(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, char **dirlist, uint32_t count) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_dlopen(WT_SESSION_IMPL *session, const char *path, WT_DLH **dlhp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_dlsym(WT_SESSION_IMPL *session, WT_DLH *dlh, const char *name, bool fail, void *sym_ret) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_dlclose(WT_SESSION_IMPL *session, WT_DLH *dlh) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_win_fs_size(WT_FILE_SYSTEM *file_system, WT_SESSION *wt_session, const char *name, wt_off_t *sizep) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_os_win(WT_SESSION_IMPL *session) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_getenv(WT_SESSION_IMPL *session, const char *variable, const char **envp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_win_map(WT_FILE_HANDLE *file_handle, WT_SESSION *wt_session, void *mapped_regionp, size_t *lenp, void *mapped_cookiep) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_win_unmap(WT_FILE_HANDLE *file_handle, WT_SESSION *wt_session, void *mapped_region, size_t length, void *mapped_cookie) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_cond_alloc(WT_SESSION_IMPL *session, const char *name, WT_CONDVAR **condp) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern void __wt_cond_wait_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond, uint64_t usecs, bool (*run_func)(WT_SESSION_IMPL *), bool *signalled);
+extern void __wt_cond_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond);
+extern void __wt_cond_destroy(WT_SESSION_IMPL *session, WT_CONDVAR **condp);
+extern int __wt_once(void (*init_routine)(void)) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_get_vm_pagesize(void) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern bool __wt_absolute_path(const char *path);
+extern const char *__wt_path_separator(void);
+extern bool __wt_has_priv(void);
+extern void __wt_stream_set_line_buffer(FILE *fp);
+extern void __wt_stream_set_no_buffer(FILE *fp);
+extern void __wt_sleep(uint64_t seconds, uint64_t micro_seconds);
+extern int __wt_vsnprintf_len_incr( char *buf, size_t size, size_t *retsizep, const char *fmt, va_list ap) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_thread_create(WT_SESSION_IMPL *session, wt_thread_t *tidret, WT_THREAD_CALLBACK(*func)(void *), void *arg) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_thread_join(WT_SESSION_IMPL *session, wt_thread_t tid) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_thread_id(char *buf, size_t buflen) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern void __wt_epoch(WT_SESSION_IMPL *session, struct timespec *tsp);
+extern int __wt_to_utf16_string( WT_SESSION_IMPL *session, const char*utf8, WT_ITEM **outbuf) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern int __wt_to_utf8_string( WT_SESSION_IMPL *session, const wchar_t*wide, WT_ITEM **outbuf) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern DWORD __wt_getlasterror(void);
+extern int __wt_map_windows_error(DWORD windows_error) WT_GCC_FUNC_DECL_ATTRIBUTE((warn_unused_result));
+extern const char *__wt_formatmessage(WT_SESSION_IMPL *session, DWORD windows_error);
+extern void __wt_yield(void);

src/include/mutex.h

@@ -37,17 +37,21 @@ struct __wt_condvar {
  * Don't modify this structure without understanding the read/write locking
  * functions.
  */
-union __wt_rwlock {			/* Read/write lock */
-	uint64_t u;
-	struct {
-		uint32_t wr;		/* Writers and readers */
-	} i;
-	struct {
-		uint16_t writers;	/* Now serving for writers */
-		uint16_t readers;	/* Now serving for readers */
-		uint16_t next;		/* Next available ticket number */
-		uint16_t writers_active;/* Count of active writers */
-	} s;
+struct __wt_rwlock {			/* Read/write lock */
+	volatile union {
+		uint64_t v;			/* Full 64-bit value */
+		struct {
+			uint8_t current;	/* Current ticket */
+			uint8_t next;		/* Next available ticket */
+			uint8_t reader;		/* Read queue ticket */
+			uint8_t __notused;	/* Padding */
+			uint16_t readers_active;/* Count of active readers */
+			uint16_t readers_queued;/* Count of queued readers */
+		} s;
+	} u;
+
+	WT_CONDVAR *cond_readers;	/* Blocking readers */
+	WT_CONDVAR *cond_writers;	/* Blocking writers */
 };
 
 /*
@@ -63,8 +67,8 @@ union __wt_rwlock {			/* Read/write lock */
 #define	SPINLOCK_PTHREAD_MUTEX_ADAPTIVE	3
 
 struct __wt_spinlock {
-	WT_CACHE_LINE_PAD_BEGIN
 #if SPINLOCK_TYPE == SPINLOCK_GCC
+	WT_CACHE_LINE_PAD_BEGIN
 	volatile int lock;
 #elif SPINLOCK_TYPE == SPINLOCK_PTHREAD_MUTEX ||\
 	SPINLOCK_TYPE == SPINLOCK_PTHREAD_MUTEX_ADAPTIVE ||\
@@ -87,5 +91,8 @@ struct __wt_spinlock {
 	int16_t stat_int_usecs_off;	/* waiting server threads offset */
 
 	int8_t initialized;		/* Lock initialized, for cleanup */
+
+#if SPINLOCK_TYPE == SPINLOCK_GCC
 	WT_CACHE_LINE_PAD_END
+#endif
 };

src/include/serial.i

@@ -154,7 +154,7 @@ __col_append_serial_func(WT_SESSION_IMPL *session, WT_INSERT_HEAD *ins_head,
 static inline int
 __wt_col_append_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
     WT_INSERT_HEAD *ins_head, WT_INSERT ***ins_stack, WT_INSERT **new_insp,
-    size_t new_ins_size, uint64_t *recnop, u_int skipdepth)
+    size_t new_ins_size, uint64_t *recnop, u_int skipdepth, bool exclusive)
 {
 	WT_INSERT *new_ins = *new_insp;
 	WT_DECL_RET;
@@ -165,11 +165,16 @@ __wt_col_append_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
 	/* Clear references to memory we now own and must free on error. */
 	*new_insp = NULL;
 
-	/* Acquire the page's spinlock, call the worker function. */
-	WT_PAGE_LOCK(session, page);
+	/*
+	 * Acquire the page's spinlock unless we already have exclusive access.
+	 * Then call the worker function.
+	 */
+	if (!exclusive)
+		WT_PAGE_LOCK(session, page);
 	ret = __col_append_serial_func(
 	    session, ins_head, ins_stack, new_ins, recnop, skipdepth);
-	WT_PAGE_UNLOCK(session, page);
+	if (!exclusive)
+		WT_PAGE_UNLOCK(session, page);
 
 	if (ret != 0) {
 		/* Free unused memory on error. */
@@ -198,7 +203,7 @@ __wt_col_append_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
 static inline int
 __wt_insert_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
     WT_INSERT_HEAD *ins_head, WT_INSERT ***ins_stack, WT_INSERT **new_insp,
-    size_t new_ins_size, u_int skipdepth)
+    size_t new_ins_size, u_int skipdepth, bool exclusive)
 {
 	WT_INSERT *new_ins = *new_insp;
 	WT_DECL_RET;
@@ -220,10 +225,12 @@ __wt_insert_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
 		ret = __insert_simple_func(
 		    session, ins_stack, new_ins, skipdepth);
 	else {
-		WT_PAGE_LOCK(session, page);
+		if (!exclusive)
+			WT_PAGE_LOCK(session, page);
 		ret = __insert_serial_func(
 		    session, ins_head, ins_stack, new_ins, skipdepth);
-		WT_PAGE_UNLOCK(session, page);
+		if (!exclusive)
+			WT_PAGE_UNLOCK(session, page);
 	}
 
 	if (ret != 0) {
@@ -252,7 +259,8 @@ __wt_insert_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
  */
 static inline int
 __wt_update_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
-    WT_UPDATE **srch_upd, WT_UPDATE **updp, size_t upd_size)
+    WT_UPDATE **srch_upd, WT_UPDATE **updp, size_t upd_size,
+    bool exclusive)
 {
 	WT_DECL_RET;
 	WT_UPDATE *obsolete, *upd = *updp;
@@ -295,7 +303,7 @@ __wt_update_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
 	/*
 	 * If there are no subsequent WT_UPDATE structures we are done here.
 	 */
-	if (upd->next == NULL)
+	if (upd->next == NULL || exclusive)
 		return (0);
 
 	/*
@@ -316,11 +324,11 @@ __wt_update_serial(WT_SESSION_IMPL *session, WT_PAGE *page,
 	}
 
 	/* If we can't lock it, don't scan, that's okay. */
-	if (__wt_try_writelock(session, &page->page_lock) != 0)
+	if (WT_PAGE_TRYLOCK(session, page) != 0)
 		return (0);
 
 	obsolete = __wt_update_obsolete_check(session, page, upd->next);
-	__wt_writeunlock(session, &page->page_lock);
+	WT_PAGE_UNLOCK(session, page);
 	if (obsolete != NULL)
 		__wt_update_obsolete_free(session, page, obsolete);
 

src/include/session.h

@@ -97,6 +97,10 @@ struct __wt_session_impl {
 	 */
 	TAILQ_HEAD(__tables, __wt_table) tables;
 
+	/* Current rwlock for callback. */
+	WT_RWLOCK *current_rwlock;
+	uint8_t current_rwticket;
+
 	WT_ITEM	**scratch;		/* Temporary memory for any function */
 	u_int	  scratch_alloc;	/* Currently allocated */
 	size_t	  scratch_cached;	/* Scratch bytes cached */

src/include/verify_build.h

@@ -59,7 +59,6 @@ __wt_verify_build(void)
 	    sizeof(s) > WT_CACHE_LINE_ALIGNMENT ||			\
 	    sizeof(s) % WT_CACHE_LINE_ALIGNMENT == 0)
 	WT_PADDING_CHECK(WT_LOGSLOT);
-	WT_PADDING_CHECK(WT_SPINLOCK);
 	WT_PADDING_CHECK(WT_TXN_STATE);
 
 	/*

src/include/wt_internal.h

@@ -268,6 +268,8 @@ struct __wt_ref;
     typedef struct __wt_ref WT_REF;
 struct __wt_row;
     typedef struct __wt_row WT_ROW;
+struct __wt_rwlock;
+    typedef struct __wt_rwlock WT_RWLOCK;
 struct __wt_salvage_cookie;
     typedef struct __wt_salvage_cookie WT_SALVAGE_COOKIE;
 struct __wt_save_upd;
@@ -302,8 +304,6 @@ union __wt_lsn;
     typedef union __wt_lsn WT_LSN;
 union __wt_rand_state;
     typedef union __wt_rand_state WT_RAND_STATE;
-union __wt_rwlock;
-    typedef union __wt_rwlock WT_RWLOCK;
 /*
  * Forward type declarations for internal types: END
  * DO NOT EDIT: automatically built by dist/s_typedef.

src/lsm/lsm_manager.c

@@ -334,7 +334,7 @@ __wt_lsm_manager_destroy(WT_SESSION_IMPL *session)
 	__wt_spin_destroy(session, &manager->switch_lock);
 	__wt_spin_destroy(session, &manager->app_lock);
 	__wt_spin_destroy(session, &manager->manager_lock);
-	WT_TRET(__wt_cond_destroy(session, &manager->work_cond));
+	__wt_cond_destroy(session, &manager->work_cond);
 
 	return (ret);
 }

src/lsm/lsm_tree.c

@@ -471,7 +471,7 @@ __lsm_tree_open(WT_SESSION_IMPL *session,
 
 	/* Try to open the tree. */
 	WT_RET(__wt_calloc_one(session, &lsm_tree));
-	__wt_rwlock_init(session, &lsm_tree->rwlock);
+	WT_ERR(__wt_rwlock_init(session, &lsm_tree->rwlock));
 
 	WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
 

src/lsm/lsm_work_unit.c

@@ -328,8 +328,9 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session,
 	 */
 	saved_isolation = session->txn.isolation;
 	session->txn.isolation = WT_ISO_READ_UNCOMMITTED;
-	WT_ERR(__wt_cache_op(session, WT_SYNC_WRITE_LEAVES));
+	ret = __wt_cache_op(session, WT_SYNC_WRITE_LEAVES);
 	session->txn.isolation = saved_isolation;
+	WT_ERR(ret);
 
 	__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointing %s",
 	    chunk->uri);

src/os_posix/os_mtx_cond.c

@@ -153,7 +153,7 @@ err:
  * __wt_cond_destroy --
  *	Destroy a condition variable.
  */
-int
+void
 __wt_cond_destroy(WT_SESSION_IMPL *session, WT_CONDVAR **condp)
 {
 	WT_CONDVAR *cond;
@@ -161,11 +161,15 @@ __wt_cond_destroy(WT_SESSION_IMPL *session, WT_CONDVAR **condp)
 
 	cond = *condp;
 	if (cond == NULL)
-		return (0);
+		return;
+
+	if ((ret = pthread_cond_destroy(&cond->cond)) != 0)
+		WT_PANIC_MSG(
+		    session, ret, "pthread_cond_destroy: %s", cond->name);
+
+	if ((ret = pthread_mutex_destroy(&cond->mtx)) != 0)
+		WT_PANIC_MSG(
+		    session, ret, "pthread_mutex_destroy: %s", cond->name);
 
-	ret = pthread_cond_destroy(&cond->cond);
-	WT_TRET(pthread_mutex_destroy(&cond->mtx));
 	__wt_free(session, *condp);
-
-	return (ret);
 }

src/os_win/os_mtx_cond.c

@@ -163,18 +163,16 @@ __wt_cond_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond)
  * __wt_cond_destroy --
  *	Destroy a condition variable.
  */
-int
+void
 __wt_cond_destroy(WT_SESSION_IMPL *session, WT_CONDVAR **condp)
 {
 	WT_CONDVAR *cond;
 
 	cond = *condp;
 	if (cond == NULL)
-		return (0);
+		return;
 
 	/* Do nothing to delete Condition Variable */
 	DeleteCriticalSection(&cond->mtx);
 	__wt_free(session, *condp);
-
-	return (0);
 }

src/reconcile/rec_write.c

@@ -386,7 +386,7 @@ __wt_reconcile(WT_SESSION_IMPL *session, WT_REF *ref,
 	 *    In-memory splits: reconciliation of an internal page cannot handle
 	 * a child page splitting during the reconciliation.
 	 */
-	__wt_writelock(session, &page->page_lock);
+	WT_PAGE_LOCK(session, page);
 
 	oldest_id = __wt_txn_oldest_id(session);
 	if (LF_ISSET(WT_EVICTING))
@@ -405,7 +405,7 @@ __wt_reconcile(WT_SESSION_IMPL *session, WT_REF *ref,
 	/* Initialize the reconciliation structure for each new run. */
 	if ((ret = __rec_write_init(
 	    session, ref, flags, salvage, &session->reconcile)) != 0) {
-		__wt_writeunlock(session, &page->page_lock);
+		WT_PAGE_UNLOCK(session, page);
 		return (ret);
 	}
 	r = session->reconcile;
@@ -446,7 +446,7 @@ __wt_reconcile(WT_SESSION_IMPL *session, WT_REF *ref,
 		WT_TRET(__rec_write_wrapup_err(session, r, page));
 
 	/* Release the reconciliation lock. */
-	__wt_writeunlock(session, &page->page_lock);
+	WT_PAGE_UNLOCK(session, page);
 
 	/*
 	 * If our caller can configure lookaside table reconciliation, flag if

src/session/session_dhandle.c

@@ -261,8 +261,8 @@ __wt_session_release_btree(WT_SESSION_IMPL *session)
 	 * can get a handle without special flags.
 	 */
 	if (F_ISSET(dhandle, WT_DHANDLE_DISCARD | WT_DHANDLE_DISCARD_FORCE)) {
-		__session_find_dhandle(session,
-		    dhandle->name, dhandle->checkpoint, &dhandle_cache);
+		WT_SAVE_DHANDLE(session, __session_find_dhandle(session,
+		    dhandle->name, dhandle->checkpoint, &dhandle_cache));
 		if (dhandle_cache != NULL)
 			__session_discard_dhandle(session, dhandle_cache);
 	}

src/support/mtx_rw.c

@@ -27,7 +27,7 @@
  */
 
 /*
- * Based on "Spinlocks and Read-Write Locks" by Dr. Steven Fuerst:
+ * Inspired by "Spinlocks and Read-Write Locks" by Dr. Steven Fuerst:
  *	http://locklessinc.com/articles/locks/
  *
  * Dr. Fuerst further credits:
@@ -39,77 +39,46 @@
  * by John Mellor-Crummey and Michael Scott in their landmark paper "Scalable
  * Reader-Writer Synchronization for Shared-Memory Multiprocessors".
  *
- * The following is an explanation of this code. First, the underlying lock
- * structure.
+ * The following is an explanation of our interpretation and implementation.
+ * First, the underlying lock structure.
  *
+ * volatile union {
+ *	uint64_t v;				// Full 64-bit value
  *	struct {
- *		uint16_t writers;	Now serving for writers
- *		uint16_t readers;	Now serving for readers
- *		uint16_t next;		Next available ticket number
- *		uint16_t __notused;	Padding
- *	}
+ *		uint8_t current;		// Current ticket
+ *		uint8_t next;			// Next available ticket
+ *		uint8_t reader;			// Read queue ticket
+ *		uint8_t __notused;		// Padding
+ *		uint16_t readers_active;	// Count of active readers
+ *		uint16_t readers_queued;	// Count of queued readers
+ *	} s;
+ * } u;
  *
  * First, imagine a store's 'take a number' ticket algorithm. A customer takes
  * a unique ticket number and customers are served in ticket order. In the data
- * structure, 'writers' is the next writer to be served, 'readers' is the next
- * reader to be served, and 'next' is the next available ticket number.
+ * structure, 'next' is the ticket that will be allocated next, and 'current'
+ * is the ticket being served.
  *
- * Next, consider exclusive (write) locks. The 'now serving' number for writers
- * is 'writers'. To lock, 'take a number' and wait until that number is being
- * served; more specifically, atomically copy and increment the current value of
- * 'next', and then wait until 'writers' equals that copied number.
+ * Next, consider exclusive (write) locks.  To lock, 'take a number' and wait
+ * until that number is being served; more specifically, atomically increment
+ * 'next', and then wait until 'current' equals that allocated ticket.
  *
- * Shared (read) locks are similar. Like writers, readers atomically get the
- * next number available. However, instead of waiting for 'writers' to equal
- * their number, they wait for 'readers' to equal their number.
+ * Shared (read) locks are similar, except that readers can share a ticket
+ * (both with each other and with a single writer).  Readers with a given
+ * ticket execute before the writer with that ticket.  In other words, writers
+ * wait for both their ticket to become current and for all readers to exit
+ * the lock.
  *
- * This has the effect of queuing lock requests in the order they arrive
- * (incidentally avoiding starvation).
+ * If there are no active writers (indicated by 'current' == 'next'), readers
+ * can immediately enter the lock by atomically incrementing 'readers_active'.
+ * When there are writers active, readers form a new queue by first setting
+ * 'reader' to 'next' (i.e. readers are scheduled after any queued writers,
+ * avoiding starvation), then atomically incrementing 'readers_queued'.
  *
- * Each lock/unlock pair requires incrementing both 'readers' and 'writers'.
- * In the case of a reader, the 'readers' increment happens when the reader
- * acquires the lock (to allow read-lock sharing), and the 'writers' increment
- * happens when the reader releases the lock. In the case of a writer, both
- * 'readers' and 'writers' are incremented when the writer releases the lock.
- *
- * For example, consider the following read (R) and write (W) lock requests:
- *
- *						writers	readers	next
- *						0	0	0
- *	R: ticket 0, readers match	OK	0	1	1
- *	R: ticket 1, readers match	OK	0	2	2
- *	R: ticket 2, readers match	OK	0	3	3
- *	W: ticket 3, writers no match	block	0	3	4
- *	R: ticket 2, unlock			1	3	4
- *	R: ticket 0, unlock			2	3	4
- *	R: ticket 1, unlock			3	3	4
- *	W: ticket 3, writers match	OK	3	3	4
- *
- * Note the writer blocks until 'writers' equals its ticket number and it does
- * not matter if readers unlock in order or not.
- *
- * Readers or writers entering the system after the write lock is queued block,
- * and the next ticket holder (reader or writer) will unblock when the writer
- * unlocks. An example, continuing from the last line of the above example:
- *
- *						writers	readers	next
- *	W: ticket 3, writers match	OK	3	3	4
- *	R: ticket 4, readers no match	block	3	3	5
- *	R: ticket 5, readers no match	block	3	3	6
- *	W: ticket 6, writers no match	block	3	3	7
- *	W: ticket 3, unlock			4	4	7
- *	R: ticket 4, readers match	OK	4	5	7
- *	R: ticket 5, readers match	OK	4	6	7
- *
- * The 'next' field is a 2-byte value so the available ticket number wraps at
- * 64K requests. If a thread's lock request is not granted until the 'next'
- * field cycles and the same ticket is taken by another thread, we could grant
- * a lock to two separate threads at the same time, and bad things happen: two
- * writer threads or a reader thread and a writer thread would run in parallel,
- * and lock waiters could be skipped if the unlocks race. This is unlikely, it
- * only happens if a lock request is blocked by 64K other requests. The fix is
- * to grow the lock structure fields, but the largest atomic instruction we have
- * is 8 bytes, the structure has no room to grow.
+ * The 'next' field is a 1-byte value so the available ticket number wraps
+ * after 256 requests. If a thread's write lock request would cause the 'next'
+ * field to catch up with 'current', instead it waits to avoid the same ticket
+ * being allocated to multiple threads.
  */
 
 #include "wt_internal.h"
@@ -118,12 +87,14 @@
  * __wt_rwlock_init --
  *	Initialize a read/write lock.
  */
-void
+int
 __wt_rwlock_init(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 {
-	WT_UNUSED(session);
+	l->u.v = 0;
 
-	l->u = 0;
+	WT_RET(__wt_cond_alloc(session, "rwlock wait", &l->cond_readers));
+	WT_RET(__wt_cond_alloc(session, "rwlock wait", &l->cond_writers));
+	return (0);
 }
 
 /*
@@ -133,9 +104,10 @@ __wt_rwlock_init(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 void
 __wt_rwlock_destroy(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 {
-	WT_UNUSED(session);
+	l->u.v = 0;
 
-	l->u = 0;
+	__wt_cond_destroy(session, &l->cond_readers);
+	__wt_cond_destroy(session, &l->cond_writers);
 }
 
 /*
@@ -149,46 +121,34 @@ __wt_try_readlock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 
 	WT_STAT_CONN_INCR(session, rwlock_read);
 
-	new = old = *l;
+	old.u.v = l->u.v;
 
-	/*
-	 * This read lock can only be granted if the lock was last granted to
-	 * a reader and there are no readers or writers blocked on the lock,
-	 * that is, if this thread's ticket would be the next ticket granted.
-	 * Do the cheap test to see if this can possibly succeed (and confirm
-	 * the lock is in the correct state to grant this read lock).
-	 */
-	if (old.s.readers != old.s.next)
+	/* This read lock can only be granted if there are no active writers. */
+	if (old.u.s.current != old.u.s.next)
 		return (EBUSY);
 
 	/*
-	 * The replacement lock value is a result of allocating a new ticket and
-	 * incrementing the reader value to match it.
+	 * The replacement lock value is a result of adding an active reader.
+	 * Check for overflow: if the maximum number of readers are already
+	 * active, no new readers can enter the lock.
 	 */
-	new.s.readers = new.s.next = old.s.next + 1;
-	return (__wt_atomic_cas64(&l->u, old.u, new.u) ? 0 : EBUSY);
+	new.u.v = old.u.v;
+	if (++new.u.s.readers_active == 0)
+		return (EBUSY);
+
+	/* We rely on this atomic operation to provide a barrier. */
+	return (__wt_atomic_casv64(&l->u.v, old.u.v, new.u.v) ? 0 : EBUSY);
 }
 
 /*
- * __wt_readlock_spin --
- *	Spin to get a read lock: only yield the CPU if the lock is held
- *	exclusive.
+ * __read_blocked --
+ *	Check whether the current read lock request should keep waiting.
  */
-void
-__wt_readlock_spin(WT_SESSION_IMPL *session, WT_RWLOCK *l)
+static bool
+__read_blocked(WT_SESSION_IMPL *session)
 {
-	/*
-	 * Try to get the lock in a single operation if it is available to
-	 * readers.  This avoids the situation where multiple readers arrive
-	 * concurrently and have to line up in order to enter the lock.  For
-	 * read-heavy workloads it can make a significant difference.
-	 */
-	while (__wt_try_readlock(session, l) != 0) {
-		if (l->s.writers_active > 0)
-			__wt_yield();
-		else
-			WT_PAUSE();
-	}
+	return (session->current_rwticket !=
+	    session->current_rwlock->u.s.current);
 }
 
 /*
@@ -198,43 +158,95 @@ __wt_readlock_spin(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 void
 __wt_readlock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 {
-	uint16_t ticket;
+	WT_RWLOCK new, old;
 	int pause_cnt;
+	int16_t writers_active;
+	uint8_t ticket;
 
 	WT_STAT_CONN_INCR(session, rwlock_read);
 
 	WT_DIAGNOSTIC_YIELD;
 
-	/*
-	 * Possibly wrap: if we have more than 64K lockers waiting, the ticket
-	 * value will wrap and two lockers will simultaneously be granted the
-	 * lock.
-	 */
-	ticket = __wt_atomic_fetch_add16(&l->s.next, 1);
-	for (pause_cnt = 0; ticket != l->s.readers;) {
+	for (;;) {
 		/*
-		 * We failed to get the lock; pause before retrying and if we've
-		 * paused enough, yield so we don't burn CPU to no purpose. This
-		 * situation happens if there are more threads than cores in the
-		 * system and we're thrashing on shared resources.
+		 * Fast path: if there is no active writer, join the current
+		 * group.
 		 */
-		if (++pause_cnt < WT_THOUSAND)
+		for (old.u.v = l->u.v;
+		    old.u.s.current == old.u.s.next;
+		    old.u.v = l->u.v) {
+			new.u.v = old.u.v;
+			/*
+			 * Check for overflow: if the maximum number of readers
+			 * are already active, no new readers can enter the
+			 * lock.
+			 */
+			if (++new.u.s.readers_active == 0)
+				goto stall;
+			if (__wt_atomic_casv64(&l->u.v, old.u.v, new.u.v))
+				return;
 			WT_PAUSE();
-		else
+		}
+
+		/*
+		 * There is an active writer: join the next group.
+		 *
+		 * Limit how many readers can queue: don't allow more readers
+		 * to queue than there are active writers (calculated as
+		 * `next - current`): otherwise, in write-heavy workloads,
+		 * readers can keep queuing up in front of writers and
+		 * throughput is unstable.
+		 *
+		 * If the maximum number of readers are already queued, wait
+		 * until we can get a valid ticket.
+		 */
+		writers_active = old.u.s.next - old.u.s.current;
+		if (old.u.s.readers_queued > writers_active) {
+stall:			__wt_cond_wait(session,
+			    l->cond_readers, 10 * WT_THOUSAND, NULL);
+			continue;
+		}
+
+		/*
+		 * If we are the first reader to queue, set the next read
+		 * group.  Note: don't re-read from the lock or we could race
+		 * with a writer unlocking.
+		 */
+		new.u.v = old.u.v;
+		if (new.u.s.readers_queued++ == 0)
+			new.u.s.reader = new.u.s.next;
+		ticket = new.u.s.reader;
+
+		if (__wt_atomic_casv64(&l->u.v, old.u.v, new.u.v))
+			break;
+	}
+
+	/* Wait for our group to start. */
+	for (pause_cnt = 0; ticket != l->u.s.current; pause_cnt++) {
+		if (pause_cnt < 1000)
+			WT_PAUSE();
+		else if (pause_cnt < 1200)
 			__wt_yield();
+		else {
+			session->current_rwlock = l;
+			session->current_rwticket = ticket;
+			__wt_cond_wait(session,
+			    l->cond_readers, 10 * WT_THOUSAND, __read_blocked);
+		}
 	}
 
-	/*
-	 * We're the only writer of the readers field, so the update does not
-	 * need to be atomic.
-	 */
-	++l->s.readers;
-
 	/*
 	 * Applications depend on a barrier here so that operations holding the
-	 * lock see consistent data.
+	 * lock see consistent data.  The atomic operation above isn't
+	 * sufficient here because we don't own the lock until our ticket comes
+	 * up and whatever data we are protecting may have changed in the
+	 * meantime.
 	 */
 	WT_READ_BARRIER();
+
+	/* Sanity check that we (still) have the lock. */
+	WT_ASSERT(session,
+	    ticket == l->u.s.current && l->u.s.readers_active > 0);
 }
 
 /*
@@ -244,13 +256,22 @@ __wt_readlock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 void
 __wt_readunlock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 {
-	WT_UNUSED(session);
+	WT_RWLOCK new, old;
 
-	/*
-	 * Increment the writers value (other readers are doing the same, make
-	 * sure we don't race).
-	 */
-	(void)__wt_atomic_add16(&l->s.writers, 1);
+	do {
+		old.u.v = l->u.v;
+		WT_ASSERT(session, old.u.s.readers_active > 0);
+
+		/*
+		 * Decrement the active reader count (other readers are doing
+		 * the same, make sure we don't race).
+		 */
+		new.u.v = old.u.v;
+		--new.u.s.readers_active;
+	} while (!__wt_atomic_casv64(&l->u.v, old.u.v, new.u.v));
+
+	if (new.u.s.readers_active == 0 && new.u.s.current != new.u.s.next)
+		__wt_cond_signal(session, l->cond_writers);
 }
 
 /*
@@ -264,22 +285,44 @@ __wt_try_writelock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 
 	WT_STAT_CONN_INCR(session, rwlock_write);
 
-	old = new = *l;
-
 	/*
-	 * This write lock can only be granted if the lock was last granted to
-	 * a writer and there are no readers or writers blocked on the lock,
-	 * that is, if this thread's ticket would be the next ticket granted.
-	 * Do the cheap test to see if this can possibly succeed (and confirm
-	 * the lock is in the correct state to grant this write lock).
+	 * This write lock can only be granted if no readers or writers blocked
+	 * on the lock, that is, if this thread's ticket would be the next
+	 * ticket granted.  Check if this can possibly succeed (and confirm the
+	 * lock is in the correct state to grant this write lock).
 	 */
-	if (old.s.writers != old.s.next)
+	old.u.v = l->u.v;
+	if (old.u.s.current != old.u.s.next || old.u.s.readers_active != 0)
 		return (EBUSY);
 
-	/* The replacement lock value is a result of allocating a new ticket. */
-	++new.s.next;
-	++new.s.writers_active;
-	return (__wt_atomic_cas64(&l->u, old.u, new.u) ? 0 : EBUSY);
+	/*
+	 * We've checked above that there is no writer active (since
+	 * `current == next`), so there should be no readers queued.
+	 */
+	WT_ASSERT(session, old.u.s.readers_queued == 0);
+
+	/*
+	 * The replacement lock value is a result of allocating a new ticket.
+	 *
+	 * We rely on this atomic operation to provide a barrier.
+	 */
+	new.u.v = old.u.v;
+	new.u.s.next++;
+	return (__wt_atomic_casv64(&l->u.v, old.u.v, new.u.v) ? 0 : EBUSY);
+}
+
+/*
+ * __write_blocked --
+ *	Check whether the current write lock request should keep waiting.
+ */
+static bool
+__write_blocked(WT_SESSION_IMPL *session)
+{
+	WT_RWLOCK *l;
+
+	l = session->current_rwlock;
+	return (session->current_rwticket != l->u.s.current ||
+	    l->u.s.readers_active != 0);
 }
 
 /*
@@ -289,36 +332,68 @@ __wt_try_writelock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 void
 __wt_writelock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 {
-	uint16_t ticket;
+	WT_RWLOCK new, old;
 	int pause_cnt;
+	uint8_t ticket;
 
 	WT_STAT_CONN_INCR(session, rwlock_write);
 
-	/*
-	 * Possibly wrap: if we have more than 64K lockers waiting, the ticket
-	 * value will wrap and two lockers will simultaneously be granted the
-	 * lock.
-	 */
-	ticket = __wt_atomic_fetch_add16(&l->s.next, 1);
-	(void)__wt_atomic_add16(&l->s.writers_active, 1);
-	for (pause_cnt = 0; ticket != l->s.writers;) {
+	for (;;) {
+		old.u.v = l->u.v;
+
+		/* Allocate a ticket. */
+		new.u.v = old.u.v;
+		ticket = new.u.s.next++;
+
 		/*
-		 * We failed to get the lock; pause before retrying and if we've
-		 * paused enough, sleep so we don't burn CPU to no purpose. This
-		 * situation happens if there are more threads than cores in the
-		 * system and we're thrashing on shared resources.
+		 * Check for overflow: if the next ticket is allowed to catch
+		 * up with the current batch, two writers could be granted the
+		 * lock simultaneously.
 		 */
-		if (++pause_cnt < WT_THOUSAND)
+		if (new.u.s.current == new.u.s.next) {
+			__wt_cond_wait(session,
+			    l->cond_writers, 10 * WT_THOUSAND, NULL);
+			continue;
+		}
+		if (__wt_atomic_casv64(&l->u.v, old.u.v, new.u.v))
+			break;
+	}
+
+	/*
+	 * Wait for our group to start and any readers to drain.
+	 *
+	 * We take care here to do an atomic read of the full 64-bit lock
+	 * value.  Otherwise, reads are not guaranteed to be ordered and we
+	 * could see no readers active from a different batch and decide that
+	 * we have the lock.
+	 */
+	for (pause_cnt = 0, old.u.v = l->u.v;
+	    ticket != old.u.s.current || old.u.s.readers_active != 0;
+	    pause_cnt++, old.u.v = l->u.v) {
+		if (pause_cnt < 1000)
 			WT_PAUSE();
-		else
-			__wt_sleep(0, 10);
+		else if (pause_cnt < 1200)
+			__wt_yield();
+		else {
+			session->current_rwlock = l;
+			session->current_rwticket = ticket;
+			__wt_cond_wait(session,
+			    l->cond_writers, 10 * WT_THOUSAND, __write_blocked);
+		}
 	}
 
 	/*
 	 * Applications depend on a barrier here so that operations holding the
-	 * lock see consistent data.
+	 * lock see consistent data.  The atomic operation above isn't
+	 * sufficient here because we don't own the lock until our ticket comes
+	 * up and whatever data we are protecting may have changed in the
+	 * meantime.
 	 */
 	WT_READ_BARRIER();
+
+	/* Sanity check that we (still) have the lock. */
+	WT_ASSERT(session,
+	    ticket == l->u.s.current && l->u.s.readers_active == 0);
 }
 
 /*
@@ -328,29 +403,35 @@ __wt_writelock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 void
 __wt_writeunlock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 {
-	WT_RWLOCK new;
+	WT_RWLOCK new, old;
 
-	WT_UNUSED(session);
+	do {
+		old.u.v = l->u.v;
 
-	(void)__wt_atomic_sub16(&l->s.writers_active, 1);
+		/*
+		 * We're holding the lock exclusive, there shouldn't be any
+		 * active readers.
+		 */
+		WT_ASSERT(session, old.u.s.readers_active == 0);
 
-	/*
-	 * Ensure that all updates made while the lock was held are visible to
-	 * the next thread to acquire the lock.
-	 */
-	WT_WRITE_BARRIER();
+		/*
+		 * Allow the next batch to start.
+		 *
+		 * If there are readers in the next group, swap queued readers
+		 * to active: this could race with new readlock requests, so we
+		 * have to spin.
+		 */
+		new.u.v = old.u.v;
+		if (++new.u.s.current == new.u.s.reader) {
+			new.u.s.readers_active = new.u.s.readers_queued;
+			new.u.s.readers_queued = 0;
+		}
+	} while (!__wt_atomic_casv64(&l->u.v, old.u.v, new.u.v));
 
-	new = *l;
-
-	/*
-	 * We're the only writer of the writers/readers fields, so the update
-	 * does not need to be atomic; we have to update both values at the
-	 * same time though, otherwise we'd potentially race with the thread
-	 * next granted the lock.
-	 */
-	++new.s.writers;
-	++new.s.readers;
-	l->i.wr = new.i.wr;
+	if (new.u.s.readers_active != 0)
+		__wt_cond_signal(session, l->cond_readers);
+	else if (new.u.s.current != new.u.s.next)
+		__wt_cond_signal(session, l->cond_writers);
 
 	WT_DIAGNOSTIC_YIELD;
 }
@@ -363,8 +444,11 @@ __wt_writeunlock(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 bool
 __wt_rwlock_islocked(WT_SESSION_IMPL *session, WT_RWLOCK *l)
 {
+	WT_RWLOCK old;
+
 	WT_UNUSED(session);
 
-	return (l->s.writers != l->s.next || l->s.readers != l->s.next);
+	old.u.v = l->u.v;
+	return (old.u.s.current != old.u.s.next || old.u.s.readers_active != 0);
 }
 #endif

src/support/thread_group.c

@@ -257,7 +257,7 @@ __wt_thread_group_create(
 	__wt_verbose(session, WT_VERB_THREAD_GROUP,
 	    "Creating thread group: %p", (void *)group);
 
-	__wt_rwlock_init(session, &group->lock);
+	WT_RET(__wt_rwlock_init(session, &group->lock));
 	WT_ERR(__wt_cond_alloc(
 	    session, "thread group cond", &group->wait_cond));
 	cond_alloced = true;
@@ -272,7 +272,7 @@ __wt_thread_group_create(
 	/* Cleanup on error to avoid leaking resources */
 err:	if (ret != 0) {
 		if (cond_alloced)
-			WT_TRET(__wt_cond_destroy(session, &group->wait_cond));
+			__wt_cond_destroy(session, &group->wait_cond);
 		__wt_rwlock_destroy(session, &group->lock);
 	}
 	return (ret);
@@ -297,7 +297,7 @@ __wt_thread_group_destroy(WT_SESSION_IMPL *session, WT_THREAD_GROUP *group)
 
 	__wt_free(session, group->threads);
 
-	WT_TRET(__wt_cond_destroy(session, &group->wait_cond));
+	__wt_cond_destroy(session, &group->wait_cond);
 	__wt_rwlock_destroy(session, &group->lock);
 
 	/*

src/txn/txn.c

@@ -126,7 +126,7 @@ __wt_txn_get_snapshot(WT_SESSION_IMPL *session)
 	n = 0;
 
 	/* We're going to scan the table: wait for the lock. */
-	__wt_readlock_spin(session, &txn_global->scan_rwlock);
+	__wt_readlock(session, &txn_global->scan_rwlock);
 
 	current_id = pinned_id = txn_global->current;
 	prev_oldest_id = txn_global->oldest_id;
@@ -293,7 +293,7 @@ __wt_txn_update_oldest(WT_SESSION_IMPL *session, uint32_t flags)
 
 	/* First do a read-only scan. */
 	if (wait)
-		__wt_readlock_spin(session, &txn_global->scan_rwlock);
+		__wt_readlock(session, &txn_global->scan_rwlock);
 	else if ((ret =
 	    __wt_try_readlock(session, &txn_global->scan_rwlock)) != 0)
 		return (ret == EBUSY ? 0 : ret);
@@ -768,8 +768,8 @@ __wt_txn_global_init(WT_SESSION_IMPL *session, const char *cfg[])
 
 	WT_RET(__wt_spin_init(session,
 	    &txn_global->id_lock, "transaction id lock"));
-	__wt_rwlock_init(session, &txn_global->scan_rwlock);
-	__wt_rwlock_init(session, &txn_global->nsnap_rwlock);
+	WT_RET(__wt_rwlock_init(session, &txn_global->scan_rwlock));
+	WT_RET(__wt_rwlock_init(session, &txn_global->nsnap_rwlock));
 	txn_global->nsnap_oldest_id = WT_TXN_NONE;
 	TAILQ_INIT(&txn_global->nsnaph);
 

test/csuite/Makefile.am

@@ -57,6 +57,9 @@ noinst_PROGRAMS += test_wt3135_search_near_collator
 test_wt3184_dup_index_collator_SOURCES = wt3184_dup_index_collator/main.c
 noinst_PROGRAMS += test_wt3184_dup_index_collator
 
+test_rwlock_SOURCES = rwlock/main.c
+noinst_PROGRAMS += test_rwlock
+
 # Run this during a "make check" smoke test.
 TESTS = $(noinst_PROGRAMS)
 LOG_COMPILER = $(TEST_WRAPPER)

test/csuite/rwlock/main.c

@@ -0,0 +1,184 @@
+/*-
+ * Public Domain 2014-2017 MongoDB, Inc.
+ * Public Domain 2008-2014 WiredTiger, Inc.
+ *
+ * This is free and unencumbered software released into the public domain.
+ *
+ * Anyone is free to copy, modify, publish, use, compile, sell, or
+ * distribute this software, either in source code form or as a compiled
+ * binary, for any purpose, commercial or non-commercial, and by any
+ * means.
+ *
+ * In jurisdictions that recognize copyright laws, the author or authors
+ * of this software dedicate any and all copyright interest in the
+ * software to the public domain. We make this dedication for the benefit
+ * of the public at large and to the detriment of our heirs and
+ * successors. We intend this dedication to be an overt act of
+ * relinquishment in perpetuity of all present and future rights to this
+ * software under copyright law.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+ * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+#include "test_util.h"
+
+/*
+ * JIRA ticket reference: HELP-4355
+ * Test rwlock collapse under load.
+ */
+#define	MAX_THREADS 1000
+#define	READS_PER_WRITE	10000
+//#define	READS_PER_WRITE	1000000
+//#define	READS_PER_WRITE	100
+
+#define	CHECK_CORRECTNESS 1
+//#define	USE_POSIX	1
+
+static WT_RWLOCK rwlock;
+static pthread_rwlock_t p_rwlock;
+static bool running;
+static uint64_t shared_counter;
+
+void *thread_rwlock(void *);
+void *thread_dump(void *);
+
+int
+main(int argc, char *argv[])
+{
+	TEST_OPTS *opts, _opts;
+	struct timespec te, ts;
+	pthread_t dump_id, id[MAX_THREADS];
+	int i;
+
+	if (!testutil_enable_long_tests())	/* Ignore unless requested */
+		return (EXIT_SUCCESS);
+
+	opts = &_opts;
+	memset(opts, 0, sizeof(*opts));
+	opts->nthreads = 100;
+	opts->nops = 1000000; /* per thread */
+	testutil_check(testutil_parse_opts(argc, argv, opts));
+	running = true;
+
+	testutil_make_work_dir(opts->home);
+	testutil_check(wiredtiger_open(opts->home, NULL,
+	    "create,session_max=1000,statistics=(fast)", &opts->conn));
+
+	testutil_check(__wt_rwlock_init(NULL, &rwlock));
+	testutil_check(pthread_rwlock_init(&p_rwlock, NULL));
+
+	testutil_check(pthread_create(
+	    &dump_id, NULL, thread_dump, (void *)opts));
+
+	__wt_epoch(NULL, &ts);
+	for (i = 0; i < (int)opts->nthreads; ++i)
+		testutil_check(pthread_create(
+		    &id[i], NULL, thread_rwlock, (void *)opts));
+
+	while (--i >= 0)
+		testutil_check(pthread_join(id[i], NULL));
+	__wt_epoch(NULL, &te);
+	printf("%.2lf\n", WT_TIMEDIFF_MS(te, ts) / 1000.0);
+
+	running = false;
+	testutil_check(pthread_join(dump_id, NULL));
+
+	testutil_check(pthread_rwlock_destroy(&p_rwlock));
+	testutil_cleanup(opts);
+	return (EXIT_SUCCESS);
+}
+
+/*
+ * Acquire a rwlock, every Nth operation, acquire exclusive.
+ */
+void *
+thread_rwlock(void *arg)
+{
+	TEST_OPTS *opts;
+	WT_SESSION *wt_session;
+	WT_SESSION_IMPL *session;
+	uint64_t i, counter;
+	bool writelock;
+
+	opts = (TEST_OPTS *)arg;
+	testutil_check(
+	    opts->conn->open_session(opts->conn, NULL, NULL, &wt_session));
+	session = (WT_SESSION_IMPL *)wt_session;
+
+	printf("Running rwlock thread\n");
+	for (i = 1; i <= opts->nops; ++i) {
+		writelock = (i % READS_PER_WRITE == 0);
+
+#ifdef USE_POSIX
+		if (writelock)
+			testutil_check(pthread_rwlock_wrlock(&p_rwlock));
+		else
+			testutil_check(pthread_rwlock_rdlock(&p_rwlock));
+#else
+		if (writelock)
+			__wt_writelock(session, &rwlock);
+		else
+			__wt_readlock(session, &rwlock);
+#endif
+
+		/*
+		 * Do a tiny amount of work inside the lock so the compiler
+		 * can't optimize everything away.
+		 */
+		(void)__wt_atomic_add64(&counter, 1);
+
+#ifdef CHECK_CORRECTNESS
+		if (writelock)
+			counter = ++shared_counter;
+		else
+			counter = shared_counter;
+
+		__wt_yield();
+
+		testutil_assert(counter == shared_counter);
+#endif
+
+#ifdef USE_POSIX
+		testutil_check(pthread_rwlock_unlock(&p_rwlock));
+#else
+		if (writelock)
+			__wt_writeunlock(session, &rwlock);
+		else
+			__wt_readunlock(session, &rwlock);
+#endif
+
+		if (i % 10000 == 0) {
+			printf("%s", session->id == 20 ? ".\n" : ".");
+			fflush(stdout);
+		}
+	}
+
+	opts->running = false;
+
+	return (NULL);
+}
+
+void *
+thread_dump(void *arg) {
+	WT_UNUSED(arg);
+
+	while (running) {
+		sleep(1);
+		printf("\n"
+		    "rwlock { current %" PRIu8 ", next %" PRIu8
+		    ", reader %" PRIu8 ", readers_active %" PRIu16
+		    ", readers_queued %" PRIu16 " }\n",
+		    rwlock.u.s.current,
+		    rwlock.u.s.next,
+		    rwlock.u.s.reader,
+		    rwlock.u.s.readers_active,
+		    rwlock.u.s.readers_queued);
+	}
+
+	return (NULL);
+}