Merge pull request #1501 from wiredtiger/page-yield

Page yield

4 files changed, 77 insertions, 27 deletions

diff --git a/src/btree/bt_page.c b/src/btree/bt_page.c
index 799f0cca3ee..db1ccdabc4f 100644
--- a/src/btree/bt_page.c
+++ b/src/btree/bt_page.c
@@ -67,9 +67,10 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 {
 	WT_DECL_RET;
 	WT_PAGE *page;
+	u_int sleep_cnt, wait_cnt;
 	int busy, force_attempts, oldgen;
 
-	for (force_attempts = oldgen = 0;;) {
+	for (force_attempts = oldgen = 0, wait_cnt = 0;;) {
 		switch (ref->state) {
 		case WT_REF_DISK:
 		case WT_REF_DELETED:
@@ -88,11 +89,14 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 		case WT_REF_READING:
 			if (LF_ISSET(WT_READ_CACHE))
 				return (WT_NOTFOUND);
-			/* FALLTHROUGH */
+			if (LF_ISSET(WT_READ_NO_WAIT))
+				return (WT_NOTFOUND);
+			WT_STAT_FAST_CONN_INCR(session, page_read_blocked);
+			break;
 		case WT_REF_LOCKED:
 			if (LF_ISSET(WT_READ_NO_WAIT))
 				return (WT_NOTFOUND);
-			/* The page is busy -- wait. */
+			WT_STAT_FAST_CONN_INCR(session, page_locked_blocked);
 			break;
 		case WT_REF_SPLIT:
 			return (WT_RESTART);
@@ -109,8 +113,11 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 #else
 			WT_RET(__wt_hazard_set(session, ref, &busy));
 #endif
-			if (busy)
+			if (busy) {
+				WT_STAT_FAST_CONN_INCR(
+				    session, page_busy_blocked);
 				break;
+			}
 
 			page = ref->page;
 			WT_ASSERT(session, page != NULL);
@@ -121,6 +128,8 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 			    __evict_force_check(session, page)) {
 				++force_attempts;
 				WT_RET(__wt_page_release(session, ref, flags));
+				WT_STAT_FAST_CONN_INCR(
+				    session, page_forcible_evict_blocked);
 				break;
 			}
 
@@ -148,8 +157,19 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 		WT_ILLEGAL_VALUE(session);
 		}
 
-		/* We failed to get the page -- yield before retrying. */
-		__wt_yield();
+		/*
+		 * We failed to get the page -- yield before retrying, and if
+		 * we've yielded enough times, start sleeping so we don't burn
+		 * CPU to no purpose.
+		 */
+		if (++wait_cnt < 1000)
+			__wt_yield();
+		 else {
+			sleep_cnt = WT_MIN(wait_cnt, 10000);
+			wait_cnt *= 2;
+			WT_STAT_FAST_CONN_INCRV(session, page_sleep, sleep_cnt);
+			__wt_sleep(0, sleep_cnt);
+		}
 	}
 }
 
diff --git a/src/include/stat.h b/src/include/stat.h
index 20b7ff9c1d4..8bdc377ac44 100644
--- a/src/include/stat.h
+++ b/src/include/stat.h
@@ -236,6 +236,11 @@ struct __wt_connection_stats {
 	WT_STATS memory_allocation;
 	WT_STATS memory_free;
 	WT_STATS memory_grow;
+	WT_STATS page_busy_blocked;
+	WT_STATS page_forcible_evict_blocked;
+	WT_STATS page_locked_blocked;
+	WT_STATS page_read_blocked;
+	WT_STATS page_sleep;
 	WT_STATS read_io;
 	WT_STATS rec_pages;
 	WT_STATS rec_pages_eviction;
diff --git a/src/include/wiredtiger.in b/src/include/wiredtiger.in
index 075514e02b4..f79938bb0db 100644
--- a/src/include/wiredtiger.in
+++ b/src/include/wiredtiger.in
@@ -3300,48 +3300,58 @@ extern int wiredtiger_extension_terminate(WT_CONNECTION *connection);
 #define	WT_STAT_CONN_MEMORY_FREE			1103
 /*! connection: memory re-allocations */
 #define	WT_STAT_CONN_MEMORY_GROW			1104
+/*! thread-yield: page acquire busy blocked */
+#define	WT_STAT_CONN_PAGE_BUSY_BLOCKED			1105
+/*! thread-yield: page acquire eviction blocked */
+#define	WT_STAT_CONN_PAGE_FORCIBLE_EVICT_BLOCKED	1106
+/*! thread-yield: page acquire locked blocked */
+#define	WT_STAT_CONN_PAGE_LOCKED_BLOCKED		1107
+/*! thread-yield: page acquire read blocked */
+#define	WT_STAT_CONN_PAGE_READ_BLOCKED			1108
+/*! thread-yield: page acquire time sleeping (usecs) */
+#define	WT_STAT_CONN_PAGE_SLEEP				1109
 /*! connection: total read I/Os */
-#define	WT_STAT_CONN_READ_IO				1105
+#define	WT_STAT_CONN_READ_IO				1110
 /*! reconciliation: page reconciliation calls */
-#define	WT_STAT_CONN_REC_PAGES				1106
+#define	WT_STAT_CONN_REC_PAGES				1111
 /*! reconciliation: page reconciliation calls for eviction */
-#define	WT_STAT_CONN_REC_PAGES_EVICTION			1107
+#define	WT_STAT_CONN_REC_PAGES_EVICTION			1112
 /*! reconciliation: split bytes currently awaiting free */
-#define	WT_STAT_CONN_REC_SPLIT_STASHED_BYTES		1108
+#define	WT_STAT_CONN_REC_SPLIT_STASHED_BYTES		1113
 /*! reconciliation: split objects currently awaiting free */
-#define	WT_STAT_CONN_REC_SPLIT_STASHED_OBJECTS		1109
+#define	WT_STAT_CONN_REC_SPLIT_STASHED_OBJECTS		1114
 /*! connection: pthread mutex shared lock read-lock calls */
-#define	WT_STAT_CONN_RWLOCK_READ			1110
+#define	WT_STAT_CONN_RWLOCK_READ			1115
 /*! connection: pthread mutex shared lock write-lock calls */
-#define	WT_STAT_CONN_RWLOCK_WRITE			1111
+#define	WT_STAT_CONN_RWLOCK_WRITE			1116
 /*! session: open cursor count */
-#define	WT_STAT_CONN_SESSION_CURSOR_OPEN		1112
+#define	WT_STAT_CONN_SESSION_CURSOR_OPEN		1117
 /*! session: open session count */
-#define	WT_STAT_CONN_SESSION_OPEN			1113
+#define	WT_STAT_CONN_SESSION_OPEN			1118
 /*! transaction: transaction begins */
-#define	WT_STAT_CONN_TXN_BEGIN				1114
+#define	WT_STAT_CONN_TXN_BEGIN				1119
 /*! transaction: transaction checkpoints */
-#define	WT_STAT_CONN_TXN_CHECKPOINT			1115
+#define	WT_STAT_CONN_TXN_CHECKPOINT			1120
 /*! transaction: transaction checkpoint currently running */
-#define	WT_STAT_CONN_TXN_CHECKPOINT_RUNNING		1116
+#define	WT_STAT_CONN_TXN_CHECKPOINT_RUNNING		1121
 /*! transaction: transaction checkpoint max time (msecs) */
-#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_MAX		1117
+#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_MAX		1122
 /*! transaction: transaction checkpoint min time (msecs) */
-#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_MIN		1118
+#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_MIN		1123
 /*! transaction: transaction checkpoint most recent time (msecs) */
-#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_RECENT		1119
+#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_RECENT		1124
 /*! transaction: transaction checkpoint total time (msecs) */
-#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_TOTAL		1120
+#define	WT_STAT_CONN_TXN_CHECKPOINT_TIME_TOTAL		1125
 /*! transaction: transactions committed */
-#define	WT_STAT_CONN_TXN_COMMIT				1121
+#define	WT_STAT_CONN_TXN_COMMIT				1126
 /*! transaction: transaction failures due to cache overflow */
-#define	WT_STAT_CONN_TXN_FAIL_CACHE			1122
+#define	WT_STAT_CONN_TXN_FAIL_CACHE			1127
 /*! transaction: transaction range of IDs currently pinned */
-#define	WT_STAT_CONN_TXN_PINNED_RANGE			1123
+#define	WT_STAT_CONN_TXN_PINNED_RANGE			1128
 /*! transaction: transactions rolled back */
-#define	WT_STAT_CONN_TXN_ROLLBACK			1124
+#define	WT_STAT_CONN_TXN_ROLLBACK			1129
 /*! connection: total write I/Os */
-#define	WT_STAT_CONN_WRITE_IO				1125
+#define	WT_STAT_CONN_WRITE_IO				1130
 
 /*!
  * @}
diff --git a/src/support/stat.c b/src/support/stat.c
index 265a2673d7c..c93168cd9a1 100644
--- a/src/support/stat.c
+++ b/src/support/stat.c
@@ -480,6 +480,16 @@ __wt_stat_init_connection_stats(WT_CONNECTION_STATS *stats)
 	    "reconciliation: split objects currently awaiting free";
 	stats->session_cursor_open.desc = "session: open cursor count";
 	stats->session_open.desc = "session: open session count";
+	stats->page_busy_blocked.desc =
+	    "thread-yield: page acquire busy blocked";
+	stats->page_forcible_evict_blocked.desc =
+	    "thread-yield: page acquire eviction blocked";
+	stats->page_locked_blocked.desc =
+	    "thread-yield: page acquire locked blocked";
+	stats->page_read_blocked.desc =
+	    "thread-yield: page acquire read blocked";
+	stats->page_sleep.desc =
+	    "thread-yield: page acquire time sleeping (usecs)";
 	stats->txn_begin.desc = "transaction: transaction begins";
 	stats->txn_checkpoint_running.desc =
 	    "transaction: transaction checkpoint currently running";
@@ -608,6 +618,11 @@ __wt_stat_refresh_connection_stats(void *stats_arg)
 	stats->lsm_work_queue_max.v = 0;
 	stats->rec_pages.v = 0;
 	stats->rec_pages_eviction.v = 0;
+	stats->page_busy_blocked.v = 0;
+	stats->page_forcible_evict_blocked.v = 0;
+	stats->page_locked_blocked.v = 0;
+	stats->page_read_blocked.v = 0;
+	stats->page_sleep.v = 0;
 	stats->txn_begin.v = 0;
 	stats->txn_checkpoint.v = 0;
 	stats->txn_fail_cache.v = 0;