Make the eviction walk incremental: don't spend too long in any one file, fix tracking of whether we are making progress.

11 files changed, 46 insertions, 35 deletions

diff --git a/src/btree/bt_compact.c b/src/btree/bt_compact.c
index 405410c6a1c..eeec041d57b 100644
--- a/src/btree/bt_compact.c
+++ b/src/btree/bt_compact.c
@@ -149,7 +149,7 @@ __wt_compact(WT_SESSION_IMPL *session, const char *cfg[])
 		 * read, set its generation to a low value so it is evicted
 		 * quickly.
 		 */
-		WT_ERR(__wt_tree_walk(session, &ref,
+		WT_ERR(__wt_tree_walk(session, &ref, NULL,
 		    WT_READ_COMPACT | WT_READ_NO_GEN | WT_READ_WONT_NEED));
 		if (ref == NULL)
 			break;
diff --git a/src/btree/bt_curnext.c b/src/btree/bt_curnext.c
index 6140dca1fad..d80a5f4740d 100644
--- a/src/btree/bt_curnext.c
+++ b/src/btree/bt_curnext.c
@@ -487,7 +487,7 @@ __wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
 			__wt_page_evict_soon(page);
 		cbt->page_deleted_count = 0;
 
-		WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
+		WT_ERR(__wt_tree_walk(session, &cbt->ref, NULL, flags));
 		WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
 	}
 
diff --git a/src/btree/bt_curprev.c b/src/btree/bt_curprev.c
index 880cb777954..f1ca81ee145 100644
--- a/src/btree/bt_curprev.c
+++ b/src/btree/bt_curprev.c
@@ -574,7 +574,7 @@ __wt_btcur_prev(WT_CURSOR_BTREE *cbt, int truncating)
 			__wt_page_evict_soon(page);
 		cbt->page_deleted_count = 0;
 
-		WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
+		WT_ERR(__wt_tree_walk(session, &cbt->ref, NULL, flags));
 		WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
 	}
 
diff --git a/src/btree/bt_handle.c b/src/btree/bt_handle.c
index 6a2789c909b..299849ad365 100644
--- a/src/btree/bt_handle.c
+++ b/src/btree/bt_handle.c
@@ -566,7 +566,7 @@ __btree_get_last_recno(WT_SESSION_IMPL *session)
 	btree = S2BT(session);
 
 	next_walk = NULL;
-	WT_RET(__wt_tree_walk(session, &next_walk, WT_READ_PREV));
+	WT_RET(__wt_tree_walk(session, &next_walk, NULL, WT_READ_PREV));
 	if (next_walk == NULL)
 		return (WT_NOTFOUND);
 
diff --git a/src/btree/bt_stat.c b/src/btree/bt_stat.c
index b8d56fe9d92..2e34a925f84 100644
--- a/src/btree/bt_stat.c
+++ b/src/btree/bt_stat.c
@@ -56,8 +56,8 @@ __wt_btree_stat_init(WT_SESSION_IMPL *session, WT_CURSOR_STAT *cst)
 	WT_STAT_SET(stats, btree_row_leaf, 0);
 
 	next_walk = NULL;
-	while ((ret =
-	    __wt_tree_walk(session, &next_walk, 0)) == 0 && next_walk != NULL) {
+	while ((ret = __wt_tree_walk(session, &next_walk, NULL, 0)) == 0 &&
+	    next_walk != NULL) {
 		WT_WITH_PAGE_INDEX(session,
 		    ret = __stat_page(session, next_walk->page, stats));
 		WT_RET(ret);
diff --git a/src/btree/bt_sync.c b/src/btree/bt_sync.c
index 7c6d49c8ea0..3cde2fa28a9 100644
--- a/src/btree/bt_sync.c
+++ b/src/btree/bt_sync.c
@@ -56,7 +56,7 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 
 		flags |= WT_READ_NO_WAIT | WT_READ_SKIP_INTL;
 		for (walk = NULL;;) {
-			WT_ERR(__wt_tree_walk(session, &walk, flags));
+			WT_ERR(__wt_tree_walk(session, &walk, NULL, flags));
 			if (walk == NULL)
 				break;
 
@@ -107,7 +107,7 @@ __sync_file(WT_SESSION_IMPL *session, int syncop)
 		/* Write all dirty in-cache pages. */
 		flags |= WT_READ_NO_EVICT;
 		for (walk = NULL;;) {
-			WT_ERR(__wt_tree_walk(session, &walk, flags));
+			WT_ERR(__wt_tree_walk(session, &walk, NULL, flags));
 			if (walk == NULL)
 				break;
 
@@ -175,6 +175,12 @@ err:	/* On error, clear any left-over tree walk. */
 		WT_FULL_BARRIER();
 
 		/*
+		 * If this tree was being skipped by the eviction server during
+		 * the checkpoint, clear the wait.
+		 */
+		btree->evict_walk_period = 0;
+
+		/*
 		 * Wake the eviction server, in case application threads have
 		 * stalled while the eviction server decided it couldn't make
 		 * progress.  Without this, application threads will be stalled
diff --git a/src/btree/bt_walk.c b/src/btree/bt_walk.c
index 8b2ce6d8fc1..70a5bc0b1ae 100644
--- a/src/btree/bt_walk.c
+++ b/src/btree/bt_walk.c
@@ -13,7 +13,8 @@
  *	Move to the next/previous page in the tree.
  */
 int
-__wt_tree_walk(WT_SESSION_IMPL *session, WT_REF **refp, uint32_t flags)
+__wt_tree_walk(WT_SESSION_IMPL *session,
+    WT_REF **refp, uint64_t *refcntp, uint32_t flags)
 {
 	WT_BTREE *btree;
 	WT_DECL_RET;
@@ -178,6 +179,9 @@ restart:	/*
 		else
 			++slot;
 
+		if (refcntp != NULL)
+			++*refcntp;
+
 		for (descending = 0;;) {
 			ref = pindex->index[slot];
 
diff --git a/src/evict/evict_file.c b/src/evict/evict_file.c
index 5aa85872a3b..910aef070ca 100644
--- a/src/evict/evict_file.c
+++ b/src/evict/evict_file.c
@@ -36,8 +36,8 @@ __wt_evict_file(WT_SESSION_IMPL *session, int syncop)
 
 	/* Walk the tree, discarding pages. */
 	next_ref = NULL;
-	WT_ERR(__wt_tree_walk(
-	    session, &next_ref, WT_READ_CACHE | WT_READ_NO_EVICT));
+	WT_ERR(__wt_tree_walk(session, &next_ref, NULL,
+	    WT_READ_CACHE | WT_READ_NO_EVICT));
 	while ((ref = next_ref) != NULL) {
 		page = ref->page;
 
@@ -73,8 +73,8 @@ __wt_evict_file(WT_SESSION_IMPL *session, int syncop)
 		 * the reconciliation, the next walk call could miss a page in
 		 * the tree.
 		 */
-		WT_ERR(__wt_tree_walk(
-		    session, &next_ref, WT_READ_CACHE | WT_READ_NO_EVICT));
+		WT_ERR(__wt_tree_walk(session, &next_ref, NULL,
+		    WT_READ_CACHE | WT_READ_NO_EVICT));
 
 		switch (syncop) {
 		case WT_SYNC_CLOSE:
diff --git a/src/evict/evict_lru.c b/src/evict/evict_lru.c
index 062356af637..e6985d750bd 100644
--- a/src/evict/evict_lru.c
+++ b/src/evict/evict_lru.c
@@ -839,7 +839,7 @@ __evict_walk(WT_SESSION_IMPL *session, uint32_t flags)
 	WT_CONNECTION_IMPL *conn;
 	WT_DATA_HANDLE *dhandle;
 	WT_DECL_RET;
-	u_int max_entries, old_slot, retries, slot, spins;
+	u_int max_entries, prev_slot, retries, slot, start_slot, spins;
 	int incr, dhandle_locked;
 	WT_DECL_SPINLOCK_ID(id);
 
@@ -869,7 +869,7 @@ __evict_walk(WT_SESSION_IMPL *session, uint32_t flags)
 	 * Set the starting slot in the queue and the maximum pages added
 	 * per walk.
 	 */
-	slot = cache->evict_entries;
+	start_slot = slot = cache->evict_entries;
 	max_entries = slot + WT_EVICT_WALK_INCR;
 
 retry:	while (slot < max_entries && ret == 0) {
@@ -934,10 +934,10 @@ retry:	while (slot < max_entries && ret == 0) {
 			continue;
 
 		/*
-		 * Also skip files that are configured to stick in cache until
-		 * we get aggressive.
+		 * Also skip files that are checkpointing or configured to
+		 * stick in cache until we get aggressive.
 		 */
-		if (btree->evict_priority != 0 &&
+		if ((btree->checkpointing || btree->evict_priority != 0) &&
 		    !LF_ISSET(WT_EVICT_PASS_AGGRESSIVE))
 			continue;
 
@@ -950,7 +950,7 @@ retry:	while (slot < max_entries && ret == 0) {
 		    btree->evict_walk_skips++ < btree->evict_walk_period)
 			continue;
 		btree->evict_walk_skips = 0;
-		old_slot = slot;
+		prev_slot = slot;
 
 		(void)WT_ATOMIC_ADD4(dhandle->session_inuse, 1);
 		incr = 1;
@@ -972,15 +972,14 @@ retry:	while (slot < max_entries && ret == 0) {
 		__wt_spin_unlock(session, &cache->evict_walk_lock);
 
 		/*
-		 * If we didn't find enough candidates in the file, skip it
-		 * next time.
+		 * If we didn't find any candidates in the file, skip it next
+		 * time.
 		 */
-		if (slot >= old_slot + WT_EVICT_WALK_PER_FILE ||
-		    slot >= max_entries)
-			btree->evict_walk_period = 0;
-		else
+		if (slot == prev_slot)
 			btree->evict_walk_period = WT_MIN(
-			    WT_MAX(1, 2 * btree->evict_walk_period), 1000);
+			    WT_MAX(1, 2 * btree->evict_walk_period), 100);
+		else
+			btree->evict_walk_period = 0;
 	}
 
 	if (incr) {
@@ -1003,8 +1002,9 @@ retry:	while (slot < max_entries && ret == 0) {
 	if (!F_ISSET(cache, WT_EVICT_CLEAR_WALKS) && ret == 0 &&
 	    slot < max_entries && (retries < 2 ||
 	    (!LF_ISSET(WT_EVICT_PASS_WOULD_BLOCK) &&
-	    retries < 10 && slot > 0))) {
+	    retries < 10 && slot > start_slot))) {
 		cache->evict_file_next = NULL;
+		start_slot = slot;
 		++retries;
 		goto retry;
 	}
@@ -1072,9 +1072,9 @@ __evict_walk_file(WT_SESSION_IMPL *session, u_int *slotp, uint32_t flags)
 	 * Get some more eviction candidate pages.
 	 */
 	for (evict = start, pages_walked = 0, internal_pages = restarts = 0;
-	    evict < end && (ret == 0 || ret == WT_NOTFOUND);
-	    ret = __wt_tree_walk(session, &btree->evict_ref, walk_flags),
-	    ++pages_walked) {
+	    evict < end && pages_walked < WT_EVICT_MAX_PER_FILE &&
+	    (ret == 0 || ret == WT_NOTFOUND);
+	    ret = __wt_tree_walk(session, &btree->evict_ref, &pages_walked, walk_flags)) {
 		if (btree->evict_ref == NULL) {
 			/*
 			 * Take care with terminating this loop.
@@ -1451,7 +1451,7 @@ __wt_cache_dump(WT_SESSION_IMPL *session)
 		next_walk = NULL;
 		session->dhandle = dhandle;
 		while (__wt_tree_walk(session,
-		    &next_walk, WT_READ_CACHE | WT_READ_NO_WAIT) == 0 &&
+		    &next_walk, NULL, WT_READ_CACHE | WT_READ_NO_WAIT) == 0 &&
 		    next_walk != NULL) {
 			page = next_walk->page;
 			if (page->type == WT_PAGE_COL_INT ||
diff --git a/src/include/cache.h b/src/include/cache.h
index de6faad608a..24ea14ab11d 100644
--- a/src/include/cache.h
+++ b/src/include/cache.h
@@ -13,9 +13,10 @@
 #define	WT_EVICT_INT_SKEW  (1<<20)	/* Prefer leaf pages over internal
 					   pages by this many increments of the
 					   read generation. */
-#define	WT_EVICT_WALK_PER_FILE	10	/* Pages to visit per file */
-#define	WT_EVICT_WALK_BASE     300	/* Pages tracked across file visits */
-#define	WT_EVICT_WALK_INCR     100	/* Pages added each walk */
+#define	WT_EVICT_WALK_PER_FILE	 10	/* Pages to queue per file */
+#define	WT_EVICT_MAX_PER_FILE	100	/* Max pages to visit per file */
+#define	WT_EVICT_WALK_BASE	300	/* Pages tracked across file visits */
+#define	WT_EVICT_WALK_INCR	100	/* Pages added each walk */
 
 #define	WT_EVICT_PASS_AGGRESSIVE	0x01
 #define	WT_EVICT_PASS_ALL		0x02
diff --git a/src/include/extern.h b/src/include/extern.h
index 40493b4aab4..c5909f7af2d 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -157,7 +157,7 @@ extern int __wt_upgrade(WT_SESSION_IMPL *session, const char *cfg[]);
 extern int __wt_verify(WT_SESSION_IMPL *session, const char *cfg[]);
 extern int __wt_verify_dsk_image(WT_SESSION_IMPL *session, const char *addr, const WT_PAGE_HEADER *dsk, size_t size, int empty_page_ok);
 extern int __wt_verify_dsk(WT_SESSION_IMPL *session, const char *addr, WT_ITEM *buf);
-extern int __wt_tree_walk(WT_SESSION_IMPL *session, WT_REF **refp, uint32_t flags);
+extern int __wt_tree_walk(WT_SESSION_IMPL *session, WT_REF **refp, uint64_t *refcntp, uint32_t flags);
 extern int __wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, uint64_t recno, WT_ITEM *value, WT_UPDATE *upd, int is_remove);
 extern int __wt_col_search(WT_SESSION_IMPL *session, uint64_t recno, WT_REF *leaf, WT_CURSOR_BTREE *cbt);
 extern int __wt_row_leaf_keys(WT_SESSION_IMPL *session, WT_PAGE *page);