WT-3261 add a checkpoint epoch to avoid draining the eviction queue (#3370)

8 files changed, 54 insertions, 65 deletions

diff --git a/src/btree/bt_read.c b/src/btree/bt_read.c
index fb69afb166c..72a69e8591c 100644
--- a/src/btree/bt_read.c
+++ b/src/btree/bt_read.c
@@ -585,15 +585,10 @@ __wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
 			 * if the page qualifies for forced eviction and update
 			 * the page's generation number. If eviction isn't being
 			 * done on this file, we're done.
-			 * In-memory split of large pages is allowed while
-			 * no_eviction is set on btree, whereas reconciliation
-			 * is not allowed.
 			 */
 			if (LF_ISSET(WT_READ_NO_EVICT) ||
 			    F_ISSET(session, WT_SESSION_NO_EVICTION) ||
-			    btree->lsm_primary ||
-			    (btree->evict_disabled > 0 &&
-			    !F_ISSET(btree, WT_BTREE_ALLOW_SPLITS)))
+			    btree->evict_disabled > 0 || btree->lsm_primary)
 				goto skip_evict;
 
 			/*
diff --git a/src/btree/bt_sync.c b/src/btree/bt_sync.c
index 81e9d1757bb..112f0725f94 100644
--- a/src/btree/bt_sync.c
+++ b/src/btree/bt_sync.c
@@ -180,21 +180,8 @@ __sync_file(WT_SESSION_IMPL *session, WT_CACHE_OP syncop)
 		 * any problematic eviction or page splits to complete.
 		 */
 		WT_PUBLISH(btree->checkpointing, WT_CKPT_PREPARE);
-
-		/*
-		 * Sync for checkpoint allows splits to happen while the queue
-		 * is being drained, but not reconciliation. We need to do this,
-		 * since draining the queue can take long enough for hot pages
-		 * to grow significantly larger than the configured maximum
-		 * size.
-		 */
-		F_SET(btree, WT_BTREE_ALLOW_SPLITS);
-		ret = __wt_evict_file_exclusive_on(session);
-		F_CLR(btree, WT_BTREE_ALLOW_SPLITS);
-		WT_ERR(ret);
-		__wt_evict_file_exclusive_off(session);
-
-		WT_PUBLISH(btree->checkpointing, WT_CKPT_RUNNING);
+		(void)__wt_gen_next_drain(session, WT_GEN_EVICT);
+		btree->checkpointing = WT_CKPT_RUNNING;
 
 		/* Write all dirty in-cache pages. */
 		flags |= WT_READ_NO_EVICT;
@@ -268,9 +255,8 @@ err:	/* On error, clear any left-over tree walk. */
 	    saved_pinned_id == WT_TXN_NONE)
 		__wt_txn_release_snapshot(session);
 
-	/* Clear the checkpoint flag and push the change. */
-	if (btree->checkpointing != WT_CKPT_OFF)
-		WT_PUBLISH(btree->checkpointing, WT_CKPT_OFF);
+	/* Clear the checkpoint flag. */
+	btree->checkpointing = WT_CKPT_OFF;
 
 	__wt_spin_unlock(session, &btree->flush_lock);
 
diff --git a/src/evict/evict_page.c b/src/evict/evict_page.c
index 9498e2fb313..edcd108e7e4 100644
--- a/src/evict/evict_page.c
+++ b/src/evict/evict_page.c
@@ -113,6 +113,9 @@ __wt_evict(WT_SESSION_IMPL *session, WT_REF *ref, bool closing)
 	/* Checkpoints should never do eviction. */
 	WT_ASSERT(session, !WT_SESSION_IS_CHECKPOINT(session));
 
+	/* Enter the eviction generation. */
+	__wt_session_gen_enter(session, WT_GEN_EVICT);
+
 	page = ref->page;
 	tree_dead = F_ISSET(session->dhandle, WT_DHANDLE_DEAD);
 
@@ -133,7 +136,7 @@ __wt_evict(WT_SESSION_IMPL *session, WT_REF *ref, bool closing)
 	 * we want: there is nothing more to do.
 	 */
 	if (LF_ISSET(WT_EVICT_INMEM_SPLIT))
-		return (0);
+		goto done;
 
 	/* Count evictions of internal pages during normal operation. */
 	if (!closing && WT_PAGE_IS_INTERNAL(page)) {
@@ -182,6 +185,9 @@ err:		if (!closing)
 		WT_STAT_DATA_INCR(session, cache_eviction_fail);
 	}
 
+done:	/* Leave the eviction generation. */
+	__wt_session_gen_leave(session, WT_GEN_EVICT);
+
 	return (ret);
 }
 
@@ -479,10 +485,6 @@ __evict_review(
 		 */
 		if (LF_ISSET(WT_EVICT_INMEM_SPLIT))
 			return (__wt_split_insert(session, ref));
-
-		/* If splits are the only permitted operation, we're done. */
-		if (F_ISSET(S2BT(session), WT_BTREE_ALLOW_SPLITS))
-			return (EBUSY);
 	}
 
 	/* If the page is clean, we're done and we can evict. */
diff --git a/src/include/btree.h b/src/include/btree.h
index 19db27d84a2..8ce77b5ecd3 100644
--- a/src/include/btree.h
+++ b/src/include/btree.h
@@ -151,7 +151,7 @@ struct __wt_btree {
 	volatile uint32_t evict_busy;	/* Count of threads in eviction */
 	int	    evict_start_type;	/* Start position for eviction walk
 					   (see WT_EVICT_WALK_START). */
-	enum {
+	volatile enum {
 		WT_CKPT_OFF, WT_CKPT_PREPARE, WT_CKPT_RUNNING
 	} checkpointing;		/* Checkpoint in progress */
 
@@ -163,19 +163,18 @@ struct __wt_btree {
 	WT_SPINLOCK	flush_lock;	/* Lock to flush the tree's pages */
 
 	/* Flags values up to 0xff are reserved for WT_DHANDLE_* */
-#define	WT_BTREE_ALLOW_SPLITS	0x000100 /* Allow splits, even with no evict */
-#define	WT_BTREE_BULK		0x000200 /* Bulk-load handle */
-#define	WT_BTREE_CLOSED		0x000400 /* Handle closed */
-#define	WT_BTREE_IGNORE_CACHE	0x000800 /* Cache-resident object */
-#define	WT_BTREE_IN_MEMORY	0x001000 /* Cache-resident object */
-#define	WT_BTREE_LOOKASIDE	0x002000 /* Look-aside table */
-#define	WT_BTREE_NO_CHECKPOINT	0x004000 /* Disable checkpoints */
-#define	WT_BTREE_NO_LOGGING	0x008000 /* Disable logging */
-#define	WT_BTREE_REBALANCE	0x020000 /* Handle is for rebalance */
-#define	WT_BTREE_SALVAGE	0x040000 /* Handle is for salvage */
-#define	WT_BTREE_SKIP_CKPT	0x080000 /* Handle skipped checkpoint */
-#define	WT_BTREE_UPGRADE	0x100000 /* Handle is for upgrade */
-#define	WT_BTREE_VERIFY		0x200000 /* Handle is for verify */
+#define	WT_BTREE_BULK		0x000100 /* Bulk-load handle */
+#define	WT_BTREE_CLOSED		0x000200 /* Handle closed */
+#define	WT_BTREE_IGNORE_CACHE	0x000400 /* Cache-resident object */
+#define	WT_BTREE_IN_MEMORY	0x000800 /* Cache-resident object */
+#define	WT_BTREE_LOOKASIDE	0x001000 /* Look-aside table */
+#define	WT_BTREE_NO_CHECKPOINT	0x002000 /* Disable checkpoints */
+#define	WT_BTREE_NO_LOGGING	0x004000 /* Disable logging */
+#define	WT_BTREE_REBALANCE	0x008000 /* Handle is for rebalance */
+#define	WT_BTREE_SALVAGE	0x010000 /* Handle is for salvage */
+#define	WT_BTREE_SKIP_CKPT	0x020000 /* Handle skipped checkpoint */
+#define	WT_BTREE_UPGRADE	0x040000 /* Handle is for upgrade */
+#define	WT_BTREE_VERIFY		0x080000 /* Handle is for verify */
 	uint32_t flags;
 };
 
diff --git a/src/include/btree.i b/src/include/btree.i
index 474b40bf805..d4db65b2033 100644
--- a/src/include/btree.i
+++ b/src/include/btree.i
@@ -1286,6 +1286,16 @@ __wt_page_can_evict(
 		return (true);
 
 	/*
+	 * We can't split or evict multiblock row-store pages where the parent's
+	 * key for the page is an overflow item, because the split into the
+	 * parent frees the backing blocks for any no-longer-used overflow keys,
+	 * which will corrupt the checkpoint's block management.
+	 */
+	if (btree->checkpointing != WT_CKPT_OFF &&
+	    F_ISSET_ATOMIC(ref->home, WT_PAGE_OVERFLOW_KEYS))
+		return (false);
+
+	/*
 	 * Check for in-memory splits before other eviction tests. If the page
 	 * should split in-memory, return success immediately and skip more
 	 * detailed eviction tests. We don't need further tests since the page
@@ -1312,16 +1322,6 @@ __wt_page_can_evict(
 	}
 
 	/*
-	 * We can't evict clean, multiblock row-store pages where the parent's
-	 * key for the page is an overflow item, because the split into the
-	 * parent frees the backing blocks for any no-longer-used overflow keys,
-	 * which will corrupt the checkpoint's block management.
-	 */
-	if (btree->checkpointing != WT_CKPT_OFF &&
-	    F_ISSET_ATOMIC(ref->home, WT_PAGE_OVERFLOW_KEYS))
-		return (false);
-
-	/*
 	 * If a split created new internal pages, those newly created internal
 	 * pages cannot be evicted until all threads are known to have exited
 	 * the original parent page's index, because evicting an internal page
diff --git a/src/include/extern.h b/src/include/extern.h
index ff835bacc56..4f38b7ac433 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -649,6 +649,8 @@ extern int __wt_unexpected_object_type( WT_SESSION_IMPL *session, const char *ur
 extern void __wt_gen_init(WT_SESSION_IMPL *session) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
 extern uint64_t __wt_gen(WT_SESSION_IMPL *session, int which) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
 extern uint64_t __wt_gen_next(WT_SESSION_IMPL *session, int which) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
+extern uint64_t __wt_gen_next_drain(WT_SESSION_IMPL *session, int which) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
+extern void __wt_gen_drain(WT_SESSION_IMPL *session, int which, uint64_t generation) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
 extern uint64_t __wt_gen_oldest(WT_SESSION_IMPL *session, int which) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
 extern uint64_t __wt_session_gen(WT_SESSION_IMPL *session, int which) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
 extern void __wt_session_gen_enter(WT_SESSION_IMPL *session, int which) WT_GCC_FUNC_DECL_ATTRIBUTE((visibility("hidden")));
diff --git a/src/include/session.h b/src/include/session.h
index adef5e39068..de2c1463684 100644
--- a/src/include/session.h
+++ b/src/include/session.h
@@ -169,10 +169,11 @@ struct __wt_session_impl {
 
 					/* Generations manager */
 #define	WT_GEN_CHECKPOINT	0	/* Checkpoint generation */
-#define	WT_GEN_HAZARD		1	/* Hazard pointer */
-#define	WT_GEN_SCHEMA		2	/* Schema version */
-#define	WT_GEN_SPLIT		3	/* Page splits */
-#define	WT_GENERATIONS		4	/* Total generation manager entries */
+#define	WT_GEN_EVICT		1	/* Eviction generation */
+#define	WT_GEN_HAZARD		2	/* Hazard pointer */
+#define	WT_GEN_SCHEMA		3	/* Schema version */
+#define	WT_GEN_SPLIT		4	/* Page splits */
+#define	WT_GENERATIONS		5	/* Total generation manager entries */
 	volatile uint64_t generations[WT_GENERATIONS];
 
 	/*
diff --git a/src/support/generation.c b/src/support/generation.c
index ed615d4c7cd..6e16d7e57fe 100644
--- a/src/support/generation.c
+++ b/src/support/generation.c
@@ -57,14 +57,12 @@ __wt_gen_next(WT_SESSION_IMPL *session, int which)
 	return (__wt_atomic_addv64(&S2C(session)->generations[which], 1));
 }
 
-#if 0
 /*
  * __wt_gen_next_drain --
  *	Switch the resource to its next generation, then wait for it to drain.
  */
 uint64_t
-	TABBED IN to avoid dist/ functions:
-	__wt_gen_next_drain(WT_SESSION_IMPL *session, int which)
+__wt_gen_next_drain(WT_SESSION_IMPL *session, int which)
 {
 	uint64_t v;
 
@@ -80,8 +78,7 @@ uint64_t
  *	Wait for the resource to drain.
  */
 void
-	TABBED IN to avoid dist/ functions:
-	__wt_gen_drain(WT_SESSION_IMPL *session, int which, uint64_t generation)
+__wt_gen_drain(WT_SESSION_IMPL *session, int which, uint64_t generation)
 {
 	WT_CONNECTION_IMPL *conn;
 	WT_SESSION_IMPL *s;
@@ -109,7 +106,14 @@ void
 			/* Ensure we only read the value once. */
 			WT_ORDERED_READ(v, s->generations[which]);
 
-			if (v == 0 || generation <= v)
+			/*
+			 * The generation argument is newer than the limit. Wait
+			 * for threads in generations older than the argument
+			 * generation, threads in argument generations are OK.
+			 *
+			 * The thread's generation may be 0 (that is, not set).
+			 */
+			if (v == 0 || v >= generation)
 				break;
 
 			/*
@@ -124,7 +128,6 @@ void
 		}
 	}
 }
-#endif
 
 /*
  * __wt_gen_oldest --
@@ -156,6 +159,7 @@ __wt_gen_oldest(WT_SESSION_IMPL *session, int which)
 
 		/* Ensure we only read the value once. */
 		WT_ORDERED_READ(v, s->generations[which]);
+
 		if (v != 0 && v < oldest)
 			oldest = v;
 	}
@@ -328,7 +332,7 @@ __wt_stash_discard_all(WT_SESSION_IMPL *session_safe, WT_SESSION_IMPL *session)
 
 	/*
 	 * This function is called during WT_CONNECTION.close to discard any
-	 * memory that remains.  For that reason, we take two WT_SESSION_IMPL
+	 * memory that remains. For that reason, we take two WT_SESSION_IMPL
 	 * arguments: session_safe is still linked to the WT_CONNECTION and
 	 * can be safely used for calls to other WiredTiger functions, while
 	 * session is the WT_SESSION_IMPL we're cleaning up.