Fixes #1735 Merge from tokudb.1735 to main with command:

svn merge -r11656:HEAD ../tokudb.1735 (executed in main sandbox)

git-svn-id: file:///svn/toku/tokudb@11714 c7de825b-a66e-492c-adef-691d508d4ae1

newbrt/brt.c

@@ -2333,7 +2333,7 @@ brt_merge_child (BRT t, BRTNODE node, int childnum_to_merge, BOOL *did_io, BOOL
     return r;
 }
 
-// TODO #1398  Get rid of this entire straddle_callback hack
+// TODO: #1398  Get rid of this entire straddle_callback hack
 #ifdef  BRT_LEVEL_STRADDLE_CALLBACK_LOGIC_NOT_READY
 int STRADDLE_HACK_INSIDE_CALLBACK = 0;
 #endif

newbrt/cachetable.c

@@ -1014,7 +1014,7 @@ int toku_cachetable_get_and_pin(CACHEFILE cachefile, CACHEKEY key, u_int32_t ful
 		write_pair_for_checkpoint(ct, p); // releases the pair_write_lock, but not the cachetable lock
 	    }
 	    // still have the cachetable lock
-	    // TODO 1398  kill this hack before it multiplies further
+	    // TODO: #1398  kill this hack before it multiplies further
 	    // This logic here to prevent deadlock that results when a query pins a node,
 	    // then the straddle callback creates a cursor that pins it again.  If 
 	    // toku_cachetable_end_checkpoint() is called between those two calls to pin
@@ -1087,9 +1087,15 @@ int toku_cachetable_get_and_pin(CACHEFILE cachefile, CACHEKEY key, u_int32_t ful
 }
 
 // Lookup a key in the cachetable.  If it is found and it is not being written, then
-// acquire a read lock on the pair, update the LRU list, and return sucess.  However,
-// if it is being written, then allow the writer to evict it.  This prevents writers
-// being suspended on a block that was just selected for eviction.
+// acquire a read lock on the pair, update the LRU list, and return sucess.
+//
+// However, if the page is clean or has checkpoint pending, don't return success.
+// This will minimize the number of dirty nodes.
+// Rationale:  maybe_get_and_pin is used when the system has an alternative to modifying a node.
+//  In the context of checkpointing, we don't want to gratuituously dirty a page, because it causes an I/O.
+//  For example, imagine that we can modify a bit in a dirty parent, or modify a bit in a clean child, then we should modify
+//  the dirty parent (which will have to do I/O eventually anyway) rather than incur a full block write to modify one bit.
+//  Similarly, if the checkpoint is actually pending, we don't want to block on it.
 int toku_cachetable_maybe_get_and_pin (CACHEFILE cachefile, CACHEKEY key, u_int32_t fullhash, void**value) {
     CACHETABLE ct = cachefile->cachetable;
     PAIR p;
@@ -1100,12 +1106,9 @@ int toku_cachetable_maybe_get_and_pin (CACHEFILE cachefile, CACHEKEY key, u_int3
 	count++;
 	if (p->key.b==key.b && p->cachefile==cachefile && p->state == CTPAIR_IDLE) {
 
-	    if (p->checkpoint_pending) {
-		rwlock_write_lock(&p->rwlock, ct->mutex);
-		write_pair_for_checkpoint(ct, p); // releases the pair_write_lock, but not the cachetable lock
+	    if (p->checkpoint_pending || !p->dirty) {
+		goto finish;
 	    }
-	    // still have the cachetable lock
-
 	    *value = p->value;
 	    rwlock_read_lock(&p->rwlock, ct->mutex);
 	    lru_touch(ct,p);
@@ -1114,6 +1117,7 @@ int toku_cachetable_maybe_get_and_pin (CACHEFILE cachefile, CACHEKEY key, u_int3
             break;
 	}
     }
+ finish:
     note_hash_count(count);
     cachetable_unlock(ct);
     return r;

newbrt/cachetable.h

@@ -9,8 +9,7 @@
 #include "brttypes.h"
 #include "workqueue.h"
 
-
-// TODO #1398  Get rid of this entire straddle_callback hack
+// TODO: #1398  Get rid of this entire straddle_callback hack
 // Man is this ugly.  
 #ifdef  BRT_LEVEL_STRADDLE_CALLBACK_LOGIC_NOT_READY
 extern int STRADDLE_HACK_INSIDE_CALLBACK;

newbrt/rwlock.h

@@ -71,9 +71,7 @@ static inline void rwlock_read_lock(RWLOCK rwlock, toku_pthread_mutex_t *mutex)
 }
 
 
-
-
-// TODO 1398  Get rid of this hack.
+// TODO: #1398  Get rid of this hack.
 #ifdef  BRT_LEVEL_STRADDLE_CALLBACK_LOGIC_NOT_READY
 
 // preferentially obtain a read lock (ignore request for write lock)
@@ -88,7 +86,6 @@ static inline void rwlock_prefer_read_lock(RWLOCK rwlock, toku_pthread_mutex_t *
 #endif
 
 
-
 // release a read lock
 // expects: mutex is locked