dm cache policy mq: tweak algorithm that decides when to promote a block

Rather than maintaining a separate promote_threshold variable that we
periodically update we now use the hit count of the oldest clean
block. Also add a fudge factor to discourage demoting dirty blocks.

With some tests this has a sizeable difference, because the old code
was too eager to demote blocks. For example, device-mapper-test-suite's
git_extract_cache_quick test goes from taking 190 seconds, to 142
(linear on spindle takes 250).

Signed-off-by: Joe Thornber <[e-mail]>
Signed-off-by: Mike Snitzer <[e-mail]>

Documentation/device-mapper/cache-policies.txt

···
 contiguous I/Os to try to spot when the io is in one of these sequential
 modes.
 
-Internally the mq policy maintains a promotion threshold variable.  If
-the hit count of a block not in the cache goes above this threshold it
-gets promoted to the cache.  The read, write and discard promote adjustment
+Internally the mq policy determines a promotion threshold.  If the hit
+count of a block not in the cache goes above this threshold it gets
+promoted to the cache.  The read, write and discard promote adjustment
 tunables allow you to tweak the promotion threshold by adding a small
 value based on the io type.  They default to 4, 8 and 1 respectively.
 If you're trying to quickly warm a new cache device you may wish to

drivers/md/dm-cache-policy-mq.c

···
  * Gives us the oldest entry of the lowest popoulated level.  If the first
  * level is emptied then we shift down one level.
  */
-static struct list_head *queue_pop(struct queue *q)
+static struct list_head *queue_peek(struct queue *q)
 {
 	unsigned level;
-	struct list_head *r;
 
 	for (level = 0; level < NR_QUEUE_LEVELS; level++)
-		if (!list_empty(q->qs + level)) {
-			r = q->qs[level].next;
-			list_del(r);
-
-			/* have we just emptied the bottom level? */
-			if (level == 0 && list_empty(q->qs))
-				queue_shift_down(q);
-
-			return r;
-		}
+		if (!list_empty(q->qs + level))
+			return q->qs[level].next;
 
 	return NULL;
+}
+
+static struct list_head *queue_pop(struct queue *q)
+{
+	struct list_head *r = queue_peek(q);
+
+	if (r) {
+		list_del(r);
+
+		/* have we just emptied the bottom level? */
+		if (list_empty(q->qs))
+			queue_shift_down(q);
+	}
+
+	return r;
 }
 
 static struct list_head *list_pop(struct list_head *lh)
···
 	unsigned generation;
 	unsigned generation_period; /* in lookups (will probably change) */
 
-	/*
-	 * Entries in the pre_cache whose hit count passes the promotion
-	 * threshold move to the cache proper.  Working out the correct
-	 * value for the promotion_threshold is crucial to this policy.
-	 */
-	unsigned promote_threshold;
-
 	unsigned discard_promote_adjustment;
 	unsigned read_promote_adjustment;
 	unsigned write_promote_adjustment;
···
 #define DEFAULT_DISCARD_PROMOTE_ADJUSTMENT 1
 #define DEFAULT_READ_PROMOTE_ADJUSTMENT 4
 #define DEFAULT_WRITE_PROMOTE_ADJUSTMENT 8
+#define DISCOURAGE_DEMOTING_DIRTY_THRESHOLD 128
 
 /*----------------------------------------------------------------*/
 
···
 	return e;
 }
 
+static struct entry *peek(struct queue *q)
+{
+	struct list_head *h = queue_peek(q);
+	return h ? container_of(h, struct entry, list) : NULL;
+}
+
 /*
  * Has this entry already been updated?
  */
···
 					break;
 			}
 		}
-
-		mq->promote_threshold = nr ? total / nr : 1;
-		if (mq->promote_threshold * nr < total)
-			mq->promote_threshold++;
 	}
 }
 
···
 }
 
 /*
+ * Entries in the pre_cache whose hit count passes the promotion
+ * threshold move to the cache proper.  Working out the correct
+ * value for the promotion_threshold is crucial to this policy.
+ */
+static unsigned promote_threshold(struct mq_policy *mq)
+{
+	struct entry *e;
+
+	if (any_free_cblocks(mq))
+		return 0;
+
+	e = peek(&mq->cache_clean);
+	if (e)
+		return e->hit_count;
+
+	e = peek(&mq->cache_dirty);
+	if (e)
+		return e->hit_count + DISCOURAGE_DEMOTING_DIRTY_THRESHOLD;
+
+	/* This should never happen */
+	return 0;
+}
+
+/*
  * We modify the basic promotion_threshold depending on the specific io.
  *
  * If the origin block has been discarded then there's no cost to copy it
···
 					   bool discarded_oblock, int data_dir)
 {
 	if (data_dir == READ)
-		return mq->promote_threshold + mq->read_promote_adjustment;
+		return promote_threshold(mq) + mq->read_promote_adjustment;
 
 	if (discarded_oblock && (any_free_cblocks(mq) || any_clean_cblocks(mq))) {
 		/*
···
 		return mq->discard_promote_adjustment;
 	}
 
-	return mq->promote_threshold + mq->write_promote_adjustment;
+	return promote_threshold(mq) + mq->write_promote_adjustment;
 }
 
 static bool should_promote(struct mq_policy *mq, struct entry *e,
···
 	mq->tick = 0;
 	mq->hit_count = 0;
 	mq->generation = 0;
-	mq->promote_threshold = 0;
 	mq->discard_promote_adjustment = DEFAULT_DISCARD_PROMOTE_ADJUSTMENT;
 	mq->read_promote_adjustment = DEFAULT_READ_PROMOTE_ADJUSTMENT;
 	mq->write_promote_adjustment = DEFAULT_WRITE_PROMOTE_ADJUSTMENT;