From: Nick Piggin <piggin@cyberone.com.au>

This patch includes Chris Mason's fix to only clear queue_full when all
tasks have been woken.  Previously I think starvation and unfairness could
still occur.

With this change to the blk-fair-batches patch, Chris is showing some much
improved numbers for 2.4 - 170 ms max wait vs 2700ms without
blk-fair-batches for a dbench 90 run.  He didn't indicate how much
difference his patch alone made, but it is an important fix I think.



 drivers/block/ll_rw_blk.c |   54 ++++++++++++++++++++--------------------------
 1 files changed, 24 insertions(+), 30 deletions(-)

diff -puN drivers/block/ll_rw_blk.c~blk-fair-batches-2 drivers/block/ll_rw_blk.c
--- 25/drivers/block/ll_rw_blk.c~blk-fair-batches-2	2003-06-11 22:36:24.000000000 -0700
+++ 25-akpm/drivers/block/ll_rw_blk.c	2003-06-11 22:36:24.000000000 -0700
@@ -1319,7 +1319,8 @@ static inline struct request *blk_alloc_
 /*
  * Get a free request, queue_lock must not be held
  */
-static struct request *get_request(request_queue_t *q, int rw, int gfp_mask)
+static struct request *
+get_request(request_queue_t *q, int rw, int gfp_mask, int force)
 {
 	struct request *rq = NULL;
 	struct request_list *rl = &q->rq;
@@ -1328,7 +1329,7 @@ static struct request *get_request(reque
 	if (rl->count[rw] == q->nr_requests)
 		blk_set_queue_full(q, rw);
 	
-	if (blk_queue_full(q, rw) && !elv_may_queue(q, rw)) {
+	if (blk_queue_full(q, rw) && !force && !elv_may_queue(q, rw)) {
 		spin_unlock_irq(q->queue_lock);
 		goto out;
 	}
@@ -1345,11 +1346,10 @@ static struct request *get_request(reque
                         clear_queue_congested(q, rw);
 
 		if (rl->count[rw] <= batch_requests(q)) {
-			if (rl->count[rw] == batch_requests(q))
-				blk_clear_queue_full(q, rw);
-			
 			if (waitqueue_active(&rl->wait[rw]))
 				wake_up(&rl->wait[rw]);
+			else
+				blk_clear_queue_full(q, rw);
 		}
 
 		spin_unlock_irq(q->queue_lock);
@@ -1388,26 +1388,22 @@ static struct request *get_request_wait(
 {
 	DEFINE_WAIT(wait);
 	struct request *rq;
+	int waited = 0;
 
 	generic_unplug_device(q);
 	do {
-		rq = get_request(q, rw, GFP_NOIO);
+		struct request_list *rl = &q->rq;
 
-		if (!rq) {
-			struct request_list *rl = &q->rq;
+		prepare_to_wait_exclusive(&rl->wait[rw], &wait,
+				TASK_UNINTERRUPTIBLE);
+		
+		rq = get_request(q, rw, GFP_NOIO, waited);
 
-			prepare_to_wait_exclusive(&rl->wait[rw], &wait,
-						TASK_UNINTERRUPTIBLE);
-			/*
-			 * If _all_ the requests were suddenly returned then
-			 * no wakeup will be delivered.  So now we're on the
-			 * waitqueue, go check for that.
-			 */
-			rq = get_request(q, rw, GFP_NOIO);
-			if (!rq)
-				io_schedule();
-			finish_wait(&rl->wait[rw], &wait);
+		if (!rq) {
+			io_schedule();
+			waited = 1;
 		}
+		finish_wait(&rl->wait[rw], &wait);
 	} while (!rq);
 
 	return rq;
@@ -1419,10 +1415,10 @@ struct request *blk_get_request(request_
 
 	BUG_ON(rw != READ && rw != WRITE);
 
-	rq = get_request(q, rw, gfp_mask);
-
-	if (!rq && (gfp_mask & __GFP_WAIT))
+	if (gfp_mask & __GFP_WAIT)
 		rq = get_request_wait(q, rw);
+	else
+		rq = get_request(q, rw, gfp_mask, 0);
 
 	return rq;
 }
@@ -1575,11 +1571,10 @@ void __blk_put_request(request_queue_t *
 			clear_queue_congested(q, rw);
 
 		if (rl->count[rw] <= batch_requests(q)) {
-			if (rl->count[rw] == batch_requests(q))
-				blk_clear_queue_full(q, rw);
-			
 			if (waitqueue_active(&rl->wait[rw]))
 				wake_up(&rl->wait[rw]);
+			else
+				blk_clear_queue_full(q, rw);
 		}
 	}
 }
@@ -1823,7 +1818,7 @@ get_rq:
 		freereq = NULL;
 	} else {
 		spin_unlock_irq(q->queue_lock);
-		if ((freereq = get_request(q, rw, GFP_ATOMIC)) == NULL) {
+		if ((freereq = get_request(q, rw, GFP_ATOMIC, 0)) == NULL) {
 			/*
 			 * READA bit set
 			 */
@@ -1931,8 +1926,7 @@ static inline void blk_partition_remap(s
  * bio happens to be merged with someone else, and may change bi_dev and
  * bi_sector for remaps as it sees fit.  So the values of these fields
  * should NOT be depended on after the call to generic_make_request.
- *
- * */
+ */
 void generic_make_request(struct bio *bio)
 {
 	request_queue_t *q;
@@ -2446,14 +2440,14 @@ queue_requests_store(struct request_queu
 		blk_set_queue_full(q, READ);
 	else if (rl->count[READ] <= batch_requests(q)) {
 		blk_clear_queue_full(q, READ);
-		wake_up(&rl->wait[READ]);
+		wake_up_all(&rl->wait[READ]);
 	}
 	
 	if (rl->count[WRITE] >= q->nr_requests)
 		blk_set_queue_full(q, WRITE);
 	else if (rl->count[WRITE] <= batch_requests(q)) {
 		blk_clear_queue_full(q, WRITE);
-		wake_up(&rl->wait[WRITE]);
+		wake_up_all(&rl->wait[WRITE]);
 	}
 	
 	return ret;

_