iosched: prevent aliased requests from starving other I/O

Hi, Jens,

If you recall, I posted an RFC patch for this back in July of last year:
http://lkml.org/lkml/2010/7/13/279

The basic problem is that a process can issue a never-ending stream of
async direct I/Os to the same sector on a device, thus starving out
other I/O in the system (due to the way the alias handling works in both
cfq and deadline).  The solution I proposed back then was to start
dispatching from the fifo after a certain number of aliases had been
dispatched.  Vivek asked why we had to treat aliases differently at all,
and I never had a good answer.  So, I put together a simple patch which
allows aliases to be added to the rb tree (it adds them to the right,
though that doesn't matter as the order isn't guaranteed anyway).  I
think this is the preferred solution, as it doesn't break up time slices
in CFQ or batches in deadline.  I've tested it, and it does solve the
starvation issue.  Let me know what you think.

Cheers,
Jeff

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

block/cfq-iosched.c

@@ -1502,16 +1502,11 @@ static void cfq_add_rq_rb(struct request *rq)
 {
 	struct cfq_queue *cfqq = RQ_CFQQ(rq);
 	struct cfq_data *cfqd = cfqq->cfqd;
-	struct request *__alias, *prev;
+	struct request *prev;
 
 	cfqq->queued[rq_is_sync(rq)]++;
 
-	/*
-	 * looks a little odd, but the first insert might return an alias.
-	 * if that happens, put the alias on the dispatch list
-	 */
-	while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
-		cfq_dispatch_insert(cfqd->queue, __alias);
+	elv_rb_add(&cfqq->sort_list, rq);
 
 	if (!cfq_cfqq_on_rr(cfqq))
 		cfq_add_cfqq_rr(cfqd, cfqq);

block/deadline-iosched.c

@@ -77,10 +77,8 @@ static void
 deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
 {
 	struct rb_root *root = deadline_rb_root(dd, rq);
-	struct request *__alias;
 
-	while (unlikely(__alias = elv_rb_add(root, rq)))
-		deadline_move_request(dd, __alias);
+	elv_rb_add(root, rq);
 }
 
 static inline void

block/elevator.c

@@ -353,7 +353,7 @@ static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
  * RB-tree support functions for inserting/lookup/removal of requests
  * in a sorted RB tree.
  */
-struct request *elv_rb_add(struct rb_root *root, struct request *rq)
+void elv_rb_add(struct rb_root *root, struct request *rq)
 {
 	struct rb_node **p = &root->rb_node;
 	struct rb_node *parent = NULL;
@@ -365,15 +365,12 @@ struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 
 		if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 			p = &(*p)->rb_left;
-		else if (blk_rq_pos(rq) > blk_rq_pos(__rq))
+		else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
 			p = &(*p)->rb_right;
-		else
-			return __rq;
 	}
 
 	rb_link_node(&rq->rb_node, parent, p);
 	rb_insert_color(&rq->rb_node, root);
-	return NULL;
 }
 EXPORT_SYMBOL(elv_rb_add);
 

block/vr-iosched.c

@@ -73,13 +73,7 @@ return q->elevator->elevator_data;
 static void
 vr_add_rq_rb(struct vr_data *vd, struct request *rq)
 {
-struct request *alias = elv_rb_add(&vd->sort_list, rq);
-
-if (unlikely(alias)) {
-vr_move_request(vd, alias);
-alias = elv_rb_add(&vd->sort_list, rq);
-BUG_ON(alias);
-}
+elv_rb_add(&vd->sort_list, rq);
 
 if (blk_rq_pos(rq) >= vd->last_sector) {
 if (!vd->next_rq || blk_rq_pos(vd->next_rq) > blk_rq_pos(rq))

include/linux/elevator.h

@@ -146,7 +146,7 @@ extern struct request *elv_rb_latter_request(struct request_queue *, struct requ
 /*
  * rb support functions.
  */
-extern struct request *elv_rb_add(struct rb_root *, struct request *);
+extern void elv_rb_add(struct rb_root *, struct request *);
 extern void elv_rb_del(struct rb_root *, struct request *);
 extern struct request *elv_rb_find(struct rb_root *, sector_t);