UBUNTU: Ubuntu-2.6.38-12.51
[linux-flexiantxendom0-natty.git] / block / cfq-iosched.c
index 5d0349d..ea83a4f 100644 (file)
@@ -96,7 +96,7 @@ struct cfq_rb_root {
  */
 struct cfq_queue {
        /* reference count */
-       atomic_t ref;
+       int ref;
        /* various state flags, see below */
        unsigned int flags;
        /* parent cfq_data */
@@ -207,7 +207,7 @@ struct cfq_group {
        struct blkio_group blkg;
 #ifdef CONFIG_CFQ_GROUP_IOSCHED
        struct hlist_node cfqd_node;
-       atomic_t ref;
+       int ref;
 #endif
        /* number of requests that are on the dispatch list or inside driver */
        int dispatched;
@@ -598,8 +598,8 @@ cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
        return cfq_target_latency * cfqg->weight / st->total_weight;
 }
 
-static inline void
-cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
+static inline unsigned
+cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
        unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
        if (cfqd->cfq_latency) {
@@ -625,6 +625,14 @@ cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
                                    low_slice);
                }
        }
+       return slice;
+}
+
+static inline void
+cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
+{
+       unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
+
        cfqq->slice_start = jiffies;
        cfqq->slice_end = jiffies + slice;
        cfqq->allocated_slice = slice;
@@ -1014,11 +1022,11 @@ cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
         * elevator which will be dropped by either elevator exit
         * or cgroup deletion path depending on who is exiting first.
         */
-       atomic_set(&cfqg->ref, 1);
+       cfqg->ref = 1;
 
        /*
         * Add group onto cgroup list. It might happen that bdi->dev is
-        * not initiliazed yet. Initialize this new group without major
+        * not initialized yet. Initialize this new group without major
         * and minor info and this info will be filled in once a new thread
         * comes for IO. See code above.
         */
@@ -1059,7 +1067,7 @@ static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
 
 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
 {
-       atomic_inc(&cfqg->ref);
+       cfqg->ref++;
        return cfqg;
 }
 
@@ -1071,7 +1079,7 @@ static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
 
        cfqq->cfqg = cfqg;
        /* cfqq reference on cfqg */
-       atomic_inc(&cfqq->cfqg->ref);
+       cfqq->cfqg->ref++;
 }
 
 static void cfq_put_cfqg(struct cfq_group *cfqg)
@@ -1079,8 +1087,9 @@ static void cfq_put_cfqg(struct cfq_group *cfqg)
        struct cfq_rb_root *st;
        int i, j;
 
-       BUG_ON(atomic_read(&cfqg->ref) <= 0);
-       if (!atomic_dec_and_test(&cfqg->ref))
+       BUG_ON(cfqg->ref <= 0);
+       cfqg->ref--;
+       if (cfqg->ref)
                return;
        for_each_cfqg_st(cfqg, i, j, st)
                BUG_ON(!RB_EMPTY_ROOT(&st->rb));
@@ -1188,7 +1197,7 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                        cfq_group_service_tree_del(cfqd, cfqq->cfqg);
                cfqq->orig_cfqg = cfqq->cfqg;
                cfqq->cfqg = &cfqd->root_group;
-               atomic_inc(&cfqd->root_group.ref);
+               cfqd->root_group.ref++;
                group_changed = 1;
        } else if (!cfqd->cfq_group_isolation
                   && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
@@ -1660,8 +1669,11 @@ __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        /*
         * store what was left of this slice, if the queue idled/timed out
         */
-       if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
-               cfqq->slice_resid = cfqq->slice_end - jiffies;
+       if (timed_out) {
+               if (cfq_cfqq_slice_new(cfqq))
+                       cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
+               else
+                       cfqq->slice_resid = cfqq->slice_end - jiffies;
                cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
        }
 
@@ -2025,7 +2037,7 @@ static int cfqq_process_refs(struct cfq_queue *cfqq)
        int process_refs, io_refs;
 
        io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
-       process_refs = atomic_read(&cfqq->ref) - io_refs;
+       process_refs = cfqq->ref - io_refs;
        BUG_ON(process_refs < 0);
        return process_refs;
 }
@@ -2065,10 +2077,10 @@ static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
         */
        if (new_process_refs >= process_refs) {
                cfqq->new_cfqq = new_cfqq;
-               atomic_add(process_refs, &new_cfqq->ref);
+               new_cfqq->ref += process_refs;
        } else {
                new_cfqq->new_cfqq = cfqq;
-               atomic_add(new_process_refs, &cfqq->ref);
+               cfqq->ref += new_process_refs;
        }
 }
 
@@ -2101,12 +2113,7 @@ static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
        unsigned count;
        struct cfq_rb_root *st;
        unsigned group_slice;
-
-       if (!cfqg) {
-               cfqd->serving_prio = IDLE_WORKLOAD;
-               cfqd->workload_expires = jiffies + 1;
-               return;
-       }
+       enum wl_prio_t original_prio = cfqd->serving_prio;
 
        /* Choose next priority. RT > BE > IDLE */
        if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
@@ -2119,6 +2126,9 @@ static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
                return;
        }
 
+       if (original_prio != cfqd->serving_prio)
+               goto new_workload;
+
        /*
         * For RT and BE, we have to choose also the type
         * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
@@ -2133,6 +2143,7 @@ static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
        if (count && !time_after(jiffies, cfqd->workload_expires))
                return;
 
+new_workload:
        /* otherwise select new workload type */
        cfqd->serving_type =
                cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
@@ -2533,9 +2544,10 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
        struct cfq_data *cfqd = cfqq->cfqd;
        struct cfq_group *cfqg, *orig_cfqg;
 
-       BUG_ON(atomic_read(&cfqq->ref) <= 0);
+       BUG_ON(cfqq->ref <= 0);
 
-       if (!atomic_dec_and_test(&cfqq->ref))
+       cfqq->ref--;
+       if (cfqq->ref)
                return;
 
        cfq_log_cfqq(cfqd, cfqq, "put_queue");
@@ -2838,7 +2850,7 @@ static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        RB_CLEAR_NODE(&cfqq->p_node);
        INIT_LIST_HEAD(&cfqq->fifo);
 
-       atomic_set(&cfqq->ref, 0);
+       cfqq->ref = 0;
        cfqq->cfqd = cfqd;
 
        cfq_mark_cfqq_prio_changed(cfqq);
@@ -2974,11 +2986,11 @@ cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
         * pin the queue now that it's allocated, scheduler exit will prune it
         */
        if (!is_sync && !(*async_cfqq)) {
-               atomic_inc(&cfqq->ref);
+               cfqq->ref++;
                *async_cfqq = cfqq;
        }
 
-       atomic_inc(&cfqq->ref);
+       cfqq->ref++;
        return cfqq;
 }
 
@@ -3283,10 +3295,19 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
  */
 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
+       struct cfq_queue *old_cfqq = cfqd->active_queue;
+
        cfq_log_cfqq(cfqd, cfqq, "preempt");
        cfq_slice_expired(cfqd, 1);
 
        /*
+        * workload type is changed, don't save slice, otherwise preempt
+        * doesn't happen
+        */
+       if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
+               cfqq->cfqg->saved_workload_slice = 0;
+
+       /*
         * Put the new queue at the front of the of the current list,
         * so we know that it will be selected next.
         */
@@ -3334,7 +3355,7 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                            cfqd->busy_queues > 1) {
                                cfq_del_timer(cfqd, cfqq);
                                cfq_clear_cfqq_wait_request(cfqq);
-                               __blk_run_queue(cfqd->queue);
+                               __blk_run_queue(cfqd->queue, false);
                        } else {
                                cfq_blkiocg_update_idle_time_stats(
                                                &cfqq->cfqg->blkg);
@@ -3349,7 +3370,7 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                 * this new queue is RT and the current one is BE
                 */
                cfq_preempt_queue(cfqd, cfqq);
-               __blk_run_queue(cfqd->queue);
+               __blk_run_queue(cfqd->queue, false);
        }
 }
 
@@ -3411,6 +3432,10 @@ static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
        struct cfq_io_context *cic = cfqd->active_cic;
 
+       /* If the queue already has requests, don't wait */
+       if (!RB_EMPTY_ROOT(&cfqq->sort_list))
+               return false;
+
        /* If there are other queues in the group, don't wait */
        if (cfqq->cfqg->nr_cfqq > 1)
                return false;
@@ -3680,13 +3705,13 @@ new_queue:
        }
 
        cfqq->allocated[rw]++;
-       atomic_inc(&cfqq->ref);
-
-       spin_unlock_irqrestore(q->queue_lock, flags);
-
+       cfqq->ref++;
        rq->elevator_private = cic;
        rq->elevator_private2 = cfqq;
        rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
+
+       spin_unlock_irqrestore(q->queue_lock, flags);
+
        return 0;
 
 queue_fail:
@@ -3706,7 +3731,7 @@ static void cfq_kick_queue(struct work_struct *work)
        struct request_queue *q = cfqd->queue;
 
        spin_lock_irq(q->queue_lock);
-       __blk_run_queue(cfqd->queue);
+       __blk_run_queue(cfqd->queue, false);
        spin_unlock_irq(q->queue_lock);
 }
 
@@ -3861,6 +3886,10 @@ static void *cfq_init_queue(struct request_queue *q)
        if (!cfqd)
                return NULL;
 
+       /*
+        * Don't need take queue_lock in the routine, since we are
+        * initializing the ioscheduler, and nobody is using cfqd
+        */
        cfqd->cic_index = i;
 
        /* Init root service tree */
@@ -3880,7 +3909,7 @@ static void *cfq_init_queue(struct request_queue *q)
         * Take a reference to root group which we never drop. This is just
         * to make sure that cfq_put_cfqg() does not try to kfree root group
         */
-       atomic_set(&cfqg->ref, 1);
+       cfqg->ref = 1;
        rcu_read_lock();
        cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
                                        (void *)cfqd, 0);
@@ -3900,7 +3929,7 @@ static void *cfq_init_queue(struct request_queue *q)
         * will not attempt to free it.
         */
        cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
-       atomic_inc(&cfqd->oom_cfqq.ref);
+       cfqd->oom_cfqq.ref++;
        cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
 
        INIT_LIST_HEAD(&cfqd->cic_list);