Update to 3.4-final.
[linux-flexiantxendom0-3.2.10.git] / block / cfq-iosched.c
index 3b07ce1..3c38536 100644 (file)
@@ -54,13 +54,11 @@ static const int cfq_hist_divisor = 4;
 #define CFQQ_SECT_THR_NONROT   (sector_t)(2 * 32)
 #define CFQQ_SEEKY(cfqq)       (hweight32(cfqq->seek_history) > 32/8)
 
-#define RQ_CIC(rq)             \
-       ((struct cfq_io_context *) (rq)->elevator_private[0])
-#define RQ_CFQQ(rq)            (struct cfq_queue *) ((rq)->elevator_private[1])
-#define RQ_CFQG(rq)            (struct cfq_group *) ((rq)->elevator_private[2])
+#define RQ_CIC(rq)             icq_to_cic((rq)->elv.icq)
+#define RQ_CFQQ(rq)            (struct cfq_queue *) ((rq)->elv.priv[0])
+#define RQ_CFQG(rq)            (struct cfq_group *) ((rq)->elv.priv[1])
 
 static struct kmem_cache *cfq_pool;
-static struct kmem_cache *cfq_ioc_pool;
 
 #define CFQ_PRIO_LISTS         IOPRIO_BE_NR
 #define cfq_class_idle(cfqq)   ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
@@ -69,6 +67,14 @@ static struct kmem_cache *cfq_ioc_pool;
 #define sample_valid(samples)  ((samples) > 80)
 #define rb_entry_cfqg(node)    rb_entry((node), struct cfq_group, rb_node)
 
+struct cfq_ttime {
+       unsigned long last_end_request;
+
+       unsigned long ttime_total;
+       unsigned long ttime_samples;
+       unsigned long ttime_mean;
+};
+
 /*
  * Most of our rbtree usage is for sorting with min extraction, so
  * if we cache the leftmost node we don't have to walk down the tree
@@ -210,6 +216,12 @@ struct cfq_group {
        struct cfq_ttime ttime;
 };
 
+struct cfq_io_cq {
+       struct io_cq            icq;            /* must be the first member */
+       struct cfq_queue        *cfqq[2];
+       struct cfq_ttime        ttime;
+};
+
 /*
  * Per block device queue structure
  */
@@ -261,7 +273,7 @@ struct cfq_data {
        struct work_struct unplug_work;
 
        struct cfq_queue *active_queue;
-       struct cfq_io_context *active_cic;
+       struct cfq_io_cq *active_cic;
 
        /*
         * async queue for each priority case
@@ -283,8 +295,7 @@ struct cfq_data {
        unsigned int cfq_slice_idle;
        unsigned int cfq_group_idle;
        unsigned int cfq_latency;
-
-       struct list_head cic_list;
+       unsigned int cfq_target_latency;
 
        /*
         * Fallback dummy cfqq for extreme OOM conditions
@@ -457,24 +468,35 @@ static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
                                       struct io_context *, gfp_t);
-static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
-                                               struct io_context *);
 
-static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
-                                           bool is_sync)
+static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
+{
+       /* cic->icq is the first member, %NULL will convert to %NULL */
+       return container_of(icq, struct cfq_io_cq, icq);
+}
+
+static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
+                                              struct io_context *ioc)
+{
+       if (ioc)
+               return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
+       return NULL;
+}
+
+static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 {
        return cic->cfqq[is_sync];
 }
 
-static inline void cic_set_cfqq(struct cfq_io_context *cic,
-                               struct cfq_queue *cfqq, bool is_sync)
+static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
+                               bool is_sync)
 {
        cic->cfqq[is_sync] = cfqq;
 }
 
-static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
+static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 {
-       return cic->q->elevator->elevator_data;
+       return cic->icq.q->elevator->elevator_data;
 }
 
 /*
@@ -583,7 +605,7 @@ cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
 {
        struct cfq_rb_root *st = &cfqd->grp_service_tree;
 
-       return cfq_target_latency * cfqg->weight / st->total_weight;
+       return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
 }
 
 static inline unsigned
@@ -1541,7 +1563,7 @@ static struct request *
 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 {
        struct task_struct *tsk = current;
-       struct cfq_io_context *cic;
+       struct cfq_io_cq *cic;
        struct cfq_queue *cfqq;
 
        cic = cfq_cic_lookup(cfqd, tsk->io_context);
@@ -1635,6 +1657,8 @@ cfq_merged_requests(struct request_queue *q, struct request *rq,
                    struct request *next)
 {
        struct cfq_queue *cfqq = RQ_CFQQ(rq);
+       struct cfq_data *cfqd = q->elevator->elevator_data;
+
        /*
         * reposition in fifo if next is older than rq
         */
@@ -1649,13 +1673,23 @@ cfq_merged_requests(struct request_queue *q, struct request *rq,
        cfq_remove_request(next);
        cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
                                        rq_data_dir(next), rq_is_sync(next));
+
+       cfqq = RQ_CFQQ(next);
+       /*
+        * all requests of this queue are merged to other queues, delete it
+        * from the service tree. If it's the active_queue,
+        * cfq_dispatch_requests() will choose to expire it or do idle
+        */
+       if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
+           cfqq != cfqd->active_queue)
+               cfq_del_cfqq_rr(cfqd, cfqq);
 }
 
 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
                           struct bio *bio)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
-       struct cfq_io_context *cic;
+       struct cfq_io_cq *cic;
        struct cfq_queue *cfqq;
 
        /*
@@ -1666,18 +1700,11 @@ static int cfq_allow_merge(struct request_queue *q, struct request *rq,
 
        /*
         * Lookup the cfqq that this bio will be queued with and allow
-        * merge only if rq is queued there.  This function can be called
-        * from plug merge without queue_lock.  In such cases, ioc of @rq
-        * and %current are guaranteed to be equal.  Avoid lookup which
-        * requires queue_lock by using @rq's cic.
+        * merge only if rq is queued there.
         */
-       if (current->io_context == RQ_CIC(rq)->ioc) {
-               cic = RQ_CIC(rq);
-       } else {
-               cic = cfq_cic_lookup(cfqd, current->io_context);
-               if (!cic)
-                       return false;
-       }
+       cic = cfq_cic_lookup(cfqd, current->io_context);
+       if (!cic)
+               return false;
 
        cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
        return cfqq == RQ_CFQQ(rq);
@@ -1761,7 +1788,7 @@ __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                cfqd->active_queue = NULL;
 
        if (cfqd->active_cic) {
-               put_io_context(cfqd->active_cic->ioc, cfqd->queue);
+               put_io_context(cfqd->active_cic->icq.ioc);
                cfqd->active_cic = NULL;
        }
 }
@@ -1981,7 +2008,7 @@ static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 {
        struct cfq_queue *cfqq = cfqd->active_queue;
-       struct cfq_io_context *cic;
+       struct cfq_io_cq *cic;
        unsigned long sl, group_idle = 0;
 
        /*
@@ -2016,7 +2043,7 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
         * task has exited, don't wait
         */
        cic = cfqd->active_cic;
-       if (!cic || !atomic_read(&cic->ioc->nr_tasks))
+       if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
                return;
 
        /*
@@ -2245,7 +2272,8 @@ new_workload:
                 * to have higher weight. A more accurate thing would be to
                 * calculate system wide asnc/sync ratio.
                 */
-               tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
+               tmp = cfqd->cfq_target_latency *
+                       cfqg_busy_async_queues(cfqd, cfqg);
                tmp = tmp/cfqd->busy_queues;
                slice = min_t(unsigned, slice, tmp);
 
@@ -2567,9 +2595,9 @@ static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
        cfq_dispatch_insert(cfqd->queue, rq);
 
        if (!cfqd->active_cic) {
-               struct cfq_io_context *cic = RQ_CIC(rq);
+               struct cfq_io_cq *cic = RQ_CIC(rq);
 
-               atomic_long_inc(&cic->ioc->refcount);
+               atomic_long_inc(&cic->icq.ioc->refcount);
                cfqd->active_cic = cic;
        }
 
@@ -2652,26 +2680,6 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
        cfq_put_cfqg(cfqg);
 }
 
-static void cfq_cic_free_rcu(struct rcu_head *head)
-{
-       kmem_cache_free(cfq_ioc_pool,
-                       container_of(head, struct cfq_io_context, rcu_head));
-}
-
-static void cfq_cic_free(struct cfq_io_context *cic)
-{
-       call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
-}
-
-static void cfq_release_cic(struct cfq_io_context *cic)
-{
-       struct io_context *ioc = cic->ioc;
-
-       radix_tree_delete(&ioc->radix_root, cic->q->id);
-       hlist_del(&cic->cic_list);
-       cfq_cic_free(cic);
-}
-
 static void cfq_put_cooperator(struct cfq_queue *cfqq)
 {
        struct cfq_queue *__cfqq, *next;
@@ -2705,20 +2713,17 @@ static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
        cfq_put_queue(cfqq);
 }
 
-static void cfq_exit_cic(struct cfq_io_context *cic)
+static void cfq_init_icq(struct io_cq *icq)
 {
-       struct cfq_data *cfqd = cic_to_cfqd(cic);
-       struct io_context *ioc = cic->ioc;
+       struct cfq_io_cq *cic = icq_to_cic(icq);
 
-       list_del_init(&cic->queue_list);
+       cic->ttime.last_end_request = jiffies;
+}
 
-       /*
-        * Both setting lookup hint to and clearing it from @cic are done
-        * under queue_lock.  If it's not pointing to @cic now, it never
-        * will.  Hint assignment itself can race safely.
-        */
-       if (rcu_dereference_raw(ioc->ioc_data) == cic)
-               rcu_assign_pointer(ioc->ioc_data, NULL);
+static void cfq_exit_icq(struct io_cq *icq)
+{
+       struct cfq_io_cq *cic = icq_to_cic(icq);
+       struct cfq_data *cfqd = cic_to_cfqd(cic);
 
        if (cic->cfqq[BLK_RW_ASYNC]) {
                cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
@@ -2731,24 +2736,6 @@ static void cfq_exit_cic(struct cfq_io_context *cic)
        }
 }
 
-static struct cfq_io_context *
-cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
-{
-       struct cfq_io_context *cic;
-
-       cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
-                                                       cfqd->queue->node);
-       if (cic) {
-               cic->ttime.last_end_request = jiffies;
-               INIT_LIST_HEAD(&cic->queue_list);
-               INIT_HLIST_NODE(&cic->cic_list);
-               cic->exit = cfq_exit_cic;
-               cic->release = cfq_release_cic;
-       }
-
-       return cic;
-}
-
 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
 {
        struct task_struct *tsk = current;
@@ -2791,7 +2778,7 @@ static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
        cfq_clear_cfqq_prio_changed(cfqq);
 }
 
-static void changed_ioprio(struct cfq_io_context *cic)
+static void changed_ioprio(struct cfq_io_cq *cic)
 {
        struct cfq_data *cfqd = cic_to_cfqd(cic);
        struct cfq_queue *cfqq;
@@ -2802,7 +2789,7 @@ static void changed_ioprio(struct cfq_io_context *cic)
        cfqq = cic->cfqq[BLK_RW_ASYNC];
        if (cfqq) {
                struct cfq_queue *new_cfqq;
-               new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
+               new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
                                                GFP_ATOMIC);
                if (new_cfqq) {
                        cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
@@ -2836,7 +2823,7 @@ static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 }
 
 #ifdef CONFIG_CFQ_GROUP_IOSCHED
-static void changed_cgroup(struct cfq_io_context *cic)
+static void changed_cgroup(struct cfq_io_cq *cic)
 {
        struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
        struct cfq_data *cfqd = cic_to_cfqd(cic);
@@ -2864,7 +2851,7 @@ cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
                     struct io_context *ioc, gfp_t gfp_mask)
 {
        struct cfq_queue *cfqq, *new_cfqq = NULL;
-       struct cfq_io_context *cic;
+       struct cfq_io_cq *cic;
        struct cfq_group *cfqg;
 
 retry:
@@ -2955,158 +2942,6 @@ cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
        return cfqq;
 }
 
-/**
- * cfq_cic_lookup - lookup cfq_io_context
- * @cfqd: the associated cfq_data
- * @ioc: the associated io_context
- *
- * Look up cfq_io_context associated with @cfqd - @ioc pair.  Must be
- * called with queue_lock held.
- */
-static struct cfq_io_context *
-cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
-{
-       struct request_queue *q = cfqd->queue;
-       struct cfq_io_context *cic;
-
-       lockdep_assert_held(cfqd->queue->queue_lock);
-       if (unlikely(!ioc))
-               return NULL;
-
-       /*
-        * cic's are indexed from @ioc using radix tree and hint pointer,
-        * both of which are protected with RCU.  All removals are done
-        * holding both q and ioc locks, and we're holding q lock - if we
-        * find a cic which points to us, it's guaranteed to be valid.
-        */
-       rcu_read_lock();
-       cic = rcu_dereference(ioc->ioc_data);
-       if (cic && cic->q == q)
-               goto out;
-
-       cic = radix_tree_lookup(&ioc->radix_root, cfqd->queue->id);
-       if (cic && cic->q == q)
-               rcu_assign_pointer(ioc->ioc_data, cic); /* allowed to race */
-       else
-               cic = NULL;
-out:
-       rcu_read_unlock();
-       return cic;
-}
-
-/**
- * cfq_create_cic - create and link a cfq_io_context
- * @cfqd: cfqd of interest
- * @gfp_mask: allocation mask
- *
- * Make sure cfq_io_context linking %current->io_context and @cfqd exists.
- * If ioc and/or cic doesn't exist, they will be created using @gfp_mask.
- */
-static int cfq_create_cic(struct cfq_data *cfqd, gfp_t gfp_mask)
-{
-       struct request_queue *q = cfqd->queue;
-       struct cfq_io_context *cic = NULL;
-       struct io_context *ioc;
-       int ret = -ENOMEM;
-
-       might_sleep_if(gfp_mask & __GFP_WAIT);
-
-       /* allocate stuff */
-       ioc = current_io_context(gfp_mask, q->node);
-       if (!ioc)
-               goto out;
-
-       cic = cfq_alloc_io_context(cfqd, gfp_mask);
-       if (!cic)
-               goto out;
-
-       ret = radix_tree_preload(gfp_mask);
-       if (ret)
-               goto out;
-
-       cic->ioc = ioc;
-       cic->q = cfqd->queue;
-
-       /* lock both q and ioc and try to link @cic */
-       spin_lock_irq(q->queue_lock);
-       spin_lock(&ioc->lock);
-
-       ret = radix_tree_insert(&ioc->radix_root, q->id, cic);
-       if (likely(!ret)) {
-               hlist_add_head(&cic->cic_list, &ioc->cic_list);
-               list_add(&cic->queue_list, &cfqd->cic_list);
-               cic = NULL;
-       } else if (ret == -EEXIST) {
-               /* someone else already did it */
-               ret = 0;
-       }
-
-       spin_unlock(&ioc->lock);
-       spin_unlock_irq(q->queue_lock);
-
-       radix_tree_preload_end();
-out:
-       if (ret)
-               printk(KERN_ERR "cfq: cic link failed!\n");
-       if (cic)
-               cfq_cic_free(cic);
-       return ret;
-}
-
-/**
- * cfq_get_io_context - acquire cfq_io_context and bump refcnt on io_context
- * @cfqd: cfqd to setup cic for
- * @gfp_mask: allocation mask
- *
- * Return cfq_io_context associating @cfqd and %current->io_context and
- * bump refcnt on io_context.  If ioc or cic doesn't exist, they're created
- * using @gfp_mask.
- *
- * Must be called under queue_lock which may be released and re-acquired.
- * This function also may sleep depending on @gfp_mask.
- */
-static struct cfq_io_context *
-cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
-{
-       struct request_queue *q = cfqd->queue;
-       struct cfq_io_context *cic = NULL;
-       struct io_context *ioc;
-       int err;
-
-       lockdep_assert_held(q->queue_lock);
-
-       while (true) {
-               /* fast path */
-               ioc = current->io_context;
-               if (likely(ioc)) {
-                       cic = cfq_cic_lookup(cfqd, ioc);
-                       if (likely(cic))
-                               break;
-               }
-
-               /* slow path - unlock, create missing ones and retry */
-               spin_unlock_irq(q->queue_lock);
-               err = cfq_create_cic(cfqd, gfp_mask);
-               spin_lock_irq(q->queue_lock);
-               if (err)
-                       return NULL;
-       }
-
-       /* bump @ioc's refcnt and handle changed notifications */
-       get_io_context(ioc);
-
-       if (unlikely(cic->changed)) {
-               if (test_and_clear_bit(CIC_IOPRIO_CHANGED, &cic->changed))
-                       changed_ioprio(cic);
-#ifdef CONFIG_CFQ_GROUP_IOSCHED
-               if (test_and_clear_bit(CIC_CGROUP_CHANGED, &cic->changed))
-                       changed_cgroup(cic);
-#endif
-       }
-
-       return cic;
-}
-
 static void
 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
 {
@@ -3120,7 +2955,7 @@ __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
 
 static void
 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
-       struct cfq_io_context *cic)
+                       struct cfq_io_cq *cic)
 {
        if (cfq_cfqq_sync(cfqq)) {
                __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
@@ -3158,7 +2993,7 @@ cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
  */
 static void
 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
-                      struct cfq_io_context *cic)
+                      struct cfq_io_cq *cic)
 {
        int old_idle, enable_idle;
 
@@ -3175,8 +3010,9 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 
        if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
                enable_idle = 0;
-       else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
+       else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
+                !cfqd->cfq_slice_idle ||
+                (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
                enable_idle = 0;
        else if (sample_valid(cic->ttime.ttime_samples)) {
                if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
@@ -3276,7 +3112,7 @@ 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;
+       enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
 
        cfq_log_cfqq(cfqd, cfqq, "preempt");
        cfq_slice_expired(cfqd, 1);
@@ -3285,7 +3121,7 @@ static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
         * workload type is changed, don't save slice, otherwise preempt
         * doesn't happen
         */
-       if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
+       if (old_type != cfqq_type(cfqq))
                cfqq->cfqg->saved_workload_slice = 0;
 
        /*
@@ -3308,7 +3144,7 @@ static void
 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                struct request *rq)
 {
-       struct cfq_io_context *cic = RQ_CIC(rq);
+       struct cfq_io_cq *cic = RQ_CIC(rq);
 
        cfqd->rq_queued++;
        if (rq->cmd_flags & REQ_PRIO)
@@ -3361,7 +3197,7 @@ static void cfq_insert_request(struct request_queue *q, struct request *rq)
        struct cfq_queue *cfqq = RQ_CFQQ(rq);
 
        cfq_log_cfqq(cfqd, cfqq, "insert_request");
-       cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
+       cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
 
        rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
        list_add_tail(&rq->queuelist, &cfqq->fifo);
@@ -3411,7 +3247,7 @@ static void cfq_update_hw_tag(struct cfq_data *cfqd)
 
 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
-       struct cfq_io_context *cic = cfqd->active_cic;
+       struct cfq_io_cq *cic = cfqd->active_cic;
 
        /* If the queue already has requests, don't wait */
        if (!RB_EMPTY_ROOT(&cfqq->sort_list))
@@ -3548,7 +3384,7 @@ static int cfq_may_queue(struct request_queue *q, int rw)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
        struct task_struct *tsk = current;
-       struct cfq_io_context *cic;
+       struct cfq_io_cq *cic;
        struct cfq_queue *cfqq;
 
        /*
@@ -3563,7 +3399,7 @@ static int cfq_may_queue(struct request_queue *q, int rw)
 
        cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
        if (cfqq) {
-               cfq_init_prio_data(cfqq, cic->ioc);
+               cfq_init_prio_data(cfqq, cic->icq.ioc);
 
                return __cfq_may_queue(cfqq);
        }
@@ -3584,21 +3420,17 @@ static void cfq_put_request(struct request *rq)
                BUG_ON(!cfqq->allocated[rw]);
                cfqq->allocated[rw]--;
 
-               put_io_context(RQ_CIC(rq)->ioc, cfqq->cfqd->queue);
-
-               rq->elevator_private[0] = NULL;
-               rq->elevator_private[1] = NULL;
-
                /* Put down rq reference on cfqg */
                cfq_put_cfqg(RQ_CFQG(rq));
-               rq->elevator_private[2] = NULL;
+               rq->elv.priv[0] = NULL;
+               rq->elv.priv[1] = NULL;
 
                cfq_put_queue(cfqq);
        }
 }
 
 static struct cfq_queue *
-cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
+cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
                struct cfq_queue *cfqq)
 {
        cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
@@ -3613,7 +3445,7 @@ cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
  * was the last process referring to said cfqq.
  */
 static struct cfq_queue *
-split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
+split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
 {
        if (cfqq_process_refs(cfqq) == 1) {
                cfqq->pid = current->pid;
@@ -3636,22 +3468,29 @@ static int
 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
-       struct cfq_io_context *cic;
+       struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
        const int rw = rq_data_dir(rq);
        const bool is_sync = rq_is_sync(rq);
        struct cfq_queue *cfqq;
+       unsigned int changed;
 
        might_sleep_if(gfp_mask & __GFP_WAIT);
 
        spin_lock_irq(q->queue_lock);
-       cic = cfq_get_io_context(cfqd, gfp_mask);
-       if (!cic)
-               goto queue_fail;
+
+       /* handle changed notifications */
+       changed = icq_get_changed(&cic->icq);
+       if (unlikely(changed & ICQ_IOPRIO_CHANGED))
+               changed_ioprio(cic);
+#ifdef CONFIG_CFQ_GROUP_IOSCHED
+       if (unlikely(changed & ICQ_CGROUP_CHANGED))
+               changed_cgroup(cic);
+#endif
 
 new_queue:
        cfqq = cic_to_cfqq(cic, is_sync);
        if (!cfqq || cfqq == &cfqd->oom_cfqq) {
-               cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
+               cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
                cic_set_cfqq(cic, cfqq, is_sync);
        } else {
                /*
@@ -3677,17 +3516,10 @@ new_queue:
        cfqq->allocated[rw]++;
 
        cfqq->ref++;
-       rq->elevator_private[0] = cic;
-       rq->elevator_private[1] = cfqq;
-       rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
+       rq->elv.priv[0] = cfqq;
+       rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
        spin_unlock_irq(q->queue_lock);
        return 0;
-
-queue_fail:
-       cfq_schedule_dispatch(cfqd);
-       spin_unlock_irq(q->queue_lock);
-       cfq_log(cfqd, "set_request fail");
-       return 1;
 }
 
 static void cfq_kick_queue(struct work_struct *work)
@@ -3791,18 +3623,6 @@ static void cfq_exit_queue(struct elevator_queue *e)
        if (cfqd->active_queue)
                __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
 
-       while (!list_empty(&cfqd->cic_list)) {
-               struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
-                                                       struct cfq_io_context,
-                                                       queue_list);
-               struct io_context *ioc = cic->ioc;
-
-               spin_lock(&ioc->lock);
-               cfq_exit_cic(cic);
-               cfq_release_cic(cic);
-               spin_unlock(&ioc->lock);
-       }
-
        cfq_put_async_queues(cfqd);
        cfq_release_cfq_groups(cfqd);
 
@@ -3904,8 +3724,6 @@ static void *cfq_init_queue(struct request_queue *q)
        cfqd->oom_cfqq.ref++;
        cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
 
-       INIT_LIST_HEAD(&cfqd->cic_list);
-
        cfqd->queue = q;
 
        init_timer(&cfqd->idle_slice_timer);
@@ -3921,6 +3739,7 @@ static void *cfq_init_queue(struct request_queue *q)
        cfqd->cfq_back_penalty = cfq_back_penalty;
        cfqd->cfq_slice[0] = cfq_slice_async;
        cfqd->cfq_slice[1] = cfq_slice_sync;
+       cfqd->cfq_target_latency = cfq_target_latency;
        cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
        cfqd->cfq_slice_idle = cfq_slice_idle;
        cfqd->cfq_group_idle = cfq_group_idle;
@@ -3934,34 +3753,6 @@ static void *cfq_init_queue(struct request_queue *q)
        return cfqd;
 }
 
-static void cfq_slab_kill(void)
-{
-       /*
-        * Caller already ensured that pending RCU callbacks are completed,
-        * so we should have no busy allocations at this point.
-        */
-       if (cfq_pool)
-               kmem_cache_destroy(cfq_pool);
-       if (cfq_ioc_pool)
-               kmem_cache_destroy(cfq_ioc_pool);
-}
-
-static int __init cfq_slab_setup(void)
-{
-       cfq_pool = KMEM_CACHE(cfq_queue, 0);
-       if (!cfq_pool)
-               goto fail;
-
-       cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
-       if (!cfq_ioc_pool)
-               goto fail;
-
-       return 0;
-fail:
-       cfq_slab_kill();
-       return -ENOMEM;
-}
-
 /*
  * sysfs parts below -->
  */
@@ -4000,6 +3791,7 @@ SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
+SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
 #undef SHOW_FUNCTION
 
 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                        \
@@ -4033,6 +3825,7 @@ STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
                UINT_MAX, 0);
 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
+STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
 #undef STORE_FUNCTION
 
 #define CFQ_ATTR(name) \
@@ -4050,6 +3843,7 @@ static struct elv_fs_entry cfq_attrs[] = {
        CFQ_ATTR(slice_idle),
        CFQ_ATTR(group_idle),
        CFQ_ATTR(low_latency),
+       CFQ_ATTR(target_latency),
        __ATTR_NULL
 };
 
@@ -4067,14 +3861,18 @@ static struct elevator_type iosched_cfq = {
                .elevator_completed_req_fn =    cfq_completed_request,
                .elevator_former_req_fn =       elv_rb_former_request,
                .elevator_latter_req_fn =       elv_rb_latter_request,
+               .elevator_init_icq_fn =         cfq_init_icq,
+               .elevator_exit_icq_fn =         cfq_exit_icq,
                .elevator_set_req_fn =          cfq_set_request,
                .elevator_put_req_fn =          cfq_put_request,
                .elevator_may_queue_fn =        cfq_may_queue,
                .elevator_init_fn =             cfq_init_queue,
                .elevator_exit_fn =             cfq_exit_queue,
        },
+       .icq_size       =       sizeof(struct cfq_io_cq),
+       .icq_align      =       __alignof__(struct cfq_io_cq),
        .elevator_attrs =       cfq_attrs,
-       .elevator_name =        "cfq",
+       .elevator_name  =       "cfq",
        .elevator_owner =       THIS_MODULE,
 };
 
@@ -4092,6 +3890,8 @@ static struct blkio_policy_type blkio_policy_cfq;
 
 static int __init cfq_init(void)
 {
+       int ret;
+
        /*
         * could be 0 on HZ < 1000 setups
         */
@@ -4106,10 +3906,16 @@ static int __init cfq_init(void)
 #else
                cfq_group_idle = 0;
 #endif
-       if (cfq_slab_setup())
+       cfq_pool = KMEM_CACHE(cfq_queue, 0);
+       if (!cfq_pool)
                return -ENOMEM;
 
-       elv_register(&iosched_cfq);
+       ret = elv_register(&iosched_cfq);
+       if (ret) {
+               kmem_cache_destroy(cfq_pool);
+               return ret;
+       }
+
        blkio_policy_register(&blkio_policy_cfq);
 
        return 0;
@@ -4119,8 +3925,7 @@ static void __exit cfq_exit(void)
 {
        blkio_policy_unregister(&blkio_policy_cfq);
        elv_unregister(&iosched_cfq);
-       rcu_barrier();  /* make sure all cic RCU frees are complete */
-       cfq_slab_kill();
+       kmem_cache_destroy(cfq_pool);
 }
 
 module_init(cfq_init);