[Bluetooth] Fix reference count when connection lookup fails
[pandora-kernel.git] / block / cfq-iosched.c
index 3c5fd9c..d3d7613 100644 (file)
@@ -4,7 +4,7 @@
  *  Based on ideas from a previously unfinished io
  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
  *
- *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
+ *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
  */
 #include <linux/module.h>
 #include <linux/blkdev.h>
@@ -17,7 +17,6 @@
  * tunables
  */
 static const int cfq_quantum = 4;              /* max queue in one round of service */
-static const int cfq_queued = 8;               /* minimum rq allocate limit per-queue*/
 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
 static const int cfq_back_max = 16 * 1024;     /* maximum backwards seek, in KiB */
 static const int cfq_back_penalty = 2;         /* penalty of a backwards seek */
@@ -32,8 +31,6 @@ static int cfq_slice_idle = HZ / 125;
 
 #define CFQ_KEY_ASYNC          (0)
 
-static DEFINE_SPINLOCK(cfq_exit_lock);
-
 /*
  * for the hash of cfqq inside the cfqd
  */
@@ -49,12 +46,11 @@ static DEFINE_SPINLOCK(cfq_exit_lock);
 static kmem_cache_t *cfq_pool;
 static kmem_cache_t *cfq_ioc_pool;
 
-static atomic_t ioc_count = ATOMIC_INIT(0);
+static DEFINE_PER_CPU(unsigned long, ioc_count);
 static struct completion *ioc_gone;
 
 #define CFQ_PRIO_LISTS         IOPRIO_BE_NR
 #define cfq_class_idle(cfqq)   ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
-#define cfq_class_be(cfqq)     ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
 #define cfq_class_rt(cfqq)     ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
 
 #define ASYNC                  (0)
@@ -85,11 +81,6 @@ struct cfq_data {
        struct list_head idle_rr;
        unsigned int busy_queues;
 
-       /*
-        * non-ordered list of empty cfqq's
-        */
-       struct list_head empty_list;
-
        /*
         * cfqq lookup hash
         */
@@ -98,9 +89,6 @@ struct cfq_data {
        int rq_in_driver;
        int hw_tag;
 
-       /*
-        * schedule slice state info
-        */
        /*
         * idle window management
         */
@@ -117,13 +105,10 @@ struct cfq_data {
        sector_t last_sector;
        unsigned long last_end_request;
 
-       unsigned int rq_starved;
-
        /*
         * tunables, see top of file
         */
        unsigned int cfq_quantum;
-       unsigned int cfq_queued;
        unsigned int cfq_fifo_expire[2];
        unsigned int cfq_back_penalty;
        unsigned int cfq_back_max;
@@ -146,7 +131,7 @@ struct cfq_queue {
        struct hlist_node cfq_hash;
        /* hash key */
        unsigned int key;
-       /* on either rr or empty list of cfqd */
+       /* member of the rr/busy/cur/idle cfqd list */
        struct list_head cfq_list;
        /* sorted list of pending requests */
        struct rb_root sort_list;
@@ -156,13 +141,14 @@ struct cfq_queue {
        int queued[2];
        /* currently allocated requests */
        int allocated[2];
+       /* pending metadata requests */
+       int meta_pending;
        /* fifo list of requests in sort_list */
        struct list_head fifo;
 
        unsigned long slice_start;
        unsigned long slice_end;
        unsigned long slice_left;
-       unsigned long service_last;
 
        /* number of requests that are on the dispatch list */
        int on_dispatch[2];
@@ -184,6 +170,7 @@ enum cfqq_state_flags {
        CFQ_CFQQ_FLAG_fifo_expire,
        CFQ_CFQQ_FLAG_idle_window,
        CFQ_CFQQ_FLAG_prio_changed,
+       CFQ_CFQQ_FLAG_queue_new,
 };
 
 #define CFQ_CFQQ_FNS(name)                                             \
@@ -208,6 +195,7 @@ CFQ_CFQQ_FNS(must_dispatch);
 CFQ_CFQQ_FNS(fifo_expire);
 CFQ_CFQQ_FNS(idle_window);
 CFQ_CFQQ_FNS(prio_changed);
+CFQ_CFQQ_FNS(queue_new);
 #undef CFQ_CFQQ_FNS
 
 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
@@ -262,6 +250,10 @@ cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
                return rq1;
        else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
                return rq2;
+       if (rq_is_meta(rq1) && !rq_is_meta(rq2))
+               return rq1;
+       else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
+               return rq2;
 
        s1 = rq1->sector;
        s2 = rq2->sector;
@@ -360,7 +352,7 @@ cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 {
        struct cfq_data *cfqd = cfqq->cfqd;
-       struct list_head *list, *entry;
+       struct list_head *list;
 
        BUG_ON(!cfq_cfqq_on_rr(cfqq));
 
@@ -385,31 +377,26 @@ static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
        }
 
        /*
-        * if queue was preempted, just add to front to be fair. busy_rr
-        * isn't sorted, but insert at the back for fairness.
+        * If this queue was preempted or is new (never been serviced), let
+        * it be added first for fairness but beind other new queues.
+        * Otherwise, just add to the back  of the list.
         */
-       if (preempted || list == &cfqd->busy_rr) {
-               if (preempted)
-                       list = list->prev;
+       if (preempted || cfq_cfqq_queue_new(cfqq)) {
+               struct list_head *n = list;
+               struct cfq_queue *__cfqq;
 
-               list_add_tail(&cfqq->cfq_list, list);
-               return;
-       }
+               while (n->next != list) {
+                       __cfqq = list_entry_cfqq(n->next);
+                       if (!cfq_cfqq_queue_new(__cfqq))
+                               break;
 
-       /*
-        * sort by when queue was last serviced
-        */
-       entry = list;
-       while ((entry = entry->prev) != list) {
-               struct cfq_queue *__cfqq = list_entry_cfqq(entry);
+                       n = n->next;
+               }
 
-               if (!__cfqq->service_last)
-                       break;
-               if (time_before(__cfqq->service_last, cfqq->service_last))
-                       break;
+               list = n;
        }
 
-       list_add(&cfqq->cfq_list, entry);
+       list_add_tail(&cfqq->cfq_list, list);
 }
 
 /*
@@ -431,7 +418,7 @@ cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
        BUG_ON(!cfq_cfqq_on_rr(cfqq));
        cfq_clear_cfqq_on_rr(cfqq);
-       list_move(&cfqq->cfq_list, &cfqd->empty_list);
+       list_del_init(&cfqq->cfq_list);
 
        BUG_ON(!cfqd->busy_queues);
        cfqd->busy_queues--;
@@ -484,12 +471,14 @@ cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 {
        struct task_struct *tsk = current;
        pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
-       sector_t sector = bio->bi_sector + bio_sectors(bio);
        struct cfq_queue *cfqq;
 
        cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
-       if (cfqq)
+       if (cfqq) {
+               sector_t sector = bio->bi_sector + bio_sectors(bio);
+
                return elv_rb_find(&cfqq->sort_list, sector);
+       }
 
        return NULL;
 }
@@ -527,6 +516,11 @@ static void cfq_remove_request(struct request *rq)
 
        list_del_init(&rq->queuelist);
        cfq_del_rq_rb(rq);
+
+       if (rq_is_meta(rq)) {
+               WARN_ON(!cfqq->meta_pending);
+               cfqq->meta_pending--;
+       }
 }
 
 static int
@@ -599,13 +593,12 @@ __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        if (cfq_cfqq_wait_request(cfqq))
                del_timer(&cfqd->idle_slice_timer);
 
-       if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
-               cfqq->service_last = now;
+       if (!preempted && !cfq_cfqq_dispatched(cfqq))
                cfq_schedule_dispatch(cfqd);
-       }
 
        cfq_clear_cfqq_must_dispatch(cfqq);
        cfq_clear_cfqq_wait_request(cfqq);
+       cfq_clear_cfqq_queue_new(cfqq);
 
        /*
         * store what was left of this slice, if the queue idled out
@@ -699,26 +692,25 @@ static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 {
        struct cfq_queue *cfqq = NULL;
 
-       /*
-        * if current list is non-empty, grab first entry. if it is empty,
-        * get next prio level and grab first entry then if any are spliced
-        */
-       if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
+       if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
+               /*
+                * if current list is non-empty, grab first entry. if it is
+                * empty, get next prio level and grab first entry then if any
+                * are spliced
+                */
                cfqq = list_entry_cfqq(cfqd->cur_rr.next);
-
-       /*
-        * If no new queues are available, check if the busy list has some
-        * before falling back to idle io.
-        */
-       if (!cfqq && !list_empty(&cfqd->busy_rr))
+       } else if (!list_empty(&cfqd->busy_rr)) {
+               /*
+                * If no new queues are available, check if the busy list has
+                * some before falling back to idle io.
+                */
                cfqq = list_entry_cfqq(cfqd->busy_rr.next);
-
-       /*
-        * if we have idle queues and no rt or be queues had pending
-        * requests, either allow immediate service if the grace period
-        * has passed or arm the idle grace timer
-        */
-       if (!cfqq && !list_empty(&cfqd->idle_rr)) {
+       } else if (!list_empty(&cfqd->idle_rr)) {
+               /*
+                * if we have idle queues and no rt or be queues had pending
+                * requests, either allow immediate service if the grace period
+                * has passed or arm the idle grace timer
+                */
                unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 
                if (time_after_eq(jiffies, end))
@@ -793,18 +785,19 @@ static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 {
        struct cfq_data *cfqd = cfqq->cfqd;
        struct request *rq;
+       int fifo;
 
        if (cfq_cfqq_fifo_expire(cfqq))
                return NULL;
+       if (list_empty(&cfqq->fifo))
+               return NULL;
 
-       if (!list_empty(&cfqq->fifo)) {
-               int fifo = cfq_cfqq_class_sync(cfqq);
+       fifo = cfq_cfqq_class_sync(cfqq);
+       rq = rq_entry_fifo(cfqq->fifo.next);
 
-               rq = rq_entry_fifo(cfqq->fifo.next);
-               if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
-                       cfq_mark_cfqq_fifo_expire(cfqq);
-                       return rq;
-               }
+       if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
+               cfq_mark_cfqq_fifo_expire(cfqq);
+               return rq;
        }
 
        return NULL;
@@ -1049,7 +1042,7 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
        kmem_cache_free(cfq_pool, cfqq);
 }
 
-static inline struct cfq_queue *
+static struct cfq_queue *
 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
                    const int hashval)
 {
@@ -1086,62 +1079,63 @@ static void cfq_free_io_context(struct io_context *ioc)
                freed++;
        }
 
-       if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
+       elv_ioc_count_mod(ioc_count, -freed);
+
+       if (ioc_gone && !elv_ioc_count_read(ioc_count))
                complete(ioc_gone);
 }
 
-static void cfq_trim(struct io_context *ioc)
+static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
-       ioc->set_ioprio = NULL;
-       cfq_free_io_context(ioc);
+       if (unlikely(cfqq == cfqd->active_queue))
+               __cfq_slice_expired(cfqd, cfqq, 0);
+
+       cfq_put_queue(cfqq);
 }
 
-/*
- * Called with interrupts disabled
- */
-static void cfq_exit_single_io_context(struct cfq_io_context *cic)
+static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
+                                        struct cfq_io_context *cic)
 {
-       struct cfq_data *cfqd = cic->key;
-       request_queue_t *q;
-
-       if (!cfqd)
-               return;
-
-       q = cfqd->queue;
-
-       WARN_ON(!irqs_disabled());
-
-       spin_lock(q->queue_lock);
+       list_del_init(&cic->queue_list);
+       smp_wmb();
+       cic->key = NULL;
 
        if (cic->cfqq[ASYNC]) {
-               if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
-                       __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
-               cfq_put_queue(cic->cfqq[ASYNC]);
+               cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
                cic->cfqq[ASYNC] = NULL;
        }
 
        if (cic->cfqq[SYNC]) {
-               if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
-                       __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
-               cfq_put_queue(cic->cfqq[SYNC]);
+               cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
                cic->cfqq[SYNC] = NULL;
        }
+}
 
-       cic->key = NULL;
-       list_del_init(&cic->queue_list);
-       spin_unlock(q->queue_lock);
+
+/*
+ * Called with interrupts disabled
+ */
+static void cfq_exit_single_io_context(struct cfq_io_context *cic)
+{
+       struct cfq_data *cfqd = cic->key;
+
+       if (cfqd) {
+               request_queue_t *q = cfqd->queue;
+
+               spin_lock_irq(q->queue_lock);
+               __cfq_exit_single_io_context(cfqd, cic);
+               spin_unlock_irq(q->queue_lock);
+       }
 }
 
 static void cfq_exit_io_context(struct io_context *ioc)
 {
        struct cfq_io_context *__cic;
-       unsigned long flags;
        struct rb_node *n;
 
        /*
         * put the reference this task is holding to the various queues
         */
-       spin_lock_irqsave(&cfq_exit_lock, flags);
 
        n = rb_first(&ioc->cic_root);
        while (n != NULL) {
@@ -1150,22 +1144,21 @@ static void cfq_exit_io_context(struct io_context *ioc)
                cfq_exit_single_io_context(__cic);
                n = rb_next(n);
        }
-
-       spin_unlock_irqrestore(&cfq_exit_lock, flags);
 }
 
 static struct cfq_io_context *
 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
 {
-       struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
+       struct cfq_io_context *cic;
 
+       cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
        if (cic) {
                memset(cic, 0, sizeof(*cic));
                cic->last_end_request = jiffies;
                INIT_LIST_HEAD(&cic->queue_list);
                cic->dtor = cfq_free_io_context;
                cic->exit = cfq_exit_io_context;
-               atomic_inc(&ioc_count);
+               elv_ioc_count_inc(ioc_count);
        }
 
        return cic;
@@ -1246,15 +1239,12 @@ static inline void changed_ioprio(struct cfq_io_context *cic)
        spin_unlock(cfqd->queue->queue_lock);
 }
 
-/*
- * callback from sys_ioprio_set, irqs are disabled
- */
-static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
+static void cfq_ioc_set_ioprio(struct io_context *ioc)
 {
        struct cfq_io_context *cic;
        struct rb_node *n;
 
-       spin_lock(&cfq_exit_lock);
+       ioc->ioprio_changed = 0;
 
        n = rb_first(&ioc->cic_root);
        while (n != NULL) {
@@ -1263,10 +1253,6 @@ static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
                changed_ioprio(cic);
                n = rb_next(n);
        }
-
-       spin_unlock(&cfq_exit_lock);
-
-       return 0;
 }
 
 static struct cfq_queue *
@@ -1286,12 +1272,18 @@ retry:
                        cfqq = new_cfqq;
                        new_cfqq = NULL;
                } else if (gfp_mask & __GFP_WAIT) {
+                       /*
+                        * Inform the allocator of the fact that we will
+                        * just repeat this allocation if it fails, to allow
+                        * the allocator to do whatever it needs to attempt to
+                        * free memory.
+                        */
                        spin_unlock_irq(cfqd->queue->queue_lock);
-                       new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
+                       new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
                        spin_lock_irq(cfqd->queue->queue_lock);
                        goto retry;
                } else {
-                       cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
+                       cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
                        if (!cfqq)
                                goto out;
                }
@@ -1306,13 +1298,13 @@ retry:
                hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
                atomic_set(&cfqq->ref, 0);
                cfqq->cfqd = cfqd;
-               cfqq->service_last = 0;
                /*
                 * set ->slice_left to allow preemption for a new process
                 */
                cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
                cfq_mark_cfqq_idle_window(cfqq);
                cfq_mark_cfqq_prio_changed(cfqq);
+               cfq_mark_cfqq_queue_new(cfqq);
                cfq_init_prio_data(cfqq);
        }
 
@@ -1328,12 +1320,10 @@ out:
 static void
 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
 {
-       spin_lock(&cfq_exit_lock);
+       WARN_ON(!list_empty(&cic->queue_list));
        rb_erase(&cic->rb_node, &ioc->cic_root);
-       list_del_init(&cic->queue_list);
-       spin_unlock(&cfq_exit_lock);
        kmem_cache_free(cfq_ioc_pool, cic);
-       atomic_dec(&ioc_count);
+       elv_ioc_count_dec(ioc_count);
 }
 
 static struct cfq_io_context *
@@ -1377,7 +1367,6 @@ cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
        cic->ioc = ioc;
        cic->key = cfqd;
 
-       ioc->set_ioprio = cfq_ioc_set_ioprio;
 restart:
        parent = NULL;
        p = &ioc->cic_root.rb_node;
@@ -1399,11 +1388,12 @@ restart:
                        BUG();
        }
 
-       spin_lock(&cfq_exit_lock);
        rb_link_node(&cic->rb_node, parent, p);
        rb_insert_color(&cic->rb_node, &ioc->cic_root);
+
+       spin_lock_irq(cfqd->queue->queue_lock);
        list_add(&cic->queue_list, &cfqd->cic_list);
-       spin_unlock(&cfq_exit_lock);
+       spin_unlock_irq(cfqd->queue->queue_lock);
 }
 
 /*
@@ -1419,7 +1409,7 @@ cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
 
        might_sleep_if(gfp_mask & __GFP_WAIT);
 
-       ioc = get_io_context(gfp_mask);
+       ioc = get_io_context(gfp_mask, cfqd->queue->node);
        if (!ioc)
                return NULL;
 
@@ -1433,6 +1423,10 @@ cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
 
        cfq_cic_link(cfqd, ioc, cic);
 out:
+       smp_read_barrier_depends();
+       if (unlikely(ioc->ioprio_changed))
+               cfq_ioc_set_ioprio(ioc);
+
        return cic;
 err:
        put_io_context(ioc);
@@ -1544,8 +1538,18 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
         */
        if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
                return 0;
+       /*
+        * if the new request is sync, but the currently running queue is
+        * not, let the sync request have priority.
+        */
        if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
                return 1;
+       /*
+        * So both queues are sync. Let the new request get disk time if
+        * it's a metadata request and the current queue is doing regular IO.
+        */
+       if (rq_is_meta(rq) && !cfqq->meta_pending)
+               return 1;
 
        return 0;
 }
@@ -1556,30 +1560,19 @@ 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 *__cfqq, *next;
-
-       list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
-               cfq_resort_rr_list(__cfqq, 1);
+       cfq_slice_expired(cfqd, 1);
 
        if (!cfqq->slice_left)
                cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
 
-       cfqq->slice_end = cfqq->slice_left + jiffies;
-       cfq_slice_expired(cfqd, 1);
-       __cfq_set_active_queue(cfqd, cfqq);
-}
-
-/*
- * should really be a ll_rw_blk.c helper
- */
-static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
-{
-       request_queue_t *q = cfqd->queue;
+       /*
+        * Put the new queue at the front of the of the current list,
+        * so we know that it will be selected next.
+        */
+       BUG_ON(!cfq_cfqq_on_rr(cfqq));
+       list_move(&cfqq->cfq_list, &cfqd->cur_rr);
 
-       if (!blk_queue_plugged(q))
-               q->request_fn(q);
-       else
-               __generic_unplug_device(q);
+       cfqq->slice_end = cfqq->slice_left + jiffies;
 }
 
 /*
@@ -1592,6 +1585,9 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 {
        struct cfq_io_context *cic = RQ_CIC(rq);
 
+       if (rq_is_meta(rq))
+               cfqq->meta_pending++;
+
        /*
         * check if this request is a better next-serve candidate)) {
         */
@@ -1610,7 +1606,7 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                if (cic == cfqd->active_cic &&
                    del_timer(&cfqd->idle_slice_timer)) {
                        cfq_slice_expired(cfqd, 0);
-                       cfq_start_queueing(cfqd, cfqq);
+                       blk_start_queueing(cfqd->queue);
                }
                return;
        }
@@ -1631,7 +1627,7 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                if (cfq_cfqq_wait_request(cfqq)) {
                        cfq_mark_cfqq_must_dispatch(cfqq);
                        del_timer(&cfqd->idle_slice_timer);
-                       cfq_start_queueing(cfqd, cfqq);
+                       blk_start_queueing(cfqd->queue);
                }
        } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
                /*
@@ -1641,7 +1637,7 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                 */
                cfq_preempt_queue(cfqd, cfqq);
                cfq_mark_cfqq_must_dispatch(cfqq);
-               cfq_start_queueing(cfqd, cfqq);
+               blk_start_queueing(cfqd->queue);
        }
 }
 
@@ -1679,12 +1675,8 @@ static void cfq_completed_request(request_queue_t *q, struct request *rq)
        if (!cfq_class_idle(cfqq))
                cfqd->last_end_request = now;
 
-       if (!cfq_cfqq_dispatched(cfqq)) {
-               if (cfq_cfqq_on_rr(cfqq)) {
-                       cfqq->service_last = now;
-                       cfq_resort_rr_list(cfqq, 0);
-               }
-       }
+       if (!cfq_cfqq_dispatched(cfqq) && cfq_cfqq_on_rr(cfqq))
+               cfq_resort_rr_list(cfqq, 0);
 
        if (sync)
                RQ_CIC(rq)->last_end_request = now;
@@ -1739,9 +1731,7 @@ static void cfq_prio_boost(struct cfq_queue *cfqq)
                cfq_resort_rr_list(cfqq, 0);
 }
 
-static inline int
-__cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
-               struct task_struct *task, int rw)
+static inline int __cfq_may_queue(struct cfq_queue *cfqq)
 {
        if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
            !cfq_cfqq_must_alloc_slice(cfqq)) {
@@ -1752,7 +1742,7 @@ __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        return ELV_MQUEUE_MAY;
 }
 
-static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
+static int cfq_may_queue(request_queue_t *q, int rw)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
        struct task_struct *tsk = current;
@@ -1769,27 +1759,12 @@ static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
                cfq_init_prio_data(cfqq);
                cfq_prio_boost(cfqq);
 
-               return __cfq_may_queue(cfqd, cfqq, tsk, rw);
+               return __cfq_may_queue(cfqq);
        }
 
        return ELV_MQUEUE_MAY;
 }
 
-static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
-{
-       struct cfq_data *cfqd = q->elevator->elevator_data;
-
-       if (unlikely(cfqd->rq_starved)) {
-               struct request_list *rl = &q->rq;
-
-               smp_mb();
-               if (waitqueue_active(&rl->wait[READ]))
-                       wake_up(&rl->wait[READ]);
-               if (waitqueue_active(&rl->wait[WRITE]))
-                       wake_up(&rl->wait[WRITE]);
-       }
-}
-
 /*
  * queue lock held here
  */
@@ -1808,7 +1783,6 @@ static void cfq_put_request(request_queue_t *q, struct request *rq)
                rq->elevator_private = NULL;
                rq->elevator_private2 = NULL;
 
-               cfq_check_waiters(q, cfqq);
                cfq_put_queue(cfqq);
        }
 }
@@ -1817,8 +1791,7 @@ static void cfq_put_request(request_queue_t *q, struct request *rq)
  * Allocate cfq data structures associated with this request.
  */
 static int
-cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
-               gfp_t gfp_mask)
+cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
        struct task_struct *tsk = current;
@@ -1849,7 +1822,6 @@ cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
 
        cfqq->allocated[rw]++;
        cfq_clear_cfqq_must_alloc(cfqq);
-       cfqd->rq_starved = 0;
        atomic_inc(&cfqq->ref);
 
        spin_unlock_irqrestore(q->queue_lock, flags);
@@ -1861,12 +1833,7 @@ cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
 queue_fail:
        if (cic)
                put_io_context(cic->ioc);
-       /*
-        * mark us rq allocation starved. we need to kickstart the process
-        * ourselves if there are no pending requests that can do it for us.
-        * that would be an extremely rare OOM situation
-        */
-       cfqd->rq_starved = 1;
+
        cfq_schedule_dispatch(cfqd);
        spin_unlock_irqrestore(q->queue_lock, flags);
        return 1;
@@ -1875,27 +1842,10 @@ queue_fail:
 static void cfq_kick_queue(void *data)
 {
        request_queue_t *q = data;
-       struct cfq_data *cfqd = q->elevator->elevator_data;
        unsigned long flags;
 
        spin_lock_irqsave(q->queue_lock, flags);
-
-       if (cfqd->rq_starved) {
-               struct request_list *rl = &q->rq;
-
-               /*
-                * we aren't guaranteed to get a request after this, but we
-                * have to be opportunistic
-                */
-               smp_mb();
-               if (waitqueue_active(&rl->wait[READ]))
-                       wake_up(&rl->wait[READ]);
-               if (waitqueue_active(&rl->wait[WRITE]))
-                       wake_up(&rl->wait[WRITE]);
-       }
-
-       blk_remove_plug(q);
-       q->request_fn(q);
+       blk_start_queueing(q);
        spin_unlock_irqrestore(q->queue_lock, flags);
 }
 
@@ -1978,7 +1928,6 @@ static void cfq_exit_queue(elevator_t *e)
 
        cfq_shutdown_timer_wq(cfqd);
 
-       spin_lock(&cfq_exit_lock);
        spin_lock_irq(q->queue_lock);
 
        if (cfqd->active_queue)
@@ -1988,20 +1937,11 @@ static void cfq_exit_queue(elevator_t *e)
                struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
                                                        struct cfq_io_context,
                                                        queue_list);
-               if (cic->cfqq[ASYNC]) {
-                       cfq_put_queue(cic->cfqq[ASYNC]);
-                       cic->cfqq[ASYNC] = NULL;
-               }
-               if (cic->cfqq[SYNC]) {
-                       cfq_put_queue(cic->cfqq[SYNC]);
-                       cic->cfqq[SYNC] = NULL;
-               }
-               cic->key = NULL;
-               list_del_init(&cic->queue_list);
+
+               __cfq_exit_single_io_context(cfqd, cic);
        }
 
        spin_unlock_irq(q->queue_lock);
-       spin_unlock(&cfq_exit_lock);
 
        cfq_shutdown_timer_wq(cfqd);
 
@@ -2014,7 +1954,7 @@ static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
        struct cfq_data *cfqd;
        int i;
 
-       cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
+       cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
        if (!cfqd)
                return NULL;
 
@@ -2026,10 +1966,9 @@ static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
        INIT_LIST_HEAD(&cfqd->busy_rr);
        INIT_LIST_HEAD(&cfqd->cur_rr);
        INIT_LIST_HEAD(&cfqd->idle_rr);
-       INIT_LIST_HEAD(&cfqd->empty_list);
        INIT_LIST_HEAD(&cfqd->cic_list);
 
-       cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
+       cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
        if (!cfqd->cfq_hash)
                goto out_free;
 
@@ -2048,7 +1987,6 @@ static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
 
        INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
 
-       cfqd->cfq_queued = cfq_queued;
        cfqd->cfq_quantum = cfq_quantum;
        cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
        cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
@@ -2120,7 +2058,6 @@ static ssize_t __FUNC(elevator_t *e, char *page)                  \
        return cfq_var_show(__data, (page));                            \
 }
 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
-SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
@@ -2148,7 +2085,6 @@ static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)      \
        return ret;                                                     \
 }
 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
-STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
@@ -2164,7 +2100,6 @@ STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX,
 
 static struct elv_fs_entry cfq_attrs[] = {
        CFQ_ATTR(quantum),
-       CFQ_ATTR(queued),
        CFQ_ATTR(fifo_expire_sync),
        CFQ_ATTR(fifo_expire_async),
        CFQ_ATTR(back_seek_max),
@@ -2194,7 +2129,7 @@ static struct elevator_type iosched_cfq = {
                .elevator_may_queue_fn =        cfq_may_queue,
                .elevator_init_fn =             cfq_init_queue,
                .elevator_exit_fn =             cfq_exit_queue,
-               .trim =                         cfq_trim,
+               .trim =                         cfq_free_io_context,
        },
        .elevator_attrs =       cfq_attrs,
        .elevator_name =        "cfq",
@@ -2225,12 +2160,12 @@ static int __init cfq_init(void)
 
 static void __exit cfq_exit(void)
 {
-       DECLARE_COMPLETION(all_gone);
+       DECLARE_COMPLETION_ONSTACK(all_gone);
        elv_unregister(&iosched_cfq);
        ioc_gone = &all_gone;
        /* ioc_gone's update must be visible before reading ioc_count */
        smp_wmb();
-       if (atomic_read(&ioc_count))
+       if (elv_ioc_count_read(ioc_count))
                wait_for_completion(ioc_gone);
        synchronize_rcu();
        cfq_slab_kill();