nfs: new subdir Documentation/filesystems/nfs
[safe/jmp/linux-2.6] / block / cfq-iosched.c
index 11efcf1..069a610 100644 (file)
@@ -48,7 +48,7 @@ static int cfq_slice_idle = HZ / 125;
 static struct kmem_cache *cfq_pool;
 static struct kmem_cache *cfq_ioc_pool;
 
-static DEFINE_PER_CPU(unsigned long, ioc_count);
+static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
 static struct completion *ioc_gone;
 static DEFINE_SPINLOCK(ioc_gone_lock);
 
@@ -56,9 +56,6 @@ static DEFINE_SPINLOCK(ioc_gone_lock);
 #define cfq_class_idle(cfqq)   ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
 #define cfq_class_rt(cfqq)     ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
 
-#define ASYNC                  (0)
-#define SYNC                   (1)
-
 #define sample_valid(samples)  ((samples) > 80)
 
 /*
@@ -74,6 +71,51 @@ struct cfq_rb_root {
 #define CFQ_RB_ROOT    (struct cfq_rb_root) { RB_ROOT, NULL, }
 
 /*
+ * Per process-grouping structure
+ */
+struct cfq_queue {
+       /* reference count */
+       atomic_t ref;
+       /* various state flags, see below */
+       unsigned int flags;
+       /* parent cfq_data */
+       struct cfq_data *cfqd;
+       /* service_tree member */
+       struct rb_node rb_node;
+       /* service_tree key */
+       unsigned long rb_key;
+       /* prio tree member */
+       struct rb_node p_node;
+       /* prio tree root we belong to, if any */
+       struct rb_root *p_root;
+       /* sorted list of pending requests */
+       struct rb_root sort_list;
+       /* if fifo isn't expired, next request to serve */
+       struct request *next_rq;
+       /* requests queued in sort_list */
+       int queued[2];
+       /* currently allocated requests */
+       int allocated[2];
+       /* fifo list of requests in sort_list */
+       struct list_head fifo;
+
+       unsigned long slice_end;
+       long slice_resid;
+       unsigned int slice_dispatch;
+
+       /* pending metadata requests */
+       int meta_pending;
+       /* number of requests that are on the dispatch list or inside driver */
+       int dispatched;
+
+       /* io prio of this group */
+       unsigned short ioprio, org_ioprio;
+       unsigned short ioprio_class, org_ioprio_class;
+
+       pid_t pid;
+};
+
+/*
  * Per block device queue structure
  */
 struct cfq_data {
@@ -83,14 +125,17 @@ struct cfq_data {
         * rr list of queues with requests and the count of them
         */
        struct cfq_rb_root service_tree;
-       unsigned int busy_queues;
+
        /*
-        * Used to track any pending rt requests so we can pre-empt current
-        * non-RT cfqq in service when this value is non-zero.
+        * Each priority tree is sorted by next_request position.  These
+        * trees are used when determining if two or more queues are
+        * interleaving requests (see cfq_close_cooperator).
         */
-       unsigned int busy_rt_queues;
+       struct rb_root prio_trees[CFQ_PRIO_LISTS];
+
+       unsigned int busy_queues;
 
-       int rq_in_driver;
+       int rq_in_driver[2];
        int sync_flight;
 
        /*
@@ -117,7 +162,6 @@ struct cfq_data {
        struct cfq_queue *async_idle_cfqq;
 
        sector_t last_position;
-       unsigned long last_end_request;
 
        /*
         * tunables, see top of file
@@ -129,61 +173,29 @@ struct cfq_data {
        unsigned int cfq_slice[2];
        unsigned int cfq_slice_async_rq;
        unsigned int cfq_slice_idle;
+       unsigned int cfq_latency;
 
        struct list_head cic_list;
-};
-
-/*
- * Per process-grouping structure
- */
-struct cfq_queue {
-       /* reference count */
-       atomic_t ref;
-       /* various state flags, see below */
-       unsigned int flags;
-       /* parent cfq_data */
-       struct cfq_data *cfqd;
-       /* service_tree member */
-       struct rb_node rb_node;
-       /* service_tree key */
-       unsigned long rb_key;
-       /* sorted list of pending requests */
-       struct rb_root sort_list;
-       /* if fifo isn't expired, next request to serve */
-       struct request *next_rq;
-       /* requests queued in sort_list */
-       int queued[2];
-       /* currently allocated requests */
-       int allocated[2];
-       /* fifo list of requests in sort_list */
-       struct list_head fifo;
 
-       unsigned long slice_end;
-       long slice_resid;
-       unsigned int slice_dispatch;
-
-       /* pending metadata requests */
-       int meta_pending;
-       /* number of requests that are on the dispatch list or inside driver */
-       int dispatched;
-
-       /* io prio of this group */
-       unsigned short ioprio, org_ioprio;
-       unsigned short ioprio_class, org_ioprio_class;
+       /*
+        * Fallback dummy cfqq for extreme OOM conditions
+        */
+       struct cfq_queue oom_cfqq;
 
-       pid_t pid;
+       unsigned long last_end_sync_rq;
 };
 
 enum cfqq_state_flags {
        CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
        CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
-       CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
+       CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
        CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
        CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
        CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
        CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
        CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
        CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
+       CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
 };
 
 #define CFQ_CFQQ_FNS(name)                                             \
@@ -202,13 +214,14 @@ static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)           \
 
 CFQ_CFQQ_FNS(on_rr);
 CFQ_CFQQ_FNS(wait_request);
-CFQ_CFQQ_FNS(must_alloc);
+CFQ_CFQQ_FNS(must_dispatch);
 CFQ_CFQQ_FNS(must_alloc_slice);
 CFQ_CFQQ_FNS(fifo_expire);
 CFQ_CFQQ_FNS(idle_window);
 CFQ_CFQQ_FNS(prio_changed);
 CFQ_CFQQ_FNS(slice_new);
 CFQ_CFQQ_FNS(sync);
+CFQ_CFQQ_FNS(coop);
 #undef CFQ_CFQQ_FNS
 
 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
@@ -217,33 +230,35 @@ CFQ_CFQQ_FNS(sync);
        blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 
 static void cfq_dispatch_insert(struct request_queue *, struct request *);
-static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
+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 int rq_in_driver(struct cfq_data *cfqd)
+{
+       return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
+}
+
 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
-                                           int is_sync)
+                                           bool is_sync)
 {
-       return cic->cfqq[!!is_sync];
+       return cic->cfqq[is_sync];
 }
 
 static inline void cic_set_cfqq(struct cfq_io_context *cic,
-                               struct cfq_queue *cfqq, int is_sync)
+                               struct cfq_queue *cfqq, bool is_sync)
 {
-       cic->cfqq[!!is_sync] = cfqq;
+       cic->cfqq[is_sync] = cfqq;
 }
 
 /*
  * We regard a request as SYNC, if it's either a read or has the SYNC bit
  * set (in which case it could also be direct WRITE).
  */
-static inline int cfq_bio_sync(struct bio *bio)
+static inline bool cfq_bio_sync(struct bio *bio)
 {
-       if (bio_data_dir(bio) == READ || bio_sync(bio))
-               return 1;
-
-       return 0;
+       return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
 }
 
 /*
@@ -270,7 +285,7 @@ static int cfq_queue_empty(struct request_queue *q)
  * if a queue is marked sync and has sync io queued. A sync queue with async
  * io only, should not get full sync slice length.
  */
-static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
+static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
                                 unsigned short prio)
 {
        const int base_slice = cfqd->cfq_slice[sync];
@@ -298,7 +313,7 @@ cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
  * isn't valid until the first request from the dispatch is activated
  * and the slice time set.
  */
-static inline int cfq_slice_used(struct cfq_queue *cfqq)
+static inline bool cfq_slice_used(struct cfq_queue *cfqq)
 {
        if (cfq_cfqq_slice_new(cfqq))
                return 0;
@@ -336,8 +351,8 @@ cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
        else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
                return rq2;
 
-       s1 = rq1->sector;
-       s2 = rq2->sector;
+       s1 = blk_rq_pos(rq1);
+       s2 = blk_rq_pos(rq2);
 
        last = cfqd->last_position;
 
@@ -417,13 +432,17 @@ static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
        return NULL;
 }
 
+static void rb_erase_init(struct rb_node *n, struct rb_root *root)
+{
+       rb_erase(n, root);
+       RB_CLEAR_NODE(n);
+}
+
 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 {
        if (root->left == n)
                root->left = NULL;
-
-       rb_erase(n, &root->rb);
-       RB_CLEAR_NODE(n);
+       rb_erase_init(n, &root->rb);
 }
 
 /*
@@ -468,8 +487,8 @@ static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
  * requests waiting to be processed. It is sorted in the order that
  * we will service the queues.
  */
-static void cfq_service_tree_add(struct cfq_data *cfqd,
-                                   struct cfq_queue *cfqq, int add_front)
+static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
+                                bool add_front)
 {
        struct rb_node **p, *parent;
        struct cfq_queue *__cfqq;
@@ -485,11 +504,20 @@ static void cfq_service_tree_add(struct cfq_data *cfqd,
                } else
                        rb_key += jiffies;
        } else if (!add_front) {
+               /*
+                * Get our rb key offset. Subtract any residual slice
+                * value carried from last service. A negative resid
+                * count indicates slice overrun, and this should position
+                * the next service time further away in the tree.
+                */
                rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
-               rb_key += cfqq->slice_resid;
+               rb_key -= cfqq->slice_resid;
                cfqq->slice_resid = 0;
-       } else
-               rb_key = 0;
+       } else {
+               rb_key = -HZ;
+               __cfqq = cfq_rb_first(&cfqd->service_tree);
+               rb_key += __cfqq ? __cfqq->rb_key : jiffies;
+       }
 
        if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
                /*
@@ -523,7 +551,7 @@ static void cfq_service_tree_add(struct cfq_data *cfqd,
                        n = &(*p)->rb_left;
                else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
                        n = &(*p)->rb_right;
-               else if (rb_key < __cfqq->rb_key)
+               else if (time_before(rb_key, __cfqq->rb_key))
                        n = &(*p)->rb_left;
                else
                        n = &(*p)->rb_right;
@@ -542,6 +570,67 @@ static void cfq_service_tree_add(struct cfq_data *cfqd,
        rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 }
 
+static struct cfq_queue *
+cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
+                    sector_t sector, struct rb_node **ret_parent,
+                    struct rb_node ***rb_link)
+{
+       struct rb_node **p, *parent;
+       struct cfq_queue *cfqq = NULL;
+
+       parent = NULL;
+       p = &root->rb_node;
+       while (*p) {
+               struct rb_node **n;
+
+               parent = *p;
+               cfqq = rb_entry(parent, struct cfq_queue, p_node);
+
+               /*
+                * Sort strictly based on sector.  Smallest to the left,
+                * largest to the right.
+                */
+               if (sector > blk_rq_pos(cfqq->next_rq))
+                       n = &(*p)->rb_right;
+               else if (sector < blk_rq_pos(cfqq->next_rq))
+                       n = &(*p)->rb_left;
+               else
+                       break;
+               p = n;
+               cfqq = NULL;
+       }
+
+       *ret_parent = parent;
+       if (rb_link)
+               *rb_link = p;
+       return cfqq;
+}
+
+static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
+{
+       struct rb_node **p, *parent;
+       struct cfq_queue *__cfqq;
+
+       if (cfqq->p_root) {
+               rb_erase(&cfqq->p_node, cfqq->p_root);
+               cfqq->p_root = NULL;
+       }
+
+       if (cfq_class_idle(cfqq))
+               return;
+       if (!cfqq->next_rq)
+               return;
+
+       cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
+       __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
+                                     blk_rq_pos(cfqq->next_rq), &parent, &p);
+       if (!__cfqq) {
+               rb_link_node(&cfqq->p_node, parent, p);
+               rb_insert_color(&cfqq->p_node, cfqq->p_root);
+       } else
+               cfqq->p_root = NULL;
+}
+
 /*
  * Update cfqq's position in the service tree.
  */
@@ -550,8 +639,10 @@ static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
        /*
         * Resorting requires the cfqq to be on the RR list already.
         */
-       if (cfq_cfqq_on_rr(cfqq))
+       if (cfq_cfqq_on_rr(cfqq)) {
                cfq_service_tree_add(cfqd, cfqq, 0);
+               cfq_prio_tree_add(cfqd, cfqq);
+       }
 }
 
 /*
@@ -564,8 +655,6 @@ static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
        BUG_ON(cfq_cfqq_on_rr(cfqq));
        cfq_mark_cfqq_on_rr(cfqq);
        cfqd->busy_queues++;
-       if (cfq_class_rt(cfqq))
-               cfqd->busy_rt_queues++;
 
        cfq_resort_rr_list(cfqd, cfqq);
 }
@@ -582,11 +671,13 @@ static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 
        if (!RB_EMPTY_NODE(&cfqq->rb_node))
                cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
+       if (cfqq->p_root) {
+               rb_erase(&cfqq->p_node, cfqq->p_root);
+               cfqq->p_root = NULL;
+       }
 
        BUG_ON(!cfqd->busy_queues);
        cfqd->busy_queues--;
-       if (cfq_class_rt(cfqq))
-               cfqd->busy_rt_queues--;
 }
 
 /*
@@ -611,7 +702,7 @@ 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;
+       struct request *__alias, *prev;
 
        cfqq->queued[rq_is_sync(rq)]++;
 
@@ -628,7 +719,15 @@ static void cfq_add_rq_rb(struct request *rq)
        /*
         * check if this request is a better next-serve candidate
         */
+       prev = cfqq->next_rq;
        cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
+
+       /*
+        * adjust priority tree position, if ->next_rq changes
+        */
+       if (prev != cfqq->next_rq)
+               cfq_prio_tree_add(cfqd, cfqq);
+
        BUG_ON(!cfqq->next_rq);
 }
 
@@ -664,21 +763,22 @@ static void cfq_activate_request(struct request_queue *q, struct request *rq)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
 
-       cfqd->rq_in_driver++;
+       cfqd->rq_in_driver[rq_is_sync(rq)]++;
        cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
-                                               cfqd->rq_in_driver);
+                                               rq_in_driver(cfqd));
 
-       cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
+       cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
 }
 
 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
+       const int sync = rq_is_sync(rq);
 
-       WARN_ON(!cfqd->rq_in_driver);
-       cfqd->rq_in_driver--;
+       WARN_ON(!cfqd->rq_in_driver[sync]);
+       cfqd->rq_in_driver[sync]--;
        cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
-                                               cfqd->rq_in_driver);
+                                               rq_in_driver(cfqd));
 }
 
 static void cfq_remove_request(struct request *rq)
@@ -731,8 +831,10 @@ cfq_merged_requests(struct request_queue *q, struct request *rq,
         * reposition in fifo if next is older than rq
         */
        if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
-           time_before(next->start_time, rq->start_time))
+           time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
                list_move(&rq->queuelist, &next->queuelist);
+               rq_set_fifo_time(rq, rq_fifo_time(next));
+       }
 
        cfq_remove_request(next);
 }
@@ -748,7 +850,7 @@ static int cfq_allow_merge(struct request_queue *q, struct request *rq,
         * Disallow merge of a sync bio into an async request.
         */
        if (cfq_bio_sync(bio) && !rq_is_sync(rq))
-               return 0;
+               return false;
 
        /*
         * Lookup the cfqq that this bio will be queued with. Allow
@@ -756,13 +858,10 @@ static int cfq_allow_merge(struct request_queue *q, struct request *rq,
         */
        cic = cfq_cic_lookup(cfqd, current->io_context);
        if (!cic)
-               return 0;
+               return false;
 
        cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
-       if (cfqq == RQ_CFQQ(rq))
-               return 1;
-
-       return 0;
+       return cfqq == RQ_CFQQ(rq);
 }
 
 static void __cfq_set_active_queue(struct cfq_data *cfqd,
@@ -774,6 +873,7 @@ static void __cfq_set_active_queue(struct cfq_data *cfqd,
                cfqq->slice_dispatch = 0;
 
                cfq_clear_cfqq_wait_request(cfqq);
+               cfq_clear_cfqq_must_dispatch(cfqq);
                cfq_clear_cfqq_must_alloc_slice(cfqq);
                cfq_clear_cfqq_fifo_expire(cfqq);
                cfq_mark_cfqq_slice_new(cfqq);
@@ -789,7 +889,7 @@ static void __cfq_set_active_queue(struct cfq_data *cfqd,
  */
 static void
 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
-                   int timed_out)
+                   bool timed_out)
 {
        cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
 
@@ -817,7 +917,7 @@ __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        }
 }
 
-static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
+static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
 {
        struct cfq_queue *cfqq = cfqd->active_queue;
 
@@ -840,11 +940,15 @@ static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 /*
  * Get and set a new active queue for service.
  */
-static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
+static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
+                                             struct cfq_queue *cfqq)
 {
-       struct cfq_queue *cfqq;
+       if (!cfqq) {
+               cfqq = cfq_get_next_queue(cfqd);
+               if (cfqq)
+                       cfq_clear_cfqq_coop(cfqq);
+       }
 
-       cfqq = cfq_get_next_queue(cfqd);
        __cfq_set_active_queue(cfqd, cfqq);
        return cfqq;
 }
@@ -852,34 +956,106 @@ static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
                                          struct request *rq)
 {
-       if (rq->sector >= cfqd->last_position)
-               return rq->sector - cfqd->last_position;
+       if (blk_rq_pos(rq) >= cfqd->last_position)
+               return blk_rq_pos(rq) - cfqd->last_position;
        else
-               return cfqd->last_position - rq->sector;
+               return cfqd->last_position - blk_rq_pos(rq);
 }
 
+#define CIC_SEEK_THR   8 * 1024
+#define CIC_SEEKY(cic) ((cic)->seek_mean > CIC_SEEK_THR)
+
 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 {
        struct cfq_io_context *cic = cfqd->active_cic;
+       sector_t sdist = cic->seek_mean;
 
        if (!sample_valid(cic->seek_samples))
-               return 0;
+               sdist = CIC_SEEK_THR;
+
+       return cfq_dist_from_last(cfqd, rq) <= sdist;
+}
+
+static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
+                                   struct cfq_queue *cur_cfqq)
+{
+       struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
+       struct rb_node *parent, *node;
+       struct cfq_queue *__cfqq;
+       sector_t sector = cfqd->last_position;
+
+       if (RB_EMPTY_ROOT(root))
+               return NULL;
+
+       /*
+        * First, if we find a request starting at the end of the last
+        * request, choose it.
+        */
+       __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
+       if (__cfqq)
+               return __cfqq;
+
+       /*
+        * If the exact sector wasn't found, the parent of the NULL leaf
+        * will contain the closest sector.
+        */
+       __cfqq = rb_entry(parent, struct cfq_queue, p_node);
+       if (cfq_rq_close(cfqd, __cfqq->next_rq))
+               return __cfqq;
+
+       if (blk_rq_pos(__cfqq->next_rq) < sector)
+               node = rb_next(&__cfqq->p_node);
+       else
+               node = rb_prev(&__cfqq->p_node);
+       if (!node)
+               return NULL;
+
+       __cfqq = rb_entry(node, struct cfq_queue, p_node);
+       if (cfq_rq_close(cfqd, __cfqq->next_rq))
+               return __cfqq;
 
-       return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
+       return NULL;
 }
 
-static int cfq_close_cooperator(struct cfq_data *cfq_data,
-                               struct cfq_queue *cfqq)
+/*
+ * cfqd - obvious
+ * cur_cfqq - passed in so that we don't decide that the current queue is
+ *           closely cooperating with itself.
+ *
+ * So, basically we're assuming that that cur_cfqq has dispatched at least
+ * one request, and that cfqd->last_position reflects a position on the disk
+ * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
+ * assumption.
+ */
+static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
+                                             struct cfq_queue *cur_cfqq,
+                                             bool probe)
 {
+       struct cfq_queue *cfqq;
+
+       /*
+        * A valid cfq_io_context is necessary to compare requests against
+        * the seek_mean of the current cfqq.
+        */
+       if (!cfqd->active_cic)
+               return NULL;
+
        /*
         * We should notice if some of the queues are cooperating, eg
         * working closely on the same area of the disk. In that case,
         * we can group them together and don't waste time idling.
         */
-       return 0;
-}
+       cfqq = cfqq_close(cfqd, cur_cfqq);
+       if (!cfqq)
+               return NULL;
+
+       if (cfq_cfqq_coop(cfqq))
+               return NULL;
 
-#define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
+       if (!probe)
+               cfq_mark_cfqq_coop(cfqq);
+       return cfqq;
+}
 
 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 {
@@ -907,7 +1083,7 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
        /*
         * still requests with the driver, don't idle
         */
-       if (cfqd->rq_in_driver)
+       if (rq_in_driver(cfqd))
                return;
 
        /*
@@ -918,10 +1094,12 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
                return;
 
        /*
-        * See if this prio level has a good candidate
+        * If our average think time is larger than the remaining time
+        * slice, then don't idle. This avoids overrunning the allotted
+        * time slice.
         */
-       if (cfq_close_cooperator(cfqd, cfqq) &&
-           (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
+       if (sample_valid(cic->ttime_samples) &&
+           (cfqq->slice_end - jiffies < cic->ttime_mean))
                return;
 
        cfq_mark_cfqq_wait_request(cfqq);
@@ -936,7 +1114,7 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
                sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
 
        mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
-       cfq_log(cfqd, "arm_idle: %lu", sl);
+       cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
 }
 
 /*
@@ -949,6 +1127,7 @@ static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
 
        cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
 
+       cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
        cfq_remove_request(rq);
        cfqq->dispatched++;
        elv_dispatch_sort(q, rq);
@@ -962,9 +1141,7 @@ static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
  */
 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 {
-       struct cfq_data *cfqd = cfqq->cfqd;
-       struct request *rq;
-       int fifo;
+       struct request *rq = NULL;
 
        if (cfq_cfqq_fifo_expire(cfqq))
                return NULL;
@@ -974,13 +1151,11 @@ static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
        if (list_empty(&cfqq->fifo))
                return NULL;
 
-       fifo = cfq_cfqq_sync(cfqq);
        rq = rq_entry_fifo(cfqq->fifo.next);
-
-       if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
+       if (time_before(jiffies, rq_fifo_time(rq)))
                rq = NULL;
 
-       cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
+       cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
        return rq;
 }
 
@@ -1000,7 +1175,7 @@ cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
  */
 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 {
-       struct cfq_queue *cfqq;
+       struct cfq_queue *cfqq, *new_cfqq = NULL;
 
        cfqq = cfqd->active_queue;
        if (!cfqq)
@@ -1009,24 +1184,10 @@ static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
        /*
         * The active queue has run out of time, expire it and select new.
         */
-       if (cfq_slice_used(cfqq))
+       if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
                goto expire;
 
        /*
-        * If we have a RT cfqq waiting, then we pre-empt the current non-rt
-        * cfqq.
-        */
-       if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
-               /*
-                * We simulate this as cfqq timed out so that it gets to bank
-                * the remaining of its time slice.
-                */
-               cfq_log_cfqq(cfqd, cfqq, "preempt");
-               cfq_slice_expired(cfqd, 1);
-               goto new_queue;
-       }
-
-       /*
         * The active queue has requests and isn't expired, allow it to
         * dispatch.
         */
@@ -1034,6 +1195,16 @@ static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
                goto keep_queue;
 
        /*
+        * If another queue has a request waiting within our mean seek
+        * distance, let it run.  The expire code will check for close
+        * cooperators and put the close queue at the front of the service
+        * tree.
+        */
+       new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
+       if (new_cfqq)
+               goto expire;
+
+       /*
         * No requests pending. If the active queue still has requests in
         * flight or is idling for a new request, allow either of these
         * conditions to happen (or time out) before selecting a new queue.
@@ -1047,7 +1218,7 @@ static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 expire:
        cfq_slice_expired(cfqd, 0);
 new_queue:
-       cfqq = cfq_set_active_queue(cfqd);
+       cfqq = cfq_set_active_queue(cfqd, new_cfqq);
 keep_queue:
        return cfqq;
 }
@@ -1081,20 +1252,87 @@ static int cfq_forced_dispatch(struct cfq_data *cfqd)
 
        BUG_ON(cfqd->busy_queues);
 
-       cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
+       cfq_log(cfqd, "forced_dispatch=%d", dispatched);
        return dispatched;
 }
 
+static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
+{
+       unsigned int max_dispatch;
+
+       /*
+        * Drain async requests before we start sync IO
+        */
+       if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
+               return false;
+
+       /*
+        * If this is an async queue and we have sync IO in flight, let it wait
+        */
+       if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
+               return false;
+
+       max_dispatch = cfqd->cfq_quantum;
+       if (cfq_class_idle(cfqq))
+               max_dispatch = 1;
+
+       /*
+        * Does this cfqq already have too much IO in flight?
+        */
+       if (cfqq->dispatched >= max_dispatch) {
+               /*
+                * idle queue must always only have a single IO in flight
+                */
+               if (cfq_class_idle(cfqq))
+                       return false;
+
+               /*
+                * We have other queues, don't allow more IO from this one
+                */
+               if (cfqd->busy_queues > 1)
+                       return false;
+
+               /*
+                * Sole queue user, allow bigger slice
+                */
+               max_dispatch *= 4;
+       }
+
+       /*
+        * Async queues must wait a bit before being allowed dispatch.
+        * We also ramp up the dispatch depth gradually for async IO,
+        * based on the last sync IO we serviced
+        */
+       if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
+               unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
+               unsigned int depth;
+
+               depth = last_sync / cfqd->cfq_slice[1];
+               if (!depth && !cfqq->dispatched)
+                       depth = 1;
+               if (depth < max_dispatch)
+                       max_dispatch = depth;
+       }
+
+       /*
+        * If we're below the current max, allow a dispatch
+        */
+       return cfqq->dispatched < max_dispatch;
+}
+
 /*
  * Dispatch a request from cfqq, moving them to the request queue
  * dispatch list.
  */
-static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
+static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
        struct request *rq;
 
        BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
 
+       if (!cfq_may_dispatch(cfqd, cfqq))
+               return false;
+
        /*
         * follow expired path, else get first next available
         */
@@ -1110,9 +1348,11 @@ static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
        if (!cfqd->active_cic) {
                struct cfq_io_context *cic = RQ_CIC(rq);
 
-               atomic_inc(&cic->ioc->refcount);
+               atomic_long_inc(&cic->ioc->refcount);
                cfqd->active_cic = cic;
        }
+
+       return true;
 }
 
 /*
@@ -1123,7 +1363,6 @@ static int cfq_dispatch_requests(struct request_queue *q, int force)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
        struct cfq_queue *cfqq;
-       unsigned int max_dispatch;
 
        if (!cfqd->busy_queues)
                return 0;
@@ -1136,43 +1375,13 @@ static int cfq_dispatch_requests(struct request_queue *q, int force)
                return 0;
 
        /*
-        * If this is an async queue and we have sync IO in flight, let it wait
+        * Dispatch a request from this cfqq, if it is allowed
         */
-       if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
+       if (!cfq_dispatch_request(cfqd, cfqq))
                return 0;
 
-       max_dispatch = cfqd->cfq_quantum;
-       if (cfq_class_idle(cfqq))
-               max_dispatch = 1;
-
-       /*
-        * Does this cfqq already have too much IO in flight?
-        */
-       if (cfqq->dispatched >= max_dispatch) {
-               /*
-                * idle queue must always only have a single IO in flight
-                */
-               if (cfq_class_idle(cfqq))
-                       return 0;
-
-               /*
-                * We have other queues, don't allow more IO from this one
-                */
-               if (cfqd->busy_queues > 1)
-                       return 0;
-
-               /*
-                * we are the only queue, allow up to 4 times of 'quantum'
-                */
-               if (cfqq->dispatched >= 4 * max_dispatch)
-                       return 0;
-       }
-
-       /*
-        * Dispatch a request from this cfqq
-        */
-       cfq_dispatch_request(cfqd, cfqq);
        cfqq->slice_dispatch++;
+       cfq_clear_cfqq_must_dispatch(cfqq);
 
        /*
         * expire an async queue immediately if it has used up its slice. idle
@@ -1185,7 +1394,7 @@ static int cfq_dispatch_requests(struct request_queue *q, int force)
                cfq_slice_expired(cfqd, 0);
        }
 
-       cfq_log(cfqd, "dispatched a request");
+       cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
        return 1;
 }
 
@@ -1250,7 +1459,7 @@ static void cfq_cic_free_rcu(struct rcu_head *head)
        cic = container_of(head, struct cfq_io_context, rcu_head);
 
        kmem_cache_free(cfq_ioc_pool, cic);
-       elv_ioc_count_dec(ioc_count);
+       elv_ioc_count_dec(cfq_ioc_count);
 
        if (ioc_gone) {
                /*
@@ -1259,7 +1468,7 @@ static void cfq_cic_free_rcu(struct rcu_head *head)
                 * complete ioc_gone and set it back to NULL
                 */
                spin_lock(&ioc_gone_lock);
-               if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
+               if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
                        complete(ioc_gone);
                        ioc_gone = NULL;
                }
@@ -1329,14 +1538,14 @@ static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
        if (ioc->ioc_data == cic)
                rcu_assign_pointer(ioc->ioc_data, NULL);
 
-       if (cic->cfqq[ASYNC]) {
-               cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
-               cic->cfqq[ASYNC] = NULL;
+       if (cic->cfqq[BLK_RW_ASYNC]) {
+               cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
+               cic->cfqq[BLK_RW_ASYNC] = NULL;
        }
 
-       if (cic->cfqq[SYNC]) {
-               cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
-               cic->cfqq[SYNC] = NULL;
+       if (cic->cfqq[BLK_RW_SYNC]) {
+               cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
+               cic->cfqq[BLK_RW_SYNC] = NULL;
        }
 }
 
@@ -1385,7 +1594,7 @@ cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
                INIT_HLIST_NODE(&cic->cic_list);
                cic->dtor = cfq_free_io_context;
                cic->exit = cfq_exit_io_context;
-               elv_ioc_count_inc(ioc_count);
+               elv_ioc_count_inc(cfq_ioc_count);
        }
 
        return cic;
@@ -1445,17 +1654,18 @@ static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
 
        spin_lock_irqsave(cfqd->queue->queue_lock, flags);
 
-       cfqq = cic->cfqq[ASYNC];
+       cfqq = cic->cfqq[BLK_RW_ASYNC];
        if (cfqq) {
                struct cfq_queue *new_cfqq;
-               new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
+               new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
+                                               GFP_ATOMIC);
                if (new_cfqq) {
-                       cic->cfqq[ASYNC] = new_cfqq;
+                       cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
                        cfq_put_queue(cfqq);
                }
        }
 
-       cfqq = cic->cfqq[SYNC];
+       cfqq = cic->cfqq[BLK_RW_SYNC];
        if (cfqq)
                cfq_mark_cfqq_prio_changed(cfqq);
 
@@ -1468,8 +1678,28 @@ static void cfq_ioc_set_ioprio(struct io_context *ioc)
        ioc->ioprio_changed = 0;
 }
 
+static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
+                         pid_t pid, bool is_sync)
+{
+       RB_CLEAR_NODE(&cfqq->rb_node);
+       RB_CLEAR_NODE(&cfqq->p_node);
+       INIT_LIST_HEAD(&cfqq->fifo);
+
+       atomic_set(&cfqq->ref, 0);
+       cfqq->cfqd = cfqd;
+
+       cfq_mark_cfqq_prio_changed(cfqq);
+
+       if (is_sync) {
+               if (!cfq_class_idle(cfqq))
+                       cfq_mark_cfqq_idle_window(cfqq);
+               cfq_mark_cfqq_sync(cfqq);
+       }
+       cfqq->pid = pid;
+}
+
 static struct cfq_queue *
-cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
+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;
@@ -1480,55 +1710,40 @@ retry:
        /* cic always exists here */
        cfqq = cic_to_cfqq(cic, is_sync);
 
-       if (!cfqq) {
+       /*
+        * Always try a new alloc if we fell back to the OOM cfqq
+        * originally, since it should just be a temporary situation.
+        */
+       if (!cfqq || cfqq == &cfqd->oom_cfqq) {
+               cfqq = NULL;
                if (new_cfqq) {
                        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_node(cfq_pool,
-                                       gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
+                                       gfp_mask | __GFP_ZERO,
                                        cfqd->queue->node);
                        spin_lock_irq(cfqd->queue->queue_lock);
-                       goto retry;
+                       if (new_cfqq)
+                               goto retry;
                } else {
                        cfqq = kmem_cache_alloc_node(cfq_pool,
                                        gfp_mask | __GFP_ZERO,
                                        cfqd->queue->node);
-                       if (!cfqq)
-                               goto out;
                }
 
-               RB_CLEAR_NODE(&cfqq->rb_node);
-               INIT_LIST_HEAD(&cfqq->fifo);
-
-               atomic_set(&cfqq->ref, 0);
-               cfqq->cfqd = cfqd;
-
-               cfq_mark_cfqq_prio_changed(cfqq);
-
-               cfq_init_prio_data(cfqq, ioc);
-
-               if (is_sync) {
-                       if (!cfq_class_idle(cfqq))
-                               cfq_mark_cfqq_idle_window(cfqq);
-                       cfq_mark_cfqq_sync(cfqq);
-               }
-               cfqq->pid = current->pid;
-               cfq_log_cfqq(cfqd, cfqq, "alloced");
+               if (cfqq) {
+                       cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
+                       cfq_init_prio_data(cfqq, ioc);
+                       cfq_log_cfqq(cfqd, cfqq, "alloced");
+               } else
+                       cfqq = &cfqd->oom_cfqq;
        }
 
        if (new_cfqq)
                kmem_cache_free(cfq_pool, new_cfqq);
 
-out:
-       WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
        return cfqq;
 }
 
@@ -1548,7 +1763,7 @@ cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
 }
 
 static struct cfq_queue *
-cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
+cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
              gfp_t gfp_mask)
 {
        const int ioprio = task_ioprio(ioc);
@@ -1561,11 +1776,8 @@ cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
                cfqq = *async_cfqq;
        }
 
-       if (!cfqq) {
+       if (!cfqq)
                cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
-               if (!cfqq)
-                       return NULL;
-       }
 
        /*
         * pin the queue now that it's allocated, scheduler exit will prune it
@@ -1741,10 +1953,12 @@ cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
        sector_t sdist;
        u64 total;
 
-       if (cic->last_request_pos < rq->sector)
-               sdist = rq->sector - cic->last_request_pos;
+       if (!cic->last_request_pos)
+               sdist = 0;
+       else if (cic->last_request_pos < blk_rq_pos(rq))
+               sdist = blk_rq_pos(rq) - cic->last_request_pos;
        else
-               sdist = cic->last_request_pos - rq->sector;
+               sdist = cic->last_request_pos - blk_rq_pos(rq);
 
        /*
         * Don't allow the seek distance to get too large from the
@@ -1781,10 +1995,13 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
 
        if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
-           (cfqd->hw_tag && CIC_SEEKY(cic)))
+           (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
                enable_idle = 0;
        else if (sample_valid(cic->ttime_samples)) {
-               if (cic->ttime_mean > cfqd->cfq_slice_idle)
+               unsigned int slice_idle = cfqd->cfq_slice_idle;
+               if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
+                       slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
+               if (cic->ttime_mean > slice_idle)
                        enable_idle = 0;
                else
                        enable_idle = 1;
@@ -1803,7 +2020,7 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
  * Check if new_cfqq should preempt the currently active queue. Return 0 for
  * no or if we aren't sure, a 1 will cause a preempt.
  */
-static int
+static bool
 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
                   struct request *rq)
 {
@@ -1811,48 +2028,48 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
 
        cfqq = cfqd->active_queue;
        if (!cfqq)
-               return 0;
+               return false;
 
        if (cfq_slice_used(cfqq))
-               return 1;
+               return true;
 
        if (cfq_class_idle(new_cfqq))
-               return 0;
+               return false;
 
        if (cfq_class_idle(cfqq))
-               return 1;
+               return true;
 
        /*
         * 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;
+               return true;
 
        /*
         * 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 false;
 
        /*
         * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
         */
        if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
-               return 1;
+               return true;
 
        if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
-               return 0;
+               return false;
 
        /*
         * if this request is as-good as one we would expect from the
         * current cfqq, let it preempt
         */
        if (cfq_rq_close(cfqd, rq))
-               return 1;
+               return true;
 
-       return 0;
+       return false;
 }
 
 /*
@@ -1894,17 +2111,26 @@ cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        cfq_update_io_seektime(cfqd, cic, rq);
        cfq_update_idle_window(cfqd, cfqq, cic);
 
-       cic->last_request_pos = rq->sector + rq->nr_sectors;
+       cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
 
        if (cfqq == cfqd->active_queue) {
                /*
-                * if we are waiting for a request for this queue, let it rip
-                * immediately and flag that we must not expire this queue
-                * just now
+                * Remember that we saw a request from this process, but
+                * don't start queuing just yet. Otherwise we risk seeing lots
+                * of tiny requests, because we disrupt the normal plugging
+                * and merging. If the request is already larger than a single
+                * page, let it rip immediately. For that case we assume that
+                * merging is already done. Ditto for a busy system that
+                * has other work pending, don't risk delaying until the
+                * idle timer unplug to continue working.
                 */
                if (cfq_cfqq_wait_request(cfqq)) {
-                       del_timer(&cfqd->idle_slice_timer);
-                       blk_start_queueing(cfqd->queue);
+                       if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
+                           cfqd->busy_queues > 1) {
+                               del_timer(&cfqd->idle_slice_timer);
+                       __blk_run_queue(cfqd->queue);
+                       }
+                       cfq_mark_cfqq_must_dispatch(cfqq);
                }
        } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
                /*
@@ -1914,7 +2140,7 @@ 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_start_queueing(cfqd->queue);
+               __blk_run_queue(cfqd->queue);
        }
 }
 
@@ -1928,6 +2154,7 @@ static void cfq_insert_request(struct request_queue *q, struct request *rq)
 
        cfq_add_rq_rb(rq);
 
+       rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
        list_add_tail(&rq->queuelist, &cfqq->fifo);
 
        cfq_rq_enqueued(cfqd, cfqq, rq);
@@ -1939,11 +2166,11 @@ static void cfq_insert_request(struct request_queue *q, struct request *rq)
  */
 static void cfq_update_hw_tag(struct cfq_data *cfqd)
 {
-       if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
-               cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
+       if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
+               cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
 
        if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
-           cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
+           rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
                return;
 
        if (cfqd->hw_tag_samples++ < 50)
@@ -1970,38 +2197,45 @@ static void cfq_completed_request(struct request_queue *q, struct request *rq)
 
        cfq_update_hw_tag(cfqd);
 
-       WARN_ON(!cfqd->rq_in_driver);
+       WARN_ON(!cfqd->rq_in_driver[sync]);
        WARN_ON(!cfqq->dispatched);
-       cfqd->rq_in_driver--;
+       cfqd->rq_in_driver[sync]--;
        cfqq->dispatched--;
 
        if (cfq_cfqq_sync(cfqq))
                cfqd->sync_flight--;
 
-       if (!cfq_class_idle(cfqq))
-               cfqd->last_end_request = now;
-
-       if (sync)
+       if (sync) {
                RQ_CIC(rq)->last_end_request = now;
+               cfqd->last_end_sync_rq = now;
+       }
 
        /*
         * If this is the active queue, check if it needs to be expired,
         * or if we want to idle in case it has no pending requests.
         */
        if (cfqd->active_queue == cfqq) {
+               const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
+
                if (cfq_cfqq_slice_new(cfqq)) {
                        cfq_set_prio_slice(cfqd, cfqq);
                        cfq_clear_cfqq_slice_new(cfqq);
                }
+               /*
+                * If there are no requests waiting in this queue, and
+                * there are other queues ready to issue requests, AND
+                * those other queues are issuing requests within our
+                * mean seek distance, give them a chance to run instead
+                * of idling.
+                */
                if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
                        cfq_slice_expired(cfqd, 1);
-               else if (sync && !rq_noidle(rq) &&
-                        RB_EMPTY_ROOT(&cfqq->sort_list)) {
+               else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
+                        sync && !rq_noidle(rq))
                        cfq_arm_slice_timer(cfqd);
-               }
        }
 
-       if (!cfqd->rq_in_driver)
+       if (!rq_in_driver(cfqd))
                cfq_schedule_dispatch(cfqd);
 }
 
@@ -2033,8 +2267,7 @@ static void cfq_prio_boost(struct cfq_queue *cfqq)
 
 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)) {
+       if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
                cfq_mark_cfqq_must_alloc_slice(cfqq);
                return ELV_MQUEUE_MUST;
        }
@@ -2059,7 +2292,7 @@ static int cfq_may_queue(struct request_queue *q, int rw)
        if (!cic)
                return ELV_MQUEUE_MAY;
 
-       cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
+       cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
        if (cfqq) {
                cfq_init_prio_data(cfqq, cic->ioc);
                cfq_prio_boost(cfqq);
@@ -2101,7 +2334,7 @@ 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;
        const int rw = rq_data_dir(rq);
-       const int is_sync = rq_is_sync(rq);
+       const bool is_sync = rq_is_sync(rq);
        struct cfq_queue *cfqq;
        unsigned long flags;
 
@@ -2115,17 +2348,12 @@ cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
                goto queue_fail;
 
        cfqq = cic_to_cfqq(cic, is_sync);
-       if (!cfqq) {
+       if (!cfqq || cfqq == &cfqd->oom_cfqq) {
                cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
-
-               if (!cfqq)
-                       goto queue_fail;
-
                cic_set_cfqq(cic, cfqq, is_sync);
        }
 
        cfqq->allocated[rw]++;
-       cfq_clear_cfqq_must_alloc(cfqq);
        atomic_inc(&cfqq->ref);
 
        spin_unlock_irqrestore(q->queue_lock, flags);
@@ -2149,11 +2377,10 @@ static void cfq_kick_queue(struct work_struct *work)
        struct cfq_data *cfqd =
                container_of(work, struct cfq_data, unplug_work);
        struct request_queue *q = cfqd->queue;
-       unsigned long flags;
 
-       spin_lock_irqsave(q->queue_lock, flags);
-       blk_start_queueing(q);
-       spin_unlock_irqrestore(q->queue_lock, flags);
+       spin_lock_irq(q->queue_lock);
+       __blk_run_queue(cfqd->queue);
+       spin_unlock_irq(q->queue_lock);
 }
 
 /*
@@ -2175,6 +2402,12 @@ static void cfq_idle_slice_timer(unsigned long data)
                timed_out = 0;
 
                /*
+                * We saw a request before the queue expired, let it through
+                */
+               if (cfq_cfqq_must_dispatch(cfqq))
+                       goto out_kick;
+
+               /*
                 * expired
                 */
                if (cfq_slice_used(cfqq))
@@ -2254,12 +2487,30 @@ static void cfq_exit_queue(struct elevator_queue *e)
 static void *cfq_init_queue(struct request_queue *q)
 {
        struct cfq_data *cfqd;
+       int i;
 
        cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
        if (!cfqd)
                return NULL;
 
        cfqd->service_tree = CFQ_RB_ROOT;
+
+       /*
+        * Not strictly needed (since RB_ROOT just clears the node and we
+        * zeroed cfqd on alloc), but better be safe in case someone decides
+        * to add magic to the rb code
+        */
+       for (i = 0; i < CFQ_PRIO_LISTS; i++)
+               cfqd->prio_trees[i] = RB_ROOT;
+
+       /*
+        * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
+        * Grab a permanent reference to it, so that the normal code flow
+        * will not attempt to free it.
+        */
+       cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
+       atomic_inc(&cfqd->oom_cfqq.ref);
+
        INIT_LIST_HEAD(&cfqd->cic_list);
 
        cfqd->queue = q;
@@ -2270,7 +2521,6 @@ static void *cfq_init_queue(struct request_queue *q)
 
        INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
 
-       cfqd->last_end_request = jiffies;
        cfqd->cfq_quantum = cfq_quantum;
        cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
        cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
@@ -2280,8 +2530,9 @@ static void *cfq_init_queue(struct request_queue *q)
        cfqd->cfq_slice[1] = cfq_slice_sync;
        cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
        cfqd->cfq_slice_idle = cfq_slice_idle;
+       cfqd->cfq_latency = 1;
        cfqd->hw_tag = 1;
-
+       cfqd->last_end_sync_rq = jiffies;
        return cfqd;
 }
 
@@ -2349,6 +2600,7 @@ SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
 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);
 #undef SHOW_FUNCTION
 
 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                        \
@@ -2380,6 +2632,7 @@ STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
 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);
 #undef STORE_FUNCTION
 
 #define CFQ_ATTR(name) \
@@ -2395,6 +2648,7 @@ static struct elv_fs_entry cfq_attrs[] = {
        CFQ_ATTR(slice_async),
        CFQ_ATTR(slice_async_rq),
        CFQ_ATTR(slice_idle),
+       CFQ_ATTR(low_latency),
        __ATTR_NULL
 };
 
@@ -2454,7 +2708,7 @@ static void __exit cfq_exit(void)
         * this also protects us from entering cfq_slab_kill() with
         * pending RCU callbacks
         */
-       if (elv_ioc_count_read(ioc_count))
+       if (elv_ioc_count_read(cfq_ioc_count))
                wait_for_completion(&all_gone);
        cfq_slab_kill();
 }