[PATCH] CFQ io scheduler updates
Jens Axboe [Mon, 27 Jun 2005 08:56:24 +0000 (10:56 +0200)]
- Adjust slice values

- Instead of one async queue, one is defined per priority level. This
  prevents kernel threads (such as reiserfs/x and others) that run at
  higher io priority from conflicting with others. Previously, it was a
  coin toss what io prio the async queue got, it was defined by who
  first set up the queue.

- Let a time slice only begin, when the previous slice is completely
  done. Previously we could be somewhat unfair to a new sync slice, if
  the previous slice was async and had several ios queued. This might
  need a little tweaking if throughput suffers a little due to this,
  allowing perhaps an overlap of a single request or so.

- Optimize the calling of kblockd_schedule_work() by doing it only when
  it is strictly necessary (no requests in driver and work left to do).

- Correct sync vs async logic. A 'normal' process can be purely async as
  well, and a flusher can be purely sync as well. Sync or async is now a
  property of the class defined and requests pending. Previously writers
  could be considered sync, when they were really async.

- Get rid of the bit fields in cfqq and crq, use flags instead.

- Various other cleanups and fixes

Signed-off-by: Jens Axboe <axboe@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

drivers/block/cfq-iosched.c
include/linux/ioprio.h

index baa3e26..1ecb179 100644 (file)
@@ -34,14 +34,15 @@ static int cfq_back_max = 16 * 1024;        /* maximum backwards seek, in KiB */
 static int cfq_back_penalty = 2;       /* penalty of a backwards seek */
 
 static int cfq_slice_sync = HZ / 10;
-static int cfq_slice_async = HZ / 50;
+static int cfq_slice_async = HZ / 25;
 static int cfq_slice_async_rq = 2;
-static int cfq_slice_idle = HZ / 50;
+static int cfq_slice_idle = HZ / 100;
 
 #define CFQ_IDLE_GRACE         (HZ / 10)
 #define CFQ_SLICE_SCALE                (5)
 
 #define CFQ_KEY_ASYNC          (0)
+#define CFQ_KEY_ANY            (0xffff)
 
 /*
  * disable queueing at the driver/hardware level
@@ -96,7 +97,16 @@ static kmem_cache_t *cfq_ioc_pool;
 #define cfq_class_be(cfqq)     ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
 #define cfq_class_rt(cfqq)     ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
 
-#define cfq_cfqq_sync(cfqq)    ((cfqq)->key != CFQ_KEY_ASYNC)
+#define ASYNC                  (0)
+#define SYNC                   (1)
+
+#define cfq_cfqq_dispatched(cfqq)      \
+       ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
+
+#define cfq_cfqq_class_sync(cfqq)      ((cfqq)->key != CFQ_KEY_ASYNC)
+
+#define cfq_cfqq_sync(cfqq)            \
+       (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
 
 /*
  * Per block device queue structure
@@ -200,28 +210,15 @@ struct cfq_queue {
        unsigned long slice_left;
        unsigned long service_last;
 
-       /* number of requests that have been handed to the driver */
-       int in_flight;
+       /* number of requests that are on the dispatch list */
+       int on_dispatch[2];
 
        /* io prio of this group */
        unsigned short ioprio, org_ioprio;
        unsigned short ioprio_class, org_ioprio_class;
 
-       /* whether queue is on rr (or empty) list */
-       unsigned on_rr : 1;
-       /* idle slice, waiting for new request submission */
-       unsigned wait_request : 1;
-       /* set when wait_request gets set, reset on first rq alloc */
-       unsigned must_alloc : 1;
-       /* only gets one must_alloc per slice */
-       unsigned must_alloc_slice : 1;
-       /* idle slice, request added, now waiting to dispatch it */
-       unsigned must_dispatch : 1;
-       /* fifo expire per-slice */
-       unsigned fifo_expire : 1;
-
-       unsigned idle_window : 1;
-       unsigned prio_changed : 1;
+       /* various state flags, see below */
+       unsigned int flags;
 };
 
 struct cfq_rq {
@@ -233,15 +230,77 @@ struct cfq_rq {
        struct cfq_queue *cfq_queue;
        struct cfq_io_context *io_context;
 
-       unsigned in_flight : 1;
-       unsigned accounted : 1;
-       unsigned is_sync   : 1;
-       unsigned requeued  : 1;
+       unsigned int crq_flags;
 };
 
-static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int);
+enum cfqq_state_flags {
+       CFQ_CFQQ_FLAG_on_rr = 0,
+       CFQ_CFQQ_FLAG_wait_request,
+       CFQ_CFQQ_FLAG_must_alloc,
+       CFQ_CFQQ_FLAG_must_alloc_slice,
+       CFQ_CFQQ_FLAG_must_dispatch,
+       CFQ_CFQQ_FLAG_fifo_expire,
+       CFQ_CFQQ_FLAG_idle_window,
+       CFQ_CFQQ_FLAG_prio_changed,
+       CFQ_CFQQ_FLAG_expired,
+};
+
+#define CFQ_CFQQ_FNS(name)                                             \
+static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)                \
+{                                                                      \
+       cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
+}                                                                      \
+static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)       \
+{                                                                      \
+       cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
+}                                                                      \
+static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)                \
+{                                                                      \
+       return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
+}
+
+CFQ_CFQQ_FNS(on_rr);
+CFQ_CFQQ_FNS(wait_request);
+CFQ_CFQQ_FNS(must_alloc);
+CFQ_CFQQ_FNS(must_alloc_slice);
+CFQ_CFQQ_FNS(must_dispatch);
+CFQ_CFQQ_FNS(fifo_expire);
+CFQ_CFQQ_FNS(idle_window);
+CFQ_CFQQ_FNS(prio_changed);
+CFQ_CFQQ_FNS(expired);
+#undef CFQ_CFQQ_FNS
+
+enum cfq_rq_state_flags {
+       CFQ_CRQ_FLAG_in_flight = 0,
+       CFQ_CRQ_FLAG_in_driver,
+       CFQ_CRQ_FLAG_is_sync,
+       CFQ_CRQ_FLAG_requeued,
+};
+
+#define CFQ_CRQ_FNS(name)                                              \
+static inline void cfq_mark_crq_##name(struct cfq_rq *crq)             \
+{                                                                      \
+       crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
+}                                                                      \
+static inline void cfq_clear_crq_##name(struct cfq_rq *crq)            \
+{                                                                      \
+       crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
+}                                                                      \
+static inline int cfq_crq_##name(const struct cfq_rq *crq)             \
+{                                                                      \
+       return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
+}
+
+CFQ_CRQ_FNS(in_flight);
+CFQ_CRQ_FNS(in_driver);
+CFQ_CRQ_FNS(is_sync);
+CFQ_CRQ_FNS(requeued);
+#undef CFQ_CRQ_FNS
+
+static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 static void cfq_dispatch_sort(request_queue_t *, struct cfq_rq *);
 static void cfq_put_cfqd(struct cfq_data *cfqd);
+static inline int cfq_pending_requests(struct cfq_data *cfqd);
 
 #define process_sync(tsk)      ((tsk)->flags & PF_SYNCWRITE)
 
@@ -305,9 +364,9 @@ cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
                return crq2;
        if (crq2 == NULL)
                return crq1;
-       if (crq1->requeued)
+       if (cfq_crq_requeued(crq1))
                return crq1;
-       if (crq2->requeued)
+       if (cfq_crq_requeued(crq2))
                return crq2;
 
        s1 = crq1->request->sector;
@@ -407,7 +466,7 @@ static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
        struct cfq_data *cfqd = cfqq->cfqd;
        struct list_head *list, *entry;
 
-       BUG_ON(!cfqq->on_rr);
+       BUG_ON(!cfq_cfqq_on_rr(cfqq));
 
        list_del(&cfqq->cfq_list);
 
@@ -423,7 +482,7 @@ static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
                 * has lots of io pending vs one that only generates one
                 * sporadically or synchronously
                 */
-               if (cfqq->in_flight)
+               if (cfq_cfqq_dispatched(cfqq))
                        list = &cfqd->busy_rr;
                else
                        list = &cfqd->rr_list[cfqq->ioprio];
@@ -461,8 +520,8 @@ static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 static inline void
 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq, int requeue)
 {
-       BUG_ON(cfqq->on_rr);
-       cfqq->on_rr = 1;
+       BUG_ON(cfq_cfqq_on_rr(cfqq));
+       cfq_mark_cfqq_on_rr(cfqq);
        cfqd->busy_queues++;
 
        cfq_resort_rr_list(cfqq, requeue);
@@ -471,8 +530,8 @@ cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq, int requeue)
 static inline void
 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
-       BUG_ON(!cfqq->on_rr);
-       cfqq->on_rr = 0;
+       BUG_ON(!cfq_cfqq_on_rr(cfqq));
+       cfq_clear_cfqq_on_rr(cfqq);
        list_move(&cfqq->cfq_list, &cfqd->empty_list);
 
        BUG_ON(!cfqd->busy_queues);
@@ -488,7 +547,7 @@ static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 
        if (ON_RB(&crq->rb_node)) {
                struct cfq_data *cfqd = cfqq->cfqd;
-               const int sync = crq->is_sync;
+               const int sync = cfq_crq_is_sync(crq);
 
                BUG_ON(!cfqq->queued[sync]);
                cfqq->queued[sync]--;
@@ -498,7 +557,7 @@ static inline void cfq_del_crq_rb(struct cfq_rq *crq)
                rb_erase(&crq->rb_node, &cfqq->sort_list);
                RB_CLEAR_COLOR(&crq->rb_node);
 
-               if (cfqq->on_rr && RB_EMPTY(&cfqq->sort_list))
+               if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
                        cfq_del_cfqq_rr(cfqd, cfqq);
        }
 }
@@ -534,7 +593,7 @@ static void cfq_add_crq_rb(struct cfq_rq *crq)
        struct cfq_rq *__alias;
 
        crq->rb_key = rq_rb_key(rq);
-       cfqq->queued[crq->is_sync]++;
+       cfqq->queued[cfq_crq_is_sync(crq)]++;
 
        /*
         * looks a little odd, but the first insert might return an alias.
@@ -545,8 +604,8 @@ static void cfq_add_crq_rb(struct cfq_rq *crq)
 
        rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 
-       if (!cfqq->on_rr)
-               cfq_add_cfqq_rr(cfqd, cfqq, crq->requeued);
+       if (!cfq_cfqq_on_rr(cfqq))
+               cfq_add_cfqq_rr(cfqd, cfqq, cfq_crq_requeued(crq));
 
        /*
         * check if this request is a better next-serve candidate
@@ -559,7 +618,7 @@ cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 {
        if (ON_RB(&crq->rb_node)) {
                rb_erase(&crq->rb_node, &cfqq->sort_list);
-               cfqq->queued[crq->is_sync]--;
+               cfqq->queued[cfq_crq_is_sync(crq)]--;
        }
 
        cfq_add_crq_rb(crq);
@@ -568,7 +627,7 @@ cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
 
 {
-       struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid);
+       struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
        struct rb_node *n;
 
        if (!cfqq)
@@ -598,17 +657,19 @@ static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
        if (crq) {
                struct cfq_queue *cfqq = crq->cfq_queue;
 
-               if (crq->accounted) {
-                       crq->accounted = 0;
+               if (cfq_crq_in_driver(crq)) {
+                       cfq_clear_crq_in_driver(crq);
                        WARN_ON(!cfqd->rq_in_driver);
                        cfqd->rq_in_driver--;
                }
-               if (crq->in_flight) {
-                       crq->in_flight = 0;
-                       WARN_ON(!cfqq->in_flight);
-                       cfqq->in_flight--;
+               if (cfq_crq_in_flight(crq)) {
+                       const int sync = cfq_crq_is_sync(crq);
+
+                       cfq_clear_crq_in_flight(crq);
+                       WARN_ON(!cfqq->on_dispatch[sync]);
+                       cfqq->on_dispatch[sync]--;
                }
-               crq->requeued = 1;
+               cfq_mark_crq_requeued(crq);
        }
 }
 
@@ -712,8 +773,9 @@ __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
                cfqq->slice_start = jiffies;
                cfqq->slice_end = 0;
                cfqq->slice_left = 0;
-               cfqq->must_alloc_slice = 0;
-               cfqq->fifo_expire = 0;
+               cfq_clear_cfqq_must_alloc_slice(cfqq);
+               cfq_clear_cfqq_fifo_expire(cfqq);
+               cfq_clear_cfqq_expired(cfqq);
        }
 
        cfqd->active_queue = cfqq;
@@ -776,9 +838,18 @@ static int cfq_get_next_prio_level(struct cfq_data *cfqd)
        return prio;
 }
 
-static void cfq_set_active_queue(struct cfq_data *cfqd)
+static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 {
-       struct cfq_queue *cfqq = NULL;
+       struct cfq_queue *cfqq;
+
+       /*
+        * if current queue is expired but not done with its requests yet,
+        * wait for that to happen
+        */
+       if ((cfqq = cfqd->active_queue) != NULL) {
+               if (cfq_cfqq_expired(cfqq) && cfq_cfqq_dispatched(cfqq))
+                       return NULL;
+       }
 
        /*
         * if current list is non-empty, grab first entry. if it is empty,
@@ -802,50 +873,66 @@ static void cfq_set_active_queue(struct cfq_data *cfqd)
        }
 
        __cfq_set_active_queue(cfqd, cfqq);
+       return cfqq;
 }
 
 /*
  * current cfqq expired its slice (or was too idle), select new one
  */
-static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
+static void
+__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
+                   int preempted)
 {
-       struct cfq_queue *cfqq = cfqd->active_queue;
-
-       if (cfqq) {
-               unsigned long now = jiffies;
+       unsigned long now = jiffies;
 
-               if (cfqq->wait_request)
-                       del_timer(&cfqd->idle_slice_timer);
+       if (cfq_cfqq_wait_request(cfqq))
+               del_timer(&cfqd->idle_slice_timer);
 
-               if (!preempted && !cfqq->in_flight)
-                       cfqq->service_last = now;
+       if (!preempted && !cfq_cfqq_dispatched(cfqq))
+               cfqq->service_last = now;
 
-               cfqq->must_dispatch = 0;
-               cfqq->wait_request = 0;
+       cfq_clear_cfqq_must_dispatch(cfqq);
+       cfq_clear_cfqq_wait_request(cfqq);
 
-               /*
-                * store what was left of this slice, if the queue idled out
-                * or was preempted
-                */
-               if (time_after(now, cfqq->slice_end))
-                       cfqq->slice_left = now - cfqq->slice_end;
-               else
-                       cfqq->slice_left = 0;
+       /*
+        * store what was left of this slice, if the queue idled out
+        * or was preempted
+        */
+       if (time_after(now, cfqq->slice_end))
+               cfqq->slice_left = now - cfqq->slice_end;
+       else
+               cfqq->slice_left = 0;
 
-               if (cfqq->on_rr)
-                       cfq_resort_rr_list(cfqq, preempted);
+       if (cfq_cfqq_on_rr(cfqq))
+               cfq_resort_rr_list(cfqq, preempted);
 
+       if (cfqq == cfqd->active_queue)
                cfqd->active_queue = NULL;
 
-               if (cfqd->active_cic) {
-                       put_io_context(cfqd->active_cic->ioc);
-                       cfqd->active_cic = NULL;
-               }
+       if (cfqd->active_cic) {
+               put_io_context(cfqd->active_cic->ioc);
+               cfqd->active_cic = NULL;
        }
 
        cfqd->dispatch_slice = 0;
 }
 
+static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
+{
+       struct cfq_queue *cfqq = cfqd->active_queue;
+
+       if (cfqq) {
+               /*
+                * use deferred expiry, if there are requests in progress as
+                * not to disturb the slice of the next queue
+                */
+               if (cfq_cfqq_dispatched(cfqq))
+                       cfq_mark_cfqq_expired(cfqq);
+               else
+                       __cfq_slice_expired(cfqd, cfqq, preempted);
+       }
+}
+
 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 
 {
@@ -857,7 +944,7 @@ static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
         */
        if (!cfqd->cfq_slice_idle)
                return 0;
-       if (!cfqq->idle_window)
+       if (!cfq_cfqq_idle_window(cfqq))
                return 0;
        /*
         * task has exited, don't wait
@@ -865,13 +952,13 @@ static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
        if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
                return 0;
 
-       cfqq->wait_request = 1;
-       cfqq->must_alloc = 1;
+       cfq_mark_cfqq_must_dispatch(cfqq);
+       cfq_mark_cfqq_wait_request(cfqq);
 
        if (!timer_pending(&cfqd->idle_slice_timer)) {
-               unsigned long slice_left = cfqq->slice_end - 1;
+               unsigned long slice_left = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 
-               cfqd->idle_slice_timer.expires = min(jiffies + cfqd->cfq_slice_idle, slice_left);
+               cfqd->idle_slice_timer.expires = jiffies + slice_left;
                add_timer(&cfqd->idle_slice_timer);
        }
 
@@ -901,7 +988,7 @@ static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
                        break;
                if (!blk_fs_request(__rq))
                        break;
-               if (__crq->requeued)
+               if (cfq_crq_requeued(__crq))
                        break;
 
                if (__rq->sector <= crq->request->sector)
@@ -920,9 +1007,10 @@ static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
        cfq_del_crq_rb(crq);
        cfq_remove_merge_hints(q, crq);
 
-       crq->in_flight = 1;
-       crq->requeued = 0;
-       cfqq->in_flight++;
+       cfq_mark_crq_in_flight(crq);
+       cfq_clear_crq_requeued(crq);
+
+       cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
        list_add_tail(&crq->request->queuelist, entry);
 }
 
@@ -935,16 +1023,16 @@ static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
        struct request *rq;
        struct cfq_rq *crq;
 
-       if (cfqq->fifo_expire)
+       if (cfq_cfqq_fifo_expire(cfqq))
                return NULL;
 
        if (!list_empty(&cfqq->fifo)) {
-               int fifo = cfq_cfqq_sync(cfqq);
+               int fifo = cfq_cfqq_class_sync(cfqq);
 
                crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
                rq = crq->request;
                if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
-                       cfqq->fifo_expire = 1;
+                       cfq_mark_cfqq_fifo_expire(cfqq);
                        return crq;
                }
        }
@@ -953,7 +1041,9 @@ static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 }
 
 /*
- * Scale schedule slice based on io priority
+ * Scale schedule slice based on io priority. Use the sync time slice only
+ * 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_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
@@ -982,6 +1072,16 @@ cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 }
 
 /*
+ * scheduler run of queue, if there are requests pending and no one in the
+ * driver that will restart queueing
+ */
+static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
+{
+       if (!cfqd->rq_in_driver && cfq_pending_requests(cfqd))
+               kblockd_schedule_work(&cfqd->unplug_work);
+}
+
+/*
  * get next queue for service
  */
 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
@@ -993,11 +1093,14 @@ static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
        if (!cfqq)
                goto new_queue;
 
+       if (cfq_cfqq_expired(cfqq))
+               goto new_queue;
+
        /*
         * slice has expired
         */
-       if (!cfqq->must_dispatch && time_after(jiffies, cfqq->slice_end))
-               goto new_queue;
+       if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
+               goto expire;
 
        /*
         * if queue has requests, dispatch one. if not, check if
@@ -1005,17 +1108,18 @@ static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
         */
        if (!RB_EMPTY(&cfqq->sort_list))
                goto keep_queue;
-       else if (!force && cfq_cfqq_sync(cfqq) &&
+       else if (!force && cfq_cfqq_class_sync(cfqq) &&
                 time_before(now, cfqq->slice_end)) {
                if (cfq_arm_slice_timer(cfqd, cfqq))
                        return NULL;
        }
 
-new_queue:
+expire:
        cfq_slice_expired(cfqd, 0);
-       cfq_set_active_queue(cfqd);
+new_queue:
+       cfqq = cfq_set_active_queue(cfqd);
 keep_queue:
-       return cfqd->active_queue;
+       return cfqq;
 }
 
 static int
@@ -1083,8 +1187,8 @@ cfq_dispatch_requests(request_queue_t *q, int max_dispatch, int force)
 
        cfqq = cfq_select_queue(cfqd, force);
        if (cfqq) {
-               cfqq->wait_request = 0;
-               cfqq->must_dispatch = 0;
+               cfq_clear_cfqq_must_dispatch(cfqq);
+               cfq_clear_cfqq_wait_request(cfqq);
                del_timer(&cfqd->idle_slice_timer);
 
                if (cfq_class_idle(cfqq))
@@ -1108,10 +1212,10 @@ static inline void cfq_account_dispatch(struct cfq_rq *crq)
         * accounted bit is necessary since some drivers will call
         * elv_next_request() many times for the same request (eg ide)
         */
-       if (crq->accounted)
+       if (cfq_crq_in_driver(crq))
                return;
 
-       crq->accounted = 1;
+       cfq_mark_crq_in_driver(crq);
        cfqd->rq_in_driver++;
 }
 
@@ -1121,7 +1225,7 @@ cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq)
        struct cfq_data *cfqd = cfqq->cfqd;
        unsigned long now;
 
-       if (!crq->accounted)
+       if (!cfq_crq_in_driver(crq))
                return;
 
        now = jiffies;
@@ -1132,12 +1236,18 @@ cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq)
        if (!cfq_class_idle(cfqq))
                cfqd->last_end_request = now;
 
-       if (!cfqq->in_flight && cfqq->on_rr) {
-               cfqq->service_last = now;
-               cfq_resort_rr_list(cfqq, 0);
+       if (!cfq_cfqq_dispatched(cfqq)) {
+               if (cfq_cfqq_on_rr(cfqq)) {
+                       cfqq->service_last = now;
+                       cfq_resort_rr_list(cfqq, 0);
+               }
+               if (cfq_cfqq_expired(cfqq)) {
+                       __cfq_slice_expired(cfqd, cfqq, 0);
+                       cfq_schedule_dispatch(cfqd);
+               }
        }
 
-       if (crq->is_sync)
+       if (cfq_crq_is_sync(crq))
                crq->io_context->last_end_request = now;
 }
 
@@ -1153,10 +1263,13 @@ dispatch:
 
                crq = RQ_DATA(rq);
                if (crq) {
+                       struct cfq_queue *cfqq = crq->cfq_queue;
+
                        /*
                         * if idle window is disabled, allow queue buildup
                         */
-                       if (!crq->in_flight && !crq->cfq_queue->idle_window &&
+                       if (!cfq_crq_in_driver(crq) &&
+                           !cfq_cfqq_idle_window(cfqq) &&
                            cfqd->rq_in_driver >= cfqd->cfq_max_depth)
                                return NULL;
 
@@ -1190,11 +1303,11 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
 
        BUG_ON(rb_first(&cfqq->sort_list));
        BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
-       BUG_ON(cfqq->on_rr);
+       BUG_ON(cfq_cfqq_on_rr(cfqq));
 
        if (unlikely(cfqd->active_queue == cfqq)) {
-               cfq_slice_expired(cfqd, 0);
-               kblockd_schedule_work(&cfqd->unplug_work);
+               __cfq_slice_expired(cfqd, cfqq, 0);
+               cfq_schedule_dispatch(cfqd);
        }
 
        cfq_put_cfqd(cfqq->cfqd);
@@ -1208,15 +1321,17 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
 }
 
 static inline struct cfq_queue *
-__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, const int hashval)
+__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
+                   const int hashval)
 {
        struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
        struct hlist_node *entry, *next;
 
        hlist_for_each_safe(entry, next, hash_list) {
                struct cfq_queue *__cfqq = list_entry_qhash(entry);
+               const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
 
-               if (__cfqq->key == key)
+               if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
                        return __cfqq;
        }
 
@@ -1224,9 +1339,9 @@ __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, const int hashval)
 }
 
 static struct cfq_queue *
-cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key)
+cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
 {
-       return __cfq_find_cfq_hash(cfqd, key, hash_long(key, CFQ_QHASH_SHIFT));
+       return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
 }
 
 static void cfq_free_io_context(struct cfq_io_context *cic)
@@ -1255,8 +1370,8 @@ static void cfq_exit_single_io_context(struct cfq_io_context *cic)
        spin_lock(q->queue_lock);
 
        if (unlikely(cic->cfqq == cfqd->active_queue)) {
-               cfq_slice_expired(cfqd, 0);
-               kblockd_schedule_work(&cfqd->unplug_work);
+               __cfq_slice_expired(cfqd, cic->cfqq, 0);
+               cfq_schedule_dispatch(cfqd);
        }
 
        cfq_put_queue(cic->cfqq);
@@ -1313,7 +1428,7 @@ static void cfq_init_prio_data(struct cfq_queue *cfqq)
        struct task_struct *tsk = current;
        int ioprio_class;
 
-       if (!cfqq->prio_changed)
+       if (!cfq_cfqq_prio_changed(cfqq))
                return;
 
        ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
@@ -1338,7 +1453,7 @@ static void cfq_init_prio_data(struct cfq_queue *cfqq)
                case IOPRIO_CLASS_IDLE:
                        cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
                        cfqq->ioprio = 7;
-                       cfqq->idle_window = 0;
+                       cfq_clear_cfqq_idle_window(cfqq);
                        break;
        }
 
@@ -1349,10 +1464,10 @@ static void cfq_init_prio_data(struct cfq_queue *cfqq)
        cfqq->org_ioprio = cfqq->ioprio;
        cfqq->org_ioprio_class = cfqq->ioprio_class;
 
-       if (cfqq->on_rr)
+       if (cfq_cfqq_on_rr(cfqq))
                cfq_resort_rr_list(cfqq, 0);
 
-       cfqq->prio_changed = 0;
+       cfq_clear_cfqq_prio_changed(cfqq);
 }
 
 static inline void changed_ioprio(struct cfq_queue *cfqq)
@@ -1361,7 +1476,7 @@ static inline void changed_ioprio(struct cfq_queue *cfqq)
                struct cfq_data *cfqd = cfqq->cfqd;
 
                spin_lock(cfqd->queue->queue_lock);
-               cfqq->prio_changed = 1;
+               cfq_mark_cfqq_prio_changed(cfqq);
                cfq_init_prio_data(cfqq);
                spin_unlock(cfqd->queue->queue_lock);
        }
@@ -1383,13 +1498,14 @@ static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
 }
 
 static struct cfq_queue *
-cfq_get_queue(struct cfq_data *cfqd, unsigned int key, int gfp_mask)
+cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
+             int gfp_mask)
 {
        const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
        struct cfq_queue *cfqq, *new_cfqq = NULL;
 
 retry:
-       cfqq = __cfq_find_cfq_hash(cfqd, key, hashval);
+       cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
 
        if (!cfqq) {
                if (new_cfqq) {
@@ -1423,10 +1539,9 @@ retry:
                 * set ->slice_left to allow preemption for a new process
                 */
                cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
-               cfqq->idle_window = 1;
-               cfqq->ioprio = -1;
-               cfqq->ioprio_class = -1;
-               cfqq->prio_changed = 1;
+               cfq_mark_cfqq_idle_window(cfqq);
+               cfq_mark_cfqq_prio_changed(cfqq);
+               cfq_init_prio_data(cfqq);
        }
 
        if (new_cfqq)
@@ -1553,7 +1668,7 @@ static void
 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                       struct cfq_io_context *cic)
 {
-       int enable_idle = cfqq->idle_window;
+       int enable_idle = cfq_cfqq_idle_window(cfqq);
 
        if (!cic->ioc->task || !cfqd->cfq_slice_idle)
                enable_idle = 0;
@@ -1564,7 +1679,10 @@ cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                        enable_idle = 1;
        }
 
-       cfqq->idle_window = enable_idle;
+       if (enable_idle)
+               cfq_mark_cfqq_idle_window(cfqq);
+       else
+               cfq_clear_cfqq_idle_window(cfqq);
 }
 
 
@@ -1586,14 +1704,14 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
 
        if (cfq_class_idle(cfqq))
                return 1;
-       if (!new_cfqq->wait_request)
+       if (!cfq_cfqq_wait_request(new_cfqq))
                return 0;
        /*
         * if it doesn't have slice left, forget it
         */
        if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
                return 0;
-       if (crq->is_sync && !cfq_cfqq_sync(cfqq))
+       if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
                return 1;
 
        return 0;
@@ -1614,7 +1732,7 @@ static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
                cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
 
        cfqq->slice_end = cfqq->slice_left + jiffies;
-       cfq_slice_expired(cfqd, 1);
+       __cfq_slice_expired(cfqd, cfqq, 1);
        __cfq_set_active_queue(cfqd, cfqq);
 }
 
@@ -1639,7 +1757,7 @@ static void
 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                 struct cfq_rq *crq)
 {
-       const int sync = crq->is_sync;
+       const int sync = cfq_crq_is_sync(crq);
 
        cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 
@@ -1658,8 +1776,8 @@ cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                 * immediately and flag that we must not expire this queue
                 * just now
                 */
-               if (cfqq->wait_request) {
-                       cfqq->must_dispatch = 1;
+               if (cfq_cfqq_wait_request(cfqq)) {
+                       cfq_mark_cfqq_must_dispatch(cfqq);
                        del_timer(&cfqd->idle_slice_timer);
                        cfq_start_queueing(cfqd, cfqq);
                }
@@ -1670,7 +1788,7 @@ cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                 * has some old slice time left and is of higher priority
                 */
                cfq_preempt_queue(cfqd, cfqq);
-               cfqq->must_dispatch = 1;
+               cfq_mark_cfqq_must_dispatch(cfqq);
                cfq_start_queueing(cfqd, cfqq);
        }
 }
@@ -1713,7 +1831,7 @@ cfq_insert_request(request_queue_t *q, struct request *rq, int where)
                         * be kicked by __make_request() afterward.
                         * Kick it here.
                         */
-                       kblockd_schedule_work(&cfqd->unplug_work);
+                       cfq_schedule_dispatch(cfqd);
                        break;
                case ELEVATOR_INSERT_FRONT:
                        list_add(&rq->queuelist, &q->queue_head);
@@ -1750,9 +1868,11 @@ static void cfq_completed_request(request_queue_t *q, struct request *rq)
 
        cfqq = crq->cfq_queue;
 
-       if (crq->in_flight) {
-               WARN_ON(!cfqq->in_flight);
-               cfqq->in_flight--;
+       if (cfq_crq_in_flight(crq)) {
+               const int sync = cfq_crq_is_sync(crq);
+
+               WARN_ON(!cfqq->on_dispatch[sync]);
+               cfqq->on_dispatch[sync]--;
        }
 
        cfq_account_completion(cfqq, crq);
@@ -1814,7 +1934,7 @@ static void cfq_prio_boost(struct cfq_queue *cfqq)
         * refile between round-robin lists if we moved the priority class
         */
        if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
-           cfqq->on_rr)
+           cfq_cfqq_on_rr(cfqq))
                cfq_resort_rr_list(cfqq, 0);
 }
 
@@ -1830,23 +1950,27 @@ static inline int
 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                struct task_struct *task, int rw)
 {
-       if (cfqq->wait_request && cfqq->must_alloc)
+#if 1
+       if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
+           !cfq_cfqq_must_alloc_slice) {
+               cfq_mark_cfqq_must_alloc_slice(cfqq);
                return ELV_MQUEUE_MUST;
+       }
 
        return ELV_MQUEUE_MAY;
-#if 0
+#else
        if (!cfqq || task->flags & PF_MEMALLOC)
                return ELV_MQUEUE_MAY;
-       if (!cfqq->allocated[rw] || cfqq->must_alloc) {
-               if (cfqq->wait_request)
+       if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
+               if (cfq_cfqq_wait_request(cfqq))
                        return ELV_MQUEUE_MUST;
 
                /*
                 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
                 * can quickly flood the queue with writes from a single task
                 */
-               if (rw == READ || !cfqq->must_alloc_slice) {
-                       cfqq->must_alloc_slice = 1;
+               if (rw == READ || !cfq_cfqq_must_alloc_slice) {
+                       cfq_mark_cfqq_must_alloc_slice(cfqq);
                        return ELV_MQUEUE_MUST;
                }
 
@@ -1881,7 +2005,7 @@ static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
         * so just lookup a possibly existing queue, or return 'may queue'
         * if that fails
         */
-       cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw));
+       cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
        if (cfqq) {
                cfq_init_prio_data(cfqq);
                cfq_prio_boost(cfqq);
@@ -1943,15 +2067,17 @@ cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
                int gfp_mask)
 {
        struct cfq_data *cfqd = q->elevator->elevator_data;
+       struct task_struct *tsk = current;
        struct cfq_io_context *cic;
        const int rw = rq_data_dir(rq);
+       pid_t key = cfq_queue_pid(tsk, rw);
        struct cfq_queue *cfqq;
        struct cfq_rq *crq;
        unsigned long flags;
 
        might_sleep_if(gfp_mask & __GFP_WAIT);
 
-       cic = cfq_get_io_context(cfqd, cfq_queue_pid(current, rw), gfp_mask);
+       cic = cfq_get_io_context(cfqd, key, gfp_mask);
 
        spin_lock_irqsave(q->queue_lock, flags);
 
@@ -1959,7 +2085,7 @@ cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
                goto queue_fail;
 
        if (!cic->cfqq) {
-               cfqq = cfq_get_queue(cfqd, current->pid, gfp_mask);
+               cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
                if (!cfqq)
                        goto queue_fail;
 
@@ -1968,7 +2094,7 @@ cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
                cfqq = cic->cfqq;
 
        cfqq->allocated[rw]++;
-       cfqq->must_alloc = 0;
+       cfq_clear_cfqq_must_alloc(cfqq);
        cfqd->rq_starved = 0;
        atomic_inc(&cfqq->ref);
        spin_unlock_irqrestore(q->queue_lock, flags);
@@ -1981,9 +2107,15 @@ cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
                INIT_HLIST_NODE(&crq->hash);
                crq->cfq_queue = cfqq;
                crq->io_context = cic;
-               crq->in_flight = crq->accounted = 0;
-               crq->is_sync = (rw == READ || process_sync(current));
-               crq->requeued = 0;
+               cfq_clear_crq_in_flight(crq);
+               cfq_clear_crq_in_driver(crq);
+               cfq_clear_crq_requeued(crq);
+
+               if (rw == READ || process_sync(tsk))
+                       cfq_mark_crq_is_sync(crq);
+               else
+                       cfq_clear_crq_is_sync(crq);
+
                rq->elevator_private = crq;
                return 0;
        }
@@ -1991,7 +2123,7 @@ cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
        spin_lock_irqsave(q->queue_lock, flags);
        cfqq->allocated[rw]--;
        if (!(cfqq->allocated[0] + cfqq->allocated[1]))
-               cfqq->must_alloc = 1;
+               cfq_mark_cfqq_must_alloc(cfqq);
        cfq_put_queue(cfqq);
 queue_fail:
        if (cic)
@@ -2002,7 +2134,7 @@ queue_fail:
         * that would be an extremely rare OOM situation
         */
        cfqd->rq_starved = 1;
-       kblockd_schedule_work(&cfqd->unplug_work);
+       cfq_schedule_dispatch(cfqd);
        spin_unlock_irqrestore(q->queue_lock, flags);
        return 1;
 }
@@ -2068,15 +2200,14 @@ static void cfq_idle_slice_timer(unsigned long data)
                 * not expired and it has a request pending, let it dispatch
                 */
                if (!RB_EMPTY(&cfqq->sort_list)) {
-                       cfqq->must_dispatch = 1;
+                       cfq_mark_cfqq_must_dispatch(cfqq);
                        goto out_kick;
                }
        }
 expire:
        cfq_slice_expired(cfqd, 0);
 out_kick:
-       if (cfq_pending_requests(cfqd))
-               kblockd_schedule_work(&cfqd->unplug_work);
+       cfq_schedule_dispatch(cfqd);
 out_cont:
        spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
 }
@@ -2099,11 +2230,17 @@ static void cfq_idle_class_timer(unsigned long data)
                cfqd->idle_class_timer.expires = end;
                add_timer(&cfqd->idle_class_timer);
        } else
-               kblockd_schedule_work(&cfqd->unplug_work);
+               cfq_schedule_dispatch(cfqd);
 
        spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
 }
 
+static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
+{
+       del_timer_sync(&cfqd->idle_slice_timer);
+       del_timer_sync(&cfqd->idle_class_timer);
+       blk_sync_queue(cfqd->queue);
+}
 
 static void cfq_put_cfqd(struct cfq_data *cfqd)
 {
@@ -2112,7 +2249,7 @@ static void cfq_put_cfqd(struct cfq_data *cfqd)
        if (!atomic_dec_and_test(&cfqd->ref))
                return;
 
-       blk_sync_queue(q);
+       cfq_shutdown_timer_wq(cfqd);
 
        blk_put_queue(q);
 
@@ -2126,8 +2263,7 @@ static void cfq_exit_queue(elevator_t *e)
 {
        struct cfq_data *cfqd = e->elevator_data;
 
-       del_timer_sync(&cfqd->idle_slice_timer);
-       del_timer_sync(&cfqd->idle_class_timer);
+       cfq_shutdown_timer_wq(cfqd);
        cfq_put_cfqd(cfqd);
 }
 
@@ -2198,6 +2334,7 @@ static int cfq_init_queue(request_queue_t *q, elevator_t *e)
        cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
        cfqd->cfq_slice_idle = cfq_slice_idle;
        cfqd->cfq_max_depth = cfq_max_depth;
+
        return 0;
 out_crqpool:
        kfree(cfqd->cfq_hash);
@@ -2369,6 +2506,7 @@ static struct cfq_fs_entry cfq_max_depth_entry = {
        .show = cfq_max_depth_show,
        .store = cfq_max_depth_store,
 };
+
 static struct attribute *default_attrs[] = {
        &cfq_quantum_entry.attr,
        &cfq_queued_entry.attr,
index 7811300..8a453a0 100644 (file)
@@ -12,6 +12,7 @@
 
 #define IOPRIO_PRIO_CLASS(mask)        ((mask) >> IOPRIO_CLASS_SHIFT)
 #define IOPRIO_PRIO_DATA(mask) ((mask) & IOPRIO_PRIO_MASK)
+#define IOPRIO_PRIO_VALUE(class, data) (((class) << IOPRIO_CLASS_SHIFT) | data)
 
 #define ioprio_valid(mask)     (IOPRIO_PRIO_CLASS((mask)) != IOPRIO_CLASS_NONE)