cfq-iosched: fairness for sync no-idle queues
authorCorrado Zoccolo <czoccolo@gmail.com>
Mon, 26 Oct 2009 21:45:29 +0000 (22:45 +0100)
committerJens Axboe <jens.axboe@oracle.com>
Wed, 28 Oct 2009 08:23:26 +0000 (09:23 +0100)
Currently no-idle queues in cfq are not serviced fairly:
even if they can only dispatch a small number of requests at a time,
they have to compete with idling queues to be serviced, experiencing
large latencies.

We should notice, instead, that no-idle queues are the ones that would
benefit most from having low latency, in fact they are any of:
* processes with large think times (e.g. interactive ones like file
  managers)
* seeky (e.g. programs faulting in their code at startup)
* or marked as no-idle from upper levels, to improve latencies of those
  requests.

This patch improves the fairness and latency for those queues, by:
* separating sync idle, sync no-idle and async queues in separate
  service_trees, for each priority
* service all no-idle queues together
* and idling when the last no-idle queue has been serviced, to
  anticipate for more no-idle work
* the timeslices allotted for idle and no-idle service_trees are
  computed proportionally to the number of processes in each set.

Servicing all no-idle queues together should have a performance boost
for NCQ-capable drives, without compromising fairness.

Signed-off-by: Corrado Zoccolo <czoccolo@gmail.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
block/cfq-iosched.c

index 76afa36968943d53089462eb22971008ac88fe61..859f534ae9ef4b1210eccd4f5ac4e004ae7fe370 100644 (file)
@@ -134,7 +134,7 @@ struct cfq_queue {
 };
 
 /*
- * Index in the service_trees.
+ * First index in the service_trees.
  * IDLE is handled separately, so it has negative index
  */
 enum wl_prio_t {
@@ -143,6 +143,16 @@ enum wl_prio_t {
        RT_WORKLOAD = 1
 };
 
+/*
+ * Second index in the service_trees.
+ */
+enum wl_type_t {
+       ASYNC_WORKLOAD = 0,
+       SYNC_NOIDLE_WORKLOAD = 1,
+       SYNC_WORKLOAD = 2
+};
+
+
 /*
  * Per block device queue structure
  */
@@ -153,12 +163,14 @@ struct cfq_data {
         * rr lists of queues with requests, onle rr for each priority class.
         * Counts are embedded in the cfq_rb_root
         */
-       struct cfq_rb_root service_trees[2];
+       struct cfq_rb_root service_trees[2][3];
        struct cfq_rb_root service_tree_idle;
        /*
         * The priority currently being served
         */
        enum wl_prio_t serving_prio;
+       enum wl_type_t serving_type;
+       unsigned long workload_expires;
 
        /*
         * Each priority tree is sorted by next_request position.  These
@@ -221,12 +233,13 @@ struct cfq_data {
 };
 
 static struct cfq_rb_root *service_tree_for(enum wl_prio_t prio,
+                                           enum wl_type_t type,
                                            struct cfq_data *cfqd)
 {
        if (prio == IDLE_WORKLOAD)
                return &cfqd->service_tree_idle;
 
-       return &cfqd->service_trees[prio];
+       return &cfqd->service_trees[prio][type];
 }
 
 enum cfqq_state_flags {
@@ -282,12 +295,24 @@ static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
        return BE_WORKLOAD;
 }
 
+
+static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
+{
+       if (!cfq_cfqq_sync(cfqq))
+               return ASYNC_WORKLOAD;
+       if (!cfq_cfqq_idle_window(cfqq))
+               return SYNC_NOIDLE_WORKLOAD;
+       return SYNC_WORKLOAD;
+}
+
 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
 {
        if (wl == IDLE_WORKLOAD)
                return cfqd->service_tree_idle.count;
 
-       return cfqd->service_trees[wl].count;
+       return cfqd->service_trees[wl][ASYNC_WORKLOAD].count
+               + cfqd->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
+               + cfqd->service_trees[wl][SYNC_WORKLOAD].count;
 }
 
 static void cfq_dispatch_insert(struct request_queue *, struct request *);
@@ -597,7 +622,7 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        struct cfq_rb_root *service_tree;
        int left;
 
-       service_tree = service_tree_for(cfqq_prio(cfqq), cfqd);
+       service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd);
        if (cfq_class_idle(cfqq)) {
                rb_key = CFQ_IDLE_DELAY;
                parent = rb_last(&service_tree->rb);
@@ -1030,7 +1055,7 @@ static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 {
        struct cfq_rb_root *service_tree =
-               service_tree_for(cfqd->serving_prio, cfqd);
+               service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd);
 
        if (RB_EMPTY_ROOT(&service_tree->rb))
                return NULL;
@@ -1167,7 +1192,7 @@ static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
        enum wl_prio_t prio = cfqq_prio(cfqq);
-       struct cfq_rb_root *service_tree;
+       struct cfq_rb_root *service_tree = cfqq->service_tree;
 
        /* We never do for idle class queues. */
        if (prio == IDLE_WORKLOAD)
@@ -1181,7 +1206,9 @@ static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
         * Otherwise, we do only if they are the last ones
         * in their service tree.
         */
-       service_tree = service_tree_for(prio, cfqd);
+       if (!service_tree)
+               service_tree = service_tree_for(prio, cfqq_type(cfqq), cfqd);
+
        if (service_tree->count == 0)
                return true;
 
@@ -1235,14 +1262,20 @@ static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 
        cfq_mark_cfqq_wait_request(cfqq);
 
-       /*
-        * we don't want to idle for seeks, but we do want to allow
-        * fair distribution of slice time for a process doing back-to-back
-        * seeks. so allow a little bit of time for him to submit a new rq
-        */
        sl = cfqd->cfq_slice_idle;
-       if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
+       /* are we servicing noidle tree, and there are more queues?
+        * non-rotational or NCQ: no idle
+        * non-NCQ rotational : very small idle, to allow
+        *     fair distribution of slice time for a process doing back-to-back
+        *     seeks.
+        */
+       if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
+           service_tree_for(cfqd->serving_prio, SYNC_NOIDLE_WORKLOAD, cfqd)
+               ->count > 0) {
+               if (blk_queue_nonrot(cfqd->queue) || cfqd->hw_tag)
+                       return;
                sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
+       }
 
        mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
        cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
@@ -1346,6 +1379,106 @@ static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
        }
 }
 
+static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, enum wl_prio_t prio,
+                                   bool prio_changed)
+{
+       struct cfq_queue *queue;
+       int i;
+       bool key_valid = false;
+       unsigned long lowest_key = 0;
+       enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
+
+       if (prio_changed) {
+               /*
+                * When priorities switched, we prefer starting
+                * from SYNC_NOIDLE (first choice), or just SYNC
+                * over ASYNC
+                */
+               if (service_tree_for(prio, cur_best, cfqd)->count)
+                       return cur_best;
+               cur_best = SYNC_WORKLOAD;
+               if (service_tree_for(prio, cur_best, cfqd)->count)
+                       return cur_best;
+
+               return ASYNC_WORKLOAD;
+       }
+
+       for (i = 0; i < 3; ++i) {
+               /* otherwise, select the one with lowest rb_key */
+               queue = cfq_rb_first(service_tree_for(prio, i, cfqd));
+               if (queue &&
+                   (!key_valid || time_before(queue->rb_key, lowest_key))) {
+                       lowest_key = queue->rb_key;
+                       cur_best = i;
+                       key_valid = true;
+               }
+       }
+
+       return cur_best;
+}
+
+static void choose_service_tree(struct cfq_data *cfqd)
+{
+       enum wl_prio_t previous_prio = cfqd->serving_prio;
+       bool prio_changed;
+       unsigned slice;
+       unsigned count;
+
+       /* Choose next priority. RT > BE > IDLE */
+       if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
+               cfqd->serving_prio = RT_WORKLOAD;
+       else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
+               cfqd->serving_prio = BE_WORKLOAD;
+       else {
+               cfqd->serving_prio = IDLE_WORKLOAD;
+               cfqd->workload_expires = jiffies + 1;
+               return;
+       }
+
+       /*
+        * For RT and BE, we have to choose also the type
+        * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
+        * expiration time
+        */
+       prio_changed = (cfqd->serving_prio != previous_prio);
+       count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
+               ->count;
+
+       /*
+        * If priority didn't change, check workload expiration,
+        * and that we still have other queues ready
+        */
+       if (!prio_changed && count &&
+           !time_after(jiffies, cfqd->workload_expires))
+               return;
+
+       /* otherwise select new workload type */
+       cfqd->serving_type =
+               cfq_choose_wl(cfqd, cfqd->serving_prio, prio_changed);
+       count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
+               ->count;
+
+       /*
+        * the workload slice is computed as a fraction of target latency
+        * proportional to the number of queues in that workload, over
+        * all the queues in the same priority class
+        */
+       slice = cfq_target_latency * count /
+               max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio],
+                     cfq_busy_queues_wl(cfqd->serving_prio, cfqd));
+
+       if (cfqd->serving_type == ASYNC_WORKLOAD)
+               /* async workload slice is scaled down according to
+                * the sync/async slice ratio. */
+               slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
+       else
+               /* sync workload slice is at least 2 * cfq_slice_idle */
+               slice = max(slice, 2 * cfqd->cfq_slice_idle);
+
+       slice = max_t(unsigned, slice, CFQ_MIN_TT);
+       cfqd->workload_expires = jiffies + slice;
+}
+
 /*
  * Select a queue for service. If we have a current active queue,
  * check whether to continue servicing it, or retrieve and set a new one.
@@ -1398,14 +1531,13 @@ static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 expire:
        cfq_slice_expired(cfqd, 0);
 new_queue:
-       if (!new_cfqq) {
-               if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
-                       cfqd->serving_prio = RT_WORKLOAD;
-               else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
-                       cfqd->serving_prio = BE_WORKLOAD;
-               else
-                       cfqd->serving_prio = IDLE_WORKLOAD;
-       }
+       /*
+        * Current queue expired. Check if we have to switch to a new
+        * service tree
+        */
+       if (!new_cfqq)
+               choose_service_tree(cfqd);
+
        cfqq = cfq_set_active_queue(cfqd, new_cfqq);
 keep_queue:
        return cfqq;
@@ -1432,10 +1564,12 @@ static int cfq_forced_dispatch(struct cfq_data *cfqd)
 {
        struct cfq_queue *cfqq;
        int dispatched = 0;
-       int i;
+       int i, j;
        for (i = 0; i < 2; ++i)
-               while ((cfqq = cfq_rb_first(&cfqd->service_trees[i])) != NULL)
-                       dispatched += __cfq_forced_dispatch_cfqq(cfqq);
+               for (j = 0; j < 3; ++j)
+                       while ((cfqq = cfq_rb_first(&cfqd->service_trees[i][j]))
+                               != NULL)
+                               dispatched += __cfq_forced_dispatch_cfqq(cfqq);
 
        while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL)
                dispatched += __cfq_forced_dispatch_cfqq(cfqq);
@@ -2218,13 +2352,10 @@ 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->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
+           (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
                enable_idle = 0;
        else if (sample_valid(cic->ttime_samples)) {
-               unsigned int slice_idle = cfqd->cfq_slice_idle;
-               if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
-                       slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
-               if (cic->ttime_mean > slice_idle)
+               if (cic->ttime_mean > cfqd->cfq_slice_idle)
                        enable_idle = 0;
                else
                        enable_idle = 1;
@@ -2262,6 +2393,10 @@ cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
        if (cfq_class_idle(cfqq))
                return true;
 
+       if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD
+           && new_cfqq->service_tree == cfqq->service_tree)
+               return true;
+
        /*
         * if the new request is sync, but the currently running queue is
         * not, let the sync request have priority.
@@ -2778,14 +2913,15 @@ static void cfq_exit_queue(struct elevator_queue *e)
 static void *cfq_init_queue(struct request_queue *q)
 {
        struct cfq_data *cfqd;
-       int i;
+       int i, j;
 
        cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
        if (!cfqd)
                return NULL;
 
        for (i = 0; i < 2; ++i)
-               cfqd->service_trees[i] = CFQ_RB_ROOT;
+               for (j = 0; j < 3; ++j)
+                       cfqd->service_trees[i][j] = CFQ_RB_ROOT;
        cfqd->service_tree_idle = CFQ_RB_ROOT;
 
        /*