]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
block, bfq: don't change ioprio class for a bfq_queue on a service tree
authorPaolo Valente <paolo.valente@linaro.org>
Mon, 3 Jul 2017 08:00:10 +0000 (10:00 +0200)
committerJens Axboe <axboe@kernel.dk>
Mon, 3 Jul 2017 22:50:00 +0000 (16:50 -0600)
On each deactivation or re-scheduling (after being served) of a
bfq_queue, BFQ invokes the function __bfq_entity_update_weight_prio(),
to perform pending updates of ioprio, weight and ioprio class for the
bfq_queue. BFQ also invokes this function on I/O-request dispatches,
to raise or lower weights more quickly when needed, thereby improving
latency. However, the entity representing the bfq_queue may be on the
active (sub)tree of a service tree when this happens, and, although
with a very low probability, the bfq_queue may happen to also have a
pending change of its ioprio class. If both conditions hold when
__bfq_entity_update_weight_prio() is invoked, then the entity moves to
a sort of hybrid state: the new service tree for the entity, as
returned by bfq_entity_service_tree(), differs from service tree on
which the entity still is. The functions that handle activations and
deactivations of entities do not cope with such a hybrid state (and
would need to become more complex to cope).

This commit addresses this issue by just making
__bfq_entity_update_weight_prio() not perform also a possible pending
change of ioprio class, when invoked on an I/O-request dispatch for a
bfq_queue. Such a change is thus postponed to when
__bfq_entity_update_weight_prio() is invoked on deactivation or
re-scheduling of the bfq_queue.

Reported-by: Marco Piazza <mpiazza@gmail.com>
Reported-by: Laurentiu Nicola <lnicola@dend.ro>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Marco Piazza <mpiazza@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/bfq-iosched.c
block/bfq-iosched.h
block/bfq-wf2q.c

index 12bbc6b8657dd686b918f27a3d5f2d7db133d778..60a6835265fc386a229ec96338eb735c7dbea65b 100644 (file)
@@ -3483,11 +3483,17 @@ static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
                        }
                }
        }
-       /* Update weight both if it must be raised and if it must be lowered */
+       /*
+        * To improve latency (for this or other queues), immediately
+        * update weight both if it must be raised and if it must be
+        * lowered. Since, entity may be on some active tree here, and
+        * might have a pending change of its ioprio class, invoke
+        * next function with the last parameter unset (see the
+        * comments on the function).
+        */
        if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
-               __bfq_entity_update_weight_prio(
-                       bfq_entity_service_tree(entity),
-                       entity);
+               __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
+                                               entity, false);
 }
 
 /*
index 5c3bf986149215b3d98f753548cba9b4880f6e62..8fd83b885774392fc1368b728621a430ffcec519 100644 (file)
@@ -892,7 +892,8 @@ void bfq_put_idle_entity(struct bfq_service_tree *st,
                         struct bfq_entity *entity);
 struct bfq_service_tree *
 __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st,
-                               struct bfq_entity *entity);
+                               struct bfq_entity *entity,
+                               bool update_class_too);
 void bfq_bfqq_served(struct bfq_queue *bfqq, int served);
 void bfq_bfqq_charge_time(struct bfq_data *bfqd, struct bfq_queue *bfqq,
                          unsigned long time_ms);
index 8726ede19eef2c632a8e53d2bc340a7cae832e6e..5ec05cd42b80725cb297129739da2edf1372a70a 100644 (file)
@@ -694,10 +694,28 @@ struct bfq_service_tree *bfq_entity_service_tree(struct bfq_entity *entity)
        return sched_data->service_tree + idx;
 }
 
-
+/*
+ * Update weight and priority of entity. If update_class_too is true,
+ * then update the ioprio_class of entity too.
+ *
+ * The reason why the update of ioprio_class is controlled through the
+ * last parameter is as follows. Changing the ioprio class of an
+ * entity implies changing the destination service trees for that
+ * entity. If such a change occurred when the entity is already on one
+ * of the service trees for its previous class, then the state of the
+ * entity would become more complex: none of the new possible service
+ * trees for the entity, according to bfq_entity_service_tree(), would
+ * match any of the possible service trees on which the entity
+ * is. Complex operations involving these trees, such as entity
+ * activations and deactivations, should take into account this
+ * additional complexity.  To avoid this issue, this function is
+ * invoked with update_class_too unset in the points in the code where
+ * entity may happen to be on some tree.
+ */
 struct bfq_service_tree *
 __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st,
-                               struct bfq_entity *entity)
+                               struct bfq_entity *entity,
+                               bool update_class_too)
 {
        struct bfq_service_tree *new_st = old_st;
 
@@ -739,9 +757,15 @@ __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st,
                                  bfq_weight_to_ioprio(entity->orig_weight);
                }
 
-               if (bfqq)
+               if (bfqq && update_class_too)
                        bfqq->ioprio_class = bfqq->new_ioprio_class;
-               entity->prio_changed = 0;
+
+               /*
+                * Reset prio_changed only if the ioprio_class change
+                * is not pending any longer.
+                */
+               if (!bfqq || bfqq->ioprio_class == bfqq->new_ioprio_class)
+                       entity->prio_changed = 0;
 
                /*
                 * NOTE: here we may be changing the weight too early,
@@ -867,7 +891,12 @@ static void bfq_update_fin_time_enqueue(struct bfq_entity *entity,
 {
        struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
 
-       st = __bfq_entity_update_weight_prio(st, entity);
+       /*
+        * When this function is invoked, entity is not in any service
+        * tree, then it is safe to invoke next function with the last
+        * parameter set (see the comments on the function).
+        */
+       st = __bfq_entity_update_weight_prio(st, entity, true);
        bfq_calc_finish(entity, entity->budget);
 
        /*