cfq-iosched: Introduce cfq_entity for CFQ queue

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------
 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +0 -0

index f3a126e..41cef2e 100644

block/cfq-iosched.c +8 -0

+	if (RB_EMPTY_NODE(&cfqe->rb_node)) {
+		/*
+		 * If this CFQ queue moves to another group, the original
+		 * vdisktime makes no sense any more, reset the vdisktime
+		 * here.
+		 */
+		parent = rb_last(&service_tree->rb);
+		if (parent) {

CFQ group hierarchical scheduling and use_hierarchy interface.

block/cfq-iosched.c +6 -11

 static void
 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
 {
-	struct cfq_rb_root *st = &cfqd->grp_service_tree;
 	struct cfq_entity *cfqe = &cfqg->cfqe;
-	struct cfq_entity *__cfqe;
 	struct rb_node *n;
+	struct cfq_entity *entity;
+	struct cfq_rb_root *st;
+	struct cfq_group *__cfqg;
 
 	cfqg->nr_cfqq++;
+
 	if (!RB_EMPTY_NODE(&cfqe->rb_node))
 		return;
 
 	/*
-	 * Currently put the group at the end. Later implement something
-	 * so that groups get lesser vtime based on their weights, so that
-	 * if group does not loose all if it was not continously backlogged.
+	 * Enqueue this group and its ancestors onto their service tree.
 	 */
-	n = rb_last(&st->rb);
-	if (n) {
-		__cfqe = rb_entry_entity(n);
-		cfqe->vdisktime = __cfqe->vdisktime + CFQ_IDLE_DELAY;
-	} else
-		cfqe->vdisktime = st->min_vdisktime;
+	while (cfqe) {

block/cfq-iosched.c +5 -0

+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+	struct cfq_group *cfqg = NULL;

Quoted 15 lines · Gui Jianfeng

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------
 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +0 -0

index f3a126e..41cef2e 100644

block/cfq-iosched.c static const int cfq_hist_divisor = 4; · line 47

Quoted 9 lines

block/cfq-iosched.c +6 -0

@@ -39,6 +39,13 @@ static const int cfq_hist_divisor = 4;
  */
 #define CFQ_IDLE_DELAY		(HZ / 5)
 
+/* 
+ * The base boosting value.
+ */
+#define CFQ_BOOST_SYNC_BASE          (HZ / 10)
+#define CFQ_BOOST_ASYNC_BASE          (HZ / 25)
+

These are same as cfq_slice_sync and cfq_slice_async. Looking at boost logic, this is equivalent of starting a new queue/group as if it is being requeued after conuming a full slice. So may be we can divide it by some const number say 4 or something like that. This is a minor point though as this algorimthm will kind of evolve and we will learn what works best.

Secondly, I think you wanted to SYNC vs ASYNC logic seem to be reversed. We would like to give ASYNC queues higher boost (Put these farther in tree) and lesser boost to SYNC queues. Looks like above constants will do the reverse?

[..]

Quoted 8 lines

block/cfq-iosched.c +8 -0

+	if (RB_EMPTY_NODE(&cfqe->rb_node)) {
+		/*
+		 * If this CFQ queue moves to another group, the original
+		 * vdisktime makes no sense any more, reset the vdisktime
+		 * here.
+		 */
+		parent = rb_last(&service_tree->rb);
+		if (parent) {

block/cfq-iosched.c static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, · line 1310

Quoted 10 lines

block/cfq-iosched.c +10 -0

+			u64 pos_offset;
+
+			/*
+			 * Estimate the position according to its weight and
+			 * ioprio.
+			 */
+			pos_offset = cfq_get_boost(cfqd, cfqq);
+			/* Debug purpose, should remove. */
+			cfq_log_cfqq(cfqd, cfqq, "pos_offset: %llu",
+				     pos_offset);

You wanted to get rid of above debugging comment?

Thanks
Vivek

Quoted 2 lines · Gui Jianfeng

CFQ group hierarchical scheduling and use_hierarchy interface.

Hi Gui,

I have done a quick high level review. Some minor comments inline.

[..]

Quoted 6 lines

struct cfq_data {

block/cfq-iosched.c +0 -2

 	struct request_queue *queue;
-	/* Root service tree for cfq_groups */
-	struct cfq_rb_root grp_service_tree;
 	struct cfq_group root_group;

block/cfq-iosched.c struct cfq_group { · line 245

Quoted 3 lines

block/cfq-iosched.c +3 -0

+	/* cfq group schedule in flat or hierarchy manner. */
+	bool use_hierarchy;
+

This seems to be redundant now? Nobody is using it?

Quoted 8 lines

block/cfq-iosched.c +0 -0

 	/*
 	 * The priority currently being served
 	 */

@@ -246,6 +251,9 @@ struct cfq_data {
 	unsigned long workload_expires;
 	struct cfq_group *serving_group;
 

block/cfq-iosched.c struct cfq_data { · line 255

Quoted 2 lines

block/cfq-iosched.c +2 -0

+	/* Service tree for cfq group flat scheduling mode. */
+	struct cfq_rb_root grp_service_tree;

Above comment is misleading. This service tree is now used both for flat as well as hierarhical mode.

[..]

Quoted 32 lines

static void
cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) {

block/cfq-iosched.c +8 -11

-	struct cfq_rb_root *st = &cfqd->grp_service_tree;
 	struct cfq_entity *cfqe = &cfqg->cfqe;
-	struct cfq_entity *__cfqe;
 	struct rb_node *n;
+	struct cfq_entity *entity;
+	struct cfq_rb_root *st;
+	struct cfq_group *__cfqg;
 
 	cfqg->nr_cfqq++;
+
 	if (!RB_EMPTY_NODE(&cfqe->rb_node))
 		return;
 
 	/*
-	 * Currently put the group at the end. Later implement something
-	 * so that groups get lesser vtime based on their weights, so that
-	 * if group does not loose all if it was not continously backlogged.
+	 * Enqueue this group and its ancestors onto their service tree.
 	 */
-	n = rb_last(&st->rb);
-	if (n) {
-		__cfqe = rb_entry_entity(n);
-		cfqe->vdisktime = __cfqe->vdisktime + CFQ_IDLE_DELAY;
-	} else
-		cfqe->vdisktime = st->min_vdisktime;
+	while (cfqe) {
+		if (!RB_EMPTY_NODE(&cfqe->rb_node))
+			return;

block/cfq-iosched.c cfq_entity_service_tree_add(struct cfq_rb_root *st, struct cfq_entity *cfqe) · line 971

Quoted 7 lines

block/cfq-iosched.c +6 -1

-	cfq_entity_service_tree_add(st, cfqe);
+		/*
+		 * Currently put the group at the end. Later implement
+		 * something so that groups get lesser vtime based on
+		 * their weights, so that if group does not loose all
+		 * if it was not continously backlogged.
+		 */

Can we use vdisktime boost logic for groups also? I think it can be a separate patch in the series (the last one). Keeping it as a separate patch will also help you to coordinate with chad's patch.

Review

Quoted 2 lines

(quoted diff) +1 -0

+		st = cfqe->service_tree;

Group entity set their service tree when they get allocated and retain this pointer even when they get deleted from serivce tree. Queue entities seem to have it NULL when they get deleted from service tree and it gets set again when queue is getting inserted. It would be nice if we can fix this discrepancy and keep it consistent. I think clearing up cfqe->service_tree is a better idea and then calculate it again for group also.

[..]

block/cfq-iosched.c void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg, · line 1135

Quoted 5 lines

block/cfq-iosched.c +2 -2

-static struct cfq_group *
-cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+static void init_cfqe(struct blkio_cgroup *blkcg,
+				    struct cfq_group *cfqg)

As you are using this function for initializing group entity, possibly rename it to init_group_entity() or init_group_cfqe() etc.

[..]

Quoted 9 lines

block/cfq-iosched.c +9 -0

+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+	struct cfq_group *cfqg = NULL;
+	void *key = cfqd;
+	struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+	unsigned int major, minor;
+

block/cfq-iosched.c cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) · line 1348

Quoted 6 lines

block/cfq-iosched.c +6 -0

+	cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+	if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
+		sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+		cfqg->blkg.dev = MKDEV(major, minor);
+		goto done;
+	}

Should we make this updation of this info hierarhical?

Thanks
Vivek

Quoted 2 lines · Gui Jianfeng

CFQ group hierarchical scheduling and use_hierarchy interface.

Hi Gui,

I have done a quick high level review. Some minor comments inline.

[..]

block/cfq-iosched.c struct cfq_group { · line 245

Quoted 9 lines

struct cfq_data {

block/cfq-iosched.c +3 -2

 	struct request_queue *queue;
-	/* Root service tree for cfq_groups */
-	struct cfq_rb_root grp_service_tree;
 	struct cfq_group root_group;
 
+	/* cfq group schedule in flat or hierarchy manner. */
+	bool use_hierarchy;
+

This seems to be redundant now? Nobody is using it?

block/cfq-iosched.c struct cfq_data { · line 255

Quoted 9 lines

block/cfq-iosched.c +2 -0

 	/*
 	 * The priority currently being served
 	 */

@@ -246,6 +251,9 @@ struct cfq_data {
 	unsigned long workload_expires;
 	struct cfq_group *serving_group;
 
+	/* Service tree for cfq group flat scheduling mode. */
+	struct cfq_rb_root grp_service_tree;

Above comment is misleading. This service tree is now used both for flat as well as hierarhical mode.

[..]

Quoted 29 lines

block/cfq-iosched.c +6 -11

 static void
 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
 {
-	struct cfq_rb_root *st = &cfqd->grp_service_tree;
 	struct cfq_entity *cfqe = &cfqg->cfqe;
-	struct cfq_entity *__cfqe;
 	struct rb_node *n;
+	struct cfq_entity *entity;
+	struct cfq_rb_root *st;
+	struct cfq_group *__cfqg;
 
 	cfqg->nr_cfqq++;
+
 	if (!RB_EMPTY_NODE(&cfqe->rb_node))
 		return;
 
 	/*
-	 * Currently put the group at the end. Later implement something
-	 * so that groups get lesser vtime based on their weights, so that
-	 * if group does not loose all if it was not continously backlogged.
+	 * Enqueue this group and its ancestors onto their service tree.
 	 */
-	n = rb_last(&st->rb);
-	if (n) {
-		__cfqe = rb_entry_entity(n);
-		cfqe->vdisktime = __cfqe->vdisktime + CFQ_IDLE_DELAY;
-	} else
-		cfqe->vdisktime = st->min_vdisktime;
+	while (cfqe) {

block/cfq-iosched.c cfq_entity_service_tree_add(struct cfq_rb_root *st, struct cfq_entity *cfqe) · line 971

Quoted 10 lines

block/cfq-iosched.c +8 -1

+		if (!RB_EMPTY_NODE(&cfqe->rb_node))
+			return;
 
-	cfq_entity_service_tree_add(st, cfqe);
+		/*
+		 * Currently put the group at the end. Later implement
+		 * something so that groups get lesser vtime based on
+		 * their weights, so that if group does not loose all
+		 * if it was not continously backlogged.
+		 */

Can we use vdisktime boost logic for groups also? I think it can be a separate patch in the series (the last one). Keeping it as a separate patch will also help you to coordinate with chad's patch.

Review

Quoted 1 line

(quoted diff) +1 -0

+		st = cfqe->service_tree;

Group entity set their service tree when they get allocated and retain this pointer even when they get deleted from serivce tree. Queue entities seem to have it NULL when they get deleted from service tree and it gets set again when queue is getting inserted. It would be nice if we can fix this discrepancy and keep it consistent. I think clearing up cfqe->service_tree is a better idea and then calculate it again for group also.

Quoted 33 lines · Vivek Goyal

Quoted 15 lines · Gui Jianfeng

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------
 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +0 -0

index f3a126e..41cef2e 100644

block/cfq-iosched.c static const int cfq_hist_divisor = 4; · line 47

Quoted 9 lines

block/cfq-iosched.c +6 -0

@@ -39,6 +39,13 @@ static const int cfq_hist_divisor = 4;
  */
 #define CFQ_IDLE_DELAY		(HZ / 5)
 
+/* 
+ * The base boosting value.
+ */
+#define CFQ_BOOST_SYNC_BASE          (HZ / 10)
+#define CFQ_BOOST_ASYNC_BASE          (HZ / 25)
+

These are same as cfq_slice_sync and cfq_slice_async. Looking at boost logic, this is equivalent of starting a new queue/group as if it is being requeued after conuming a full slice. So may be we can divide it by some const number say 4 or something like that. This is a minor point though as this algorimthm will kind of evolve and we will learn what works best.

Quoted 4 lines

Secondly, I think you wanted to SYNC vs ASYNC logic seem to be reversed. We would like to give ASYNC queues higher boost (Put these farther in tree) and lesser boost to SYNC queues. Looks like above constants will do the reverse?

Hi Vivek,

Currently, SYNC and ASYNC queues are in different service tree, they don't impact each other. Here, I Really want use this logic.

Quoted 9 lines

Quoted 9 lines

block/cfq-iosched.c +9 -0

+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+	struct cfq_group *cfqg = NULL;
+	void *key = cfqd;
+	struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+	unsigned int major, minor;
+

Quoted 8 lines

block/cfq-iosched.c cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) · line 1348

Quoted 6 lines

block/cfq-iosched.c +6 -0

+	cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+	if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
+		sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+		cfqg->blkg.dev = MKDEV(major, minor);
+		goto done;
+	}

Should we make this updation of this info hierarhical?

IMHO, it's fine to defer the updation when we really get the cfqg.

Quoted 79 lines · Vivek Goyal

Quoted 2 lines · Gui Jianfeng

CFQ group hierarchical scheduling and use_hierarchy interface.

Hi Gui,

I have done a quick high level review. Some minor comments inline.

[..]

block/cfq-iosched.c struct cfq_group { · line 245

Quoted 9 lines

struct cfq_data {

block/cfq-iosched.c +3 -2

 	struct request_queue *queue;
-	/* Root service tree for cfq_groups */
-	struct cfq_rb_root grp_service_tree;
 	struct cfq_group root_group;
 
+	/* cfq group schedule in flat or hierarchy manner. */
+	bool use_hierarchy;
+

This seems to be redundant now? Nobody is using it?

block/cfq-iosched.c struct cfq_data { · line 255

Quoted 9 lines

block/cfq-iosched.c +2 -0

 	/*
 	 * The priority currently being served
 	 */

@@ -246,6 +251,9 @@ struct cfq_data {
 	unsigned long workload_expires;
 	struct cfq_group *serving_group;
 
+	/* Service tree for cfq group flat scheduling mode. */
+	struct cfq_rb_root grp_service_tree;

Above comment is misleading. This service tree is now used both for flat as well as hierarhical mode.

[..]

Quoted 29 lines

block/cfq-iosched.c +6 -11

 static void
 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
 {
-	struct cfq_rb_root *st = &cfqd->grp_service_tree;
 	struct cfq_entity *cfqe = &cfqg->cfqe;
-	struct cfq_entity *__cfqe;
 	struct rb_node *n;
+	struct cfq_entity *entity;
+	struct cfq_rb_root *st;
+	struct cfq_group *__cfqg;
 
 	cfqg->nr_cfqq++;
+
 	if (!RB_EMPTY_NODE(&cfqe->rb_node))
 		return;
 
 	/*
-	 * Currently put the group at the end. Later implement something
-	 * so that groups get lesser vtime based on their weights, so that
-	 * if group does not loose all if it was not continously backlogged.
+	 * Enqueue this group and its ancestors onto their service tree.
 	 */
-	n = rb_last(&st->rb);
-	if (n) {
-		__cfqe = rb_entry_entity(n);
-		cfqe->vdisktime = __cfqe->vdisktime + CFQ_IDLE_DELAY;
-	} else
-		cfqe->vdisktime = st->min_vdisktime;
+	while (cfqe) {

block/cfq-iosched.c cfq_entity_service_tree_add(struct cfq_rb_root *st, struct cfq_entity *cfqe) · line 971

Quoted 10 lines

block/cfq-iosched.c +8 -1

+		if (!RB_EMPTY_NODE(&cfqe->rb_node))
+			return;
 
-	cfq_entity_service_tree_add(st, cfqe);
+		/*
+		 * Currently put the group at the end. Later implement
+		 * something so that groups get lesser vtime based on
+		 * their weights, so that if group does not loose all
+		 * if it was not continously backlogged.
+		 */

Can we use vdisktime boost logic for groups also? I think it can be a separate patch in the series (the last one). Keeping it as a separate patch will also help you to coordinate with chad's patch.

Quoted 9 lines

Review

Quoted 1 line

(quoted diff) +1 -0

+		st = cfqe->service_tree;

Group entity set their service tree when they get allocated and retain this pointer even when they get deleted from serivce tree. Queue entities seem to have it NULL when they get deleted from service tree and it gets set again when queue is getting inserted. It would be nice if we can fix this discrepancy and keep it consistent. I think clearing up cfqe->service_tree is a better idea and then calculate it again for group also.

Vivek,

Currently, cfq queue might change workload type and io class, so we need to recalculate its service_tree. But for cfq groups, IMHO we don't need to add this complexity for the time being.
I think we can add this change as soon as different io classes or workload types are introduced. How do you think?

block/cfq-iosched.c +14 -2

+/*
+ * The time when a CFQ queue is put onto a service tree is recoreded in
+ * cfqq->reposition_time. Currently, we check the first priority CFQ queues
+ * on each service tree, and select the workload type that contains the lowest
+ * reposition_time CFQ queue among them.
+ */
 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
 				struct cfq_group *cfqg, enum wl_prio_t prio)
 {
 	struct cfq_entity *cfqe;
+	struct cfq_queue *cfqq;
+	unsigned long lowest_start_time;
 	int i;
-	bool key_valid = false;
-	unsigned long lowest_key = 0;
+	bool time_valid = false;
 	enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
 
+	/*
+	 * TODO: We may take io priority and io class into account when
+	 * choosing a workload type. But for the time being just make use of
+	 * reposition_time only.
+	 */
 	for (i = 0; i <= SYNC_WORKLOAD; ++i) {

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------

 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +8 -4

index f3a126e..41cef2e 100644
@@ -39,6 +39,13 @@ static const int cfq_hist_divisor = 4;
  */
  #define CFQ_IDLE_DELAY         (HZ / 5)
 
+/*
+ * The base boosting value.
+ */
+#define CFQ_BOOST_SYNC_BASE          (HZ / 10)
+#define CFQ_BOOST_ASYNC_BASE          (HZ / 25)
+
+
  /*
  * below this threshold, we consider thinktime immediate
  */

@@ -99,10 +106,7 @@ struct cfq_entity {
        struct cfq_rb_root *service_tree;
        /* service_tree member */
        struct rb_node rb_node;
-       /* service_tree key, represent the position on the tree */
-       unsigned long rb_key;
-
-       /* group service_tree key */
+       /* service_tree key */
        u64 vdisktime;
        bool is_group_entity;
        unsigned int weight;

block/cfq-iosched.c +18 -0

        /* reference count */
        int ref;
        /* various state flags, see below */

@@ -312,6 +318,24 @@ struct cfq_data {
        struct rcu_head rcu;
  };
 
+/*
+ * Map io priority(7 ~ 0) to io weight(100 ~ 1000) as follows
+ *     prio       0    1     2    3    4    5    6     7
+ *     weight  1000  868   740  612  484  356  228   100
+ */
+static inline unsigned int cfq_prio_to_weight(unsigned short ioprio)
+{
+       unsigned int step;
+
+       BUG_ON(ioprio >= IOPRIO_BE_NR);
+
+       step = (BLKIO_WEIGHT_MAX - BLKIO_WEIGHT_MIN) / (IOPRIO_BE_NR - 1);
+       if (ioprio == 0)
+               return BLKIO_WEIGHT_MAX;
+
+       return BLKIO_WEIGHT_MIN + (IOPRIO_BE_NR - ioprio - 1) * step;
+}
+

 static inline struct cfq_queue *
 cfqq_of_entity(struct cfq_entity *cfqe)
 {

block/cfq-iosched.c +0 -10

@@ -840,16 +864,6 @@ cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
  }
 
-static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
-                                     struct cfq_queue *cfqq)
-{
-       /*
-        * just an approximation, should be ok.
-        */
-       return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
-                      cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
-}
-

 static inline s64
 entity_key(struct cfq_rb_root *st, struct cfq_entity *entity) {

block/cfq-iosched.c +4 -0

@@ -1199,6 +1213,21 @@ static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
 
  #endif /* GROUP_IOSCHED */
 
+static inline u64 cfq_get_boost(struct cfq_data *cfqd,
+                                struct cfq_queue *cfqq)
+{
+       u64 d;

block/cfq-iosched.c +4 -2

+
  /*
  * The cfqd->service_trees holds all pending cfq_queue's that have
  * requests waiting to be processed. It is sorted in the order that

@@ -1210,13 +1239,14 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
        struct cfq_entity *cfqe;
        struct rb_node **p, *parent;
        struct cfq_entity *__cfqe;
-       unsigned long rb_key;
-       struct cfq_rb_root *service_tree;
+       struct cfq_rb_root *service_tree, *orig_st;
        int left;
        int new_cfqq = 1;
        int group_changed = 0;
+       s64 key;
 
        cfqe = &cfqq->cfqe;
+       orig_st = cfqe->service_tree;
 
  #ifdef CONFIG_CFQ_GROUP_IOSCHED

       if (!cfqd->cfq_group_isolation

block/cfq-iosched.c +63 -28

@@ -1224,8 +1254,15 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
            && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
                /* Move this cfq to root group */
                cfq_log_cfqq(cfqd, cfqq, "moving to root group");
-               if (!RB_EMPTY_NODE(&cfqe->rb_node))
+               if (!RB_EMPTY_NODE(&cfqe->rb_node)) {
                        cfq_group_service_tree_del(cfqd, cfqq->cfqg);
+                       /*
+                        * Group changed, dequeue this CFQ queue from the
+                        * original service tree.
+                        */
+                       cfq_rb_erase(&cfqe->rb_node, cfqe->service_tree);
+                       orig_st->total_weight -= cfqe->weight;
+               }
                cfqq->orig_cfqg = cfqq->cfqg;
                cfqq->cfqg = &cfqd->root_group;
                cfqd->root_group.ref++;

@@ -1234,8 +1271,15 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                   && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
                /* cfqq is sequential now needs to go to its original group */
                BUG_ON(cfqq->cfqg != &cfqd->root_group);
-               if (!RB_EMPTY_NODE(&cfqe->rb_node))
+               if (!RB_EMPTY_NODE(&cfqe->rb_node)) {
                        cfq_group_service_tree_del(cfqd, cfqq->cfqg);
+                       /*
+                        * Group changed, dequeue this CFQ queue from the
+                        * original service tree.
+                        */
+                       cfq_rb_erase(&cfqe->rb_node, cfqe->service_tree);
+                       orig_st->total_weight -= cfqe->weight;
+               }
                cfq_put_cfqg(cfqq->cfqg);
                cfqq->cfqg = cfqq->orig_cfqg;
                cfqq->orig_cfqg = NULL;

@@ -1246,47 +1290,68 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 
        service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
                                                cfqq_type(cfqq));
+       if (RB_EMPTY_NODE(&cfqe->rb_node)) {
+               /*
+                * If this CFQ queue moves to another group, the original
+                * vdisktime makes no sense any more, reset the vdisktime
+                * here.
+                */
+               parent = rb_last(&service_tree->rb);
+               if (parent) {
+                       u64 pos_offset;
+
+                       /*
+                        * Estimate the position according to its weight and
+                        * ioprio.
+                        */
+                       pos_offset = cfq_get_boost(cfqd, cfqq);
+                       /* Debug purpose, should remove. */
+                       cfq_log_cfqq(cfqd, cfqq, "pos_offset: %llu",
+                                    pos_offset);
+                       cfqe->vdisktime = service_tree->min_vdisktime +
+                                               pos_offset;
+               } else
+                       cfqe->vdisktime = service_tree->min_vdisktime;
+
+               goto insert;
+       }
+
+       /*
+        * Ok, we get here, this CFQ queue is on the service tree, dequeue it
+        * firstly.
+        */
+       cfq_rb_erase(&cfqe->rb_node, cfqe->service_tree);
+       orig_st->total_weight -= cfqe->weight;
+
+       new_cfqq = 0;
+
        if (cfq_class_idle(cfqq)) {
-               rb_key = CFQ_IDLE_DELAY;
                parent = rb_last(&service_tree->rb);
                if (parent && parent != &cfqe->rb_node) {
                        __cfqe = rb_entry(parent, struct cfq_entity, rb_node);
-                       rb_key += __cfqe->rb_key;
+                       cfqe->vdisktime = __cfqe->vdisktime + CFQ_IDLE_DELAY;
                } else
-                       rb_key += jiffies;
+                       cfqe->vdisktime = service_tree->min_vdisktime;
        } 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.
+                * We charge the CFQ queue by the time this queue runs, and
+                * repsition it on the service tree.
                 */
-               rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
-               rb_key -= cfqq->slice_resid;
+               unsigned int used_sl;
+
+               used_sl = cfq_cfqq_slice_usage(cfqq);
+               cfqe->vdisktime += cfq_scale_slice(used_sl, cfqe);
                cfqq->slice_resid = 0;
        } else {
-               rb_key = -HZ;
-               __cfqe = cfq_rb_first(service_tree);
-               rb_key += __cfqe ? __cfqe->rb_key : jiffies;
-       }
-
-       if (!RB_EMPTY_NODE(&cfqe->rb_node)) {
-               new_cfqq = 0;
-               /*
-                * same position, nothing more to do
-                */
-               if (rb_key == cfqe->rb_key &&
-                   cfqe->service_tree == service_tree)
-                       return;
-
-               cfq_rb_erase(&cfqe->rb_node, cfqe->service_tree);
-               cfqe->service_tree = NULL;
+               cfqe->vdisktime = service_tree->min_vdisktime;
        }
 
+insert:
        left = 1;
        parent = NULL;
        cfqe->service_tree = service_tree;
        p = &service_tree->rb.rb_node;
+       key = entity_key(service_tree, cfqe);
        while (*p) {
                struct rb_node **n;
 

@@ -1296,7 +1361,7 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                /*
                 * sort by key, that represents service time.
                 */
-               if (time_before(rb_key, __cfqe->rb_key))
+               if (key < entity_key(service_tree, __cfqe))
                        n = &(*p)->rb_left;
                else {
                        n = &(*p)->rb_right;

block/cfq-iosched.c +28 -7

+       cfqq->reposition_time = jiffies;
        if ((add_front || !new_cfqq) && !group_changed)
                return;
        cfq_group_service_tree_add(cfqd, cfqq->cfqg);

@@ -1414,14 +1481,18 @@ static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
  static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
  {
        struct cfq_entity *cfqe;
+       struct cfq_rb_root *service_tree;
+
        cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
        BUG_ON(!cfq_cfqq_on_rr(cfqq));
        cfq_clear_cfqq_on_rr(cfqq);
 
        cfqe = &cfqq->cfqe;
+       service_tree = cfqe->service_tree;
 
        if (!RB_EMPTY_NODE(&cfqe->rb_node)) {
                cfq_rb_erase(&cfqe->rb_node, cfqe->service_tree);
+               service_tree->total_weight -= cfqe->weight;
                cfqe->service_tree = NULL;
        }
        if (cfqq->p_root) {

@@ -2120,23 +2191,36 @@ static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
        }
  }
 
+/*
+ * The time when a CFQ queue is put onto a service tree is recoreded in
+ * cfqq->reposition_time. Currently, we check the first priority CFQ queues
+ * on each service tree, and select the workload type that contains the lowest
+ * reposition_time CFQ queue among them.
+ */
  static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
                                struct cfq_group *cfqg, enum wl_prio_t prio)
  {
        struct cfq_entity *cfqe;
+       struct cfq_queue *cfqq;
+       unsigned long lowest_start_time;
        int i;
-       bool key_valid = false;
-       unsigned long lowest_key = 0;
+       bool time_valid = false;
        enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
 
+       /*
+        * TODO: We may take io priority and io class into account when
+        * choosing a workload type. But for the time being just make use of
+        * reposition_time only.
+        */
        for (i = 0; i <= SYNC_WORKLOAD; ++i) {
-               /* select the one with lowest rb_key */
                cfqe = cfq_rb_first(service_tree_for(cfqg, prio, i));
-               if (cfqe &&
-                   (!key_valid || time_before(cfqe->rb_key, lowest_key))) {
-                       lowest_key = cfqe->rb_key;
+               cfqq = cfqq_of_entity(cfqe);
+               if (cfqe && (!time_valid ||
+                            time_before(cfqq->reposition_time,
+                                        lowest_start_time))) {
+                       lowest_start_time = cfqq->reposition_time;
                        cur_best = i;
-                       key_valid = true;
+                       time_valid = true;
                }
        }
 

@@ -2808,10 +2892,13 @@ static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
  {
        struct task_struct *tsk = current;
        int ioprio_class;
+       struct cfq_entity *cfqe;
 
        if (!cfq_cfqq_prio_changed(cfqq))
                return;
 
+       cfqe = &cfqq->cfqe;
+
        ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
        switch (ioprio_class) {

       default:

block/cfq-iosched.c +28 -1

@@ -2838,6 +2925,17 @@ static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
                break;
        }
 
+       if (!RB_EMPTY_NODE(&cfqe->rb_node)) {
+               /*
+                * If this CFQ entity is already on service tree, we need to
+                * adjust service tree's total weight accordingly.
+                */
+               cfqe->service_tree->total_weight -= cfqe->weight;
+               cfqe->weight = cfq_prio_to_weight(cfqq->ioprio);
+               cfqe->service_tree->total_weight += cfqe->weight;
+       } else
+               cfqe->weight = cfq_prio_to_weight(cfqq->ioprio);
+
        /*
         * keep track of original prio settings in case we have to temporarily
         * elevate the priority of this queue

@@ -3572,6 +3670,9 @@ static void cfq_completed_request(struct request_queue *q, struct request *rq)
  */
  static void cfq_prio_boost(struct cfq_queue *cfqq)
  {
+       struct cfq_entity *cfqe;
+
+       cfqe = &cfqq->cfqe;
        if (has_fs_excl()) {
                /*
                 * boost idle prio on transactions that would lock out other

@@ -3588,6 +3689,20 @@ static void cfq_prio_boost(struct cfq_queue *cfqq)
                cfqq->ioprio_class = cfqq->org_ioprio_class;
                cfqq->ioprio = cfqq->org_ioprio;
        }
+
+       /*
+        * update the io weight if io priority gets changed.
+        */
+       if (!RB_EMPTY_NODE(&cfqe->rb_node)) {
+               /*
+                * If this CFQ entity is already on service tree, we need to
+                * adjust service tree's total weight accordingly.
+                */
+               cfqe->service_tree->total_weight -= cfqe->weight;
+               cfqe->weight = cfq_prio_to_weight(cfqq->ioprio);
+               cfqe->service_tree->total_weight += cfqe->weight;
+       } else
+               cfqe->weight = cfq_prio_to_weight(cfqq->ioprio);
  }
 
  static inline int __cfq_may_queue(struct cfq_queue *cfqq)
--

1.7.1

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Quoted 4 lines · Gui Jianfeng

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Hi Gui,
I have a couple of questions inline.

Thanks,
Justin

Quoted 36 lines

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------
 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +8 -4

index f3a126e..41cef2e 100644
@@ -39,6 +39,13 @@ static const int cfq_hist_divisor = 4;
 */
 #define CFQ_IDLE_DELAY         (HZ / 5)
 
+/*
+ * The base boosting value.
+ */
+#define CFQ_BOOST_SYNC_BASE          (HZ / 10)
+#define CFQ_BOOST_ASYNC_BASE          (HZ / 25)
+
+
 /*
 * below this threshold, we consider thinktime immediate
 */

@@ -99,10 +106,7 @@ struct cfq_entity {
       struct cfq_rb_root *service_tree;
       /* service_tree member */
       struct rb_node rb_node;
-       /* service_tree key, represent the position on the tree */
-       unsigned long rb_key;
-
-       /* group service_tree key */
+       /* service_tree key */
       u64 vdisktime;
       bool is_group_entity;
       unsigned int weight;

@@ -114,6 +118,8 @@ struct cfq_entity {

block/cfq-iosched.c struct cfq_entity { · line 122

Quoted 5 lines

block/cfq-iosched.c +2 -0

 struct cfq_queue {
       /* The schedule entity */
       struct cfq_entity cfqe;
+       /* Reposition time */
+       unsigned long reposition_time;

Can this be addition time or something else instead? This is set, even when we are not repositioning among service trees.

Quoted 54 lines

block/cfq-iosched.c +22 -10

       /* reference count */
       int ref;
       /* various state flags, see below */

@@ -312,6 +318,24 @@ struct cfq_data {
       struct rcu_head rcu;
 };
 
+/*
+ * Map io priority(7 ~ 0) to io weight(100 ~ 1000) as follows
+ *     prio       0    1     2    3    4    5    6     7
+ *     weight  1000  868   740  612  484  356  228   100
+ */
+static inline unsigned int cfq_prio_to_weight(unsigned short ioprio)
+{
+       unsigned int step;
+
+       BUG_ON(ioprio >= IOPRIO_BE_NR);
+
+       step = (BLKIO_WEIGHT_MAX - BLKIO_WEIGHT_MIN) / (IOPRIO_BE_NR - 1);
+       if (ioprio == 0)
+               return BLKIO_WEIGHT_MAX;
+
+       return BLKIO_WEIGHT_MIN + (IOPRIO_BE_NR - ioprio - 1) * step;
+}
+
 static inline struct cfq_queue *
 cfqq_of_entity(struct cfq_entity *cfqe)
 {

@@ -840,16 +864,6 @@ cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
       return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
 }
 
-static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
-                                     struct cfq_queue *cfqq)
-{
-       /*
-        * just an approximation, should be ok.
-        */
-       return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
-                      cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
-}
-
 static inline s64
 entity_key(struct cfq_rb_root *st, struct cfq_entity *entity)
 {

@@ -1199,6 +1213,21 @@ static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
 
 #endif /* GROUP_IOSCHED */
 
+static inline u64 cfq_get_boost(struct cfq_data *cfqd,
+                                struct cfq_queue *cfqq)
+{
+       u64 d;

block/cfq-iosched.c static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} · line 1229

Quoted 10 lines

block/cfq-iosched.c +10 -0

+
+       if (cfq_cfqq_sync(cfqq))
+               d = CFQ_BOOST_SYNC_BASE << CFQ_SERVICE_SHIFT;
+       else
+               d = CFQ_BOOST_ASYNC_BASE << CFQ_SERVICE_SHIFT;
+
+       d = d * BLKIO_WEIGHT_DEFAULT;
+       do_div(d, cfqq->cfqe.weight);
+       return d;
+}

The logic for cfq_get_boost() looks a lot like cfq_scale_slice(). Instead of duplicating code, can't it just be
u64 d;
if (cfq_cfqq_sync(cfqq))

        return cfq_scale_slice(CFQ_BOOST_SYNC_BASE, cfqq->cfqe);
else
        return cfq_scale_slice(CFQ_BOOST_ASYNC_BASE, cfqq->cfqe);

Quoted 4 lines

block/cfq-iosched.c +1 -0

+
 /*
 * The cfqd->service_trees holds all pending cfq_queue's that have

I'm confused by this line. Why are we doing some adjustment for cfqe weight that we weren't doing previously? I think cfq_group_service_tree_add below will still do the total_weight adjustment.

Quoted 8 lines

block/cfq-iosched.c +1 -0

+       cfqq->reposition_time = jiffies;
       if ((add_front || !new_cfqq) && !group_changed)
               return;
       cfq_group_service_tree_add(cfqd, cfqq->cfqg);

@@ -1414,14 +1481,18 @@ static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {

On Thu, Feb 10, 2011 at 03:47:16PM +0800, Gui Jianfeng wrote:

[..]

Quoted 24 lines

block/cfq-iosched.c +14 -2

+/*
+ * The time when a CFQ queue is put onto a service tree is recoreded in
+ * cfqq->reposition_time. Currently, we check the first priority CFQ queues
+ * on each service tree, and select the workload type that contains the lowest
+ * reposition_time CFQ queue among them.
+ */
 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
 				struct cfq_group *cfqg, enum wl_prio_t prio)
 {
 	struct cfq_entity *cfqe;
+	struct cfq_queue *cfqq;
+	unsigned long lowest_start_time;
 	int i;
-	bool key_valid = false;
-	unsigned long lowest_key = 0;
+	bool time_valid = false;
 	enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
 
+	/*
+	 * TODO: We may take io priority and io class into account when
+	 * choosing a workload type. But for the time being just make use of
+	 * reposition_time only.
+	 */
 	for (i = 0; i <= SYNC_WORKLOAD; ++i) {

block/cfq-iosched.c static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) · line 2221

Quoted 10 lines

block/cfq-iosched.c +5 -4

-		/* select the one with lowest rb_key */
 		cfqe = cfq_rb_first(service_tree_for(cfqg, prio, i));
-		if (cfqe &&
-		    (!key_valid || time_before(cfqe->rb_key, lowest_key))) {
-			lowest_key = cfqe->rb_key;
+		cfqq = cfqq_of_entity(cfqe);
+		if (cfqe && (!time_valid ||
+			     time_before(cfqq->reposition_time,
+					 lowest_start_time))) {
+			lowest_start_time = cfqq->reposition_time;

Gui,

Have you had a chance to run some mixed workloads in a group (some sync, some async and some sync-idle queues), and see how latency and throughput of sync-idle workload changes due to this "resposition_time" logic. I just want to make sure that latency of sync-noidle workload does not go up as that's the workload that people care and gets noticed first.

Thanks
Vivek

Quoted 33 lines · Gui Jianfeng

Quoted 32 lines · Vivek Goyal

Quoted 15 lines · Gui Jianfeng

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------
 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +0 -0

index f3a126e..41cef2e 100644

block/cfq-iosched.c static const int cfq_hist_divisor = 4; · line 47

Quoted 9 lines

block/cfq-iosched.c +6 -0

@@ -39,6 +39,13 @@ static const int cfq_hist_divisor = 4;
  */
 #define CFQ_IDLE_DELAY		(HZ / 5)
 
+/* 
+ * The base boosting value.
+ */
+#define CFQ_BOOST_SYNC_BASE          (HZ / 10)
+#define CFQ_BOOST_ASYNC_BASE          (HZ / 25)
+

These are same as cfq_slice_sync and cfq_slice_async. Looking at boost logic, this is equivalent of starting a new queue/group as if it is being requeued after conuming a full slice. So may be we can divide it by some const number say 4 or something like that. This is a minor point though as this algorimthm will kind of evolve and we will learn what works best.

Quoted 8 lines

Quoted 4 lines

Secondly, I think you wanted to SYNC vs ASYNC logic seem to be reversed. We would like to give ASYNC queues higher boost (Put these farther in tree) and lesser boost to SYNC queues. Looks like above constants will do the reverse?

Hi Vivek,

Currently, SYNC and ASYNC queues are in different service tree, they don't impact each other. Here, I Really want use this logic.

Ok, SYNC and ASYNC are on separate service tree so their vtime are not comparable (as of today, down the line one might want to look at those for better workload selection logic).

Anyway, because two are on seprate tree so why should we have separate boosting constants for them? How does it help?

Thanks
Vivek
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

On Sat, Feb 12, 2011 at 10:21:47AM +0800, Gui Jianfeng wrote:
[..]

Quoted 16 lines

Quoted 16 lines

Quoted 5 lines

block/cfq-iosched.c +5 -0

+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+	struct cfq_group *cfqg = NULL;

block/cfq-iosched.c cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) · line 1348

Quoted 10 lines

block/cfq-iosched.c +10 -0

+	void *key = cfqd;
+	struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+	unsigned int major, minor;
+
+	cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+	if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
+		sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+		cfqg->blkg.dev = MKDEV(major, minor);
+		goto done;
+	}

Should we make this updation of this info hierarhical?

Quoted 1 line

IMHO, it's fine to defer the updation when we really get the cfqg.

But if cfqg is alrady present, we will never hit the allocation path again. So if somebody creates 2-3 level deep hierarchy and does IO only in the children cgroup, parent cgroups will potentially not get device info updated hence no visible stats?

Thanks
Vivek
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quoted 47 lines · Vivek Goyal

Quoted 33 lines · Gui Jianfeng

Quoted 32 lines · Vivek Goyal

Quoted 15 lines · Gui Jianfeng

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------
 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +0 -0

index f3a126e..41cef2e 100644

block/cfq-iosched.c static const int cfq_hist_divisor = 4; · line 47

Quoted 9 lines

block/cfq-iosched.c +6 -0

@@ -39,6 +39,13 @@ static const int cfq_hist_divisor = 4;
  */
 #define CFQ_IDLE_DELAY		(HZ / 5)
 
+/* 
+ * The base boosting value.
+ */
+#define CFQ_BOOST_SYNC_BASE          (HZ / 10)
+#define CFQ_BOOST_ASYNC_BASE          (HZ / 25)
+

These are same as cfq_slice_sync and cfq_slice_async. Looking at boost logic, this is equivalent of starting a new queue/group as if it is being requeued after conuming a full slice. So may be we can divide it by some const number say 4 or something like that. This is a minor point though as this algorimthm will kind of evolve and we will learn what works best.

Quoted 8 lines

Quoted 4 lines

Secondly, I think you wanted to SYNC vs ASYNC logic seem to be reversed. We would like to give ASYNC queues higher boost (Put these farther in tree) and lesser boost to SYNC queues. Looks like above constants will do the reverse?

Hi Vivek,

Currently, SYNC and ASYNC queues are in different service tree, they don't impact each other. Here, I Really want use this logic.

Ok, SYNC and ASYNC are on separate service tree so their vtime are not comparable (as of today, down the line one might want to look at those for better workload selection logic).

Quoted 2 lines

Anyway, because two are on seprate tree so why should we have separate boosting constants for them? How does it help?

Here if we are using CFQ_BOOST_SYNC_BASE for both, I think it might boost too much for an ASYNC cfqe as compare to others on the same service tree(async). So I make charging and boosting follow the same base.

Quoted 20 lines · Vivek Goyal

On Sat, Feb 12, 2011 at 10:21:47AM +0800, Gui Jianfeng wrote:
[..]

Quoted 17 lines

Quoted 8 lines

Quoted 8 lines

block/cfq-iosched.c +8 -0

+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+	struct cfq_group *cfqg = NULL;
+	void *key = cfqd;
+	struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+	unsigned int major, minor;

Quoted 8 lines

block/cfq-iosched.c cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) · line 1348

Quoted 7 lines

block/cfq-iosched.c +7 -0

+
+	cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+	if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
+		sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+		cfqg->blkg.dev = MKDEV(major, minor);
+		goto done;
+	}

Should we make this updation of this info hierarhical?

IMHO, it's fine to defer the updation when we really get the cfqg.

Quoted 4 lines

But if cfqg is alrady present, we will never hit the allocation path again. So if somebody creates 2-3 level deep hierarchy and does IO only in the children cgroup, parent cgroups will potentially not get device info updated hence no visible stats?

Ahh, I see your concern. But do we really need to show the stats even if a cgroup doesn't issue any IO on a given device?

Quoted 15 lines

Quoted 14 lines

block/cfq-iosched.c +14 -0

+static inline u64 cfq_get_boost(struct cfq_data *cfqd,
+                                struct cfq_queue *cfqq)
+{
+       u64 d;
+
+       if (cfq_cfqq_sync(cfqq))
+               d = CFQ_BOOST_SYNC_BASE << CFQ_SERVICE_SHIFT;
+       else
+               d = CFQ_BOOST_ASYNC_BASE << CFQ_SERVICE_SHIFT;
+
+       d = d * BLKIO_WEIGHT_DEFAULT;
+       do_div(d, cfqq->cfqe.weight);
+       return d;
+}

Quoted 8 lines

The logic for cfq_get_boost() looks a lot like cfq_scale_slice(). Instead of duplicating code, can't it just be
u64 d;
if (cfq_cfqq_sync(cfqq))

        return cfq_scale_slice(CFQ_BOOST_SYNC_BASE, cfqq->cfqe);
else
        return cfq_scale_slice(CFQ_BOOST_ASYNC_BASE, cfqq->cfqe);

Quoted 49 lines · Gui Jianfeng

Quoted 48 lines · Vivek Goyal

Quoted 33 lines · Gui Jianfeng

Quoted 32 lines · Vivek Goyal

Quoted 15 lines · Gui Jianfeng

Introduce vdisktime and io weight for CFQ queue scheduling. Currently, io priority maps to a range [100,1000]. It also gets rid of cfq_slice_offset() logic and makes use the same scheduling algorithm as CFQ group does. This helps for CFQ queue and group scheduling on the same service tree.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/cfq-iosched.c |  219 +++++++++++++++++++++++++++++++++++++++------------
 1 files changed, 167 insertions(+), 52 deletions(-)

block/cfq-iosched.c +0 -0

index f3a126e..41cef2e 100644

block/cfq-iosched.c static const int cfq_hist_divisor = 4; · line 47

Quoted 9 lines

block/cfq-iosched.c +6 -0

@@ -39,6 +39,13 @@ static const int cfq_hist_divisor = 4;
  */
 #define CFQ_IDLE_DELAY		(HZ / 5)
 
+/* 
+ * The base boosting value.
+ */
+#define CFQ_BOOST_SYNC_BASE          (HZ / 10)
+#define CFQ_BOOST_ASYNC_BASE          (HZ / 25)
+

These are same as cfq_slice_sync and cfq_slice_async. Looking at boost logic, this is equivalent of starting a new queue/group as if it is being requeued after conuming a full slice. So may be we can divide it by some const number say 4 or something like that. This is a minor point though as this algorimthm will kind of evolve and we will learn what works best.

Quoted 8 lines

Quoted 4 lines

Secondly, I think you wanted to SYNC vs ASYNC logic seem to be reversed. We would like to give ASYNC queues higher boost (Put these farther in tree) and lesser boost to SYNC queues. Looks like above constants will do the reverse?

Hi Vivek,

Currently, SYNC and ASYNC queues are in different service tree, they don't impact each other. Here, I Really want use this logic.

Ok, SYNC and ASYNC are on separate service tree so their vtime are not comparable (as of today, down the line one might want to look at those for better workload selection logic).

Anyway, because two are on seprate tree so why should we have separate boosting constants for them? How does it help?

Quoted 3 lines

Here if we are using CFQ_BOOST_SYNC_BASE for both, I think it might boost too much for an ASYNC cfqe as compare to others on the same service tree(async). So I make charging and boosting follow the same base.

Ok, that makes sense. So as suggested in other mails, lets use a even smaller base so that freshly backlogged queues get smaller vdisktimes as compared to existing queues which are using disks for longer time.

Thanks
Vivek

Quoted 24 lines · Gui Jianfeng

Quoted 23 lines · Vivek Goyal

On Sat, Feb 12, 2011 at 10:21:47AM +0800, Gui Jianfeng wrote:
[..]

Quoted 16 lines

Quoted 8 lines

Quoted 8 lines

block/cfq-iosched.c +8 -0

+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+	struct cfq_group *cfqg = NULL;
+	void *key = cfqd;
+	struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+	unsigned int major, minor;

Quoted 8 lines

block/cfq-iosched.c cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) · line 1348

Quoted 7 lines

block/cfq-iosched.c +7 -0

+
+	cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+	if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
+		sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+		cfqg->blkg.dev = MKDEV(major, minor);
+		goto done;
+	}

Should we make this updation of this info hierarhical?

Quoted 1 line

IMHO, it's fine to defer the updation when we really get the cfqg.

But if cfqg is alrady present, we will never hit the allocation path again. So if somebody creates 2-3 level deep hierarchy and does IO only in the children cgroup, parent cgroups will potentially not get device info updated hence no visible stats?

Quoted 2 lines

Ahh, I see your concern. But do we really need to show the stats even if a cgroup doesn't issue any IO on a given device?

I am assuming that once use_hierarchy=1, you are aggregating the stats in parent cgroups? So if a child services 5 IOs, these are accounted to parent group also when user_hier=1?

What happens in case of memoy cgroup controller?

Thanks
Vivek

Quoted 32 lines · Vivek Goyal

Quoted 19 lines · Gui Jianfeng

Quoted 18 lines · Vivek Goyal

On Sat, Feb 12, 2011 at 10:21:47AM +0800, Gui Jianfeng wrote:
[..]

Quoted 16 lines

Quoted 16 lines

Quoted 5 lines

(quoted diff) +5 -0

+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+	struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+	struct cfq_group *cfqg = NULL;

Review

Quoted 10 lines

(quoted diff) +10 -0

+	void *key = cfqd;
+	struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+	unsigned int major, minor;
+
+	cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+	if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
+		sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+		cfqg->blkg.dev = MKDEV(major, minor);
+		goto done;
+	}

Should we make this updation of this info hierarhical?

Quoted 7 lines

Quoted 5 lines

Quoted 1 line

IMHO, it's fine to defer the updation when we really get the cfqg.

But if cfqg is alrady present, we will never hit the allocation path again. So if somebody creates 2-3 level deep hierarchy and does IO only in the children cgroup, parent cgroups will potentially not get device info updated hence no visible stats?

Ahh, I see your concern. But do we really need to show the stats even if a cgroup doesn't issue any IO on a given device?

I am assuming that once use_hierarchy=1, you are aggregating the stats in parent cgroups? So if a child services 5 IOs, these are accounted to parent group also when user_hier=1?

Quoted 1 line

What happens in case of memoy cgroup controller?

Hmm, it seems memcg aggregating stats in parent group.
But do we really need to do that in kernel? I think it's easier to do it in userland, and it makes kernel much simpler.

Hmm, it seems memcg aggregating stats in parent group.
But do we really need to do that in kernel? I think it's easier to do it in userland, and it makes kernel much simpler.

Quoted 4 lines · Gui Jianfeng

Hmm, it seems memcg aggregating stats in parent group.
But do we really need to do that in kernel? I think it's easier to do it in userland, and it makes kernel much simpler.

I would prefer having stats aggregated up the hierarchy. One trick we used at Google earlier was to do lazy updates for most stats. So we

would accumulate stats for a given timeslice and at the end of that timeslice propagate those counts all the way up to the parent. For device info like this, the update can be instantaneous since its not a very frequent event. It would also be very useful to distinguish between stats for the cgroup itself vs total values accumulated from the subtree rooted at that cgroup like memcg does.

--
Thanks,
Divyesh

Quoted 7 lines · Divyesh Shah

Quoted 4 lines · Gui Jianfeng

Hmm, it seems memcg aggregating stats in parent group.
But do we really need to do that in kernel? I think it's easier to do it in userland, and it makes kernel much simpler.

Quoted 2 lines

I would prefer having stats aggregated up the hierarchy. One trick we used at Google earlier was to do lazy updates for most stats. So we

Note that this was on ancient version of the blkio controller :). As Vivek mentioned,
it may be ok to add the hierarchical accounting later.

CFQ group hierarchical scheduling and use_hierarchy interface.

Signed-off-by Gui Jianfeng <[email protected]>

---
 block/blk-cgroup.c  |   61 +++++-
 block/blk-cgroup.h  |    3 +
 block/cfq-iosched.c |  603 +++++++++++++++++++++++++++++++++++++--------------

 3 files changed, 500 insertions(+), 167 deletions(-)

block/blk-cgroup.c +11 -1

index 455768a..c55fecd 100644
@@ -25,7 +25,10 @@
  static DEFINE_SPINLOCK(blkio_list_lock);
  static LIST_HEAD(blkio_list);
 
-struct blkio_cgroup blkio_root_cgroup = { .weight = 2*BLKIO_WEIGHT_DEFAULT };
+struct blkio_cgroup blkio_root_cgroup = {
+       .weight = 2*BLKIO_WEIGHT_DEFAULT,
+       .use_hierarchy = 0
+};
  EXPORT_SYMBOL_GPL(blkio_root_cgroup);
 
  static struct cgroup_subsys_state *blkiocg_create(struct cgroup_subsys *,

@@ -454,6 +457,7 @@ static void __blkiocg_del_blkio_group(struct blkio_group *blkg)
        blkg->blkcg_id = 0;
  }
 
+
  /*
  * returns 0 if blkio_group was still on cgroup list. Otherwise returns 1
  * indicating that blk_group was unhashed by the time we got to it.

@@ -765,6 +769,12 @@ unsigned int blkcg_get_weight(struct blkio_cgroup *blkcg,
  }
  EXPORT_SYMBOL_GPL(blkcg_get_weight);
 
+unsigned int blkcg_get_use_hierarchy(struct blkio_cgroup *blkcg)
+{
+       return blkcg->use_hierarchy;
+}
+EXPORT_SYMBOL_GPL(blkcg_get_use_hierarchy);
+
  uint64_t blkcg_get_read_bps(struct blkio_cgroup *blkcg, dev_t dev)
  {
        struct blkio_policy_node *pn;

@@ -1202,6 +1212,8 @@ static u64 blkiocg_file_read_u64 (struct cgroup *cgrp, struct cftype *cft) {
                switch(name) {

               case BLKIO_PROP_weight:
                       return (u64)blkcg->weight;
+               case BLKIO_PROP_use_hierarchy:
+                       return (u64)blkcg->use_hierarchy; }
               break;
       default:

block/blk-cgroup.c +30 -0

@@ -1210,6 +1222,36 @@ static u64 blkiocg_file_read_u64 (struct cgroup *cgrp, struct cftype *cft) {
        return 0;
  }
 
+static int blkio_use_hierarchy_write(struct cgroup *cgrp, u64 val)
+{
+       struct cgroup *parent = cgrp->parent;
+       struct blkio_cgroup *blkcg, *parent_blkcg = NULL;
+       int ret = 0;
+
+       if (val != 0 && val != 1)
+               return -EINVAL;
+
+       blkcg = cgroup_to_blkio_cgroup(cgrp);
+       if (parent)
+               parent_blkcg = cgroup_to_blkio_cgroup(parent);
+
+       cgroup_lock();
+       /*
+        * If parent's use_hierarchy is set, we can't make any modifications
+        * in the child subtrees. If it is unset, then the change can occur,
+        * provided the current cgroup has no children.
+        */
+       if (!parent_blkcg || !parent_blkcg->use_hierarchy) {
+               if (list_empty(&cgrp->children))
+                       blkcg->use_hierarchy = val;
+               else
+                       ret = -EBUSY;
+       } else
+               ret = -EINVAL;
+       cgroup_unlock();
+       return ret;
+}
+

 static int
 blkiocg_file_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val) {

block/blk-cgroup.c +0 -0

@@ -1224,6 +1266,8 @@ blkiocg_file_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
                switch(name) {

               case BLKIO_PROP_weight:
                       return blkio_weight_write(blkcg, val);
+               case BLKIO_PROP_use_hierarchy:
+                       return blkio_use_hierarchy_write(cgrp, val); }
               break;
       default:

block/blk-cgroup.c +7 -0

@@ -1301,6 +1345,13 @@ struct cftype blkio_files[] = {
                .name = "reset_stats",
                .write_u64 = blkiocg_reset_stats,
        },
+       {
+               .name = "use_hierarchy",
+               .private = BLKIOFILE_PRIVATE(BLKIO_POLICY_PROP,
+                                            BLKIO_PROP_use_hierarchy),
+               .read_u64 = blkiocg_file_read_u64,
+               .write_u64 = blkiocg_file_write_u64,
+       },
  #ifdef CONFIG_BLK_DEV_THROTTLING
        {
                .name = "throttle.read_bps_device",

@@ -1444,7 +1495,7 @@ static void blkiocg_destroy(struct cgroup_subsys *subsys, struct cgroup *cgroup)

 static struct cgroup_subsys_state * blkiocg_create(struct cgroup_subsys *subsys, struct cgroup *cgroup) {
-       struct blkio_cgroup *blkcg;
+       struct blkio_cgroup *blkcg, *parent_blkcg = NULL; struct cgroup *parent = cgroup->parent;

       if (!parent) {

block/blk-cgroup.c +6 -0

@@ -1452,6 +1503,7 @@ blkiocg_create(struct cgroup_subsys *subsys, struct cgroup *cgroup)
                goto done;
        }
 
+       parent_blkcg = cgroup_to_blkio_cgroup(parent);
        blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
        if (!blkcg)
                return ERR_PTR(-ENOMEM);

@@ -1462,6 +1514,11 @@ done:
        INIT_HLIST_HEAD(&blkcg->blkg_list);
 
        INIT_LIST_HEAD(&blkcg->policy_list);
+       if (parent)
+               blkcg->use_hierarchy = parent_blkcg->use_hierarchy;
+       else
+               blkcg->use_hierarchy = 0;
+
        return &blkcg->css;
  }
 

block/blk-cgroup.h +2 -0

index ea4861b..5b4b351 100644
@@ -90,6 +90,7 @@ enum blkcg_file_name_prop {
        BLKIO_PROP_idle_time,
        BLKIO_PROP_empty_time,
        BLKIO_PROP_dequeue,
+       BLKIO_PROP_use_hierarchy,
  };
 
  /* cgroup files owned by throttle policy */

@@ -105,6 +106,7 @@ enum blkcg_file_name_throtl {
  struct blkio_cgroup {
        struct cgroup_subsys_state css;
        unsigned int weight;
+       bool use_hierarchy;
        spinlock_t lock;
        struct hlist_head blkg_list;

       struct list_head policy_list; /* list of blkio_policy_node */

block/blk-cgroup.h +1 -0

@@ -179,6 +181,7 @@ struct blkio_policy_node {
 
  extern unsigned int blkcg_get_weight(struct blkio_cgroup *blkcg,
                                     dev_t dev);
+extern unsigned int blkcg_get_use_hierarchy(struct blkio_cgroup *blkcg);
  extern uint64_t blkcg_get_read_bps(struct blkio_cgroup *blkcg,
                                     dev_t dev);
  extern uint64_t blkcg_get_write_bps(struct blkio_cgroup *blkcg,

block/cfq-iosched.c +26 -6

index aa3eda8..0e21d27 100644
@@ -110,6 +110,9 @@ struct cfq_entity {
        u64 vdisktime;
        bool is_group_entity;
        unsigned int weight;
+       struct cfq_entity *parent;
+       /* Reposition time */
+       unsigned long reposition_time;
  };
 
  /*

@@ -118,8 +121,6 @@ struct cfq_entity {
  struct cfq_queue {
        /* The schedule entity */
        struct cfq_entity cfqe;
-       /* Reposition time */
-       unsigned long reposition_time;
        /* reference count */
        int ref;
        /* various state flags, see below */

@@ -199,6 +200,9 @@ struct cfq_group {
        /* number of cfqq currently on this group */
        int nr_cfqq;
 
+       /* number of sub cfq groups */
+       int nr_subgp;
+
        /*
         * Per group busy queus average. Useful for workload slice calc. We
         * create the array for each prio class but at run time it is used

@@ -234,10 +238,11 @@ struct cfq_group {
  */
  struct cfq_data {
        struct request_queue *queue;
-       /* Root service tree for cfq_groups */
-       struct cfq_rb_root grp_service_tree;
        struct cfq_group root_group;
 
+       /* cfq group schedule in flat or hierarchy manner. */
+       bool use_hierarchy;
+
        /*
         * The priority currently being served
         */

@@ -246,6 +251,9 @@ struct cfq_data {
        unsigned long workload_expires;
        struct cfq_group *serving_group;
 
+       /* Service tree for cfq group flat scheduling mode. */
+       struct cfq_rb_root grp_service_tree;
+
        /*
         * Each priority tree is sorted by next_request position.  These
         * trees are used when determining if two or more queues are

@@ -355,8 +363,6 @@ cfqg_of_entity(struct cfq_entity *cfqe)
  }
 
 
-static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
-
  static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
                                            enum wl_prio_t prio,
                                            enum wl_type_t type)

@@ -643,13 +649,50 @@ static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
        return cfqg->busy_queues_avg[rt];
  }
 
+static inline unsigned int
+cfq_group_get_total_weight(struct cfq_group *cfqg)
+{
+       int i, j;
+       struct cfq_rb_root *st;
+       unsigned int total_weight = 0;
+
+       for_each_cfqg_st(cfqg, i, j, st) {
+               total_weight += st->total_weight;
+       }
+
+       return total_weight;
+}
+

 static inline unsigned
 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) {
-       struct cfq_rb_root *st = &cfqd->grp_service_tree; struct cfq_entity *cfqe = &cfqg->cfqe;
+       struct cfq_rb_root *st;
+       int group_slice = cfq_target_latency;
+       unsigned int grp_total_weight;
+       struct cfq_group *p_cfqg;

block/cfq-iosched.c +21 -1

+
+       /*
+        * Calculate group slice in a hierarchical way.
+        * Note, the calculation is cross all service trees under a group.
+        */
+       do {
+               if (cfqe->parent) {
+                       p_cfqg = cfqg_of_entity(cfqe->parent);
+                       grp_total_weight = cfq_group_get_total_weight(p_cfqg);
+                       group_slice = group_slice * cfqe->weight /
+                                       grp_total_weight;
+               } else {
+                       /* For top level groups */
+                       st = cfqe->service_tree;
+                       group_slice = group_slice * cfqe->weight /
+                                       st->total_weight;
+               }
 
-       return cfq_target_latency * cfqe->weight / st->total_weight;
+               cfqe = cfqe->parent;
+       } while (cfqe);
+
+       return group_slice;

 }

 static inline void

block/cfq-iosched.c +0 -0

@@ -672,7 +715,8 @@ cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
                        /* scale low_slice according to IO priority
                         * and sync vs async */

                       unsigned low_slice =
-                               min(slice, base_low_slice * slice / sync_slice);
+                               min(slice, base_low_slice * slice /
+                                   sync_slice);
                       /* the adapted slice value is scaled to fit all iqs
                        * into the target latency */
                       slice = max(slice * group_slice / expect_latency,

block/cfq-iosched.c +2 -11

@@ -812,17 +856,6 @@ static struct cfq_entity *cfq_rb_first(struct cfq_rb_root *root)
        return NULL;
  }
 
-static struct cfq_entity *cfq_rb_first_entity(struct cfq_rb_root *root)
-{
-       if (!root->left)
-               root->left = rb_first(&root->rb);
-
-       if (root->left)
-               return rb_entry_entity(root->left);
-
-       return NULL;
-}
-
  static void rb_erase_init(struct rb_node *n, struct rb_root *root)
  {
        rb_erase(n, root);

@@ -896,12 +929,15 @@ __cfq_entity_service_tree_add(struct cfq_rb_root *st, struct cfq_entity *cfqe)
 
        rb_link_node(&cfqe->rb_node, parent, node);
        rb_insert_color(&cfqe->rb_node, &st->rb);
+
+       update_min_vdisktime(st);
  }
 

 static void
 cfq_entity_service_tree_add(struct cfq_rb_root *st, struct cfq_entity *cfqe) {
       __cfq_entity_service_tree_add(st, cfqe);
+       cfqe->reposition_time = jiffies;
       st->count++;
       st->total_weight += cfqe->weight;
 }

block/cfq-iosched.c +0 -0

@@ -909,34 +945,52 @@ cfq_entity_service_tree_add(struct cfq_rb_root *st, struct cfq_entity *cfqe)

 static void
 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) {
-       struct cfq_rb_root *st = &cfqd->grp_service_tree; struct cfq_entity *cfqe = &cfqg->cfqe;
-       struct cfq_entity *__cfqe;
       struct rb_node *n;
+       struct cfq_entity *entity;
+       struct cfq_rb_root *st;
+       struct cfq_group *__cfqg;

       cfqg->nr_cfqq++;
+
       if (!RB_EMPTY_NODE(&cfqe->rb_node))
               return;

       /*
-        * Currently put the group at the end. Later implement something
-        * so that groups get lesser vtime based on their weights, so that
-        * if group does not loose all if it was not continously backlogged.
+        * Enqueue this group and its ancestors onto their service tree. */
-       n = rb_last(&st->rb);
-       if (n) {
-               __cfqe = rb_entry_entity(n);
-               cfqe->vdisktime = __cfqe->vdisktime + CFQ_IDLE_DELAY;
-       } else
-               cfqe->vdisktime = st->min_vdisktime;
+       while (cfqe) {
+               if (!RB_EMPTY_NODE(&cfqe->rb_node))
+                       return;

-       cfq_entity_service_tree_add(st, cfqe);
+               /*
+                * Currently put the group at the end. Later implement
+                * something so that groups get lesser vtime based on
+                * their weights, so that if group does not loose all
+                * if it was not continously backlogged.
+                */
+               st = cfqe->service_tree;
+               n = rb_last(&st->rb);
+               if (n) {
+                       entity = rb_entry_entity(n);
+                       cfqe->vdisktime = entity->vdisktime +
+                               CFQ_IDLE_DELAY;
+               } else
+                       cfqe->vdisktime = st->min_vdisktime; +
+               cfq_entity_service_tree_add(st, cfqe);
+               cfqe = cfqe->parent;
+               __cfqg = cfqg_of_entity(cfqe);
+               if (__cfqg)
+                       __cfqg->nr_subgp++;
+       }
 }

 static void
 __cfq_entity_service_tree_del(struct cfq_rb_root *st, struct cfq_entity *cfqe) {
       cfq_rb_erase(&cfqe->rb_node, st);
+       update_min_vdisktime(st);
 }

 static void

block/cfq-iosched.c +0 -1

@@ -945,27 +999,43 @@ cfq_entity_service_tree_del(struct cfq_rb_root *st, struct cfq_entity *cfqe)
        if (!RB_EMPTY_NODE(&cfqe->rb_node)) {
                __cfq_entity_service_tree_del(st, cfqe);
                st->total_weight -= cfqe->weight;
-               cfqe->service_tree = NULL;
        }
  }
 

 static void
 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) {
-       struct cfq_rb_root *st = &cfqd->grp_service_tree; struct cfq_entity *cfqe = &cfqg->cfqe;
+       struct cfq_group *__cfqg, *p_cfqg;

       BUG_ON(cfqg->nr_cfqq < 1);
       cfqg->nr_cfqq--;

-       /* If there are other cfq queues under this group, don't delete it */
-       if (cfqg->nr_cfqq)
+       /*
+        * If there are other cfq queues under this group, or there are other
+        * cfq groups under this group, don't delete it.
+        */
+       if (cfqg->nr_cfqq || cfqg->nr_subgp)
               return;

-       cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
-       cfq_entity_service_tree_del(st, cfqe);
-       cfqg->saved_workload_slice = 0;
-       cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
+       /*
+        * Dequeue this group and its ancestors from their service
+        * tree.
+        */
+       while (cfqe) {
+               __cfqg = cfqg_of_entity(cfqe);
+               p_cfqg = cfqg_of_entity(cfqe->parent);
+               cfq_entity_service_tree_del(cfqe->service_tree, cfqe);
+               cfq_blkiocg_update_dequeue_stats(&__cfqg->blkg, 1);
+               cfq_log_cfqg(cfqd, __cfqg, "del_from_rr group");
+               __cfqg->saved_workload_slice = 0;
+               cfqe = cfqe->parent;
+               if (p_cfqg) {
+                       p_cfqg->nr_subgp--;
+                       if (p_cfqg->nr_cfqq || p_cfqg->nr_subgp)
+                               return;
+               }
+       }
 }

 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)

block/cfq-iosched.c +19 -6

@@ -997,7 +1067,6 @@ static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
  static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
                                struct cfq_queue *cfqq)
  {
-       struct cfq_rb_root *st = &cfqd->grp_service_tree;
        unsigned int used_sl, charge;
        int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
                        - cfqg->service_tree_idle.count;

@@ -1011,10 +1080,23 @@ static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
        else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
                charge = cfqq->allocated_slice;
 
-       /* Can't update vdisktime while group is on service tree */
-       __cfq_entity_service_tree_del(st, cfqe);
-       cfqe->vdisktime += cfq_scale_slice(charge, cfqe);
-       __cfq_entity_service_tree_add(st, cfqe);
+       /*
+        * Update the vdisktime on the whole chain.
+        */
+       while (cfqe) {
+               struct cfq_rb_root *st = cfqe->service_tree;
+
+               /*
+                * Can't update vdisktime while group is on service
+                * tree.
+                */
+               __cfq_entity_service_tree_del(st, cfqe);
+               cfqe->vdisktime += cfq_scale_slice(charge, cfqe);
+               __cfq_entity_service_tree_add(st, cfqe);
+               st->count++;
+               cfqe->reposition_time = jiffies;
+               cfqe = cfqe->parent;
+       }
 
        /* This group is being expired. Save the context */
        if (time_after(cfqd->workload_expires, jiffies)) {

@@ -1026,7 +1108,8 @@ static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
                cfqg->saved_workload_slice = 0;
 
        cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu",
-                    cfqe->vdisktime, st->min_vdisktime);
+                    cfqg->cfqe.vdisktime,
+                    cfqg->cfqe.service_tree->min_vdisktime);

       cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u" " sect=%u", used_sl, cfqq->slice_dispatch, charge, iops_mode(cfqd), cfqq->nr_sectors);

block/cfq-iosched.c +203 -36

@@ -1048,35 +1131,27 @@ void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
        cfqg_of_blkg(blkg)->cfqe.weight = weight;
  }
 
-static struct cfq_group *
-cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+static void init_cfqe(struct blkio_cgroup *blkcg,
+                                   struct cfq_group *cfqg)
+{
+       struct cfq_entity *cfqe = &cfqg->cfqe;
+
+       cfqe->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
+       RB_CLEAR_NODE(&cfqe->rb_node);
+       cfqe->is_group_entity = true;
+       cfqe->parent = NULL;
+}
+
+static void init_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg,
+                     struct cfq_group *cfqg)
  {
-       struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
-       struct cfq_group *cfqg = NULL;
-       void *key = cfqd;
        int i, j;
        struct cfq_rb_root *st;
-       struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
        unsigned int major, minor;
-
-       cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
-       if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
-               sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
-               cfqg->blkg.dev = MKDEV(major, minor);
-               goto done;
-       }
-       if (cfqg || !create)
-               goto done;
-
-       cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
-       if (!cfqg)
-               goto done;
+       struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
 
        for_each_cfqg_st(cfqg, i, j, st)
                *st = CFQ_RB_ROOT;
-       RB_CLEAR_NODE(&cfqg->cfqe.rb_node);
-
-       cfqg->cfqe.is_group_entity = true;
 
        /*
         * Take the initial reference that will be released on destroy

@@ -1086,24 +1161,199 @@ cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
         */
        cfqg->ref = 1;
 
+       /* Add group onto cgroup list */
+       sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+       cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
+                                   MKDEV(major, minor));
+       /* Initiate group entity */
+       init_cfqe(blkcg, cfqg);
+       /* Add group on cfqd list */
+       hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
+}
+
+static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg);
+
+static void uninit_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
+{
+       if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
+               cfq_destroy_cfqg(cfqd, cfqg);
+}
+
+static void cfqg_set_parent(struct cfq_data *cfqd, struct cfq_group *cfqg,
+                           struct cfq_group *p_cfqg)
+{
+       struct cfq_entity *cfqe, *p_cfqe;
+
+       cfqe = &cfqg->cfqe;
+
        /*
-        * Add group onto cgroup list. It might happen that bdi->dev is
-        * not initiliazed yet. Initialize this new group without major
-        * and minor info and this info will be filled in once a new thread
-        * comes for IO. See code above.
+        * 1. If use_hierarchy of the CGroup where cfqg's parent stays is not
+        *    set, we put this cfqg onto global service tree.
+        * 2. If cfqg is root cfqg, put it onto global service tree.
         */
-       if (bdi->dev) {
-               sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
-               cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
-                                       MKDEV(major, minor));
-       } else
-               cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
-                                       0);
+       if (!p_cfqg) {
+               cfqe->service_tree = &cfqd->grp_service_tree;
+               cfqe->parent = NULL;
+               return;
+       }
 
-       cfqg->cfqe.weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
+       p_cfqe = &p_cfqg->cfqe;
 
-       /* Add group on cfqd list */
-       hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
+       cfqe->parent = p_cfqe;
+
+       /*
+        * Currently, just put cfq group entity on "BE:SYNC" workload
+        * service tree.
+        */
+       cfqe->service_tree = service_tree_for(p_cfqg, BE_WORKLOAD,
+                                                     SYNC_WORKLOAD);
+       /* child reference */
+       p_cfqg->ref++;
+}
+
+static struct cfq_group *cfqg_get_parent(struct cfq_group * cfqg)
+{
+       struct cfq_entity *cfqe, *p_cfqe;
+
+       if (!cfqg)
+               return NULL;
+
+       cfqe = &cfqg->cfqe;
+       p_cfqe = cfqe->parent;
+       if (!p_cfqe)
+               return NULL;
+
+       return cfqg_of_entity(p_cfqe);
+}
+
+static struct cfq_group *
+cfqg_chain_alloc(struct cfq_data *cfqd, struct cgroup *cgroup)
+{
+       struct blkio_cgroup *blkcg;
+       struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+       unsigned int major, minor;
+       struct cfq_group *cfqg, *leaf_cfqg, *child_cfqg, *tmp_cfqg;
+       void *key = cfqd;
+
+       /*
+        * If CGroup's use_hierarchy is unset, we just need to allocate only
+        * one CFQ group, and this group will put onto the "grp_service_tree".
+        * We don't need to check whether the cfqg exists, the caller has
+        * already checked it.
+        */
+       blkcg = cgroup_to_blkio_cgroup(cgroup);
+       if (!blkcg_get_use_hierarchy(blkcg)) {
+               cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC,
+                                   cfqd->queue->node);
+               if (!cfqg)
+                       return NULL;
+
+               init_cfqg(cfqd, blkcg, cfqg);
+               cfqg_set_parent(cfqd, cfqg, NULL);
+               return cfqg;
+       }
+
+       /*
+        * Allocate the CFQ group chain until we meet the group we'v already
+        * allocated before, or to the CGroup whose use_hierarchy is not set.
+        */
+       leaf_cfqg = NULL;
+       child_cfqg = NULL;
+       for (; cgroup != NULL; cgroup = cgroup->parent) {
+               blkcg = cgroup_to_blkio_cgroup(cgroup);
+               cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+               if (cfqg) {
+                       if (!cfqg->blkg.dev && bdi->dev &&
+                           dev_name(bdi->dev)) {
+                               sscanf(dev_name(bdi->dev), "%u:%u",
+                                      &major, &minor);
+                               cfqg->blkg.dev = MKDEV(major, minor);
+                       }
+
+                       /*
+                        * Initialization of parent doesn't finish yet, get
+                        * it done.
+                        */
+                       if (child_cfqg) {
+                               if (blkcg_get_use_hierarchy(blkcg))
+                                       cfqg_set_parent(cfqd, child_cfqg,
+                                                       cfqg);
+                               else
+                                       cfqg_set_parent(cfqd, child_cfqg,
+                                                       NULL);
+                       }
+
+                       /* chain has already been built */
+                       break;
+               }
+
+               /*
+                * We only allocate a cfqg that the corresponding cgroup's
+                * use_hierarchy is set.
+                */
+               if (blkcg_get_use_hierarchy(blkcg)) {
+                       cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC,
+                                           cfqd->queue->node);
+                       if (!cfqg)
+                               goto clean_up;
+
+                       if (!leaf_cfqg)
+                               leaf_cfqg = cfqg;
+
+                       init_cfqg(cfqd, blkcg, cfqg);
+               } else {
+                       cfqg = NULL;
+               }
+
+               if (child_cfqg)
+                       cfqg_set_parent(cfqd, child_cfqg, cfqg);
+
+               /*
+                * This CGroup's use_hierarchy isn't set, this means the CFQ
+                * group chain has been built.
+                */
+               if (!blkcg_get_use_hierarchy(blkcg))
+                       break;
+
+               child_cfqg = cfqg;
+       }
+
+       return leaf_cfqg;
+
+clean_up:
+       /* clean up the allocated cfq groups. */
+       while (leaf_cfqg) {
+               tmp_cfqg = leaf_cfqg;
+               leaf_cfqg = cfqg_get_parent(leaf_cfqg);
+               uninit_cfqg(cfqd, tmp_cfqg);
+       }
+
+       return NULL;
+}
+
+static struct cfq_group *
+cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
+{
+       struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
+       struct cfq_group *cfqg = NULL;
+       void *key = cfqd;
+       struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
+       unsigned int major, minor;
+
+       cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
+       if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
+               sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
+               cfqg->blkg.dev = MKDEV(major, minor);
+               goto done;
+       }
+       if (cfqg || !create)
+               goto done;
+
+       /*
+        * Allocate CFQ group chain to the root group or we meet the CGroup
+        * with use_hierarchy disabled.
+        */
+       cfqg = cfqg_chain_alloc(cfqd, cgroup);
 

 done:
       return cfqg;

block/cfq-iosched.c +17 -3

@@ -1148,6 +1398,7 @@ static void cfq_put_cfqg(struct cfq_group *cfqg)
  {
        struct cfq_rb_root *st;
        int i, j;
+       struct cfq_group *p_cfqg;
 
        BUG_ON(cfqg->ref <= 0);
        cfqg->ref--;

@@ -1155,6 +1406,22 @@ static void cfq_put_cfqg(struct cfq_group *cfqg)
                return;
        for_each_cfqg_st(cfqg, i, j, st)
                BUG_ON(!RB_EMPTY_ROOT(&st->rb));
+
+       do {
+               p_cfqg = cfqg_get_parent(cfqg);
+               kfree(cfqg);
+               cfqg = NULL;
+               /*
+                * Drop the reference taken by children, if nobody references
+                * parent group, we need delete the parent also.
+                */
+               if (p_cfqg) {
+                       p_cfqg->ref--;
+                       if (p_cfqg->ref == 0)
+                               cfqg = p_cfqg;
+               }
+       } while (cfqg);
+
        kfree(cfqg);
  }
 

@@ -1321,9 +1588,6 @@ static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
                         * ioprio.
                         */
                        pos_offset = cfq_get_boost(cfqd, cfqq);
-                       /* Debug purpose, should remove. */
-                       cfq_log_cfqq(cfqd, cfqq, "pos_offset: %llu",
-                                    pos_offset);

                       cfqe->vdisktime = service_tree->min_vdisktime + pos_offset; } else

block/cfq-iosched.c +27 -14

@@ -1365,9 +1629,8 @@ insert:
        cfqe->service_tree = service_tree;
 
        /* Add cfqq onto service tree. */
+
        cfq_entity_service_tree_add(service_tree, cfqe);
-       update_min_vdisktime(service_tree);
-       cfqq->reposition_time = jiffies;
        if ((add_front || !new_cfqq) && !group_changed)
                return;
        cfq_group_service_tree_add(cfqd, cfqq->cfqg);

@@ -1810,28 +2073,43 @@ static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
        return cfqq_of_entity(cfq_rb_first(service_tree));
  }
 
-static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
+struct cfq_rb_root *choose_service_tree_forced(struct cfq_group *cfqg)
  {
-       struct cfq_group *cfqg;
-       struct cfq_entity *cfqe;
        int i, j;
        struct cfq_rb_root *st;
 
-       if (!cfqd->rq_queued)
-               return NULL;
+       for_each_cfqg_st(cfqg, i, j, st) {
+               if (st->count != 0)
+                       return st;
+       }
 
-       cfqg = cfq_get_next_cfqg(cfqd);
-       if (!cfqg)
+       return NULL;
+}
+
+static struct cfq_entity *
+cfq_get_next_entity_forced(struct cfq_data *cfqd)
+{
+       struct cfq_entity *cfqe;
+       struct cfq_rb_root *st = &cfqd->grp_service_tree;
+       struct cfq_group *cfqg;
+
+       if (!cfqd->rq_queued)
                return NULL;
 
-       for_each_cfqg_st(cfqg, i, j, st) {
+       do {
                cfqe = cfq_rb_first(st);
-               if (cfqe != NULL)
-                       return cfqq_of_entity(cfqe);
-       }
+               if (cfqe && !cfqe->is_group_entity)
+                       return cfqe;
+               else if (cfqe && cfqe->is_group_entity)
+                       cfqg = cfqg_of_entity(cfqe);
+
+               st = choose_service_tree_forced(cfqg);
+       } while (st);
+
        return NULL;
  }
 
+
  /*
  * Get and set a new active qu
 
 
--

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/