[PATCHSET] block: implement per-blkg request allocation, take#2

[PATCHSET] block: implement per-blkg request allocation, take#2

[Tejun Heo]

Hello,

(The posting yesterday went out w/o lkml cc'd and old head message,
 resending w/ Vivek's suggestions applied.  Sorry about the noise.)

This is the second take of "implement per-blkg request allocation"
patchset.  Changes from the last take[L] are

* 0001-fix-blkg_alloc-failure-path patch is separated from this series
  and merged to block/for-linus.

* Updated patches posted incrementally merged into the series.

* Rebased on top of the current block/for-linus.

* Documentation/block/queue-sysfs.txt updated to note that nr_requests
  is per-blkcg.

* Acked/Reviewed-by's added.

The original description follows.

Currently block layer shares a single request_list (@q->rq) for all
IOs regardless of their blkcg associations.  This means that once the
shared pool is exhausted, blkcg limits don't mean much.  Whoever grabs
the requests being freed the first grabs the next IO slot.

This priority inversion can be easily demonstrated by creating a blkio
cgroup w/ very low weight, put a program which can issue a lot of
random direct IOs there and running a sequential IO from a different
cgroup.  As soon as the request pool is used up, the sequential IO
bandwidth crashes.

This patchset implements per-blkg request allocation so that each
blkcg-request_queue pair has its own request pool to allocate from.
This isolates different blkcgs in terms of request allocation.

Most changes are straight-forward; unfortunately, bdi isn't
blkcg-aware yet so it currently just propagates the congestion state
from root blkcg.  As writeback currently is always on the root blkcg,
this kinda works for write congestion but readahead may behave
non-optimally under congestion for now.  This needs to be improved but
the situation is still way better than blkcg completely collapsing.

 0001-blkcg-__blkg_lookup_create-doesn-t-need-radix-preloa.patch
 0002-blkcg-make-root-blkcg-allocation-use-GFP_KERNEL.patch
 0003-mempool-add-gfp_mask-to-mempool_create_node.patch
 0004-block-drop-custom-queue-draining-used-by-scsi_transp.patch
 0005-block-refactor-get_request-_wait.patch
 0006-block-allocate-io_context-upfront.patch
 0007-blkcg-inline-bio_blkcg-and-friends.patch
 0008-block-add-q-nr_rqs-and-move-q-rq.elvpriv-to-q-nr_rqs.patch
 0009-block-prepare-for-multiple-request_lists.patch
 0010-blkcg-implement-per-blkg-request-allocation.patch

0001-0002 are misc preps.

0003 adds @gfp_mask to mempool_create_node().  This is necessary
because blkg allocation is on the IO path and now blkg contains
mempool for request_list.  Note that blkg allocation failure doesn't
lead to catastrophic failure.  It just hinders blkcg enforcement.

0004 drops custom queue draining which I dont't think is necessary and
hinders with further updates.

0005-0009 are prep patches and 0010 implements per-blkg request
allocation.

This patchset is on top of the current block/for-linus - 9b2ea86bc9e
"blkcg: fix blkg_alloc() failure path" and is also available in the
following git branch.

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

 Documentation/block/queue-sysfs.txt |    7 +
 block/blk-cgroup.c                  |  139 ++++++++++++++++---------
 block/blk-cgroup.h                  |  121 +++++++++++++++++++++
 block/blk-core.c                    |  200 ++++++++++++++++++------------------
 block/blk-sysfs.c                   |   34 +++---
 block/blk-throttle.c                |    3 
 block/blk.h                         |    3 
 block/bsg-lib.c                     |   53 ---------
 drivers/scsi/scsi_transport_fc.c    |   38 ------
 drivers/scsi/scsi_transport_iscsi.c |    2 
 include/linux/blkdev.h              |   53 +++++----
 include/linux/bsg-lib.h             |    1 
 include/linux/mempool.h             |    3 
 mm/mempool.c                        |   12 +-
 14 files changed, 382 insertions(+), 287 deletions(-)

Thanks.

--
tejun

[L] http://thread.gmane.org/gmane.linux.kernel.containers/23159

[PATCHSET] block: implement per-blkg request allocation, take#2

[Tejun Heo]

Hello,

(The posting yesterday went out w/o lkml cc'd and old head message,
 resending w/ Vivek's suggestions applied.  Sorry about the noise.)

This is the second take of "implement per-blkg request allocation"
patchset.  Changes from the last take[L] are

* 0001-fix-blkg_alloc-failure-path patch is separated from this series
  and merged to block/for-linus.

* Updated patches posted incrementally merged into the series.

* Rebased on top of the current block/for-linus.

* Documentation/block/queue-sysfs.txt updated to note that nr_requests
  is per-blkcg.

* Acked/Reviewed-by's added.

The original description follows.

Currently block layer shares a single request_list (@q->rq) for all
IOs regardless of their blkcg associations.  This means that once the
shared pool is exhausted, blkcg limits don't mean much.  Whoever grabs
the requests being freed the first grabs the next IO slot.

This priority inversion can be easily demonstrated by creating a blkio
cgroup w/ very low weight, put a program which can issue a lot of
random direct IOs there and running a sequential IO from a different
cgroup.  As soon as the request pool is used up, the sequential IO
bandwidth crashes.

This patchset implements per-blkg request allocation so that each
blkcg-request_queue pair has its own request pool to allocate from.
This isolates different blkcgs in terms of request allocation.

Most changes are straight-forward; unfortunately, bdi isn't
blkcg-aware yet so it currently just propagates the congestion state
from root blkcg.  As writeback currently is always on the root blkcg,
this kinda works for write congestion but readahead may behave
non-optimally under congestion for now.  This needs to be improved but
the situation is still way better than blkcg completely collapsing.

 0001-blkcg-__blkg_lookup_create-doesn-t-need-radix-preloa.patch
 0002-blkcg-make-root-blkcg-allocation-use-GFP_KERNEL.patch
 0003-mempool-add-gfp_mask-to-mempool_create_node.patch
 0004-block-drop-custom-queue-draining-used-by-scsi_transp.patch
 0005-block-refactor-get_request-_wait.patch
 0006-block-allocate-io_context-upfront.patch
 0007-blkcg-inline-bio_blkcg-and-friends.patch
 0008-block-add-q-nr_rqs-and-move-q-rq.elvpriv-to-q-nr_rqs.patch
 0009-block-prepare-for-multiple-request_lists.patch
 0010-blkcg-implement-per-blkg-request-allocation.patch

0001-0002 are misc preps.

0003 adds @gfp_mask to mempool_create_node().  This is necessary
because blkg allocation is on the IO path and now blkg contains
mempool for request_list.  Note that blkg allocation failure doesn't
lead to catastrophic failure.  It just hinders blkcg enforcement.

0004 drops custom queue draining which I dont't think is necessary and
hinders with further updates.

0005-0009 are prep patches and 0010 implements per-blkg request
allocation.

This patchset is on top of the current block/for-linus - 9b2ea86bc9e
"blkcg: fix blkg_alloc() failure path" and is also available in the
following git branch.

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

 Documentation/block/queue-sysfs.txt |    7 +
 block/blk-cgroup.c                  |  139 ++++++++++++++++---------
 block/blk-cgroup.h                  |  121 +++++++++++++++++++++
 block/blk-core.c                    |  200 ++++++++++++++++++------------------
 block/blk-sysfs.c                   |   34 +++---
 block/blk-throttle.c                |    3 
 block/blk.h                         |    3 
 block/bsg-lib.c                     |   53 ---------
 drivers/scsi/scsi_transport_fc.c    |   38 ------
 drivers/scsi/scsi_transport_iscsi.c |    2 
 include/linux/blkdev.h              |   53 +++++----
 include/linux/bsg-lib.h             |    1 
 include/linux/mempool.h             |    3 
 mm/mempool.c                        |   12 +-
 14 files changed, 382 insertions(+), 287 deletions(-)

Thanks.

--
tejun

[L] http://thread.gmane.org/gmane.linux.kernel.containers/23159

[PATCH 01/10] blkcg: __blkg_lookup_create() doesn't need radix preload

[Tejun Heo]

There's no point in calling radix_tree_preload() if preloading doesn't
use more permissible GFP mask.  Drop preloading from
__blkg_lookup_create().

While at it, drop sparse locking annotation which no longer applies.

v2: Vivek pointed out the odd preload usage.  Instead of updating,
    just drop it.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 block/blk-cgroup.c |   10 +---------
 1 files changed, 1 insertions(+), 9 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 4ab7420..af61db0 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -177,7 +177,6 @@ EXPORT_SYMBOL_GPL(blkg_lookup);
 
 static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 					     struct request_queue *q)
-	__releases(q->queue_lock) __acquires(q->queue_lock)
 {
 	struct blkcg_gq *blkg;
 	int ret;
@@ -203,10 +202,6 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 		goto err_put;
 
 	/* insert */
-	ret = radix_tree_preload(GFP_ATOMIC);
-	if (ret)
-		goto err_free;
-
 	spin_lock(&blkcg->lock);
 	ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
 	if (likely(!ret)) {
@@ -215,14 +210,11 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 	}
 	spin_unlock(&blkcg->lock);
 
-	radix_tree_preload_end();
-
 	if (!ret)
 		return blkg;
-err_free:
-	blkg_free(blkg);
 err_put:
 	css_put(&blkcg->css);
+	blkg_free(blkg);
 	return ERR_PTR(ret);
 }
 
-- 
1.7.7.3

[PATCH 01/10] blkcg: __blkg_lookup_create() doesn't need radix preload

[Tejun Heo]

There's no point in calling radix_tree_preload() if preloading doesn't
use more permissible GFP mask.  Drop preloading from
__blkg_lookup_create().

While at it, drop sparse locking annotation which no longer applies.

v2: Vivek pointed out the odd preload usage.  Instead of updating,
    just drop it.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-cgroup.c |   10 +---------
 1 files changed, 1 insertions(+), 9 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 4ab7420..af61db0 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -177,7 +177,6 @@ EXPORT_SYMBOL_GPL(blkg_lookup);
 
 static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 					     struct request_queue *q)
-	__releases(q->queue_lock) __acquires(q->queue_lock)
 {
 	struct blkcg_gq *blkg;
 	int ret;
@@ -203,10 +202,6 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 		goto err_put;
 
 	/* insert */
-	ret = radix_tree_preload(GFP_ATOMIC);
-	if (ret)
-		goto err_free;
-
 	spin_lock(&blkcg->lock);
 	ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
 	if (likely(!ret)) {
@@ -215,14 +210,11 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 	}
 	spin_unlock(&blkcg->lock);
 
-	radix_tree_preload_end();
-
 	if (!ret)
 		return blkg;
-err_free:
-	blkg_free(blkg);
 err_put:
 	css_put(&blkcg->css);
+	blkg_free(blkg);
 	return ERR_PTR(ret);
 }
 
-- 
1.7.7.3

[PATCH 02/10] blkcg: make root blkcg allocation use %GFP_KERNEL

[Tejun Heo]

Currently, blkcg_activate_policy() depends on %GFP_ATOMIC allocation
from __blkg_lookup_create() for root blkcg creation.  This could make
policy fail unnecessarily.

Make blkg_alloc() take @gfp_mask, __blkg_lookup_create() take an
optional @new_blkg for preallocated blkg, and blkcg_activate_policy()
preload radix tree and preallocate blkg with %GFP_KERNEL before trying
to create the root blkg.

v2: __blkg_lookup_create() was returning %NULL on blkg alloc failure
   instead of ERR_PTR() value.  Fixed.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 block/blk-cgroup.c |   59 +++++++++++++++++++++++++++++++++++++--------------
 1 files changed, 43 insertions(+), 16 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index af61db0..cbeeb54 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -91,16 +91,18 @@ static void blkg_free(struct blkcg_gq *blkg)
  * blkg_alloc - allocate a blkg
  * @blkcg: block cgroup the new blkg is associated with
  * @q: request_queue the new blkg is associated with
+ * @gfp_mask: allocation mask to use
  *
  * Allocate a new blkg assocating @blkcg and @q.
  */
-static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q)
+static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
+				   gfp_t gfp_mask)
 {
 	struct blkcg_gq *blkg;
 	int i;
 
 	/* alloc and init base part */
-	blkg = kzalloc_node(sizeof(*blkg), GFP_ATOMIC, q->node);
+	blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
 	if (!blkg)
 		return NULL;
 
@@ -117,7 +119,7 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q)
 			continue;
 
 		/* alloc per-policy data and attach it to blkg */
-		pd = kzalloc_node(pol->pd_size, GFP_ATOMIC, q->node);
+		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
 		if (!pd) {
 			blkg_free(blkg);
 			return NULL;
@@ -175,8 +177,13 @@ struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
 }
 EXPORT_SYMBOL_GPL(blkg_lookup);
 
+/*
+ * If @new_blkg is %NULL, this function tries to allocate a new one as
+ * necessary using %GFP_ATOMIC.  @new_blkg is always consumed on return.
+ */
 static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
-					     struct request_queue *q)
+					     struct request_queue *q,
+					     struct blkcg_gq *new_blkg)
 {
 	struct blkcg_gq *blkg;
 	int ret;
@@ -188,18 +195,24 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 	blkg = __blkg_lookup(blkcg, q);
 	if (blkg) {
 		rcu_assign_pointer(blkcg->blkg_hint, blkg);
-		return blkg;
+		goto out_free;
 	}
 
 	/* blkg holds a reference to blkcg */
-	if (!css_tryget(&blkcg->css))
-		return ERR_PTR(-EINVAL);
+	if (!css_tryget(&blkcg->css)) {
+		blkg = ERR_PTR(-EINVAL);
+		goto out_free;
+	}
 
 	/* allocate */
-	ret = -ENOMEM;
-	blkg = blkg_alloc(blkcg, q);
-	if (unlikely(!blkg))
-		goto err_put;
+	if (!new_blkg) {
+		new_blkg = blkg_alloc(blkcg, q, GFP_ATOMIC);
+		if (unlikely(!new_blkg)) {
+			blkg = ERR_PTR(-ENOMEM);
+			goto out_put;
+		}
+	}
+	blkg = new_blkg;
 
 	/* insert */
 	spin_lock(&blkcg->lock);
@@ -212,10 +225,13 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 
 	if (!ret)
 		return blkg;
-err_put:
+
+	blkg = ERR_PTR(ret);
+out_put:
 	css_put(&blkcg->css);
-	blkg_free(blkg);
-	return ERR_PTR(ret);
+out_free:
+	blkg_free(new_blkg);
+	return blkg;
 }
 
 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
@@ -227,7 +243,7 @@ struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
 	 */
 	if (unlikely(blk_queue_bypass(q)))
 		return ERR_PTR(blk_queue_dead(q) ? -EINVAL : -EBUSY);
-	return __blkg_lookup_create(blkcg, q);
+	return __blkg_lookup_create(blkcg, q, NULL);
 }
 EXPORT_SYMBOL_GPL(blkg_lookup_create);
 
@@ -727,19 +743,30 @@ int blkcg_activate_policy(struct request_queue *q,
 	struct blkcg_gq *blkg;
 	struct blkg_policy_data *pd, *n;
 	int cnt = 0, ret;
+	bool preloaded;
 
 	if (blkcg_policy_enabled(q, pol))
 		return 0;
 
+	/* preallocations for root blkg */
+	blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
+	if (!blkg)
+		return -ENOMEM;
+
+	preloaded = !radix_tree_preload(GFP_KERNEL);
+
 	blk_queue_bypass_start(q);
 
 	/* make sure the root blkg exists and count the existing blkgs */
 	spin_lock_irq(q->queue_lock);
 
 	rcu_read_lock();
-	blkg = __blkg_lookup_create(&blkcg_root, q);
+	blkg = __blkg_lookup_create(&blkcg_root, q, blkg);
 	rcu_read_unlock();
 
+	if (preloaded)
+		radix_tree_preload_end();
+
 	if (IS_ERR(blkg)) {
 		ret = PTR_ERR(blkg);
 		goto out_unlock;
-- 
1.7.7.3

[PATCH 02/10] blkcg: make root blkcg allocation use %GFP_KERNEL

[Tejun Heo]

Currently, blkcg_activate_policy() depends on %GFP_ATOMIC allocation
from __blkg_lookup_create() for root blkcg creation.  This could make
policy fail unnecessarily.

Make blkg_alloc() take @gfp_mask, __blkg_lookup_create() take an
optional @new_blkg for preallocated blkg, and blkcg_activate_policy()
preload radix tree and preallocate blkg with %GFP_KERNEL before trying
to create the root blkg.

v2: __blkg_lookup_create() was returning %NULL on blkg alloc failure
   instead of ERR_PTR() value.  Fixed.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-cgroup.c |   59 +++++++++++++++++++++++++++++++++++++--------------
 1 files changed, 43 insertions(+), 16 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index af61db0..cbeeb54 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -91,16 +91,18 @@ static void blkg_free(struct blkcg_gq *blkg)
  * blkg_alloc - allocate a blkg
  * @blkcg: block cgroup the new blkg is associated with
  * @q: request_queue the new blkg is associated with
+ * @gfp_mask: allocation mask to use
  *
  * Allocate a new blkg assocating @blkcg and @q.
  */
-static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q)
+static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
+				   gfp_t gfp_mask)
 {
 	struct blkcg_gq *blkg;
 	int i;
 
 	/* alloc and init base part */
-	blkg = kzalloc_node(sizeof(*blkg), GFP_ATOMIC, q->node);
+	blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
 	if (!blkg)
 		return NULL;
 
@@ -117,7 +119,7 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q)
 			continue;
 
 		/* alloc per-policy data and attach it to blkg */
-		pd = kzalloc_node(pol->pd_size, GFP_ATOMIC, q->node);
+		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
 		if (!pd) {
 			blkg_free(blkg);
 			return NULL;
@@ -175,8 +177,13 @@ struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
 }
 EXPORT_SYMBOL_GPL(blkg_lookup);
 
+/*
+ * If @new_blkg is %NULL, this function tries to allocate a new one as
+ * necessary using %GFP_ATOMIC.  @new_blkg is always consumed on return.
+ */
 static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
-					     struct request_queue *q)
+					     struct request_queue *q,
+					     struct blkcg_gq *new_blkg)
 {
 	struct blkcg_gq *blkg;
 	int ret;
@@ -188,18 +195,24 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 	blkg = __blkg_lookup(blkcg, q);
 	if (blkg) {
 		rcu_assign_pointer(blkcg->blkg_hint, blkg);
-		return blkg;
+		goto out_free;
 	}
 
 	/* blkg holds a reference to blkcg */
-	if (!css_tryget(&blkcg->css))
-		return ERR_PTR(-EINVAL);
+	if (!css_tryget(&blkcg->css)) {
+		blkg = ERR_PTR(-EINVAL);
+		goto out_free;
+	}
 
 	/* allocate */
-	ret = -ENOMEM;
-	blkg = blkg_alloc(blkcg, q);
-	if (unlikely(!blkg))
-		goto err_put;
+	if (!new_blkg) {
+		new_blkg = blkg_alloc(blkcg, q, GFP_ATOMIC);
+		if (unlikely(!new_blkg)) {
+			blkg = ERR_PTR(-ENOMEM);
+			goto out_put;
+		}
+	}
+	blkg = new_blkg;
 
 	/* insert */
 	spin_lock(&blkcg->lock);
@@ -212,10 +225,13 @@ static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
 
 	if (!ret)
 		return blkg;
-err_put:
+
+	blkg = ERR_PTR(ret);
+out_put:
 	css_put(&blkcg->css);
-	blkg_free(blkg);
-	return ERR_PTR(ret);
+out_free:
+	blkg_free(new_blkg);
+	return blkg;
 }
 
 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
@@ -227,7 +243,7 @@ struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
 	 */
 	if (unlikely(blk_queue_bypass(q)))
 		return ERR_PTR(blk_queue_dead(q) ? -EINVAL : -EBUSY);
-	return __blkg_lookup_create(blkcg, q);
+	return __blkg_lookup_create(blkcg, q, NULL);
 }
 EXPORT_SYMBOL_GPL(blkg_lookup_create);
 
@@ -727,19 +743,30 @@ int blkcg_activate_policy(struct request_queue *q,
 	struct blkcg_gq *blkg;
 	struct blkg_policy_data *pd, *n;
 	int cnt = 0, ret;
+	bool preloaded;
 
 	if (blkcg_policy_enabled(q, pol))
 		return 0;
 
+	/* preallocations for root blkg */
+	blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
+	if (!blkg)
+		return -ENOMEM;
+
+	preloaded = !radix_tree_preload(GFP_KERNEL);
+
 	blk_queue_bypass_start(q);
 
 	/* make sure the root blkg exists and count the existing blkgs */
 	spin_lock_irq(q->queue_lock);
 
 	rcu_read_lock();
-	blkg = __blkg_lookup_create(&blkcg_root, q);
+	blkg = __blkg_lookup_create(&blkcg_root, q, blkg);
 	rcu_read_unlock();
 
+	if (preloaded)
+		radix_tree_preload_end();
+
 	if (IS_ERR(blkg)) {
 		ret = PTR_ERR(blkg);
 		goto out_unlock;
-- 
1.7.7.3

[PATCH 03/10] mempool: add @gfp_mask to mempool_create_node()

[Tejun Heo]

mempool_create_node() currently assumes %GFP_KERNEL.  Its only user,
blk_init_free_list(), is about to be updated to use other allocation
flags - add @gfp_mask argument to the function.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
Cc: Hugh Dickins <hughd-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>
---
 block/blk-core.c        |    4 ++--
 include/linux/mempool.h |    3 ++-
 mm/mempool.c            |   12 +++++++-----
 3 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 3c923a7..7465d89 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -518,8 +518,8 @@ static int blk_init_free_list(struct request_queue *q)
 	init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]);
 
 	rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
-				mempool_free_slab, request_cachep, q->node);
-
+					  mempool_free_slab, request_cachep,
+					  GFP_KERNEL, q->node);
 	if (!rl->rq_pool)
 		return -ENOMEM;
 
diff --git a/include/linux/mempool.h b/include/linux/mempool.h
index 7c08052..39ed62a 100644
--- a/include/linux/mempool.h
+++ b/include/linux/mempool.h
@@ -26,7 +26,8 @@ typedef struct mempool_s {
 extern mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn,
 			mempool_free_t *free_fn, void *pool_data);
 extern mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
-			mempool_free_t *free_fn, void *pool_data, int nid);
+			mempool_free_t *free_fn, void *pool_data,
+			gfp_t gfp_mask, int nid);
 
 extern int mempool_resize(mempool_t *pool, int new_min_nr, gfp_t gfp_mask);
 extern void mempool_destroy(mempool_t *pool);
diff --git a/mm/mempool.c b/mm/mempool.c
index d904981..5499047 100644
--- a/mm/mempool.c
+++ b/mm/mempool.c
@@ -63,19 +63,21 @@ EXPORT_SYMBOL(mempool_destroy);
 mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn,
 				mempool_free_t *free_fn, void *pool_data)
 {
-	return  mempool_create_node(min_nr,alloc_fn,free_fn, pool_data,-1);
+	return mempool_create_node(min_nr,alloc_fn,free_fn, pool_data,
+				   GFP_KERNEL, NUMA_NO_NODE);
 }
 EXPORT_SYMBOL(mempool_create);
 
 mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
-			mempool_free_t *free_fn, void *pool_data, int node_id)
+			       mempool_free_t *free_fn, void *pool_data,
+			       gfp_t gfp_mask, int node_id)
 {
 	mempool_t *pool;
-	pool = kmalloc_node(sizeof(*pool), GFP_KERNEL | __GFP_ZERO, node_id);
+	pool = kmalloc_node(sizeof(*pool), gfp_mask | __GFP_ZERO, node_id);
 	if (!pool)
 		return NULL;
 	pool->elements = kmalloc_node(min_nr * sizeof(void *),
-					GFP_KERNEL, node_id);
+				      gfp_mask, node_id);
 	if (!pool->elements) {
 		kfree(pool);
 		return NULL;
@@ -93,7 +95,7 @@ mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
 	while (pool->curr_nr < pool->min_nr) {
 		void *element;
 
-		element = pool->alloc(GFP_KERNEL, pool->pool_data);
+		element = pool->alloc(gfp_mask, pool->pool_data);
 		if (unlikely(!element)) {
 			mempool_destroy(pool);
 			return NULL;
-- 
1.7.7.3

[PATCH 03/10] mempool: add @gfp_mask to mempool_create_node()

[Tejun Heo]

mempool_create_node() currently assumes %GFP_KERNEL.  Its only user,
blk_init_free_list(), is about to be updated to use other allocation
flags - add @gfp_mask argument to the function.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Hugh Dickins <hughd@google.com>
---
 block/blk-core.c        |    4 ++--
 include/linux/mempool.h |    3 ++-
 mm/mempool.c            |   12 +++++++-----
 3 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 3c923a7..7465d89 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -518,8 +518,8 @@ static int blk_init_free_list(struct request_queue *q)
 	init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]);
 
 	rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
-				mempool_free_slab, request_cachep, q->node);
-
+					  mempool_free_slab, request_cachep,
+					  GFP_KERNEL, q->node);
 	if (!rl->rq_pool)
 		return -ENOMEM;
 
diff --git a/include/linux/mempool.h b/include/linux/mempool.h
index 7c08052..39ed62a 100644
--- a/include/linux/mempool.h
+++ b/include/linux/mempool.h
@@ -26,7 +26,8 @@ typedef struct mempool_s {
 extern mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn,
 			mempool_free_t *free_fn, void *pool_data);
 extern mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
-			mempool_free_t *free_fn, void *pool_data, int nid);
+			mempool_free_t *free_fn, void *pool_data,
+			gfp_t gfp_mask, int nid);
 
 extern int mempool_resize(mempool_t *pool, int new_min_nr, gfp_t gfp_mask);
 extern void mempool_destroy(mempool_t *pool);
diff --git a/mm/mempool.c b/mm/mempool.c
index d904981..5499047 100644
--- a/mm/mempool.c
+++ b/mm/mempool.c
@@ -63,19 +63,21 @@ EXPORT_SYMBOL(mempool_destroy);
 mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn,
 				mempool_free_t *free_fn, void *pool_data)
 {
-	return  mempool_create_node(min_nr,alloc_fn,free_fn, pool_data,-1);
+	return mempool_create_node(min_nr,alloc_fn,free_fn, pool_data,
+				   GFP_KERNEL, NUMA_NO_NODE);
 }
 EXPORT_SYMBOL(mempool_create);
 
 mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
-			mempool_free_t *free_fn, void *pool_data, int node_id)
+			       mempool_free_t *free_fn, void *pool_data,
+			       gfp_t gfp_mask, int node_id)
 {
 	mempool_t *pool;
-	pool = kmalloc_node(sizeof(*pool), GFP_KERNEL | __GFP_ZERO, node_id);
+	pool = kmalloc_node(sizeof(*pool), gfp_mask | __GFP_ZERO, node_id);
 	if (!pool)
 		return NULL;
 	pool->elements = kmalloc_node(min_nr * sizeof(void *),
-					GFP_KERNEL, node_id);
+				      gfp_mask, node_id);
 	if (!pool->elements) {
 		kfree(pool);
 		return NULL;
@@ -93,7 +95,7 @@ mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
 	while (pool->curr_nr < pool->min_nr) {
 		void *element;
 
-		element = pool->alloc(GFP_KERNEL, pool->pool_data);
+		element = pool->alloc(gfp_mask, pool->pool_data);
 		if (unlikely(!element)) {
 			mempool_destroy(pool);
 			return NULL;
-- 
1.7.7.3

[PATCH 04/10] block: drop custom queue draining used by scsi_transport_{iscsi|fc}

[Tejun Heo]

iscsi_remove_host() uses bsg_remove_queue() which implements custom
queue draining.  fc_bsg_remove() open-codes mostly identical logic.

The draining logic isn't correct in that blk_stop_queue() doesn't
prevent new requests from being queued - it just stops processing, so
nothing prevents new requests to be queued after the logic determines
that the queue is drained.

blk_cleanup_queue() now implements proper queue draining and these
custom draining logics aren't necessary.  Drop them and use
bsg_unregister_queue() + blk_cleanup_queue() instead.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Reviewed-by: Mike Christie <michaelc-hcNo3dDEHLuVc3sceRu5cw@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: James Bottomley <James.Bottomley-d9PhHud1JfjCXq6kfMZ53/egYHeGw8Jk@public.gmane.org>
Cc: James Smart <james.smart-laKkSmNT4hbQT0dZR+AlfA@public.gmane.org>
---
 block/bsg-lib.c                     |   53 -----------------------------------
 drivers/scsi/scsi_transport_fc.c    |   38 -------------------------
 drivers/scsi/scsi_transport_iscsi.c |    2 +-
 include/linux/bsg-lib.h             |    1 -
 4 files changed, 1 insertions(+), 93 deletions(-)

diff --git a/block/bsg-lib.c b/block/bsg-lib.c
index 7ad49c8..deee61f 100644
--- a/block/bsg-lib.c
+++ b/block/bsg-lib.c
@@ -243,56 +243,3 @@ int bsg_setup_queue(struct device *dev, struct request_queue *q,
 	return 0;
 }
 EXPORT_SYMBOL_GPL(bsg_setup_queue);
-
-/**
- * bsg_remove_queue - Deletes the bsg dev from the q
- * @q:	the request_queue that is to be torn down.
- *
- * Notes:
- *   Before unregistering the queue empty any requests that are blocked
- */
-void bsg_remove_queue(struct request_queue *q)
-{
-	struct request *req; /* block request */
-	int counts; /* totals for request_list count and starved */
-
-	if (!q)
-		return;
-
-	/* Stop taking in new requests */
-	spin_lock_irq(q->queue_lock);
-	blk_stop_queue(q);
-
-	/* drain all requests in the queue */
-	while (1) {
-		/* need the lock to fetch a request
-		 * this may fetch the same reqeust as the previous pass
-		 */
-		req = blk_fetch_request(q);
-		/* save requests in use and starved */
-		counts = q->rq.count[0] + q->rq.count[1] +
-			 q->rq.starved[0] + q->rq.starved[1];
-		spin_unlock_irq(q->queue_lock);
-		/* any requests still outstanding? */
-		if (counts == 0)
-			break;
-
-		/* This may be the same req as the previous iteration,
-		 * always send the blk_end_request_all after a prefetch.
-		 * It is not okay to not end the request because the
-		 * prefetch started the request.
-		 */
-		if (req) {
-			/* return -ENXIO to indicate that this queue is
-			 * going away
-			 */
-			req->errors = -ENXIO;
-			blk_end_request_all(req, -ENXIO);
-		}
-
-		msleep(200); /* allow bsg to possibly finish */
-		spin_lock_irq(q->queue_lock);
-	}
-	bsg_unregister_queue(q);
-}
-EXPORT_SYMBOL_GPL(bsg_remove_queue);
diff --git a/drivers/scsi/scsi_transport_fc.c b/drivers/scsi/scsi_transport_fc.c
index 5797604..a9617ad 100644
--- a/drivers/scsi/scsi_transport_fc.c
+++ b/drivers/scsi/scsi_transport_fc.c
@@ -4130,45 +4130,7 @@ fc_bsg_rportadd(struct Scsi_Host *shost, struct fc_rport *rport)
 static void
 fc_bsg_remove(struct request_queue *q)
 {
-	struct request *req; /* block request */
-	int counts; /* totals for request_list count and starved */
-
 	if (q) {
-		/* Stop taking in new requests */
-		spin_lock_irq(q->queue_lock);
-		blk_stop_queue(q);
-
-		/* drain all requests in the queue */
-		while (1) {
-			/* need the lock to fetch a request
-			 * this may fetch the same reqeust as the previous pass
-			 */
-			req = blk_fetch_request(q);
-			/* save requests in use and starved */
-			counts = q->rq.count[0] + q->rq.count[1] +
-				q->rq.starved[0] + q->rq.starved[1];
-			spin_unlock_irq(q->queue_lock);
-			/* any requests still outstanding? */
-			if (counts == 0)
-				break;
-
-			/* This may be the same req as the previous iteration,
-			 * always send the blk_end_request_all after a prefetch.
-			 * It is not okay to not end the request because the
-			 * prefetch started the request.
-			 */
-			if (req) {
-				/* return -ENXIO to indicate that this queue is
-				 * going away
-				 */
-				req->errors = -ENXIO;
-				blk_end_request_all(req, -ENXIO);
-			}
-
-			msleep(200); /* allow bsg to possibly finish */
-			spin_lock_irq(q->queue_lock);
-		}
-
 		bsg_unregister_queue(q);
 		blk_cleanup_queue(q);
 	}
diff --git a/drivers/scsi/scsi_transport_iscsi.c b/drivers/scsi/scsi_transport_iscsi.c
index 1cf640e..c737a16 100644
--- a/drivers/scsi/scsi_transport_iscsi.c
+++ b/drivers/scsi/scsi_transport_iscsi.c
@@ -575,7 +575,7 @@ static int iscsi_remove_host(struct transport_container *tc,
 	struct iscsi_cls_host *ihost = shost->shost_data;
 
 	if (ihost->bsg_q) {
-		bsg_remove_queue(ihost->bsg_q);
+		bsg_unregister_queue(ihost->bsg_q);
 		blk_cleanup_queue(ihost->bsg_q);
 	}
 	return 0;
diff --git a/include/linux/bsg-lib.h b/include/linux/bsg-lib.h
index f55ab8c..4d0fb3d 100644
--- a/include/linux/bsg-lib.h
+++ b/include/linux/bsg-lib.h
@@ -67,7 +67,6 @@ void bsg_job_done(struct bsg_job *job, int result,
 int bsg_setup_queue(struct device *dev, struct request_queue *q, char *name,
 		    bsg_job_fn *job_fn, int dd_job_size);
 void bsg_request_fn(struct request_queue *q);
-void bsg_remove_queue(struct request_queue *q);
 void bsg_goose_queue(struct request_queue *q);
 
 #endif
-- 
1.7.7.3

[PATCH 04/10] block: drop custom queue draining used by scsi_transport_{iscsi|fc}

[Tejun Heo]

iscsi_remove_host() uses bsg_remove_queue() which implements custom
queue draining.  fc_bsg_remove() open-codes mostly identical logic.

The draining logic isn't correct in that blk_stop_queue() doesn't
prevent new requests from being queued - it just stops processing, so
nothing prevents new requests to be queued after the logic determines
that the queue is drained.

blk_cleanup_queue() now implements proper queue draining and these
custom draining logics aren't necessary.  Drop them and use
bsg_unregister_queue() + blk_cleanup_queue() instead.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Mike Christie <michaelc@cs.wisc.edu>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: James Smart <james.smart@emulex.com>
---
 block/bsg-lib.c                     |   53 -----------------------------------
 drivers/scsi/scsi_transport_fc.c    |   38 -------------------------
 drivers/scsi/scsi_transport_iscsi.c |    2 +-
 include/linux/bsg-lib.h             |    1 -
 4 files changed, 1 insertions(+), 93 deletions(-)

diff --git a/block/bsg-lib.c b/block/bsg-lib.c
index 7ad49c8..deee61f 100644
--- a/block/bsg-lib.c
+++ b/block/bsg-lib.c
@@ -243,56 +243,3 @@ int bsg_setup_queue(struct device *dev, struct request_queue *q,
 	return 0;
 }
 EXPORT_SYMBOL_GPL(bsg_setup_queue);
-
-/**
- * bsg_remove_queue - Deletes the bsg dev from the q
- * @q:	the request_queue that is to be torn down.
- *
- * Notes:
- *   Before unregistering the queue empty any requests that are blocked
- */
-void bsg_remove_queue(struct request_queue *q)
-{
-	struct request *req; /* block request */
-	int counts; /* totals for request_list count and starved */
-
-	if (!q)
-		return;
-
-	/* Stop taking in new requests */
-	spin_lock_irq(q->queue_lock);
-	blk_stop_queue(q);
-
-	/* drain all requests in the queue */
-	while (1) {
-		/* need the lock to fetch a request
-		 * this may fetch the same reqeust as the previous pass
-		 */
-		req = blk_fetch_request(q);
-		/* save requests in use and starved */
-		counts = q->rq.count[0] + q->rq.count[1] +
-			 q->rq.starved[0] + q->rq.starved[1];
-		spin_unlock_irq(q->queue_lock);
-		/* any requests still outstanding? */
-		if (counts == 0)
-			break;
-
-		/* This may be the same req as the previous iteration,
-		 * always send the blk_end_request_all after a prefetch.
-		 * It is not okay to not end the request because the
-		 * prefetch started the request.
-		 */
-		if (req) {
-			/* return -ENXIO to indicate that this queue is
-			 * going away
-			 */
-			req->errors = -ENXIO;
-			blk_end_request_all(req, -ENXIO);
-		}
-
-		msleep(200); /* allow bsg to possibly finish */
-		spin_lock_irq(q->queue_lock);
-	}
-	bsg_unregister_queue(q);
-}
-EXPORT_SYMBOL_GPL(bsg_remove_queue);
diff --git a/drivers/scsi/scsi_transport_fc.c b/drivers/scsi/scsi_transport_fc.c
index 5797604..a9617ad 100644
--- a/drivers/scsi/scsi_transport_fc.c
+++ b/drivers/scsi/scsi_transport_fc.c
@@ -4130,45 +4130,7 @@ fc_bsg_rportadd(struct Scsi_Host *shost, struct fc_rport *rport)
 static void
 fc_bsg_remove(struct request_queue *q)
 {
-	struct request *req; /* block request */
-	int counts; /* totals for request_list count and starved */
-
 	if (q) {
-		/* Stop taking in new requests */
-		spin_lock_irq(q->queue_lock);
-		blk_stop_queue(q);
-
-		/* drain all requests in the queue */
-		while (1) {
-			/* need the lock to fetch a request
-			 * this may fetch the same reqeust as the previous pass
-			 */
-			req = blk_fetch_request(q);
-			/* save requests in use and starved */
-			counts = q->rq.count[0] + q->rq.count[1] +
-				q->rq.starved[0] + q->rq.starved[1];
-			spin_unlock_irq(q->queue_lock);
-			/* any requests still outstanding? */
-			if (counts == 0)
-				break;
-
-			/* This may be the same req as the previous iteration,
-			 * always send the blk_end_request_all after a prefetch.
-			 * It is not okay to not end the request because the
-			 * prefetch started the request.
-			 */
-			if (req) {
-				/* return -ENXIO to indicate that this queue is
-				 * going away
-				 */
-				req->errors = -ENXIO;
-				blk_end_request_all(req, -ENXIO);
-			}
-
-			msleep(200); /* allow bsg to possibly finish */
-			spin_lock_irq(q->queue_lock);
-		}
-
 		bsg_unregister_queue(q);
 		blk_cleanup_queue(q);
 	}
diff --git a/drivers/scsi/scsi_transport_iscsi.c b/drivers/scsi/scsi_transport_iscsi.c
index 1cf640e..c737a16 100644
--- a/drivers/scsi/scsi_transport_iscsi.c
+++ b/drivers/scsi/scsi_transport_iscsi.c
@@ -575,7 +575,7 @@ static int iscsi_remove_host(struct transport_container *tc,
 	struct iscsi_cls_host *ihost = shost->shost_data;
 
 	if (ihost->bsg_q) {
-		bsg_remove_queue(ihost->bsg_q);
+		bsg_unregister_queue(ihost->bsg_q);
 		blk_cleanup_queue(ihost->bsg_q);
 	}
 	return 0;
diff --git a/include/linux/bsg-lib.h b/include/linux/bsg-lib.h
index f55ab8c..4d0fb3d 100644
--- a/include/linux/bsg-lib.h
+++ b/include/linux/bsg-lib.h
@@ -67,7 +67,6 @@ void bsg_job_done(struct bsg_job *job, int result,
 int bsg_setup_queue(struct device *dev, struct request_queue *q, char *name,
 		    bsg_job_fn *job_fn, int dd_job_size);
 void bsg_request_fn(struct request_queue *q);
-void bsg_remove_queue(struct request_queue *q);
 void bsg_goose_queue(struct request_queue *q);
 
 #endif
-- 
1.7.7.3

[PATCH 05/10] block: refactor get_request[_wait]()

[Tejun Heo]

Currently, there are two request allocation functions - get_request()
and get_request_wait().  The former tries to allocate a request once
and the latter keeps retrying until it succeeds.  The latter wraps the
former and keeps retrying until allocation succeeds.

The combination of two functions deliver fallible non-wait allocation,
fallible wait allocation and unfailing wait allocation.  However,
given that forward progress is guaranteed, fallible wait allocation
isn't all that useful and in fact nobody uses it.

This patch simplifies the interface as follows.

* get_request() is renamed to __get_request() and is only used by the
  wrapper function.

* get_request_wait() is renamed to get_request().  It now takes
  @gfp_mask and retries iff it contains %__GFP_WAIT.

This patch doesn't introduce any functional change and is to prepare
for further updates to request allocation path.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 block/blk-core.c |   74 +++++++++++++++++++++++++----------------------------
 1 files changed, 35 insertions(+), 39 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 7465d89..9b25d73 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -824,7 +824,7 @@ static struct io_context *rq_ioc(struct bio *bio)
 }
 
 /**
- * get_request - get a free request
+ * __get_request - get a free request
  * @q: request_queue to allocate request from
  * @rw_flags: RW and SYNC flags
  * @bio: bio to allocate request for (can be %NULL)
@@ -837,8 +837,8 @@ static struct io_context *rq_ioc(struct bio *bio)
  * Returns %NULL on failure, with @q->queue_lock held.
  * Returns !%NULL on success, with @q->queue_lock *not held*.
  */
-static struct request *get_request(struct request_queue *q, int rw_flags,
-				   struct bio *bio, gfp_t gfp_mask)
+static struct request *__get_request(struct request_queue *q, int rw_flags,
+				     struct bio *bio, gfp_t gfp_mask)
 {
 	struct request *rq;
 	struct request_list *rl = &q->rq;
@@ -1016,56 +1016,55 @@ rq_starved:
 }
 
 /**
- * get_request_wait - get a free request with retry
+ * get_request - get a free request
  * @q: request_queue to allocate request from
  * @rw_flags: RW and SYNC flags
  * @bio: bio to allocate request for (can be %NULL)
+ * @gfp_mask: allocation mask
  *
- * Get a free request from @q.  This function keeps retrying under memory
- * pressure and fails iff @q is dead.
+ * Get a free request from @q.  If %__GFP_WAIT is set in @gfp_mask, this
+ * function keeps retrying under memory pressure and fails iff @q is dead.
  *
  * Must be callled with @q->queue_lock held and,
  * Returns %NULL on failure, with @q->queue_lock held.
  * Returns !%NULL on success, with @q->queue_lock *not held*.
  */
-static struct request *get_request_wait(struct request_queue *q, int rw_flags,
-					struct bio *bio)
+static struct request *get_request(struct request_queue *q, int rw_flags,
+				   struct bio *bio, gfp_t gfp_mask)
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
+	DEFINE_WAIT(wait);
+	struct request_list *rl = &q->rq;
 	struct request *rq;
+retry:
+	rq = __get_request(q, rw_flags, bio, gfp_mask);
+	if (rq)
+		return rq;
 
-	rq = get_request(q, rw_flags, bio, GFP_NOIO);
-	while (!rq) {
-		DEFINE_WAIT(wait);
-		struct request_list *rl = &q->rq;
-
-		if (unlikely(blk_queue_dead(q)))
-			return NULL;
-
-		prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
-				TASK_UNINTERRUPTIBLE);
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+		return NULL;
 
-		trace_block_sleeprq(q, bio, rw_flags & 1);
+	/* wait on @rl and retry */
+	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
+				  TASK_UNINTERRUPTIBLE);
 
-		spin_unlock_irq(q->queue_lock);
-		io_schedule();
+	trace_block_sleeprq(q, bio, rw_flags & 1);
 
-		/*
-		 * After sleeping, we become a "batching" process and
-		 * will be able to allocate at least one request, and
-		 * up to a big batch of them for a small period time.
-		 * See ioc_batching, ioc_set_batching
-		 */
-		create_io_context(GFP_NOIO, q->node);
-		ioc_set_batching(q, current->io_context);
+	spin_unlock_irq(q->queue_lock);
+	io_schedule();
 
-		spin_lock_irq(q->queue_lock);
-		finish_wait(&rl->wait[is_sync], &wait);
+	/*
+	 * After sleeping, we become a "batching" process and will be able
+	 * to allocate at least one request, and up to a big batch of them
+	 * for a small period time.  See ioc_batching, ioc_set_batching
+	 */
+	create_io_context(GFP_NOIO, q->node);
+	ioc_set_batching(q, current->io_context);
 
-		rq = get_request(q, rw_flags, bio, GFP_NOIO);
-	};
+	spin_lock_irq(q->queue_lock);
+	finish_wait(&rl->wait[is_sync], &wait);
 
-	return rq;
+	goto retry;
 }
 
 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
@@ -1075,10 +1074,7 @@ struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
 	BUG_ON(rw != READ && rw != WRITE);
 
 	spin_lock_irq(q->queue_lock);
-	if (gfp_mask & __GFP_WAIT)
-		rq = get_request_wait(q, rw, NULL);
-	else
-		rq = get_request(q, rw, NULL, gfp_mask);
+	rq = get_request(q, rw, NULL, gfp_mask);
 	if (!rq)
 		spin_unlock_irq(q->queue_lock);
 	/* q->queue_lock is unlocked at this point */
@@ -1468,7 +1464,7 @@ get_rq:
 	 * Grab a free request. This is might sleep but can not fail.
 	 * Returns with the queue unlocked.
 	 */
-	req = get_request_wait(q, rw_flags, bio);
+	req = get_request(q, rw_flags, bio, GFP_NOIO);
 	if (unlikely(!req)) {
 		bio_endio(bio, -ENODEV);	/* @q is dead */
 		goto out_unlock;
-- 
1.7.7.3

[PATCH 05/10] block: refactor get_request[_wait]()

[Tejun Heo]

Currently, there are two request allocation functions - get_request()
and get_request_wait().  The former tries to allocate a request once
and the latter keeps retrying until it succeeds.  The latter wraps the
former and keeps retrying until allocation succeeds.

The combination of two functions deliver fallible non-wait allocation,
fallible wait allocation and unfailing wait allocation.  However,
given that forward progress is guaranteed, fallible wait allocation
isn't all that useful and in fact nobody uses it.

This patch simplifies the interface as follows.

* get_request() is renamed to __get_request() and is only used by the
  wrapper function.

* get_request_wait() is renamed to get_request().  It now takes
  @gfp_mask and retries iff it contains %__GFP_WAIT.

This patch doesn't introduce any functional change and is to prepare
for further updates to request allocation path.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-core.c |   74 +++++++++++++++++++++++++----------------------------
 1 files changed, 35 insertions(+), 39 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 7465d89..9b25d73 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -824,7 +824,7 @@ static struct io_context *rq_ioc(struct bio *bio)
 }
 
 /**
- * get_request - get a free request
+ * __get_request - get a free request
  * @q: request_queue to allocate request from
  * @rw_flags: RW and SYNC flags
  * @bio: bio to allocate request for (can be %NULL)
@@ -837,8 +837,8 @@ static struct io_context *rq_ioc(struct bio *bio)
  * Returns %NULL on failure, with @q->queue_lock held.
  * Returns !%NULL on success, with @q->queue_lock *not held*.
  */
-static struct request *get_request(struct request_queue *q, int rw_flags,
-				   struct bio *bio, gfp_t gfp_mask)
+static struct request *__get_request(struct request_queue *q, int rw_flags,
+				     struct bio *bio, gfp_t gfp_mask)
 {
 	struct request *rq;
 	struct request_list *rl = &q->rq;
@@ -1016,56 +1016,55 @@ rq_starved:
 }
 
 /**
- * get_request_wait - get a free request with retry
+ * get_request - get a free request
  * @q: request_queue to allocate request from
  * @rw_flags: RW and SYNC flags
  * @bio: bio to allocate request for (can be %NULL)
+ * @gfp_mask: allocation mask
  *
- * Get a free request from @q.  This function keeps retrying under memory
- * pressure and fails iff @q is dead.
+ * Get a free request from @q.  If %__GFP_WAIT is set in @gfp_mask, this
+ * function keeps retrying under memory pressure and fails iff @q is dead.
  *
  * Must be callled with @q->queue_lock held and,
  * Returns %NULL on failure, with @q->queue_lock held.
  * Returns !%NULL on success, with @q->queue_lock *not held*.
  */
-static struct request *get_request_wait(struct request_queue *q, int rw_flags,
-					struct bio *bio)
+static struct request *get_request(struct request_queue *q, int rw_flags,
+				   struct bio *bio, gfp_t gfp_mask)
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
+	DEFINE_WAIT(wait);
+	struct request_list *rl = &q->rq;
 	struct request *rq;
+retry:
+	rq = __get_request(q, rw_flags, bio, gfp_mask);
+	if (rq)
+		return rq;
 
-	rq = get_request(q, rw_flags, bio, GFP_NOIO);
-	while (!rq) {
-		DEFINE_WAIT(wait);
-		struct request_list *rl = &q->rq;
-
-		if (unlikely(blk_queue_dead(q)))
-			return NULL;
-
-		prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
-				TASK_UNINTERRUPTIBLE);
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+		return NULL;
 
-		trace_block_sleeprq(q, bio, rw_flags & 1);
+	/* wait on @rl and retry */
+	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
+				  TASK_UNINTERRUPTIBLE);
 
-		spin_unlock_irq(q->queue_lock);
-		io_schedule();
+	trace_block_sleeprq(q, bio, rw_flags & 1);
 
-		/*
-		 * After sleeping, we become a "batching" process and
-		 * will be able to allocate at least one request, and
-		 * up to a big batch of them for a small period time.
-		 * See ioc_batching, ioc_set_batching
-		 */
-		create_io_context(GFP_NOIO, q->node);
-		ioc_set_batching(q, current->io_context);
+	spin_unlock_irq(q->queue_lock);
+	io_schedule();
 
-		spin_lock_irq(q->queue_lock);
-		finish_wait(&rl->wait[is_sync], &wait);
+	/*
+	 * After sleeping, we become a "batching" process and will be able
+	 * to allocate at least one request, and up to a big batch of them
+	 * for a small period time.  See ioc_batching, ioc_set_batching
+	 */
+	create_io_context(GFP_NOIO, q->node);
+	ioc_set_batching(q, current->io_context);
 
-		rq = get_request(q, rw_flags, bio, GFP_NOIO);
-	};
+	spin_lock_irq(q->queue_lock);
+	finish_wait(&rl->wait[is_sync], &wait);
 
-	return rq;
+	goto retry;
 }
 
 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
@@ -1075,10 +1074,7 @@ struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
 	BUG_ON(rw != READ && rw != WRITE);
 
 	spin_lock_irq(q->queue_lock);
-	if (gfp_mask & __GFP_WAIT)
-		rq = get_request_wait(q, rw, NULL);
-	else
-		rq = get_request(q, rw, NULL, gfp_mask);
+	rq = get_request(q, rw, NULL, gfp_mask);
 	if (!rq)
 		spin_unlock_irq(q->queue_lock);
 	/* q->queue_lock is unlocked at this point */
@@ -1468,7 +1464,7 @@ get_rq:
 	 * Grab a free request. This is might sleep but can not fail.
 	 * Returns with the queue unlocked.
 	 */
-	req = get_request_wait(q, rw_flags, bio);
+	req = get_request(q, rw_flags, bio, GFP_NOIO);
 	if (unlikely(!req)) {
 		bio_endio(bio, -ENODEV);	/* @q is dead */
 		goto out_unlock;
-- 
1.7.7.3

[PATCH 06/10] block: allocate io_context upfront

[Tejun Heo]

Block layer very lazy allocation of ioc.  It waits until the moment
ioc is absolutely necessary; unfortunately, that time could be inside
queue lock and __get_request() performs unlock - try alloc - retry
dancing.

Just allocate it up-front on entry to block layer.  We're not saving
the rain forest by deferring it to the last possible moment and
complicating things unnecessarily.

This patch is to prepare for further updates to request allocation
path.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 block/blk-core.c     |   42 +++++++++++++++---------------------------
 block/blk-throttle.c |    3 ---
 2 files changed, 15 insertions(+), 30 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 9b25d73..ada4bc0 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -842,15 +842,11 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 {
 	struct request *rq;
 	struct request_list *rl = &q->rq;
-	struct elevator_type *et;
-	struct io_context *ioc;
+	struct elevator_type *et = q->elevator->type;
+	struct io_context *ioc = rq_ioc(bio);
 	struct io_cq *icq = NULL;
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
-	bool retried = false;
 	int may_queue;
-retry:
-	et = q->elevator->type;
-	ioc = rq_ioc(bio);
 
 	if (unlikely(blk_queue_dead(q)))
 		return NULL;
@@ -862,20 +858,6 @@ retry:
 	if (rl->count[is_sync]+1 >= queue_congestion_on_threshold(q)) {
 		if (rl->count[is_sync]+1 >= q->nr_requests) {
 			/*
-			 * We want ioc to record batching state.  If it's
-			 * not already there, creating a new one requires
-			 * dropping queue_lock, which in turn requires
-			 * retesting conditions to avoid queue hang.
-			 */
-			if (!ioc && !retried) {
-				spin_unlock_irq(q->queue_lock);
-				create_io_context(gfp_mask, q->node);
-				spin_lock_irq(q->queue_lock);
-				retried = true;
-				goto retry;
-			}
-
-			/*
 			 * The queue will fill after this allocation, so set
 			 * it as full, and mark this process as "batching".
 			 * This process will be allowed to complete a batch of
@@ -942,12 +924,8 @@ retry:
 	/* init elvpriv */
 	if (rw_flags & REQ_ELVPRIV) {
 		if (unlikely(et->icq_cache && !icq)) {
-			create_io_context(gfp_mask, q->node);
-			ioc = rq_ioc(bio);
-			if (!ioc)
-				goto fail_elvpriv;
-
-			icq = ioc_create_icq(ioc, q, gfp_mask);
+			if (ioc)
+				icq = ioc_create_icq(ioc, q, gfp_mask);
 			if (!icq)
 				goto fail_elvpriv;
 		}
@@ -1058,7 +1036,6 @@ retry:
 	 * to allocate at least one request, and up to a big batch of them
 	 * for a small period time.  See ioc_batching, ioc_set_batching
 	 */
-	create_io_context(GFP_NOIO, q->node);
 	ioc_set_batching(q, current->io_context);
 
 	spin_lock_irq(q->queue_lock);
@@ -1073,6 +1050,9 @@ struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
 
 	BUG_ON(rw != READ && rw != WRITE);
 
+	/* create ioc upfront */
+	create_io_context(gfp_mask, q->node);
+
 	spin_lock_irq(q->queue_lock);
 	rq = get_request(q, rw, NULL, gfp_mask);
 	if (!rq)
@@ -1685,6 +1665,14 @@ generic_make_request_checks(struct bio *bio)
 		goto end_io;
 	}
 
+	/*
+	 * Various block parts want %current->io_context and lazy ioc
+	 * allocation ends up trading a lot of pain for a small amount of
+	 * memory.  Just allocate it upfront.  This may fail and block
+	 * layer knows how to live with it.
+	 */
+	create_io_context(GFP_ATOMIC, q->node);
+
 	if (blk_throtl_bio(q, bio))
 		return false;	/* throttled, will be resubmitted later */
 
diff --git a/block/blk-throttle.c b/block/blk-throttle.c
index 5b06595..e287c19 100644
--- a/block/blk-throttle.c
+++ b/block/blk-throttle.c
@@ -1123,9 +1123,6 @@ bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
 		goto out;
 	}
 
-	/* bio_associate_current() needs ioc, try creating */
-	create_io_context(GFP_ATOMIC, q->node);
-
 	/*
 	 * A throtl_grp pointer retrieved under rcu can be used to access
 	 * basic fields like stats and io rates. If a group has no rules,
-- 
1.7.7.3

[PATCH 06/10] block: allocate io_context upfront

[Tejun Heo]

Block layer very lazy allocation of ioc.  It waits until the moment
ioc is absolutely necessary; unfortunately, that time could be inside
queue lock and __get_request() performs unlock - try alloc - retry
dancing.

Just allocate it up-front on entry to block layer.  We're not saving
the rain forest by deferring it to the last possible moment and
complicating things unnecessarily.

This patch is to prepare for further updates to request allocation
path.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-core.c     |   42 +++++++++++++++---------------------------
 block/blk-throttle.c |    3 ---
 2 files changed, 15 insertions(+), 30 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 9b25d73..ada4bc0 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -842,15 +842,11 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 {
 	struct request *rq;
 	struct request_list *rl = &q->rq;
-	struct elevator_type *et;
-	struct io_context *ioc;
+	struct elevator_type *et = q->elevator->type;
+	struct io_context *ioc = rq_ioc(bio);
 	struct io_cq *icq = NULL;
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
-	bool retried = false;
 	int may_queue;
-retry:
-	et = q->elevator->type;
-	ioc = rq_ioc(bio);
 
 	if (unlikely(blk_queue_dead(q)))
 		return NULL;
@@ -862,20 +858,6 @@ retry:
 	if (rl->count[is_sync]+1 >= queue_congestion_on_threshold(q)) {
 		if (rl->count[is_sync]+1 >= q->nr_requests) {
 			/*
-			 * We want ioc to record batching state.  If it's
-			 * not already there, creating a new one requires
-			 * dropping queue_lock, which in turn requires
-			 * retesting conditions to avoid queue hang.
-			 */
-			if (!ioc && !retried) {
-				spin_unlock_irq(q->queue_lock);
-				create_io_context(gfp_mask, q->node);
-				spin_lock_irq(q->queue_lock);
-				retried = true;
-				goto retry;
-			}
-
-			/*
 			 * The queue will fill after this allocation, so set
 			 * it as full, and mark this process as "batching".
 			 * This process will be allowed to complete a batch of
@@ -942,12 +924,8 @@ retry:
 	/* init elvpriv */
 	if (rw_flags & REQ_ELVPRIV) {
 		if (unlikely(et->icq_cache && !icq)) {
-			create_io_context(gfp_mask, q->node);
-			ioc = rq_ioc(bio);
-			if (!ioc)
-				goto fail_elvpriv;
-
-			icq = ioc_create_icq(ioc, q, gfp_mask);
+			if (ioc)
+				icq = ioc_create_icq(ioc, q, gfp_mask);
 			if (!icq)
 				goto fail_elvpriv;
 		}
@@ -1058,7 +1036,6 @@ retry:
 	 * to allocate at least one request, and up to a big batch of them
 	 * for a small period time.  See ioc_batching, ioc_set_batching
 	 */
-	create_io_context(GFP_NOIO, q->node);
 	ioc_set_batching(q, current->io_context);
 
 	spin_lock_irq(q->queue_lock);
@@ -1073,6 +1050,9 @@ struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
 
 	BUG_ON(rw != READ && rw != WRITE);
 
+	/* create ioc upfront */
+	create_io_context(gfp_mask, q->node);
+
 	spin_lock_irq(q->queue_lock);
 	rq = get_request(q, rw, NULL, gfp_mask);
 	if (!rq)
@@ -1685,6 +1665,14 @@ generic_make_request_checks(struct bio *bio)
 		goto end_io;
 	}
 
+	/*
+	 * Various block parts want %current->io_context and lazy ioc
+	 * allocation ends up trading a lot of pain for a small amount of
+	 * memory.  Just allocate it upfront.  This may fail and block
+	 * layer knows how to live with it.
+	 */
+	create_io_context(GFP_ATOMIC, q->node);
+
 	if (blk_throtl_bio(q, bio))
 		return false;	/* throttled, will be resubmitted later */
 
diff --git a/block/blk-throttle.c b/block/blk-throttle.c
index 5b06595..e287c19 100644
--- a/block/blk-throttle.c
+++ b/block/blk-throttle.c
@@ -1123,9 +1123,6 @@ bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
 		goto out;
 	}
 
-	/* bio_associate_current() needs ioc, try creating */
-	create_io_context(GFP_ATOMIC, q->node);
-
 	/*
 	 * A throtl_grp pointer retrieved under rcu can be used to access
 	 * basic fields like stats and io rates. If a group has no rules,
-- 
1.7.7.3

[PATCH 07/10] blkcg: inline bio_blkcg() and friends

[Tejun Heo]

Make bio_blkcg() and friends inline.  They all are very simple and
used only in few places.

This patch is to prepare for further updates to request allocation
path.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 block/blk-cgroup.c |   21 ---------------------
 block/blk-cgroup.h |   26 ++++++++++++++++++++++----
 2 files changed, 22 insertions(+), 25 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index cbeeb54..a48f35b 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -31,27 +31,6 @@ EXPORT_SYMBOL_GPL(blkcg_root);
 
 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
 
-struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup)
-{
-	return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
-			    struct blkcg, css);
-}
-EXPORT_SYMBOL_GPL(cgroup_to_blkcg);
-
-static struct blkcg *task_blkcg(struct task_struct *tsk)
-{
-	return container_of(task_subsys_state(tsk, blkio_subsys_id),
-			    struct blkcg, css);
-}
-
-struct blkcg *bio_blkcg(struct bio *bio)
-{
-	if (bio && bio->bi_css)
-		return container_of(bio->bi_css, struct blkcg, css);
-	return task_blkcg(current);
-}
-EXPORT_SYMBOL_GPL(bio_blkcg);
-
 static bool blkcg_policy_enabled(struct request_queue *q,
 				 const struct blkcg_policy *pol)
 {
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index 8ac457c..e74cce1 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -120,8 +120,6 @@ struct blkcg_policy {
 
 extern struct blkcg blkcg_root;
 
-struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup);
-struct blkcg *bio_blkcg(struct bio *bio);
 struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q);
 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
 				    struct request_queue *q);
@@ -160,6 +158,25 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
 void blkg_conf_finish(struct blkg_conf_ctx *ctx);
 
 
+static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup)
+{
+	return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
+			    struct blkcg, css);
+}
+
+static inline struct blkcg *task_blkcg(struct task_struct *tsk)
+{
+	return container_of(task_subsys_state(tsk, blkio_subsys_id),
+			    struct blkcg, css);
+}
+
+static inline struct blkcg *bio_blkcg(struct bio *bio)
+{
+	if (bio && bio->bi_css)
+		return container_of(bio->bi_css, struct blkcg, css);
+	return task_blkcg(current);
+}
+
 /**
  * blkg_to_pdata - get policy private data
  * @blkg: blkg of interest
@@ -351,6 +368,7 @@ static inline void blkg_rwstat_reset(struct blkg_rwstat *rwstat)
 #else	/* CONFIG_BLK_CGROUP */
 
 struct cgroup;
+struct blkcg;
 
 struct blkg_policy_data {
 };
@@ -361,8 +379,6 @@ struct blkcg_gq {
 struct blkcg_policy {
 };
 
-static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
-static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
 static inline struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, void *key) { return NULL; }
 static inline int blkcg_init_queue(struct request_queue *q) { return 0; }
 static inline void blkcg_drain_queue(struct request_queue *q) { }
@@ -374,6 +390,8 @@ static inline int blkcg_activate_policy(struct request_queue *q,
 static inline void blkcg_deactivate_policy(struct request_queue *q,
 					   const struct blkcg_policy *pol) { }
 
+static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
+static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
-- 
1.7.7.3

[PATCH 07/10] blkcg: inline bio_blkcg() and friends

[Tejun Heo]

Make bio_blkcg() and friends inline.  They all are very simple and
used only in few places.

This patch is to prepare for further updates to request allocation
path.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-cgroup.c |   21 ---------------------
 block/blk-cgroup.h |   26 ++++++++++++++++++++++----
 2 files changed, 22 insertions(+), 25 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index cbeeb54..a48f35b 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -31,27 +31,6 @@ EXPORT_SYMBOL_GPL(blkcg_root);
 
 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
 
-struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup)
-{
-	return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
-			    struct blkcg, css);
-}
-EXPORT_SYMBOL_GPL(cgroup_to_blkcg);
-
-static struct blkcg *task_blkcg(struct task_struct *tsk)
-{
-	return container_of(task_subsys_state(tsk, blkio_subsys_id),
-			    struct blkcg, css);
-}
-
-struct blkcg *bio_blkcg(struct bio *bio)
-{
-	if (bio && bio->bi_css)
-		return container_of(bio->bi_css, struct blkcg, css);
-	return task_blkcg(current);
-}
-EXPORT_SYMBOL_GPL(bio_blkcg);
-
 static bool blkcg_policy_enabled(struct request_queue *q,
 				 const struct blkcg_policy *pol)
 {
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index 8ac457c..e74cce1 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -120,8 +120,6 @@ struct blkcg_policy {
 
 extern struct blkcg blkcg_root;
 
-struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup);
-struct blkcg *bio_blkcg(struct bio *bio);
 struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q);
 struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
 				    struct request_queue *q);
@@ -160,6 +158,25 @@ int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
 void blkg_conf_finish(struct blkg_conf_ctx *ctx);
 
 
+static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup)
+{
+	return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
+			    struct blkcg, css);
+}
+
+static inline struct blkcg *task_blkcg(struct task_struct *tsk)
+{
+	return container_of(task_subsys_state(tsk, blkio_subsys_id),
+			    struct blkcg, css);
+}
+
+static inline struct blkcg *bio_blkcg(struct bio *bio)
+{
+	if (bio && bio->bi_css)
+		return container_of(bio->bi_css, struct blkcg, css);
+	return task_blkcg(current);
+}
+
 /**
  * blkg_to_pdata - get policy private data
  * @blkg: blkg of interest
@@ -351,6 +368,7 @@ static inline void blkg_rwstat_reset(struct blkg_rwstat *rwstat)
 #else	/* CONFIG_BLK_CGROUP */
 
 struct cgroup;
+struct blkcg;
 
 struct blkg_policy_data {
 };
@@ -361,8 +379,6 @@ struct blkcg_gq {
 struct blkcg_policy {
 };
 
-static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
-static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
 static inline struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, void *key) { return NULL; }
 static inline int blkcg_init_queue(struct request_queue *q) { return 0; }
 static inline void blkcg_drain_queue(struct request_queue *q) { }
@@ -374,6 +390,8 @@ static inline int blkcg_activate_policy(struct request_queue *q,
 static inline void blkcg_deactivate_policy(struct request_queue *q,
 					   const struct blkcg_policy *pol) { }
 
+static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
+static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
-- 
1.7.7.3

[PATCH 08/10] block: add q->nr_rqs[] and move q->rq.elvpriv to q->nr_rqs_elvpriv

[Tejun Heo]

Add q->nr_rqs[] which currently behaves the same as q->rq.count[] and
move q->rq.elvpriv to q->nr_rqs_elvpriv.  blk_drain_queue() is updated
to use q->nr_rqs[] instead of q->rq.count[].

These counters separates queue-wide request statistics from the
request list and allow implementation of per-queue request allocation.

While at it, properly indent fields of struct request_list.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 block/blk-core.c       |   13 +++++++------
 include/linux/blkdev.h |   11 ++++++-----
 2 files changed, 13 insertions(+), 11 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index ada4bc0..8149c76 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -386,7 +386,7 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 		if (!list_empty(&q->queue_head) && q->request_fn)
 			__blk_run_queue(q);
 
-		drain |= q->rq.elvpriv;
+		drain |= q->nr_rqs_elvpriv;
 
 		/*
 		 * Unfortunately, requests are queued at and tracked from
@@ -396,7 +396,7 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 		if (drain_all) {
 			drain |= !list_empty(&q->queue_head);
 			for (i = 0; i < 2; i++) {
-				drain |= q->rq.count[i];
+				drain |= q->nr_rqs[i];
 				drain |= q->in_flight[i];
 				drain |= !list_empty(&q->flush_queue[i]);
 			}
@@ -513,7 +513,6 @@ static int blk_init_free_list(struct request_queue *q)
 
 	rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0;
 	rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0;
-	rl->elvpriv = 0;
 	init_waitqueue_head(&rl->wait[BLK_RW_SYNC]);
 	init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]);
 
@@ -778,9 +777,10 @@ static void freed_request(struct request_queue *q, unsigned int flags)
 	struct request_list *rl = &q->rq;
 	int sync = rw_is_sync(flags);
 
+	q->nr_rqs[sync]--;
 	rl->count[sync]--;
 	if (flags & REQ_ELVPRIV)
-		rl->elvpriv--;
+		q->nr_rqs_elvpriv--;
 
 	__freed_request(q, sync);
 
@@ -889,6 +889,7 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 	if (rl->count[is_sync] >= (3 * q->nr_requests / 2))
 		return NULL;
 
+	q->nr_rqs[is_sync]++;
 	rl->count[is_sync]++;
 	rl->starved[is_sync] = 0;
 
@@ -904,7 +905,7 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 	 */
 	if (blk_rq_should_init_elevator(bio) && !blk_queue_bypass(q)) {
 		rw_flags |= REQ_ELVPRIV;
-		rl->elvpriv++;
+		q->nr_rqs_elvpriv++;
 		if (et->icq_cache && ioc)
 			icq = ioc_lookup_icq(ioc, q);
 	}
@@ -965,7 +966,7 @@ fail_elvpriv:
 	rq->elv.icq = NULL;
 
 	spin_lock_irq(q->queue_lock);
-	rl->elvpriv--;
+	q->nr_rqs_elvpriv--;
 	spin_unlock_irq(q->queue_lock);
 	goto out;
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index ba43f40..0c6f527 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,11 +51,10 @@ struct request_list {
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
 	 */
-	int count[2];
-	int starved[2];
-	int elvpriv;
-	mempool_t *rq_pool;
-	wait_queue_head_t wait[2];
+	int			count[2];
+	int			starved[2];
+	mempool_t		*rq_pool;
+	wait_queue_head_t	wait[2];
 };
 
 /*
@@ -282,6 +281,8 @@ struct request_queue {
 	struct list_head	queue_head;
 	struct request		*last_merge;
 	struct elevator_queue	*elevator;
+	int			nr_rqs[2];	/* # allocated [a]sync rqs */
+	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
 	 * the queue request freelist, one for reads and one for writes
-- 
1.7.7.3

[PATCH 08/10] block: add q->nr_rqs[] and move q->rq.elvpriv to q->nr_rqs_elvpriv

[Tejun Heo]

Add q->nr_rqs[] which currently behaves the same as q->rq.count[] and
move q->rq.elvpriv to q->nr_rqs_elvpriv.  blk_drain_queue() is updated
to use q->nr_rqs[] instead of q->rq.count[].

These counters separates queue-wide request statistics from the
request list and allow implementation of per-queue request allocation.

While at it, properly indent fields of struct request_list.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-core.c       |   13 +++++++------
 include/linux/blkdev.h |   11 ++++++-----
 2 files changed, 13 insertions(+), 11 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index ada4bc0..8149c76 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -386,7 +386,7 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 		if (!list_empty(&q->queue_head) && q->request_fn)
 			__blk_run_queue(q);
 
-		drain |= q->rq.elvpriv;
+		drain |= q->nr_rqs_elvpriv;
 
 		/*
 		 * Unfortunately, requests are queued at and tracked from
@@ -396,7 +396,7 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 		if (drain_all) {
 			drain |= !list_empty(&q->queue_head);
 			for (i = 0; i < 2; i++) {
-				drain |= q->rq.count[i];
+				drain |= q->nr_rqs[i];
 				drain |= q->in_flight[i];
 				drain |= !list_empty(&q->flush_queue[i]);
 			}
@@ -513,7 +513,6 @@ static int blk_init_free_list(struct request_queue *q)
 
 	rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0;
 	rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0;
-	rl->elvpriv = 0;
 	init_waitqueue_head(&rl->wait[BLK_RW_SYNC]);
 	init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]);
 
@@ -778,9 +777,10 @@ static void freed_request(struct request_queue *q, unsigned int flags)
 	struct request_list *rl = &q->rq;
 	int sync = rw_is_sync(flags);
 
+	q->nr_rqs[sync]--;
 	rl->count[sync]--;
 	if (flags & REQ_ELVPRIV)
-		rl->elvpriv--;
+		q->nr_rqs_elvpriv--;
 
 	__freed_request(q, sync);
 
@@ -889,6 +889,7 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 	if (rl->count[is_sync] >= (3 * q->nr_requests / 2))
 		return NULL;
 
+	q->nr_rqs[is_sync]++;
 	rl->count[is_sync]++;
 	rl->starved[is_sync] = 0;
 
@@ -904,7 +905,7 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 	 */
 	if (blk_rq_should_init_elevator(bio) && !blk_queue_bypass(q)) {
 		rw_flags |= REQ_ELVPRIV;
-		rl->elvpriv++;
+		q->nr_rqs_elvpriv++;
 		if (et->icq_cache && ioc)
 			icq = ioc_lookup_icq(ioc, q);
 	}
@@ -965,7 +966,7 @@ fail_elvpriv:
 	rq->elv.icq = NULL;
 
 	spin_lock_irq(q->queue_lock);
-	rl->elvpriv--;
+	q->nr_rqs_elvpriv--;
 	spin_unlock_irq(q->queue_lock);
 	goto out;
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index ba43f40..0c6f527 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,11 +51,10 @@ struct request_list {
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
 	 */
-	int count[2];
-	int starved[2];
-	int elvpriv;
-	mempool_t *rq_pool;
-	wait_queue_head_t wait[2];
+	int			count[2];
+	int			starved[2];
+	mempool_t		*rq_pool;
+	wait_queue_head_t	wait[2];
 };
 
 /*
@@ -282,6 +281,8 @@ struct request_queue {
 	struct list_head	queue_head;
 	struct request		*last_merge;
 	struct elevator_queue	*elevator;
+	int			nr_rqs[2];	/* # allocated [a]sync rqs */
+	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
 	 * the queue request freelist, one for reads and one for writes
-- 
1.7.7.3

[PATCH 09/10] block: prepare for multiple request_lists

[Tejun Heo]

Request allocation is about to be made per-blkg meaning that there'll
be multiple request lists.

* Make queue full state per request_list.  blk_*queue_full() functions
  are renamed to blk_*rl_full() and takes @rl instead of @q.

* Rename blk_init_free_list() to blk_init_rl() and make it take @rl
  instead of @q.  Also add @gfp_mask parameter.

* Add blk_exit_rl() instead of destroying rl directly from
  blk_release_queue().

* Add request_list->q and make request alloc/free functions -
  blk_free_request(), [__]freed_request(), __get_request() - take @rl
  instead of @q.

This patch doesn't introduce any functional difference.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 block/blk-core.c       |   56 ++++++++++++++++++++++++++---------------------
 block/blk-sysfs.c      |   12 ++++------
 block/blk.h            |    3 ++
 include/linux/blkdev.h |   32 +++++++++++++++------------
 4 files changed, 57 insertions(+), 46 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 8149c76..0fbd233 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -504,13 +504,13 @@ void blk_cleanup_queue(struct request_queue *q)
 }
 EXPORT_SYMBOL(blk_cleanup_queue);
 
-static int blk_init_free_list(struct request_queue *q)
+int blk_init_rl(struct request_list *rl, struct request_queue *q,
+		gfp_t gfp_mask)
 {
-	struct request_list *rl = &q->rq;
-
 	if (unlikely(rl->rq_pool))
 		return 0;
 
+	rl->q = q;
 	rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0;
 	rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0;
 	init_waitqueue_head(&rl->wait[BLK_RW_SYNC]);
@@ -518,13 +518,19 @@ static int blk_init_free_list(struct request_queue *q)
 
 	rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
 					  mempool_free_slab, request_cachep,
-					  GFP_KERNEL, q->node);
+					  gfp_mask, q->node);
 	if (!rl->rq_pool)
 		return -ENOMEM;
 
 	return 0;
 }
 
+void blk_exit_rl(struct request_list *rl)
+{
+	if (rl->rq_pool)
+		mempool_destroy(rl->rq_pool);
+}
+
 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
 {
 	return blk_alloc_queue_node(gfp_mask, -1);
@@ -666,7 +672,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_free_list(q))
+	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -708,15 +714,15 @@ bool blk_get_queue(struct request_queue *q)
 }
 EXPORT_SYMBOL(blk_get_queue);
 
-static inline void blk_free_request(struct request_queue *q, struct request *rq)
+static inline void blk_free_request(struct request_list *rl, struct request *rq)
 {
 	if (rq->cmd_flags & REQ_ELVPRIV) {
-		elv_put_request(q, rq);
+		elv_put_request(rl->q, rq);
 		if (rq->elv.icq)
 			put_io_context(rq->elv.icq->ioc);
 	}
 
-	mempool_free(rq, q->rq.rq_pool);
+	mempool_free(rq, rl->rq_pool);
 }
 
 /*
@@ -753,9 +759,9 @@ static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
 	ioc->last_waited = jiffies;
 }
 
-static void __freed_request(struct request_queue *q, int sync)
+static void __freed_request(struct request_list *rl, int sync)
 {
-	struct request_list *rl = &q->rq;
+	struct request_queue *q = rl->q;
 
 	if (rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
@@ -764,7 +770,7 @@ static void __freed_request(struct request_queue *q, int sync)
 		if (waitqueue_active(&rl->wait[sync]))
 			wake_up(&rl->wait[sync]);
 
-		blk_clear_queue_full(q, sync);
+		blk_clear_rl_full(rl, sync);
 	}
 }
 
@@ -772,9 +778,9 @@ static void __freed_request(struct request_queue *q, int sync)
  * A request has just been released.  Account for it, update the full and
  * congestion status, wake up any waiters.   Called under q->queue_lock.
  */
-static void freed_request(struct request_queue *q, unsigned int flags)
+static void freed_request(struct request_list *rl, unsigned int flags)
 {
-	struct request_list *rl = &q->rq;
+	struct request_queue *q = rl->q;
 	int sync = rw_is_sync(flags);
 
 	q->nr_rqs[sync]--;
@@ -782,10 +788,10 @@ static void freed_request(struct request_queue *q, unsigned int flags)
 	if (flags & REQ_ELVPRIV)
 		q->nr_rqs_elvpriv--;
 
-	__freed_request(q, sync);
+	__freed_request(rl, sync);
 
 	if (unlikely(rl->starved[sync ^ 1]))
-		__freed_request(q, sync ^ 1);
+		__freed_request(rl, sync ^ 1);
 }
 
 /*
@@ -825,7 +831,7 @@ static struct io_context *rq_ioc(struct bio *bio)
 
 /**
  * __get_request - get a free request
- * @q: request_queue to allocate request from
+ * @rl: request list to allocate from
  * @rw_flags: RW and SYNC flags
  * @bio: bio to allocate request for (can be %NULL)
  * @gfp_mask: allocation mask
@@ -837,11 +843,11 @@ static struct io_context *rq_ioc(struct bio *bio)
  * Returns %NULL on failure, with @q->queue_lock held.
  * Returns !%NULL on success, with @q->queue_lock *not held*.
  */
-static struct request *__get_request(struct request_queue *q, int rw_flags,
+static struct request *__get_request(struct request_list *rl, int rw_flags,
 				     struct bio *bio, gfp_t gfp_mask)
 {
+	struct request_queue *q = rl->q;
 	struct request *rq;
-	struct request_list *rl = &q->rq;
 	struct elevator_type *et = q->elevator->type;
 	struct io_context *ioc = rq_ioc(bio);
 	struct io_cq *icq = NULL;
@@ -863,9 +869,9 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 			 * This process will be allowed to complete a batch of
 			 * requests, others will be blocked.
 			 */
-			if (!blk_queue_full(q, is_sync)) {
+			if (!blk_rl_full(rl, is_sync)) {
 				ioc_set_batching(q, ioc);
-				blk_set_queue_full(q, is_sync);
+				blk_set_rl_full(rl, is_sync);
 			} else {
 				if (may_queue != ELV_MQUEUE_MUST
 						&& !ioc_batching(q, ioc)) {
@@ -915,7 +921,7 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 	spin_unlock_irq(q->queue_lock);
 
 	/* allocate and init request */
-	rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
+	rq = mempool_alloc(rl->rq_pool, gfp_mask);
 	if (!rq)
 		goto fail_alloc;
 
@@ -979,7 +985,7 @@ fail_alloc:
 	 * queue, but this is pretty rare.
 	 */
 	spin_lock_irq(q->queue_lock);
-	freed_request(q, rw_flags);
+	freed_request(rl, rw_flags);
 
 	/*
 	 * in the very unlikely event that allocation failed and no
@@ -1016,7 +1022,7 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 	struct request_list *rl = &q->rq;
 	struct request *rq;
 retry:
-	rq = __get_request(q, rw_flags, bio, gfp_mask);
+	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
@@ -1216,8 +1222,8 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(q, req);
-		freed_request(q, flags);
+		blk_free_request(&q->rq, req);
+		freed_request(&q->rq, flags);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index aa41b47..234ce7c 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -66,16 +66,16 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
 	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_queue_full(q, BLK_RW_SYNC);
+		blk_set_rl_full(rl, BLK_RW_SYNC);
 	} else {
-		blk_clear_queue_full(q, BLK_RW_SYNC);
+		blk_clear_rl_full(rl, BLK_RW_SYNC);
 		wake_up(&rl->wait[BLK_RW_SYNC]);
 	}
 
 	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_queue_full(q, BLK_RW_ASYNC);
+		blk_set_rl_full(rl, BLK_RW_ASYNC);
 	} else {
-		blk_clear_queue_full(q, BLK_RW_ASYNC);
+		blk_clear_rl_full(rl, BLK_RW_ASYNC);
 		wake_up(&rl->wait[BLK_RW_ASYNC]);
 	}
 	spin_unlock_irq(q->queue_lock);
@@ -476,7 +476,6 @@ static void blk_release_queue(struct kobject *kobj)
 {
 	struct request_queue *q =
 		container_of(kobj, struct request_queue, kobj);
-	struct request_list *rl = &q->rq;
 
 	blk_sync_queue(q);
 
@@ -489,8 +488,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	if (rl->rq_pool)
-		mempool_destroy(rl->rq_pool);
+	blk_exit_rl(&q->rq);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/block/blk.h b/block/blk.h
index 85f6ae4..a134231 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -18,6 +18,9 @@ static inline void __blk_get_queue(struct request_queue *q)
 	kobject_get(&q->kobj);
 }
 
+int blk_init_rl(struct request_list *rl, struct request_queue *q,
+		gfp_t gfp_mask);
+void blk_exit_rl(struct request_list *rl);
 void init_request_from_bio(struct request *req, struct bio *bio);
 void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
 			struct bio *bio);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 0c6f527..d16d29f 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -46,7 +46,12 @@ struct blkcg_gq;
 struct request;
 typedef void (rq_end_io_fn)(struct request *, int);
 
+#define BLK_RL_SYNCFULL		(1U << 0)
+#define BLK_RL_ASYNCFULL	(1U << 1)
+
 struct request_list {
+	struct request_queue	*q;	/* the queue this rl belongs to */
+
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -55,6 +60,7 @@ struct request_list {
 	int			starved[2];
 	mempool_t		*rq_pool;
 	wait_queue_head_t	wait[2];
+	unsigned int		flags;
 };
 
 /*
@@ -562,27 +568,25 @@ static inline bool rq_is_sync(struct request *rq)
 	return rw_is_sync(rq->cmd_flags);
 }
 
-static inline int blk_queue_full(struct request_queue *q, int sync)
+static inline bool blk_rl_full(struct request_list *rl, bool sync)
 {
-	if (sync)
-		return test_bit(QUEUE_FLAG_SYNCFULL, &q->queue_flags);
-	return test_bit(QUEUE_FLAG_ASYNCFULL, &q->queue_flags);
+	unsigned int flag = sync ? BLK_RL_SYNCFULL : BLK_RL_ASYNCFULL;
+
+	return rl->flags & flag;
 }
 
-static inline void blk_set_queue_full(struct request_queue *q, int sync)
+static inline void blk_set_rl_full(struct request_list *rl, bool sync)
 {
-	if (sync)
-		queue_flag_set(QUEUE_FLAG_SYNCFULL, q);
-	else
-		queue_flag_set(QUEUE_FLAG_ASYNCFULL, q);
+	unsigned int flag = sync ? BLK_RL_SYNCFULL : BLK_RL_ASYNCFULL;
+
+	rl->flags |= flag;
 }
 
-static inline void blk_clear_queue_full(struct request_queue *q, int sync)
+static inline void blk_clear_rl_full(struct request_list *rl, bool sync)
 {
-	if (sync)
-		queue_flag_clear(QUEUE_FLAG_SYNCFULL, q);
-	else
-		queue_flag_clear(QUEUE_FLAG_ASYNCFULL, q);
+	unsigned int flag = sync ? BLK_RL_SYNCFULL : BLK_RL_ASYNCFULL;
+
+	rl->flags &= ~flag;
 }
 
 
-- 
1.7.7.3

[PATCH 09/10] block: prepare for multiple request_lists

[Tejun Heo]

Request allocation is about to be made per-blkg meaning that there'll
be multiple request lists.

* Make queue full state per request_list.  blk_*queue_full() functions
  are renamed to blk_*rl_full() and takes @rl instead of @q.

* Rename blk_init_free_list() to blk_init_rl() and make it take @rl
  instead of @q.  Also add @gfp_mask parameter.

* Add blk_exit_rl() instead of destroying rl directly from
  blk_release_queue().

* Add request_list->q and make request alloc/free functions -
  blk_free_request(), [__]freed_request(), __get_request() - take @rl
  instead of @q.

This patch doesn't introduce any functional difference.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
---
 block/blk-core.c       |   56 ++++++++++++++++++++++++++---------------------
 block/blk-sysfs.c      |   12 ++++------
 block/blk.h            |    3 ++
 include/linux/blkdev.h |   32 +++++++++++++++------------
 4 files changed, 57 insertions(+), 46 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 8149c76..0fbd233 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -504,13 +504,13 @@ void blk_cleanup_queue(struct request_queue *q)
 }
 EXPORT_SYMBOL(blk_cleanup_queue);
 
-static int blk_init_free_list(struct request_queue *q)
+int blk_init_rl(struct request_list *rl, struct request_queue *q,
+		gfp_t gfp_mask)
 {
-	struct request_list *rl = &q->rq;
-
 	if (unlikely(rl->rq_pool))
 		return 0;
 
+	rl->q = q;
 	rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0;
 	rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0;
 	init_waitqueue_head(&rl->wait[BLK_RW_SYNC]);
@@ -518,13 +518,19 @@ static int blk_init_free_list(struct request_queue *q)
 
 	rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
 					  mempool_free_slab, request_cachep,
-					  GFP_KERNEL, q->node);
+					  gfp_mask, q->node);
 	if (!rl->rq_pool)
 		return -ENOMEM;
 
 	return 0;
 }
 
+void blk_exit_rl(struct request_list *rl)
+{
+	if (rl->rq_pool)
+		mempool_destroy(rl->rq_pool);
+}
+
 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
 {
 	return blk_alloc_queue_node(gfp_mask, -1);
@@ -666,7 +672,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_free_list(q))
+	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -708,15 +714,15 @@ bool blk_get_queue(struct request_queue *q)
 }
 EXPORT_SYMBOL(blk_get_queue);
 
-static inline void blk_free_request(struct request_queue *q, struct request *rq)
+static inline void blk_free_request(struct request_list *rl, struct request *rq)
 {
 	if (rq->cmd_flags & REQ_ELVPRIV) {
-		elv_put_request(q, rq);
+		elv_put_request(rl->q, rq);
 		if (rq->elv.icq)
 			put_io_context(rq->elv.icq->ioc);
 	}
 
-	mempool_free(rq, q->rq.rq_pool);
+	mempool_free(rq, rl->rq_pool);
 }
 
 /*
@@ -753,9 +759,9 @@ static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
 	ioc->last_waited = jiffies;
 }
 
-static void __freed_request(struct request_queue *q, int sync)
+static void __freed_request(struct request_list *rl, int sync)
 {
-	struct request_list *rl = &q->rq;
+	struct request_queue *q = rl->q;
 
 	if (rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
@@ -764,7 +770,7 @@ static void __freed_request(struct request_queue *q, int sync)
 		if (waitqueue_active(&rl->wait[sync]))
 			wake_up(&rl->wait[sync]);
 
-		blk_clear_queue_full(q, sync);
+		blk_clear_rl_full(rl, sync);
 	}
 }
 
@@ -772,9 +778,9 @@ static void __freed_request(struct request_queue *q, int sync)
  * A request has just been released.  Account for it, update the full and
  * congestion status, wake up any waiters.   Called under q->queue_lock.
  */
-static void freed_request(struct request_queue *q, unsigned int flags)
+static void freed_request(struct request_list *rl, unsigned int flags)
 {
-	struct request_list *rl = &q->rq;
+	struct request_queue *q = rl->q;
 	int sync = rw_is_sync(flags);
 
 	q->nr_rqs[sync]--;
@@ -782,10 +788,10 @@ static void freed_request(struct request_queue *q, unsigned int flags)
 	if (flags & REQ_ELVPRIV)
 		q->nr_rqs_elvpriv--;
 
-	__freed_request(q, sync);
+	__freed_request(rl, sync);
 
 	if (unlikely(rl->starved[sync ^ 1]))
-		__freed_request(q, sync ^ 1);
+		__freed_request(rl, sync ^ 1);
 }
 
 /*
@@ -825,7 +831,7 @@ static struct io_context *rq_ioc(struct bio *bio)
 
 /**
  * __get_request - get a free request
- * @q: request_queue to allocate request from
+ * @rl: request list to allocate from
  * @rw_flags: RW and SYNC flags
  * @bio: bio to allocate request for (can be %NULL)
  * @gfp_mask: allocation mask
@@ -837,11 +843,11 @@ static struct io_context *rq_ioc(struct bio *bio)
  * Returns %NULL on failure, with @q->queue_lock held.
  * Returns !%NULL on success, with @q->queue_lock *not held*.
  */
-static struct request *__get_request(struct request_queue *q, int rw_flags,
+static struct request *__get_request(struct request_list *rl, int rw_flags,
 				     struct bio *bio, gfp_t gfp_mask)
 {
+	struct request_queue *q = rl->q;
 	struct request *rq;
-	struct request_list *rl = &q->rq;
 	struct elevator_type *et = q->elevator->type;
 	struct io_context *ioc = rq_ioc(bio);
 	struct io_cq *icq = NULL;
@@ -863,9 +869,9 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 			 * This process will be allowed to complete a batch of
 			 * requests, others will be blocked.
 			 */
-			if (!blk_queue_full(q, is_sync)) {
+			if (!blk_rl_full(rl, is_sync)) {
 				ioc_set_batching(q, ioc);
-				blk_set_queue_full(q, is_sync);
+				blk_set_rl_full(rl, is_sync);
 			} else {
 				if (may_queue != ELV_MQUEUE_MUST
 						&& !ioc_batching(q, ioc)) {
@@ -915,7 +921,7 @@ static struct request *__get_request(struct request_queue *q, int rw_flags,
 	spin_unlock_irq(q->queue_lock);
 
 	/* allocate and init request */
-	rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
+	rq = mempool_alloc(rl->rq_pool, gfp_mask);
 	if (!rq)
 		goto fail_alloc;
 
@@ -979,7 +985,7 @@ fail_alloc:
 	 * queue, but this is pretty rare.
 	 */
 	spin_lock_irq(q->queue_lock);
-	freed_request(q, rw_flags);
+	freed_request(rl, rw_flags);
 
 	/*
 	 * in the very unlikely event that allocation failed and no
@@ -1016,7 +1022,7 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 	struct request_list *rl = &q->rq;
 	struct request *rq;
 retry:
-	rq = __get_request(q, rw_flags, bio, gfp_mask);
+	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
@@ -1216,8 +1222,8 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(q, req);
-		freed_request(q, flags);
+		blk_free_request(&q->rq, req);
+		freed_request(&q->rq, flags);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index aa41b47..234ce7c 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -66,16 +66,16 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
 	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_queue_full(q, BLK_RW_SYNC);
+		blk_set_rl_full(rl, BLK_RW_SYNC);
 	} else {
-		blk_clear_queue_full(q, BLK_RW_SYNC);
+		blk_clear_rl_full(rl, BLK_RW_SYNC);
 		wake_up(&rl->wait[BLK_RW_SYNC]);
 	}
 
 	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_queue_full(q, BLK_RW_ASYNC);
+		blk_set_rl_full(rl, BLK_RW_ASYNC);
 	} else {
-		blk_clear_queue_full(q, BLK_RW_ASYNC);
+		blk_clear_rl_full(rl, BLK_RW_ASYNC);
 		wake_up(&rl->wait[BLK_RW_ASYNC]);
 	}
 	spin_unlock_irq(q->queue_lock);
@@ -476,7 +476,6 @@ static void blk_release_queue(struct kobject *kobj)
 {
 	struct request_queue *q =
 		container_of(kobj, struct request_queue, kobj);
-	struct request_list *rl = &q->rq;
 
 	blk_sync_queue(q);
 
@@ -489,8 +488,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	if (rl->rq_pool)
-		mempool_destroy(rl->rq_pool);
+	blk_exit_rl(&q->rq);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/block/blk.h b/block/blk.h
index 85f6ae4..a134231 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -18,6 +18,9 @@ static inline void __blk_get_queue(struct request_queue *q)
 	kobject_get(&q->kobj);
 }
 
+int blk_init_rl(struct request_list *rl, struct request_queue *q,
+		gfp_t gfp_mask);
+void blk_exit_rl(struct request_list *rl);
 void init_request_from_bio(struct request *req, struct bio *bio);
 void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
 			struct bio *bio);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 0c6f527..d16d29f 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -46,7 +46,12 @@ struct blkcg_gq;
 struct request;
 typedef void (rq_end_io_fn)(struct request *, int);
 
+#define BLK_RL_SYNCFULL		(1U << 0)
+#define BLK_RL_ASYNCFULL	(1U << 1)
+
 struct request_list {
+	struct request_queue	*q;	/* the queue this rl belongs to */
+
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -55,6 +60,7 @@ struct request_list {
 	int			starved[2];
 	mempool_t		*rq_pool;
 	wait_queue_head_t	wait[2];
+	unsigned int		flags;
 };
 
 /*
@@ -562,27 +568,25 @@ static inline bool rq_is_sync(struct request *rq)
 	return rw_is_sync(rq->cmd_flags);
 }
 
-static inline int blk_queue_full(struct request_queue *q, int sync)
+static inline bool blk_rl_full(struct request_list *rl, bool sync)
 {
-	if (sync)
-		return test_bit(QUEUE_FLAG_SYNCFULL, &q->queue_flags);
-	return test_bit(QUEUE_FLAG_ASYNCFULL, &q->queue_flags);
+	unsigned int flag = sync ? BLK_RL_SYNCFULL : BLK_RL_ASYNCFULL;
+
+	return rl->flags & flag;
 }
 
-static inline void blk_set_queue_full(struct request_queue *q, int sync)
+static inline void blk_set_rl_full(struct request_list *rl, bool sync)
 {
-	if (sync)
-		queue_flag_set(QUEUE_FLAG_SYNCFULL, q);
-	else
-		queue_flag_set(QUEUE_FLAG_ASYNCFULL, q);
+	unsigned int flag = sync ? BLK_RL_SYNCFULL : BLK_RL_ASYNCFULL;
+
+	rl->flags |= flag;
 }
 
-static inline void blk_clear_queue_full(struct request_queue *q, int sync)
+static inline void blk_clear_rl_full(struct request_list *rl, bool sync)
 {
-	if (sync)
-		queue_flag_clear(QUEUE_FLAG_SYNCFULL, q);
-	else
-		queue_flag_clear(QUEUE_FLAG_ASYNCFULL, q);
+	unsigned int flag = sync ? BLK_RL_SYNCFULL : BLK_RL_ASYNCFULL;
+
+	rl->flags &= ~flag;
 }
 
 
-- 
1.7.7.3

[PATCH 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Wu Fengguang <fengguang.wu-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 +++++++++++++++++--
 block/blk-cgroup.h                  |   95 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   33 +++++++++---
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 202 insertions(+), 28 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index a48f35b..82883b5 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -301,6 +311,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -751,6 +793,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..93da70b 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,88 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root) {
+		rcu_read_unlock();
+		return &q->root_rl;
+	}
+
+	blkg = blkg_lookup_create(blkcg, q);
+	blkg_get(blkg);
+
+	rcu_read_unlock();
+
+	return &blkg->rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +477,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +485,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index 0fbd233..6aff69e 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -672,7 +672,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -763,7 +763,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up root
+	 * blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -884,7 +889,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -926,6 +936,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1019,15 +1030,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1218,12 +1233,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index d16d29f..1ff9a6b 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3

[PATCH 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
---
 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 +++++++++++++++++--
 block/blk-cgroup.h                  |   95 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   33 +++++++++---
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 202 insertions(+), 28 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index a48f35b..82883b5 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -301,6 +311,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -751,6 +793,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..93da70b 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,88 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root) {
+		rcu_read_unlock();
+		return &q->root_rl;
+	}
+
+	blkg = blkg_lookup_create(blkcg, q);
+	blkg_get(blkg);
+
+	rcu_read_unlock();
+
+	return &blkg->rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +477,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +485,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index 0fbd233..6aff69e 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -672,7 +672,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -763,7 +763,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up root
+	 * blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -884,7 +889,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -926,6 +936,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1019,15 +1030,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1218,12 +1233,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index d16d29f..1ff9a6b 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3

Re: [PATCHSET] block: implement per-blkg request allocation, take#2

[Tejun Heo]

Jens, ping?

-- 
tejun

Re: [PATCHSET] block: implement per-blkg request allocation, take#2

[Tejun Heo]

Jens, ping?

-- 
tejun

[PATCH UPDATED 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

From 6a8958a0ba2313d224aabe01a58275eff52fec23 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Date: Thu, 21 Jun 2012 15:20:19 -0700

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

v3: blk_get_rl() wasn't checking error return from
    blkg_lookup_create() and may cause oops on lookup failure.  Fix it
    by falling back to root_rl on blkg lookup failures.  This problem
    was spotted by Rakesh Iyer <rni-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Wu Fengguang <fengguang.wu-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---

The git branch is udpated such that it's on top of 3.5-rc3 and this
updated patch is included.

  git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 ++++++++++++++++--
 block/blk-cgroup.h                  |  102 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   33 +++++++++---
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 209 insertions(+), 28 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index a48f35b..82883b5 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -301,6 +311,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -751,6 +793,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..2459730 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,95 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root)
+		goto root_rl;
+
+	/*
+	 * Try to use blkg->rl.  blkg lookup may fail under memory pressure
+	 * or if either the blkcg or queue is going away.  Fall back to
+	 * root_rl in such cases.
+	 */
+	blkg = blkg_lookup_create(blkcg, q);
+	if (unlikely(IS_ERR(blkg)))
+		goto root_rl;
+
+	blkg_get(blkg);
+	rcu_read_unlock();
+	return &blkg->rl;
+root_rl:
+	rcu_read_unlock();
+	return &q->root_rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +484,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +492,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index 0fbd233..6aff69e 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -672,7 +672,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -763,7 +763,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up root
+	 * blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -884,7 +889,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -926,6 +936,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1019,15 +1030,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1218,12 +1233,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index d16d29f..1ff9a6b 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3

[PATCH UPDATED 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

>From 6a8958a0ba2313d224aabe01a58275eff52fec23 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj@kernel.org>
Date: Thu, 21 Jun 2012 15:20:19 -0700

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

v3: blk_get_rl() wasn't checking error return from
    blkg_lookup_create() and may cause oops on lookup failure.  Fix it
    by falling back to root_rl on blkg lookup failures.  This problem
    was spotted by Rakesh Iyer <rni@google.com>.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
---

The git branch is udpated such that it's on top of 3.5-rc3 and this
updated patch is included.

  git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 ++++++++++++++++--
 block/blk-cgroup.h                  |  102 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   33 +++++++++---
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 209 insertions(+), 28 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index a48f35b..82883b5 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -301,6 +311,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -751,6 +793,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..2459730 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,95 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root)
+		goto root_rl;
+
+	/*
+	 * Try to use blkg->rl.  blkg lookup may fail under memory pressure
+	 * or if either the blkcg or queue is going away.  Fall back to
+	 * root_rl in such cases.
+	 */
+	blkg = blkg_lookup_create(blkcg, q);
+	if (unlikely(IS_ERR(blkg)))
+		goto root_rl;
+
+	blkg_get(blkg);
+	rcu_read_unlock();
+	return &blkg->rl;
+root_rl:
+	rcu_read_unlock();
+	return &q->root_rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +484,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +492,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index 0fbd233..6aff69e 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -672,7 +672,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -763,7 +763,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up root
+	 * blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -884,7 +889,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -926,6 +936,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1019,15 +1030,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1218,12 +1233,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index d16d29f..1ff9a6b 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3

Re: [PATCHSET] block: implement per-blkg request allocation, take#2

[Jens Axboe]

On 06/19/2012 12:34 AM, Tejun Heo wrote:
> Jens, ping?

Applied to for-3.6/core, thanks!

-- 
Jens Axboe

Re: [PATCHSET] block: implement per-blkg request allocation, take#2

[Jens Axboe]

On 06/19/2012 12:34 AM, Tejun Heo wrote:
> Jens, ping?

Applied to for-3.6/core, thanks!

-- 
Jens Axboe

[PATCH UPDATED v4 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

From 866f8c15fafc3ae461463d5937f448eee2631ac5 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Date: Tue, 26 Jun 2012 15:05:44 -0700

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

v3: blk_get_rl() wasn't checking error return from
    blkg_lookup_create() and may cause oops on lookup failure.  Fix it
    by falling back to root_rl on blkg lookup failures.  This problem
    was spotted by Rakesh Iyer <rni-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>.

v4: Updated to accomodate 458f27a982 "block: Avoid missed wakeup in
    request waitqueue".  blk_drain_queue() now wakes up waiters on all
    blkg->rl on the target queue.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Wu Fengguang <fengguang.wu-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
v4 to fix build failure on top of block/for-3.6/core.  git branch has
been udpated too.

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

Only the top commit has changed.

Thanks.

 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 ++++++++++++++++--
 block/blk-cgroup.h                  |  102 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   42 +++++++++++----
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 216 insertions(+), 30 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 63b31eb..f3b44a6 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -300,6 +310,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -750,6 +792,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..2459730 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,95 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root)
+		goto root_rl;
+
+	/*
+	 * Try to use blkg->rl.  blkg lookup may fail under memory pressure
+	 * or if either the blkcg or queue is going away.  Fall back to
+	 * root_rl in such cases.
+	 */
+	blkg = blkg_lookup_create(blkcg, q);
+	if (unlikely(IS_ERR(blkg)))
+		goto root_rl;
+
+	blkg_get(blkg);
+	rcu_read_unlock();
+	return &blkg->rl;
+root_rl:
+	rcu_read_unlock();
+	return &q->root_rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +484,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +492,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index f392a2e..dd134d8 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -416,9 +416,14 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 	 * left with hung waiters. We need to wake up those waiters.
 	 */
 	if (q->request_fn) {
+		struct request_list *rl;
+
 		spin_lock_irq(q->queue_lock);
-		for (i = 0; i < ARRAY_SIZE(q->rq.wait); i++)
-			wake_up_all(&q->rq.wait[i]);
+
+		blk_queue_for_each_rl(rl, q)
+			for (i = 0; i < ARRAY_SIZE(rl->wait); i++)
+				wake_up_all(&rl->wait[i]);
+
 		spin_unlock_irq(q->queue_lock);
 	}
 }
@@ -685,7 +690,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -776,7 +781,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up root
+	 * blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -897,7 +907,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -939,6 +954,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1032,15 +1048,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1231,12 +1251,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f2385ee..3816ce8 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3

[PATCH UPDATED v4 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

>From 866f8c15fafc3ae461463d5937f448eee2631ac5 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj@kernel.org>
Date: Tue, 26 Jun 2012 15:05:44 -0700

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

v3: blk_get_rl() wasn't checking error return from
    blkg_lookup_create() and may cause oops on lookup failure.  Fix it
    by falling back to root_rl on blkg lookup failures.  This problem
    was spotted by Rakesh Iyer <rni@google.com>.

v4: Updated to accomodate 458f27a982 "block: Avoid missed wakeup in
    request waitqueue".  blk_drain_queue() now wakes up waiters on all
    blkg->rl on the target queue.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
---
v4 to fix build failure on top of block/for-3.6/core.  git branch has
been udpated too.

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

Only the top commit has changed.

Thanks.

 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 ++++++++++++++++--
 block/blk-cgroup.h                  |  102 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   42 +++++++++++----
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 216 insertions(+), 30 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 63b31eb..f3b44a6 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -300,6 +310,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -750,6 +792,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..2459730 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,95 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root)
+		goto root_rl;
+
+	/*
+	 * Try to use blkg->rl.  blkg lookup may fail under memory pressure
+	 * or if either the blkcg or queue is going away.  Fall back to
+	 * root_rl in such cases.
+	 */
+	blkg = blkg_lookup_create(blkcg, q);
+	if (unlikely(IS_ERR(blkg)))
+		goto root_rl;
+
+	blkg_get(blkg);
+	rcu_read_unlock();
+	return &blkg->rl;
+root_rl:
+	rcu_read_unlock();
+	return &q->root_rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +484,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +492,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index f392a2e..dd134d8 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -416,9 +416,14 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 	 * left with hung waiters. We need to wake up those waiters.
 	 */
 	if (q->request_fn) {
+		struct request_list *rl;
+
 		spin_lock_irq(q->queue_lock);
-		for (i = 0; i < ARRAY_SIZE(q->rq.wait); i++)
-			wake_up_all(&q->rq.wait[i]);
+
+		blk_queue_for_each_rl(rl, q)
+			for (i = 0; i < ARRAY_SIZE(rl->wait); i++)
+				wake_up_all(&rl->wait[i]);
+
 		spin_unlock_irq(q->queue_lock);
 	}
 }
@@ -685,7 +690,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -776,7 +781,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up root
+	 * blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -897,7 +907,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -939,6 +954,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1032,15 +1048,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1231,12 +1251,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f2385ee..3816ce8 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3

[PATCH UPDATED v5 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

From f5a1424230fb2da7143852ce5bc4049ff9c5f618 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Date: Tue, 26 Jun 2012 15:46:31 -0700

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

v3: blk_get_rl() wasn't checking error return from
    blkg_lookup_create() and may cause oops on lookup failure.  Fix it
    by falling back to root_rl on blkg lookup failures.  This problem
    was spotted by Rakesh Iyer <rni-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>.

v4: Updated to accomodate 458f27a982 "block: Avoid missed wakeup in
    request waitqueue".  blk_drain_queue() now wakes up waiters on all
    blkg->rl on the target queue.  Breakge spotted by Wu Fengguang.

v5: Comment updated as suggested by Andrew Morton.

Signed-off-by: Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Acked-by: Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Cc: Wu Fengguang <fengguang.wu-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Cc: Andrew Morton <akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org>
---
Comments updated.  git branch updated accordingly.  Thanks.

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 ++++++++++++++++--
 block/blk-cgroup.h                  |  102 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   42 +++++++++++----
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 216 insertions(+), 30 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 63b31eb..f3b44a6 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -300,6 +310,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -750,6 +792,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..d6000fc 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,95 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root)
+		goto root_rl;
+
+	/*
+	 * Try to use blkg->rl.  blkg lookup may fail under memory pressure
+	 * or if either the blkcg or queue is going away.  Fall back to
+	 * root_rl in such cases.
+	 */
+	blkg = blkg_lookup_create(blkcg, q);
+	if (unlikely(IS_ERR(blkg)))
+		goto root_rl;
+
+	blkg_get(blkg);
+	rcu_read_unlock();
+	return &blkg->rl;
+root_rl:
+	rcu_read_unlock();
+	return &q->root_rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know where the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +484,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +492,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index f392a2e..05d3784 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -416,9 +416,14 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 	 * left with hung waiters. We need to wake up those waiters.
 	 */
 	if (q->request_fn) {
+		struct request_list *rl;
+
 		spin_lock_irq(q->queue_lock);
-		for (i = 0; i < ARRAY_SIZE(q->rq.wait); i++)
-			wake_up_all(&q->rq.wait[i]);
+
+		blk_queue_for_each_rl(rl, q)
+			for (i = 0; i < ARRAY_SIZE(rl->wait); i++)
+				wake_up_all(&rl->wait[i]);
+
 		spin_unlock_irq(q->queue_lock);
 	}
 }
@@ -685,7 +690,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -776,7 +781,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up in the
+	 * root blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -897,7 +907,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -939,6 +954,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1032,15 +1048,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1231,12 +1251,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f2385ee..3816ce8 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3

[PATCH UPDATED v5 10/10] blkcg: implement per-blkg request allocation

[Tejun Heo]

>From f5a1424230fb2da7143852ce5bc4049ff9c5f618 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj@kernel.org>
Date: Tue, 26 Jun 2012 15:46:31 -0700

Currently, request_queue has one request_list to allocate requests
from regardless of blkcg of the IO being issued.  When the unified
request pool is used up, cfq proportional IO limits become meaningless
- whoever grabs the next request being freed wins the race regardless
of the configured weights.

This can be easily demonstrated by creating a blkio cgroup w/ very low
weight, put a program which can issue a lot of random direct IOs there
and running a sequential IO from a different cgroup.  As soon as the
request pool is used up, the sequential IO bandwidth crashes.

This patch implements per-blkg request_list.  Each blkg has its own
request_list and any IO allocates its request from the matching blkg
making blkcgs completely isolated in terms of request allocation.

* Root blkcg uses the request_list embedded in each request_queue,
  which was renamed to @q->root_rl from @q->rq.  While making blkcg rl
  handling a bit harier, this enables avoiding most overhead for root
  blkcg.

* Queue fullness is properly per request_list but bdi isn't blkcg
  aware yet, so congestion state currently just follows the root
  blkcg.  As writeback isn't aware of blkcg yet, this works okay for
  async congestion but readahead may get the wrong signals.  It's
  better than blkcg completely collapsing with shared request_list but
  needs to be improved with future changes.

* After this change, each block cgroup gets a full request pool making
  resource consumption of each cgroup higher.  This makes allowing
  non-root users to create cgroups less desirable; however, note that
  allowing non-root users to directly manage cgroups is already
  severely broken regardless of this patch - each block cgroup
  consumes kernel memory and skews IO weight (IO weights are not
  hierarchical).

v2: queue-sysfs.txt updated and patch description udpated as suggested
    by Vivek.

v3: blk_get_rl() wasn't checking error return from
    blkg_lookup_create() and may cause oops on lookup failure.  Fix it
    by falling back to root_rl on blkg lookup failures.  This problem
    was spotted by Rakesh Iyer <rni@google.com>.

v4: Updated to accomodate 458f27a982 "block: Avoid missed wakeup in
    request waitqueue".  blk_drain_queue() now wakes up waiters on all
    blkg->rl on the target queue.  Breakge spotted by Wu Fengguang.

v5: Comment updated as suggested by Andrew Morton.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
---
Comments updated.  git branch updated accordingly.  Thanks.

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl

 Documentation/block/queue-sysfs.txt |    7 +++
 block/blk-cgroup.c                  |   51 ++++++++++++++++--
 block/blk-cgroup.h                  |  102 +++++++++++++++++++++++++++++++++++
 block/blk-core.c                    |   42 +++++++++++----
 block/blk-sysfs.c                   |   32 +++++++-----
 include/linux/blkdev.h              |   12 +++-
 6 files changed, 216 insertions(+), 30 deletions(-)

diff --git a/Documentation/block/queue-sysfs.txt b/Documentation/block/queue-sysfs.txt
index d8147b3..6518a55 100644
--- a/Documentation/block/queue-sysfs.txt
+++ b/Documentation/block/queue-sysfs.txt
@@ -38,6 +38,13 @@ read or write requests. Note that the total allocated number may be twice
 this amount, since it applies only to reads or writes (not the accumulated
 sum).
 
+To avoid priority inversion through request starvation, a request
+queue maintains a separate request pool per each cgroup when
+CONFIG_BLK_CGROUP is enabled, and this parameter applies to each such
+per-block-cgroup request pool.  IOW, if there are N block cgroups,
+each request queue may have upto N request pools, each independently
+regulated by nr_requests.
+
 read_ahead_kb (RW)
 ------------------
 Maximum number of kilobytes to read-ahead for filesystems on this block
diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 63b31eb..f3b44a6 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -63,6 +63,7 @@ static void blkg_free(struct blkcg_gq *blkg)
 		kfree(pd);
 	}
 
+	blk_exit_rl(&blkg->rl);
 	kfree(blkg);
 }
 
@@ -90,6 +91,13 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 	blkg->refcnt = 1;
 
+	/* root blkg uses @q->root_rl, init rl only for !root blkgs */
+	if (blkcg != &blkcg_root) {
+		if (blk_init_rl(&blkg->rl, q, gfp_mask))
+			goto err_free;
+		blkg->rl.blkg = blkg;
+	}
+
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkg_policy_data *pd;
@@ -99,10 +107,8 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 
 		/* alloc per-policy data and attach it to blkg */
 		pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
-		if (!pd) {
-			blkg_free(blkg);
-			return NULL;
-		}
+		if (!pd)
+			goto err_free;
 
 		blkg->pd[i] = pd;
 		pd->blkg = blkg;
@@ -113,6 +119,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	}
 
 	return blkg;
+
+err_free:
+	blkg_free(blkg);
+	return NULL;
 }
 
 static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
@@ -300,6 +310,38 @@ void __blkg_release(struct blkcg_gq *blkg)
 }
 EXPORT_SYMBOL_GPL(__blkg_release);
 
+/*
+ * The next function used by blk_queue_for_each_rl().  It's a bit tricky
+ * because the root blkg uses @q->root_rl instead of its own rl.
+ */
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q)
+{
+	struct list_head *ent;
+	struct blkcg_gq *blkg;
+
+	/*
+	 * Determine the current blkg list_head.  The first entry is
+	 * root_rl which is off @q->blkg_list and mapped to the head.
+	 */
+	if (rl == &q->root_rl) {
+		ent = &q->blkg_list;
+	} else {
+		blkg = container_of(rl, struct blkcg_gq, rl);
+		ent = &blkg->q_node;
+	}
+
+	/* walk to the next list_head, skip root blkcg */
+	ent = ent->next;
+	if (ent == &q->root_blkg->q_node)
+		ent = ent->next;
+	if (ent == &q->blkg_list)
+		return NULL;
+
+	blkg = container_of(ent, struct blkcg_gq, q_node);
+	return &blkg->rl;
+}
+
 static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
 			     u64 val)
 {
@@ -750,6 +792,7 @@ int blkcg_activate_policy(struct request_queue *q,
 		goto out_unlock;
 	}
 	q->root_blkg = blkg;
+	q->root_rl.blkg = blkg;
 
 	list_for_each_entry(blkg, &q->blkg_list, q_node)
 		cnt++;
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index e74cce1..d6000fc 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -17,6 +17,7 @@
 #include <linux/u64_stats_sync.h>
 #include <linux/seq_file.h>
 #include <linux/radix-tree.h>
+#include <linux/blkdev.h>
 
 /* Max limits for throttle policy */
 #define THROTL_IOPS_MAX		UINT_MAX
@@ -93,6 +94,8 @@ struct blkcg_gq {
 	struct list_head		q_node;
 	struct hlist_node		blkcg_node;
 	struct blkcg			*blkcg;
+	/* request allocation list for this blkcg-q pair */
+	struct request_list		rl;
 	/* reference count */
 	int				refcnt;
 
@@ -251,6 +254,95 @@ static inline void blkg_put(struct blkcg_gq *blkg)
 }
 
 /**
+ * blk_get_rl - get request_list to use
+ * @q: request_queue of interest
+ * @bio: bio which will be attached to the allocated request (may be %NULL)
+ *
+ * The caller wants to allocate a request from @q to use for @bio.  Find
+ * the request_list to use and obtain a reference on it.  Should be called
+ * under queue_lock.  This function is guaranteed to return non-%NULL
+ * request_list.
+ */
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio)
+{
+	struct blkcg *blkcg;
+	struct blkcg_gq *blkg;
+
+	rcu_read_lock();
+
+	blkcg = bio_blkcg(bio);
+
+	/* bypass blkg lookup and use @q->root_rl directly for root */
+	if (blkcg == &blkcg_root)
+		goto root_rl;
+
+	/*
+	 * Try to use blkg->rl.  blkg lookup may fail under memory pressure
+	 * or if either the blkcg or queue is going away.  Fall back to
+	 * root_rl in such cases.
+	 */
+	blkg = blkg_lookup_create(blkcg, q);
+	if (unlikely(IS_ERR(blkg)))
+		goto root_rl;
+
+	blkg_get(blkg);
+	rcu_read_unlock();
+	return &blkg->rl;
+root_rl:
+	rcu_read_unlock();
+	return &q->root_rl;
+}
+
+/**
+ * blk_put_rl - put request_list
+ * @rl: request_list to put
+ *
+ * Put the reference acquired by blk_get_rl().  Should be called under
+ * queue_lock.
+ */
+static inline void blk_put_rl(struct request_list *rl)
+{
+	/* root_rl may not have blkg set */
+	if (rl->blkg && rl->blkg->blkcg != &blkcg_root)
+		blkg_put(rl->blkg);
+}
+
+/**
+ * blk_rq_set_rl - associate a request with a request_list
+ * @rq: request of interest
+ * @rl: target request_list
+ *
+ * Associate @rq with @rl so that accounting and freeing can know where the
+ * request_list @rq came from.
+ */
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl)
+{
+	rq->rl = rl;
+}
+
+/**
+ * blk_rq_rl - return the request_list a request came from
+ * @rq: request of interest
+ *
+ * Return the request_list @rq is allocated from.
+ */
+static inline struct request_list *blk_rq_rl(struct request *rq)
+{
+	return rq->rl;
+}
+
+struct request_list *__blk_queue_next_rl(struct request_list *rl,
+					 struct request_queue *q);
+/**
+ * blk_queue_for_each_rl - iterate through all request_lists of a request_queue
+ *
+ * Should be used under queue_lock.
+ */
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = __blk_queue_next_rl((rl), (q)))
+
+/**
  * blkg_stat_add - add a value to a blkg_stat
  * @stat: target blkg_stat
  * @val: value to add
@@ -392,6 +484,7 @@ static inline void blkcg_deactivate_policy(struct request_queue *q,
 
 static inline struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup) { return NULL; }
 static inline struct blkcg *bio_blkcg(struct bio *bio) { return NULL; }
+
 static inline struct blkg_policy_data *blkg_to_pd(struct blkcg_gq *blkg,
 						  struct blkcg_policy *pol) { return NULL; }
 static inline struct blkcg_gq *pd_to_blkg(struct blkg_policy_data *pd) { return NULL; }
@@ -399,5 +492,14 @@ static inline char *blkg_path(struct blkcg_gq *blkg) { return NULL; }
 static inline void blkg_get(struct blkcg_gq *blkg) { }
 static inline void blkg_put(struct blkcg_gq *blkg) { }
 
+static inline struct request_list *blk_get_rl(struct request_queue *q,
+					      struct bio *bio) { return &q->root_rl; }
+static inline void blk_put_rl(struct request_list *rl) { }
+static inline void blk_rq_set_rl(struct request *rq, struct request_list *rl) { }
+static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q->root_rl; }
+
+#define blk_queue_for_each_rl(rl, q)	\
+	for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
+
 #endif	/* CONFIG_BLK_CGROUP */
 #endif	/* _BLK_CGROUP_H */
diff --git a/block/blk-core.c b/block/blk-core.c
index f392a2e..05d3784 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -416,9 +416,14 @@ void blk_drain_queue(struct request_queue *q, bool drain_all)
 	 * left with hung waiters. We need to wake up those waiters.
 	 */
 	if (q->request_fn) {
+		struct request_list *rl;
+
 		spin_lock_irq(q->queue_lock);
-		for (i = 0; i < ARRAY_SIZE(q->rq.wait); i++)
-			wake_up_all(&q->rq.wait[i]);
+
+		blk_queue_for_each_rl(rl, q)
+			for (i = 0; i < ARRAY_SIZE(rl->wait); i++)
+				wake_up_all(&rl->wait[i]);
+
 		spin_unlock_irq(q->queue_lock);
 	}
 }
@@ -685,7 +690,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	if (blk_init_rl(&q->rq, q, GFP_KERNEL))
+	if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
 		return NULL;
 
 	q->request_fn		= rfn;
@@ -776,7 +781,12 @@ static void __freed_request(struct request_list *rl, int sync)
 {
 	struct request_queue *q = rl->q;
 
-	if (rl->count[sync] < queue_congestion_off_threshold(q))
+	/*
+	 * bdi isn't aware of blkcg yet.  As all async IOs end up in the
+	 * root blkcg anyway, just use root blkcg state.
+	 */
+	if (rl == &q->root_rl &&
+	    rl->count[sync] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, sync);
 
 	if (rl->count[sync] + 1 <= q->nr_requests) {
@@ -897,7 +907,12 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 				}
 			}
 		}
-		blk_set_queue_congested(q, is_sync);
+		/*
+		 * bdi isn't aware of blkcg yet.  As all async IOs end up
+		 * root blkcg anyway, just use root blkcg state.
+		 */
+		if (rl == &q->root_rl)
+			blk_set_queue_congested(q, is_sync);
 	}
 
 	/*
@@ -939,6 +954,7 @@ static struct request *__get_request(struct request_list *rl, int rw_flags,
 		goto fail_alloc;
 
 	blk_rq_init(q, rq);
+	blk_rq_set_rl(rq, rl);
 	rq->cmd_flags = rw_flags | REQ_ALLOCED;
 
 	/* init elvpriv */
@@ -1032,15 +1048,19 @@ static struct request *get_request(struct request_queue *q, int rw_flags,
 {
 	const bool is_sync = rw_is_sync(rw_flags) != 0;
 	DEFINE_WAIT(wait);
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	struct request *rq;
+
+	rl = blk_get_rl(q, bio);	/* transferred to @rq on success */
 retry:
-	rq = __get_request(&q->rq, rw_flags, bio, gfp_mask);
+	rq = __get_request(rl, rw_flags, bio, gfp_mask);
 	if (rq)
 		return rq;
 
-	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q)))
+	if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
+		blk_put_rl(rl);
 		return NULL;
+	}
 
 	/* wait on @rl and retry */
 	prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
@@ -1231,12 +1251,14 @@ void __blk_put_request(struct request_queue *q, struct request *req)
 	 */
 	if (req->cmd_flags & REQ_ALLOCED) {
 		unsigned int flags = req->cmd_flags;
+		struct request_list *rl = blk_rq_rl(req);
 
 		BUG_ON(!list_empty(&req->queuelist));
 		BUG_ON(!hlist_unhashed(&req->hash));
 
-		blk_free_request(&q->rq, req);
-		freed_request(&q->rq, flags);
+		blk_free_request(rl, req);
+		freed_request(rl, flags);
+		blk_put_rl(rl);
 	}
 }
 EXPORT_SYMBOL_GPL(__blk_put_request);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 234ce7c..9628b29 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -40,7 +40,7 @@ static ssize_t queue_requests_show(struct request_queue *q, char *page)
 static ssize_t
 queue_requests_store(struct request_queue *q, const char *page, size_t count)
 {
-	struct request_list *rl = &q->rq;
+	struct request_list *rl;
 	unsigned long nr;
 	int ret;
 
@@ -55,6 +55,9 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	q->nr_requests = nr;
 	blk_queue_congestion_threshold(q);
 
+	/* congestion isn't cgroup aware and follows root blkcg for now */
+	rl = &q->root_rl;
+
 	if (rl->count[BLK_RW_SYNC] >= queue_congestion_on_threshold(q))
 		blk_set_queue_congested(q, BLK_RW_SYNC);
 	else if (rl->count[BLK_RW_SYNC] < queue_congestion_off_threshold(q))
@@ -65,19 +68,22 @@ queue_requests_store(struct request_queue *q, const char *page, size_t count)
 	else if (rl->count[BLK_RW_ASYNC] < queue_congestion_off_threshold(q))
 		blk_clear_queue_congested(q, BLK_RW_ASYNC);
 
-	if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_SYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_SYNC);
-		wake_up(&rl->wait[BLK_RW_SYNC]);
+	blk_queue_for_each_rl(rl, q) {
+		if (rl->count[BLK_RW_SYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_SYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_SYNC);
+			wake_up(&rl->wait[BLK_RW_SYNC]);
+		}
+
+		if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
+			blk_set_rl_full(rl, BLK_RW_ASYNC);
+		} else {
+			blk_clear_rl_full(rl, BLK_RW_ASYNC);
+			wake_up(&rl->wait[BLK_RW_ASYNC]);
+		}
 	}
 
-	if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) {
-		blk_set_rl_full(rl, BLK_RW_ASYNC);
-	} else {
-		blk_clear_rl_full(rl, BLK_RW_ASYNC);
-		wake_up(&rl->wait[BLK_RW_ASYNC]);
-	}
 	spin_unlock_irq(q->queue_lock);
 	return ret;
 }
@@ -488,7 +494,7 @@ static void blk_release_queue(struct kobject *kobj)
 		elevator_exit(q->elevator);
 	}
 
-	blk_exit_rl(&q->rq);
+	blk_exit_rl(&q->root_rl);
 
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f2385ee..3816ce8 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -51,7 +51,9 @@ typedef void (rq_end_io_fn)(struct request *, int);
 
 struct request_list {
 	struct request_queue	*q;	/* the queue this rl belongs to */
-
+#ifdef CONFIG_BLK_CGROUP
+	struct blkcg_gq		*blkg;	/* blkg this request pool belongs to */
+#endif
 	/*
 	 * count[], starved[], and wait[] are indexed by
 	 * BLK_RW_SYNC/BLK_RW_ASYNC
@@ -143,6 +145,7 @@ struct request {
 	struct hd_struct *part;
 	unsigned long start_time;
 #ifdef CONFIG_BLK_CGROUP
+	struct request_list *rl;		/* rl this rq is alloced from */
 	unsigned long long start_time_ns;
 	unsigned long long io_start_time_ns;    /* when passed to hardware */
 #endif
@@ -291,9 +294,12 @@ struct request_queue {
 	int			nr_rqs_elvpriv;	/* # allocated rqs w/ elvpriv */
 
 	/*
-	 * the queue request freelist, one for reads and one for writes
+	 * If blkcg is not used, @q->root_rl serves all requests.  If blkcg
+	 * is used, root blkg allocates from @q->root_rl and all other
+	 * blkgs from their own blkg->rl.  Which one to use should be
+	 * determined using bio_request_list().
 	 */
-	struct request_list	rq;
+	struct request_list	root_rl;
 
 	request_fn_proc		*request_fn;
 	make_request_fn		*make_request_fn;
-- 
1.7.7.3