diff mbox

[v3,2/2] virtio-ring: Allocate indirect buffers from cache when possible

Message ID 1346325718-11151-2-git-send-email-levinsasha928@gmail.com (mailing list archive)
State New, archived
Headers show

Commit Message

Sasha Levin Aug. 30, 2012, 11:21 a.m. UTC
Currently if VIRTIO_RING_F_INDIRECT_DESC is enabled we will
use indirect descriptors and allocate them using a simple
kmalloc().

This patch adds a cache which will allow indirect buffers under
a configurable size to be allocated from that cache instead.

Signed-off-by: Sasha Levin <levinsasha928@gmail.com>
---
 drivers/block/virtio_blk.c          |  4 ++++
 drivers/char/hw_random/virtio-rng.c |  4 ++++
 drivers/char/virtio_console.c       |  4 ++++
 drivers/net/virtio_net.c            |  4 ++++
 drivers/virtio/virtio_balloon.c     |  4 ++++
 drivers/virtio/virtio_ring.c        | 34 ++++++++++++++++++++++++++++++----
 include/linux/virtio.h              |  1 +
 net/9p/trans_virtio.c               |  5 +++++
 8 files changed, 56 insertions(+), 4 deletions(-)

Comments

Michael S. Tsirkin Aug. 30, 2012, 1:38 p.m. UTC | #1
On Thu, Aug 30, 2012 at 01:21:58PM +0200, Sasha Levin wrote:
> Currently if VIRTIO_RING_F_INDIRECT_DESC is enabled we will
> use indirect descriptors and allocate them using a simple
> kmalloc().
> 
> This patch adds a cache which will allow indirect buffers under
> a configurable size to be allocated from that cache instead.
> 
> Signed-off-by: Sasha Levin <levinsasha928@gmail.com>

I am not sure we need these module params.
But assuming we keep them, we need to validate values - they come from
user.



> ---
>  drivers/block/virtio_blk.c          |  4 ++++
>  drivers/char/hw_random/virtio-rng.c |  4 ++++
>  drivers/char/virtio_console.c       |  4 ++++
>  drivers/net/virtio_net.c            |  4 ++++
>  drivers/virtio/virtio_balloon.c     |  4 ++++
>  drivers/virtio/virtio_ring.c        | 34 ++++++++++++++++++++++++++++++----
>  include/linux/virtio.h              |  1 +
>  net/9p/trans_virtio.c               |  5 +++++
>  8 files changed, 56 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
> index 7c63065..e4c6c42 100644
> --- a/drivers/block/virtio_blk.c
> +++ b/drivers/block/virtio_blk.c
> @@ -25,6 +25,9 @@ struct workqueue_struct *virtblk_wq;
>  static unsigned int indirect_thresh;
>  module_param(indirect_thresh, uint, S_IRUGO);
>  
> +static unsigned int indirect_alloc_thresh;
> +module_param(indirect_alloc_thresh, uint, S_IRUGO);
> +
>  struct virtio_blk
>  {
>  	struct virtio_device *vdev;
> @@ -739,6 +742,7 @@ static int __devinit virtblk_probe(struct virtio_device *vdev)
>  	INIT_WORK(&vblk->config_work, virtblk_config_changed_work);
>  	vblk->config_enable = true;
>  	vdev->indirect_thresh = indirect_thresh;
> +	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
>  
>  	err = init_vq(vblk);
>  	if (err)
> diff --git a/drivers/char/hw_random/virtio-rng.c b/drivers/char/hw_random/virtio-rng.c
> index 3a644f1..ed22db8 100644
> --- a/drivers/char/hw_random/virtio-rng.c
> +++ b/drivers/char/hw_random/virtio-rng.c
> @@ -28,6 +28,9 @@
>  static unsigned int indirect_thresh;
>  module_param(indirect_thresh, uint, S_IRUGO);
>  
> +static unsigned int indirect_alloc_thresh;
> +module_param(indirect_alloc_thresh, uint, S_IRUGO);
> +
>  static struct virtqueue *vq;
>  static unsigned int data_avail;
>  static DECLARE_COMPLETION(have_data);
> @@ -97,6 +100,7 @@ static int probe_common(struct virtio_device *vdev)
>  
>  	/* We expect a single virtqueue. */
>  	vdev->indirect_thresh = indirect_thresh;
> +	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
>  	vq = virtio_find_single_vq(vdev, random_recv_done, "input");
>  	if (IS_ERR(vq))
>  		return PTR_ERR(vq);
> diff --git a/drivers/char/virtio_console.c b/drivers/char/virtio_console.c
> index da2e44c..8f30732 100644
> --- a/drivers/char/virtio_console.c
> +++ b/drivers/char/virtio_console.c
> @@ -42,6 +42,9 @@
>  static unsigned int indirect_thresh;
>  module_param(indirect_thresh, uint, S_IRUGO);
>  
> +static unsigned int indirect_alloc_thresh;
> +module_param(indirect_alloc_thresh, uint, S_IRUGO);
> +
>  /*
>   * This is a global struct for storing common data for all the devices
>   * this driver handles.
> @@ -1891,6 +1894,7 @@ static int __devinit virtcons_probe(struct virtio_device *vdev)
>  			      &portdev->config.max_nr_ports) == 0)
>  		multiport = true;
>  	vdev->indirect_thresh = indirect_thresh;
> +	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
>  
>  	err = init_vqs(portdev);
>  	if (err < 0) {
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 949c89e..a00e19d 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -37,6 +37,9 @@ module_param(gso, bool, 0444);
>  static unsigned int indirect_thresh = 16;
>  module_param(indirect_thresh, uint, S_IRUGO);
>  
> +static unsigned int indirect_alloc_thresh = 16;

Why 16?  Please make is MAX_SG + 1 this makes some sense.


> +module_param(indirect_alloc_thresh, uint, S_IRUGO);
> +
>  /* FIXME: MTU in config. */
>  #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
>  #define GOOD_COPY_LEN	128
> @@ -1132,6 +1135,7 @@ static int virtnet_probe(struct virtio_device *vdev)
>  	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
>  		vi->mergeable_rx_bufs = true;
>  	vdev->indirect_thresh = indirect_thresh;
> +	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
>  
>  	err = init_vqs(vi);
>  	if (err)
> diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c
> index ca5ae7a..039c4a6 100644
> --- a/drivers/virtio/virtio_balloon.c
> +++ b/drivers/virtio/virtio_balloon.c
> @@ -38,6 +38,9 @@
>  static unsigned int indirect_thresh;
>  module_param(indirect_thresh, uint, S_IRUGO);
>  
> +static unsigned int indirect_alloc_thresh;
> +module_param(indirect_alloc_thresh, uint, S_IRUGO);
> +
>  struct virtio_balloon
>  {
>  	struct virtio_device *vdev;
> @@ -360,6 +363,7 @@ static int virtballoon_probe(struct virtio_device *vdev)
>  	vb->vdev = vdev;
>  	vb->need_stats_update = 0;
>  	vdev->indirect_thresh = indirect_thresh;
> +	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
>  
>  	err = init_vqs(vb);
>  	if (err)
> diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> index 4063e03..dde867b 100644
> --- a/drivers/virtio/virtio_ring.c
> +++ b/drivers/virtio/virtio_ring.c
> @@ -93,6 +93,10 @@ struct vring_virtqueue
>  	 */
>  	unsigned int indirect_thresh;
>  
> +	/* Buffers below this size will be allocated from cache */
> +	unsigned int indirect_alloc_thresh;
> +	struct kmem_cache *indirect_cache;
> +
>  	/* Host publishes avail event idx */
>  	bool event;
>  
> @@ -135,7 +139,10 @@ static int vring_add_indirect(struct vring_virtqueue *vq,
>  	unsigned head;
>  	int i;
>  
> -	desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
> +	if ((out + in) <= vq->indirect_alloc_thresh)
> +		desc = kmem_cache_alloc(vq->indirect_cache, gfp);
> +	else
> +		desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
>  	if (!desc)
>  		return -ENOMEM;
>  
> @@ -384,8 +391,14 @@ static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
>  	i = head;
>  
>  	/* Free the indirect table */
> -	if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT)
> -		kfree(phys_to_virt(vq->vring.desc[i].addr));
> +	if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT) {
> +		u32 descs = vq->vring.desc[i].len / sizeof(struct vring_desc);
> +		if (descs > vq->indirect_alloc_thresh)
> +			kfree(phys_to_virt(vq->vring.desc[i].addr));
> +		else
> +			kmem_cache_free(vq->indirect_cache,
> +					phys_to_virt(vq->vring.desc[i].addr));
> +	}

If logic in two chunks above does not match it all
blows up. So let's add a helper is_cache(vq, buf)
and call from both places.

>  
>  	while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
>  		i = vq->vring.desc[i].next;
> @@ -654,14 +667,25 @@ struct virtqueue *vring_new_virtqueue(unsigned int num,
>  	vq->last_used_idx = 0;
>  	vq->num_added = 0;
>  	vq->indirect_thresh = 0;
> +	vq->indirect_alloc_thresh = 0;
> +	vq->indirect_cache = NULL;
>  	list_add_tail(&vq->vq.list, &vdev->vqs);
>  #ifdef DEBUG
>  	vq->in_use = false;
>  	vq->last_add_time_valid = false;
>  #endif
>  
> -	if (virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
> +	if (virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC)) {
>  		vq->indirect_thresh = vdev->indirect_thresh;
> +		vq->indirect_alloc_thresh = vdev->indirect_alloc_thresh;

This means for virtio-net we still create a cache for both
TX and RX, but it's later unused for RX unless
big packet mode is set.
Pls make this flag per vq not per device.

> +		if (vq->indirect_alloc_thresh) {
> +			vq->indirect_cache =
> +				KMEM_CACHE(vring_desc[vq->indirect_alloc_thresh], 0);
> +
> +			if (vq->indirect_cache == NULL)
> +				vq->indirect_alloc_thresh = 0;
> +		}
> +	}
>  
>  	vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
>  
> @@ -685,6 +709,8 @@ EXPORT_SYMBOL_GPL(vring_new_virtqueue);
>  void vring_del_virtqueue(struct virtqueue *vq)
>  {
>  	list_del(&vq->list);
> +	if (to_vvq(vq)->indirect_cache)
> +		kmem_cache_destroy(to_vvq(vq)->indirect_cache);
>  	kfree(to_vvq(vq));
>  }
>  EXPORT_SYMBOL_GPL(vring_del_virtqueue);
> diff --git a/include/linux/virtio.h b/include/linux/virtio.h
> index 48bc457..3261c02 100644
> --- a/include/linux/virtio.h
> +++ b/include/linux/virtio.h
> @@ -70,6 +70,7 @@ struct virtio_device {
>  	unsigned long features[1];
>  	void *priv;
>  	unsigned int indirect_thresh;
> +	unsigned int indirect_alloc_thresh;
>  };
>  
>  #define dev_to_virtio(dev) container_of(dev, struct virtio_device, dev)

So what is a reasonable value?
It would be such that most bufs have # of s/g below it.
So I think 'expected_sg' would be a better name,
add documentation explaining what it is.

> diff --git a/net/9p/trans_virtio.c b/net/9p/trans_virtio.c
> index 418f933..058b6dd 100644
> --- a/net/9p/trans_virtio.c
> +++ b/net/9p/trans_virtio.c
> @@ -55,6 +55,9 @@
>  static unsigned int indirect_thresh;
>  module_param(indirect_thresh, uint, S_IRUGO);
>  
> +static unsigned int indirect_alloc_thresh;
> +module_param(indirect_alloc_thresh, uint, S_IRUGO);
> +
>  /* a single mutex to manage channel initialization and attachment */
>  static DEFINE_MUTEX(virtio_9p_lock);
>  static DECLARE_WAIT_QUEUE_HEAD(vp_wq);
> @@ -505,6 +508,8 @@ static int p9_virtio_probe(struct virtio_device *vdev)
>  
>  	/* We expect one virtqueue, for requests. */
>  	vdev->indirect_thresh = indirect_thresh;
> +	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
> +
>  	chan->vq = virtio_find_single_vq(vdev, req_done, "requests");
>  	if (IS_ERR(chan->vq)) {
>  		err = PTR_ERR(chan->vq);
> -- 
> 1.7.12
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Sasha Levin Aug. 31, 2012, 9:36 a.m. UTC | #2
On 08/30/2012 03:38 PM, Michael S. Tsirkin wrote:
>> +static unsigned int indirect_alloc_thresh = 16;
> Why 16?  Please make is MAX_SG + 1 this makes some sense.

Wouldn't MAX_SG mean we always allocate from the cache? Isn't the memory waste
too big in this case?

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Michael S. Tsirkin Aug. 31, 2012, 9:56 a.m. UTC | #3
On Fri, Aug 31, 2012 at 11:36:07AM +0200, Sasha Levin wrote:
> On 08/30/2012 03:38 PM, Michael S. Tsirkin wrote:
> >> +static unsigned int indirect_alloc_thresh = 16;
> > Why 16?  Please make is MAX_SG + 1 this makes some sense.
> 
> Wouldn't MAX_SG mean we always allocate from the cache? Isn't the memory waste
> too big in this case?

Sorry. I really meant MAX_SKB_FRAGS + 1. MAX_SKB_FRAGS is 17 so gets us
threshold of 18. It is less than the size of an skb+shinfo itself so -
does it look too big to you? Also why do you think 16 is not too big but
18 is?  If there's a reason then I am fine with 16 too but then please
put it in code comment near where the value is set.

Yes this means virtio net always allocates from cache
but this is a good thing, isn't it? Gets us more consistent
performance.
Avi Kivity Sept. 4, 2012, 4:34 p.m. UTC | #4
On 08/31/2012 12:56 PM, Michael S. Tsirkin wrote:
> On Fri, Aug 31, 2012 at 11:36:07AM +0200, Sasha Levin wrote:
>> On 08/30/2012 03:38 PM, Michael S. Tsirkin wrote:
>> >> +static unsigned int indirect_alloc_thresh = 16;
>> > Why 16?  Please make is MAX_SG + 1 this makes some sense.
>> 
>> Wouldn't MAX_SG mean we always allocate from the cache? Isn't the memory waste
>> too big in this case?
> 
> Sorry. I really meant MAX_SKB_FRAGS + 1. MAX_SKB_FRAGS is 17 so gets us
> threshold of 18. It is less than the size of an skb+shinfo itself so -
> does it look too big to you? Also why do you think 16 is not too big but
> 18 is?  If there's a reason then I am fine with 16 too but then please
> put it in code comment near where the value is set.
> 
> Yes this means virtio net always allocates from cache
> but this is a good thing, isn't it? Gets us more consistent
> performance.

kmalloc() also goes to a cache.  Is there a measurable difference?

Ugh, there's an ugly loop in __find_general_cachep(), which really wants
to be replaced with fls().
Avi Kivity Sept. 4, 2012, 4:36 p.m. UTC | #5
On 09/04/2012 07:34 PM, Avi Kivity wrote:
> On 08/31/2012 12:56 PM, Michael S. Tsirkin wrote:
>> On Fri, Aug 31, 2012 at 11:36:07AM +0200, Sasha Levin wrote:
>>> On 08/30/2012 03:38 PM, Michael S. Tsirkin wrote:
>>> >> +static unsigned int indirect_alloc_thresh = 16;
>>> > Why 16?  Please make is MAX_SG + 1 this makes some sense.
>>> 
>>> Wouldn't MAX_SG mean we always allocate from the cache? Isn't the memory waste
>>> too big in this case?
>> 
>> Sorry. I really meant MAX_SKB_FRAGS + 1. MAX_SKB_FRAGS is 17 so gets us
>> threshold of 18. It is less than the size of an skb+shinfo itself so -
>> does it look too big to you? Also why do you think 16 is not too big but
>> 18 is?  If there's a reason then I am fine with 16 too but then please
>> put it in code comment near where the value is set.
>> 
>> Yes this means virtio net always allocates from cache
>> but this is a good thing, isn't it? Gets us more consistent
>> performance.
> 
> kmalloc() also goes to a cache.  Is there a measurable difference?
> 
> Ugh, there's an ugly loop in __find_general_cachep(), which really wants
> to be replaced with fls().
> 

Actually, not, as the loop will be very short for small sizes.
Michael S. Tsirkin Sept. 4, 2012, 6:41 p.m. UTC | #6
On Tue, Sep 04, 2012 at 07:34:19PM +0300, Avi Kivity wrote:
> On 08/31/2012 12:56 PM, Michael S. Tsirkin wrote:
> > On Fri, Aug 31, 2012 at 11:36:07AM +0200, Sasha Levin wrote:
> >> On 08/30/2012 03:38 PM, Michael S. Tsirkin wrote:
> >> >> +static unsigned int indirect_alloc_thresh = 16;
> >> > Why 16?  Please make is MAX_SG + 1 this makes some sense.
> >> 
> >> Wouldn't MAX_SG mean we always allocate from the cache? Isn't the memory waste
> >> too big in this case?
> > 
> > Sorry. I really meant MAX_SKB_FRAGS + 1. MAX_SKB_FRAGS is 17 so gets us
> > threshold of 18. It is less than the size of an skb+shinfo itself so -
> > does it look too big to you? Also why do you think 16 is not too big but
> > 18 is?  If there's a reason then I am fine with 16 too but then please
> > put it in code comment near where the value is set.
> > 
> > Yes this means virtio net always allocates from cache
> > but this is a good thing, isn't it? Gets us more consistent
> > performance.
> 
> kmalloc() also goes to a cache.  Is there a measurable difference?

Yes see 0/2 and followup discussion.

> Ugh, there's an ugly loop in __find_general_cachep(), which really wants
> to be replaced with fls().
> 
> -- 
> error compiling committee.c: too many arguments to function
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Avi Kivity Sept. 5, 2012, 2:21 p.m. UTC | #7
On 09/04/2012 09:41 PM, Michael S. Tsirkin wrote:
> On Tue, Sep 04, 2012 at 07:34:19PM +0300, Avi Kivity wrote:
>> On 08/31/2012 12:56 PM, Michael S. Tsirkin wrote:
>> > On Fri, Aug 31, 2012 at 11:36:07AM +0200, Sasha Levin wrote:
>> >> On 08/30/2012 03:38 PM, Michael S. Tsirkin wrote:
>> >> >> +static unsigned int indirect_alloc_thresh = 16;
>> >> > Why 16?  Please make is MAX_SG + 1 this makes some sense.
>> >> 
>> >> Wouldn't MAX_SG mean we always allocate from the cache? Isn't the memory waste
>> >> too big in this case?
>> > 
>> > Sorry. I really meant MAX_SKB_FRAGS + 1. MAX_SKB_FRAGS is 17 so gets us
>> > threshold of 18. It is less than the size of an skb+shinfo itself so -
>> > does it look too big to you? Also why do you think 16 is not too big but
>> > 18 is?  If there's a reason then I am fine with 16 too but then please
>> > put it in code comment near where the value is set.
>> > 
>> > Yes this means virtio net always allocates from cache
>> > but this is a good thing, isn't it? Gets us more consistent
>> > performance.
>> 
>> kmalloc() also goes to a cache.  Is there a measurable difference?
> 
> Yes see 0/2 and followup discussion.

I don't see 0/2, looks like this was not threaded properly.  What was
the subject line?
Michael S. Tsirkin Sept. 5, 2012, 2:27 p.m. UTC | #8
On Wed, Sep 05, 2012 at 05:21:12PM +0300, Avi Kivity wrote:
> On 09/04/2012 09:41 PM, Michael S. Tsirkin wrote:
> > On Tue, Sep 04, 2012 at 07:34:19PM +0300, Avi Kivity wrote:
> >> On 08/31/2012 12:56 PM, Michael S. Tsirkin wrote:
> >> > On Fri, Aug 31, 2012 at 11:36:07AM +0200, Sasha Levin wrote:
> >> >> On 08/30/2012 03:38 PM, Michael S. Tsirkin wrote:
> >> >> >> +static unsigned int indirect_alloc_thresh = 16;
> >> >> > Why 16?  Please make is MAX_SG + 1 this makes some sense.
> >> >> 
> >> >> Wouldn't MAX_SG mean we always allocate from the cache? Isn't the memory waste
> >> >> too big in this case?
> >> > 
> >> > Sorry. I really meant MAX_SKB_FRAGS + 1. MAX_SKB_FRAGS is 17 so gets us
> >> > threshold of 18. It is less than the size of an skb+shinfo itself so -
> >> > does it look too big to you? Also why do you think 16 is not too big but
> >> > 18 is?  If there's a reason then I am fine with 16 too but then please
> >> > put it in code comment near where the value is set.
> >> > 
> >> > Yes this means virtio net always allocates from cache
> >> > but this is a good thing, isn't it? Gets us more consistent
> >> > performance.
> >> 
> >> kmalloc() also goes to a cache.  Is there a measurable difference?
> > 
> > Yes see 0/2 and followup discussion.
> 
> I don't see 0/2, looks like this was not threaded properly.  What was
> the subject line?

My mistake, there is no 0/2, the resolts where in the followup thread
of the previous version:
	[PATCH v2 2/2] virtio-ring: Allocate indirect buffers from cache when possible

Sasha, could you please accompany the next version
with a cover letter 0/2) including performance results?

> 
> -- 
> error compiling committee.c: too many arguments to function
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
index 7c63065..e4c6c42 100644
--- a/drivers/block/virtio_blk.c
+++ b/drivers/block/virtio_blk.c
@@ -25,6 +25,9 @@  struct workqueue_struct *virtblk_wq;
 static unsigned int indirect_thresh;
 module_param(indirect_thresh, uint, S_IRUGO);
 
+static unsigned int indirect_alloc_thresh;
+module_param(indirect_alloc_thresh, uint, S_IRUGO);
+
 struct virtio_blk
 {
 	struct virtio_device *vdev;
@@ -739,6 +742,7 @@  static int __devinit virtblk_probe(struct virtio_device *vdev)
 	INIT_WORK(&vblk->config_work, virtblk_config_changed_work);
 	vblk->config_enable = true;
 	vdev->indirect_thresh = indirect_thresh;
+	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
 
 	err = init_vq(vblk);
 	if (err)
diff --git a/drivers/char/hw_random/virtio-rng.c b/drivers/char/hw_random/virtio-rng.c
index 3a644f1..ed22db8 100644
--- a/drivers/char/hw_random/virtio-rng.c
+++ b/drivers/char/hw_random/virtio-rng.c
@@ -28,6 +28,9 @@ 
 static unsigned int indirect_thresh;
 module_param(indirect_thresh, uint, S_IRUGO);
 
+static unsigned int indirect_alloc_thresh;
+module_param(indirect_alloc_thresh, uint, S_IRUGO);
+
 static struct virtqueue *vq;
 static unsigned int data_avail;
 static DECLARE_COMPLETION(have_data);
@@ -97,6 +100,7 @@  static int probe_common(struct virtio_device *vdev)
 
 	/* We expect a single virtqueue. */
 	vdev->indirect_thresh = indirect_thresh;
+	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
 	vq = virtio_find_single_vq(vdev, random_recv_done, "input");
 	if (IS_ERR(vq))
 		return PTR_ERR(vq);
diff --git a/drivers/char/virtio_console.c b/drivers/char/virtio_console.c
index da2e44c..8f30732 100644
--- a/drivers/char/virtio_console.c
+++ b/drivers/char/virtio_console.c
@@ -42,6 +42,9 @@ 
 static unsigned int indirect_thresh;
 module_param(indirect_thresh, uint, S_IRUGO);
 
+static unsigned int indirect_alloc_thresh;
+module_param(indirect_alloc_thresh, uint, S_IRUGO);
+
 /*
  * This is a global struct for storing common data for all the devices
  * this driver handles.
@@ -1891,6 +1894,7 @@  static int __devinit virtcons_probe(struct virtio_device *vdev)
 			      &portdev->config.max_nr_ports) == 0)
 		multiport = true;
 	vdev->indirect_thresh = indirect_thresh;
+	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
 
 	err = init_vqs(portdev);
 	if (err < 0) {
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 949c89e..a00e19d 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -37,6 +37,9 @@  module_param(gso, bool, 0444);
 static unsigned int indirect_thresh = 16;
 module_param(indirect_thresh, uint, S_IRUGO);
 
+static unsigned int indirect_alloc_thresh = 16;
+module_param(indirect_alloc_thresh, uint, S_IRUGO);
+
 /* FIXME: MTU in config. */
 #define MAX_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
 #define GOOD_COPY_LEN	128
@@ -1132,6 +1135,7 @@  static int virtnet_probe(struct virtio_device *vdev)
 	if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
 		vi->mergeable_rx_bufs = true;
 	vdev->indirect_thresh = indirect_thresh;
+	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
 
 	err = init_vqs(vi);
 	if (err)
diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c
index ca5ae7a..039c4a6 100644
--- a/drivers/virtio/virtio_balloon.c
+++ b/drivers/virtio/virtio_balloon.c
@@ -38,6 +38,9 @@ 
 static unsigned int indirect_thresh;
 module_param(indirect_thresh, uint, S_IRUGO);
 
+static unsigned int indirect_alloc_thresh;
+module_param(indirect_alloc_thresh, uint, S_IRUGO);
+
 struct virtio_balloon
 {
 	struct virtio_device *vdev;
@@ -360,6 +363,7 @@  static int virtballoon_probe(struct virtio_device *vdev)
 	vb->vdev = vdev;
 	vb->need_stats_update = 0;
 	vdev->indirect_thresh = indirect_thresh;
+	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
 
 	err = init_vqs(vb);
 	if (err)
diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 4063e03..dde867b 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -93,6 +93,10 @@  struct vring_virtqueue
 	 */
 	unsigned int indirect_thresh;
 
+	/* Buffers below this size will be allocated from cache */
+	unsigned int indirect_alloc_thresh;
+	struct kmem_cache *indirect_cache;
+
 	/* Host publishes avail event idx */
 	bool event;
 
@@ -135,7 +139,10 @@  static int vring_add_indirect(struct vring_virtqueue *vq,
 	unsigned head;
 	int i;
 
-	desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
+	if ((out + in) <= vq->indirect_alloc_thresh)
+		desc = kmem_cache_alloc(vq->indirect_cache, gfp);
+	else
+		desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
 	if (!desc)
 		return -ENOMEM;
 
@@ -384,8 +391,14 @@  static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
 	i = head;
 
 	/* Free the indirect table */
-	if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT)
-		kfree(phys_to_virt(vq->vring.desc[i].addr));
+	if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT) {
+		u32 descs = vq->vring.desc[i].len / sizeof(struct vring_desc);
+		if (descs > vq->indirect_alloc_thresh)
+			kfree(phys_to_virt(vq->vring.desc[i].addr));
+		else
+			kmem_cache_free(vq->indirect_cache,
+					phys_to_virt(vq->vring.desc[i].addr));
+	}
 
 	while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
 		i = vq->vring.desc[i].next;
@@ -654,14 +667,25 @@  struct virtqueue *vring_new_virtqueue(unsigned int num,
 	vq->last_used_idx = 0;
 	vq->num_added = 0;
 	vq->indirect_thresh = 0;
+	vq->indirect_alloc_thresh = 0;
+	vq->indirect_cache = NULL;
 	list_add_tail(&vq->vq.list, &vdev->vqs);
 #ifdef DEBUG
 	vq->in_use = false;
 	vq->last_add_time_valid = false;
 #endif
 
-	if (virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
+	if (virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC)) {
 		vq->indirect_thresh = vdev->indirect_thresh;
+		vq->indirect_alloc_thresh = vdev->indirect_alloc_thresh;
+		if (vq->indirect_alloc_thresh) {
+			vq->indirect_cache =
+				KMEM_CACHE(vring_desc[vq->indirect_alloc_thresh], 0);
+
+			if (vq->indirect_cache == NULL)
+				vq->indirect_alloc_thresh = 0;
+		}
+	}
 
 	vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
 
@@ -685,6 +709,8 @@  EXPORT_SYMBOL_GPL(vring_new_virtqueue);
 void vring_del_virtqueue(struct virtqueue *vq)
 {
 	list_del(&vq->list);
+	if (to_vvq(vq)->indirect_cache)
+		kmem_cache_destroy(to_vvq(vq)->indirect_cache);
 	kfree(to_vvq(vq));
 }
 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
diff --git a/include/linux/virtio.h b/include/linux/virtio.h
index 48bc457..3261c02 100644
--- a/include/linux/virtio.h
+++ b/include/linux/virtio.h
@@ -70,6 +70,7 @@  struct virtio_device {
 	unsigned long features[1];
 	void *priv;
 	unsigned int indirect_thresh;
+	unsigned int indirect_alloc_thresh;
 };
 
 #define dev_to_virtio(dev) container_of(dev, struct virtio_device, dev)
diff --git a/net/9p/trans_virtio.c b/net/9p/trans_virtio.c
index 418f933..058b6dd 100644
--- a/net/9p/trans_virtio.c
+++ b/net/9p/trans_virtio.c
@@ -55,6 +55,9 @@ 
 static unsigned int indirect_thresh;
 module_param(indirect_thresh, uint, S_IRUGO);
 
+static unsigned int indirect_alloc_thresh;
+module_param(indirect_alloc_thresh, uint, S_IRUGO);
+
 /* a single mutex to manage channel initialization and attachment */
 static DEFINE_MUTEX(virtio_9p_lock);
 static DECLARE_WAIT_QUEUE_HEAD(vp_wq);
@@ -505,6 +508,8 @@  static int p9_virtio_probe(struct virtio_device *vdev)
 
 	/* We expect one virtqueue, for requests. */
 	vdev->indirect_thresh = indirect_thresh;
+	vdev->indirect_alloc_thresh = indirect_alloc_thresh;
+
 	chan->vq = virtio_find_single_vq(vdev, req_done, "requests");
 	if (IS_ERR(chan->vq)) {
 		err = PTR_ERR(chan->vq);