[PATCH] block: virtio_blk: don't hold spin lock during world switch

[PATCH] block: virtio_blk: don't hold spin lock during world switch

[Ming Lei]

Firstly, it isn't necessary to hold lock of vblk->vq_lock
when notifying hypervisor about queued I/O.

Secondly, virtqueue_notify() will cause world switch and
it may take long time on some hypervisors(such as, qemu-arm),
so it isn't good to hold the lock and block other vCPUs.

On arm64 quad core VM(qemu-kvm), the patch can increase I/O
performance a lot with VIRTIO_RING_F_EVENT_IDX enabled:
	- without the patch: 14K IOPS
	- with the patch: 34K IOPS

fio script:
	[global]
	direct=1
	bsrange=4k-4k
	timeout=10
	numjobs=4
	ioengine=libaio
	iodepth=64

	filename=/dev/vdc
	group_reporting=1

	[f1]
	rw=randread

Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: virtualization@lists.linux-foundation.org
Signed-off-by: Ming Lei <ming.lei@canonical.com>
---
 drivers/block/virtio_blk.c |    9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
index 9f340fa..a6f5424 100644
--- a/drivers/block/virtio_blk.c
+++ b/drivers/block/virtio_blk.c
@@ -162,6 +162,7 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *req)
 	unsigned int num;
 	const bool last = (req->cmd_flags & REQ_END) != 0;
 	int err;
+	bool notify = false;
 
 	BUG_ON(req->nr_phys_segments + 2 > vblk->sg_elems);
 
@@ -214,10 +215,12 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *req)
 		return BLK_MQ_RQ_QUEUE_ERROR;
 	}
 
-	if (last)
-		virtqueue_kick(vblk->vq);
-
+	if (last && virtqueue_kick_prepare(vblk->vq))
+		notify = true;
 	spin_unlock_irqrestore(&vblk->vq_lock, flags);
+
+	if (notify)
+		virtqueue_notify(vblk->vq);
 	return BLK_MQ_RQ_QUEUE_OK;
 }
 
-- 
1.7.9.5

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Jens Axboe]

On 2014-05-29 20:49, Ming Lei wrote:
> Firstly, it isn't necessary to hold lock of vblk->vq_lock
> when notifying hypervisor about queued I/O.
>
> Secondly, virtqueue_notify() will cause world switch and
> it may take long time on some hypervisors(such as, qemu-arm),
> so it isn't good to hold the lock and block other vCPUs.
>
> On arm64 quad core VM(qemu-kvm), the patch can increase I/O
> performance a lot with VIRTIO_RING_F_EVENT_IDX enabled:
> 	- without the patch: 14K IOPS
> 	- with the patch: 34K IOPS

Patch looks good to me. I don't see a hit on my qemu-kvm testing, but it 
definitely makes sense and I can see it hurting in other places.

-- 
Jens Axboe

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Ming Lei]

On Fri, May 30, 2014 at 11:19 AM, Jens Axboe <axboe@kernel.dk> wrote:
> On 2014-05-29 20:49, Ming Lei wrote:
>>
>> Firstly, it isn't necessary to hold lock of vblk->vq_lock
>> when notifying hypervisor about queued I/O.
>>
>> Secondly, virtqueue_notify() will cause world switch and
>> it may take long time on some hypervisors(such as, qemu-arm),
>> so it isn't good to hold the lock and block other vCPUs.
>>
>> On arm64 quad core VM(qemu-kvm), the patch can increase I/O
>> performance a lot with VIRTIO_RING_F_EVENT_IDX enabled:
>>         - without the patch: 14K IOPS
>>         - with the patch: 34K IOPS
>
>
> Patch looks good to me. I don't see a hit on my qemu-kvm testing, but it
> definitely makes sense and I can see it hurting in other places.

It isn't easy to observe the improvement on x86 VM, especially
with few vCPUs, because qemu-system-x86_64 only takes
several microseconds to handle the notification, but on arm64, it
may take hundreds of microseconds, so the improvement is
obvious on arm VM.

I hope this patch can be merged, at least arm VM can benefit
from it.


Thanks,
--
Ming Lei

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Jens Axboe]

On 2014-05-29 21:34, Ming Lei wrote:
> On Fri, May 30, 2014 at 11:19 AM, Jens Axboe <axboe@kernel.dk> wrote:
>> On 2014-05-29 20:49, Ming Lei wrote:
>>>
>>> Firstly, it isn't necessary to hold lock of vblk->vq_lock
>>> when notifying hypervisor about queued I/O.
>>>
>>> Secondly, virtqueue_notify() will cause world switch and
>>> it may take long time on some hypervisors(such as, qemu-arm),
>>> so it isn't good to hold the lock and block other vCPUs.
>>>
>>> On arm64 quad core VM(qemu-kvm), the patch can increase I/O
>>> performance a lot with VIRTIO_RING_F_EVENT_IDX enabled:
>>>          - without the patch: 14K IOPS
>>>          - with the patch: 34K IOPS
>>
>>
>> Patch looks good to me. I don't see a hit on my qemu-kvm testing, but it
>> definitely makes sense and I can see it hurting in other places.
>
> It isn't easy to observe the improvement on x86 VM, especially
> with few vCPUs, because qemu-system-x86_64 only takes
> several microseconds to handle the notification, but on arm64, it
> may take hundreds of microseconds, so the improvement is
> obvious on arm VM.
>
> I hope this patch can be merged, at least arm VM can benefit
> from it.

If Rusty agrees, I'd like to add it for 3.16 with a stable marker.

-- 
Jens Axboe

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Ming Lei]

On Fri, May 30, 2014 at 11:35 AM, Jens Axboe <axboe@kernel.dk> wrote:
> On 2014-05-29 21:34, Ming Lei wrote:
>>
>> On Fri, May 30, 2014 at 11:19 AM, Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>> On 2014-05-29 20:49, Ming Lei wrote:
>>>>
>>>>
>>>> Firstly, it isn't necessary to hold lock of vblk->vq_lock
>>>> when notifying hypervisor about queued I/O.
>>>>
>>>> Secondly, virtqueue_notify() will cause world switch and
>>>> it may take long time on some hypervisors(such as, qemu-arm),
>>>> so it isn't good to hold the lock and block other vCPUs.
>>>>
>>>> On arm64 quad core VM(qemu-kvm), the patch can increase I/O
>>>> performance a lot with VIRTIO_RING_F_EVENT_IDX enabled:
>>>>          - without the patch: 14K IOPS
>>>>          - with the patch: 34K IOPS
>>>
>>>
>>>
>>> Patch looks good to me. I don't see a hit on my qemu-kvm testing, but it
>>> definitely makes sense and I can see it hurting in other places.
>>
>>
>> It isn't easy to observe the improvement on x86 VM, especially
>> with few vCPUs, because qemu-system-x86_64 only takes
>> several microseconds to handle the notification, but on arm64, it
>> may take hundreds of microseconds, so the improvement is
>> obvious on arm VM.
>>
>> I hope this patch can be merged, at least arm VM can benefit
>> from it.
>
>
> If Rusty agrees, I'd like to add it for 3.16 with a stable marker.

Interesting, even on x86, I still can observe the improvement
when the numjobs is set as 2 in the fio script(see commit log), but
when numjobs is set as 4, 8, 12, the difference isn't obvious between
patched kernel and non-patched kernel.

1, environment
- host: 2sockets, each CPU(4cores, 2 threads), total 16 logical cores
- guest: 16cores, 8GB ram
- guest kernel: 3.15-rc7-next with patch[1]
- fio: the script in commit log with numjobs set as 2

2, result
- without the patch: ~104K IOPS
- with the patch: ~140K IOPS


Rusty, considered the same trick has been applied in virt-scsi,
do you agree to take the same approach in virt-blk too?


[1], http://marc.info/?l=linux-kernel&m=140135041423441&w=2

Thanks,
--
Ming Lei

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Rusty Russell]

Ming Lei <ming.lei@canonical.com> writes:
> Firstly, it isn't necessary to hold lock of vblk->vq_lock
> when notifying hypervisor about queued I/O.
>
> Secondly, virtqueue_notify() will cause world switch and
> it may take long time on some hypervisors(such as, qemu-arm),
> so it isn't good to hold the lock and block other vCPUs.
>
> On arm64 quad core VM(qemu-kvm), the patch can increase I/O
> performance a lot with VIRTIO_RING_F_EVENT_IDX enabled:
> 	- without the patch: 14K IOPS
> 	- with the patch: 34K IOPS
>
> fio script:
> 	[global]
> 	direct=1
> 	bsrange=4k-4k
> 	timeout=10
> 	numjobs=4
> 	ioengine=libaio
> 	iodepth=64
>
> 	filename=/dev/vdc
> 	group_reporting=1
>
> 	[f1]
> 	rw=randread
>
> Cc: Rusty Russell <rusty@rustcorp.com.au>
> Cc: "Michael S. Tsirkin" <mst@redhat.com>
> Cc: virtualization@lists.linux-foundation.org
> Signed-off-by: Ming Lei <ming.lei@canonical.com>

Acked-by: Rusty Russell <rusty@rustcorp.com.au>

Thanks!
Rusty.

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Rusty Russell]

Jens Axboe <axboe@kernel.dk> writes:
> If Rusty agrees, I'd like to add it for 3.16 with a stable marker.

Really stable?  It improves performance, which is nice.  But every patch
which goes into the kernel fixes a bug, improves clarity, improves
performance or adds a feature.  I've now seen all four cases get CC'd
into stable.

Including some of mine explicitly not marked stable which get swept up
by enthusiastic stable maintainers :(

Is now there *any* patch short of a major rewrite which shouldn't get
cc: stable?

Cheers,
Rusty.

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Jens Axboe]

On 2014-05-30 00:10, Rusty Russell wrote:
> Jens Axboe <axboe@kernel.dk> writes:
>> If Rusty agrees, I'd like to add it for 3.16 with a stable marker.
>
> Really stable?  It improves performance, which is nice.  But every patch
> which goes into the kernel fixes a bug, improves clarity, improves
> performance or adds a feature.  I've now seen all four cases get CC'd
> into stable.
>
> Including some of mine explicitly not marked stable which get swept up
> by enthusiastic stable maintainers :(
>
> Is now there *any* patch short of a major rewrite which shouldn't get
> cc: stable?

I agree that there's sometimes an unfortunate trend there. I didn't 
check, but my assumption was that this is a regression after the blk-mq 
conversion, in which case I do think it belongs in stable.

But in any case, I think the patch is obviously correct and the wins are 
sufficiently large to warrant a stable inclusion even if it isn't a 
regression.

-- 
Jens Axboe

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Michael S. Tsirkin]

On Fri, May 30, 2014 at 10:49:29AM +0800, Ming Lei wrote:
> Firstly, it isn't necessary to hold lock of vblk->vq_lock
> when notifying hypervisor about queued I/O.
> 
> Secondly, virtqueue_notify() will cause world switch and
> it may take long time on some hypervisors(such as, qemu-arm),
> so it isn't good to hold the lock and block other vCPUs.
> 
> On arm64 quad core VM(qemu-kvm), the patch can increase I/O
> performance a lot with VIRTIO_RING_F_EVENT_IDX enabled:
> 	- without the patch: 14K IOPS
> 	- with the patch: 34K IOPS
> 
> fio script:
> 	[global]
> 	direct=1
> 	bsrange=4k-4k
> 	timeout=10
> 	numjobs=4
> 	ioengine=libaio
> 	iodepth=64
> 
> 	filename=/dev/vdc
> 	group_reporting=1
> 
> 	[f1]
> 	rw=randread
> 
> Cc: Rusty Russell <rusty@rustcorp.com.au>
> Cc: "Michael S. Tsirkin" <mst@redhat.com>
> Cc: virtualization@lists.linux-foundation.org
> Signed-off-by: Ming Lei <ming.lei@canonical.com>

Acked-by: Michael S. Tsirkin <mst@redhat.com>

> ---
>  drivers/block/virtio_blk.c |    9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
> index 9f340fa..a6f5424 100644
> --- a/drivers/block/virtio_blk.c
> +++ b/drivers/block/virtio_blk.c
> @@ -162,6 +162,7 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *req)
>  	unsigned int num;
>  	const bool last = (req->cmd_flags & REQ_END) != 0;
>  	int err;
> +	bool notify = false;
>  
>  	BUG_ON(req->nr_phys_segments + 2 > vblk->sg_elems);
>  
> @@ -214,10 +215,12 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *req)
>  		return BLK_MQ_RQ_QUEUE_ERROR;
>  	}
>  
> -	if (last)
> -		virtqueue_kick(vblk->vq);
> -
> +	if (last && virtqueue_kick_prepare(vblk->vq))
> +		notify = true;
>  	spin_unlock_irqrestore(&vblk->vq_lock, flags);
> +
> +	if (notify)
> +		virtqueue_notify(vblk->vq);
>  	return BLK_MQ_RQ_QUEUE_OK;
>  }
>  
> -- 
> 1.7.9.5

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Rusty Russell]

Jens Axboe <axboe@kernel.dk> writes:
> On 2014-05-30 00:10, Rusty Russell wrote:
>> Jens Axboe <axboe@kernel.dk> writes:
>>> If Rusty agrees, I'd like to add it for 3.16 with a stable marker.
>>
>> Really stable?  It improves performance, which is nice.  But every patch
>> which goes into the kernel fixes a bug, improves clarity, improves
>> performance or adds a feature.  I've now seen all four cases get CC'd
>> into stable.
>>
>> Including some of mine explicitly not marked stable which get swept up
>> by enthusiastic stable maintainers :(
>>
>> Is now there *any* patch short of a major rewrite which shouldn't get
>> cc: stable?
>
> I agree that there's sometimes an unfortunate trend there. I didn't 
> check, but my assumption was that this is a regression after the blk-mq 
> conversion, in which case I do think it belongs in stable.

No, it's always been that way.  In the original driver the entire "issue
requests" function was under the lock.

It was your mq conversion which made this optimization possible, and
also made it an optimization: now other the queues can continue while
this one is going.

> But in any case, I think the patch is obviously correct and the wins are 
> sufficiently large to warrant a stable inclusion even if it isn't a 
> regression.

If you're running SMP under an emulator where exits are expensive, then
this wins.  Under KVM it's marginal at best.

Locking changes which are "obviously correct" make me nervous, too :)

But IIRC last KS the argument is that not *enough* is going into stable,
not that stable isn't stable enough.  So maybe it's a non-problem?

Cheers,
Rusty.

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Ming Lei]

On Mon, Jun 2, 2014 at 9:23 AM, Rusty Russell <rusty@rustcorp.com.au> wrote:
> Jens Axboe <axboe@kernel.dk> writes:
>> On 2014-05-30 00:10, Rusty Russell wrote:
>>> Jens Axboe <axboe@kernel.dk> writes:
>>>> If Rusty agrees, I'd like to add it for 3.16 with a stable marker.
>>>
>>> Really stable?  It improves performance, which is nice.  But every patch
>>> which goes into the kernel fixes a bug, improves clarity, improves
>>> performance or adds a feature.  I've now seen all four cases get CC'd
>>> into stable.
>>>
>>> Including some of mine explicitly not marked stable which get swept up
>>> by enthusiastic stable maintainers :(
>>>
>>> Is now there *any* patch short of a major rewrite which shouldn't get
>>> cc: stable?
>>
>> I agree that there's sometimes an unfortunate trend there. I didn't
>> check, but my assumption was that this is a regression after the blk-mq
>> conversion, in which case I do think it belongs in stable.
>
> No, it's always been that way.  In the original driver the entire "issue
> requests" function was under the lock.
>
> It was your mq conversion which made this optimization possible, and
> also made it an optimization: now other the queues can continue while
> this one is going.
>
>> But in any case, I think the patch is obviously correct and the wins are
>> sufficiently large to warrant a stable inclusion even if it isn't a
>> regression.
>
> If you're running SMP under an emulator where exits are expensive, then
> this wins.  Under KVM it's marginal at best.

Both my tests on arm64 and x86 are under KVM, and looks the
patch can improve performance a lot. IMO, even though under
KVM, virtio-blk performance still depends how well hypervisor(
qemu, ...) emulates the device, and basically speaking, it is
expensive to switch from guest to host and let host handle the
notification.


Thanks,
--
Ming Lei

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Jens Axboe]

On 2014-06-01 19:23, Rusty Russell wrote:
> Jens Axboe <axboe@kernel.dk> writes:
>> On 2014-05-30 00:10, Rusty Russell wrote:
>>> Jens Axboe <axboe@kernel.dk> writes:
>>>> If Rusty agrees, I'd like to add it for 3.16 with a stable marker.
>>>
>>> Really stable?  It improves performance, which is nice.  But every patch
>>> which goes into the kernel fixes a bug, improves clarity, improves
>>> performance or adds a feature.  I've now seen all four cases get CC'd
>>> into stable.
>>>
>>> Including some of mine explicitly not marked stable which get swept up
>>> by enthusiastic stable maintainers :(
>>>
>>> Is now there *any* patch short of a major rewrite which shouldn't get
>>> cc: stable?
>>
>> I agree that there's sometimes an unfortunate trend there. I didn't
>> check, but my assumption was that this is a regression after the blk-mq
>> conversion, in which case I do think it belongs in stable.
>
> No, it's always been that way.  In the original driver the entire "issue
> requests" function was under the lock.
 >
 > It was your mq conversion which made this optimization possible, and
 > also made it an optimization: now other the queues can continue while
 > this one is going.

Ah, I stand corrected, you are right. I had this recollection that the 
prepare and kick where separate before as well, but apparently just bad 
memory.

>> But in any case, I think the patch is obviously correct and the wins are
>> sufficiently large to warrant a stable inclusion even if it isn't a
>> regression.
>
> If you're running SMP under an emulator where exits are expensive, then
> this wins.  Under KVM it's marginal at best.
>
> Locking changes which are "obviously correct" make me nervous, too :)

I tend to agree. But I think this one is simple enough to warrant doing 
it, when the performance increase is as large as it is.

> But IIRC last KS the argument is that not *enough* is going into stable,
> not that stable isn't stable enough.  So maybe it's a non-problem?

In principle, pushing the patch to stable definitely isn't an issue with 
the stable crew. And yes, they apparently do want more stuff. If you 
look at it from the distro side, having a stable(r) repository is a no 
brainer. And they'd want to pick this patch anyway, so...

-- 
Jens Axboe

Re: [PATCH] block: virtio_blk: don't hold spin lock during world switch

[Paolo Bonzini]

Il 02/06/2014 15:06, Ming Lei ha scritto:
>> >
>> > If you're running SMP under an emulator where exits are expensive, then
>> > this wins.  Under KVM it's marginal at best.
> Both my tests on arm64 and x86 are under KVM, and looks the
> patch can improve performance a lot. IMO, even though under
> KVM, virtio-blk performance still depends how well hypervisor(
> qemu, ...) emulates the device, and basically speaking, it is
> expensive to switch from guest to host and let host handle the
> notification.

The difference is that virtio-pci supports ioeventfd and virtio-mmio 
doesn't.

With ioeventfd you can tell KVM "I don't care about the value that is 
written to a memory location, only that it is accessed".  Then when the 
write happens, KVM doesn't do an expensive userspace exit; it just 
writes 1 to an eventfd.

It then returns to the guest, userspace picks up the eventfd via its 
poll() loop and services the device.

This is already useful for throughput on UP, and the small latency cost 
(because of the cost of the event loop in the I/O thread, and possibly 
the cost of waking up the thread) is usually offset by the benefit.

But on SMP you get double benefit.  Obviously, the kernel doesn't have 
to spin while userspace does its stuff.  On top of this, there is also a 
latency improvement from ioeventfd, because QEMU processes 
virtqueue_notify under its "big QEMU lock".  With ioeventfd, serialized 
virtqueue processing can be a bottleneck, but it doesn't affect latency. 
  Without ioeventfd it affects the VCPUs' latency and negates a lot of 
the benefit of Ming Lei's patch.

You can try disabling ioeventfd with "-global 
virtio-blk-pci.ioeventfd=off" on the QEMU command line.  Performance 
will plummet. :)

Paolo