[net-next,V4,5/5] vhost: access vq metadata through kernel virtual address
diff mbox series

Message ID 20190123095557.30168-6-jasowang@redhat.com
State New, archived
Headers show
Series
  • vhost: accelerate metadata access through vmap()
Related show

Commit Message

Jason Wang Jan. 23, 2019, 9:55 a.m. UTC
It was noticed that the copy_user() friends that was used to access
virtqueue metdata tends to be very expensive for dataplane
implementation like vhost since it involves lots of software checks,
speculation barrier, hardware feature toggling (e.g SMAP). The
extra cost will be more obvious when transferring small packets since
the time spent on metadata accessing become more significant.

This patch tries to eliminate those overheads by accessing them
through kernel virtual address by vmap(). To make the pages can be
migrated, instead of pinning them through GUP, we use MMU notifiers to
invalidate vmaps and re-establish vmaps during each round of metadata
prefetching if necessary. For devices that doesn't use metadata
prefetching, the memory accessors fallback to normal copy_user()
implementation gracefully. The invalidation was synchronized with
datapath through vq mutex, and in order to avoid hold vq mutex during
range checking, MMU notifier was teared down when trying to modify vq
metadata.

Another thing is kernel lacks efficient solution for tracking dirty
pages by vmap(), this will lead issues if vhost is using file backed
memory which needs care of writeback. This patch solves this issue by
just skipping the vma that is file backed and fallback to normal
copy_user() friends. This might introduce some overheads for file
backed users but consider this use case is rare we could do
optimizations on top.

Note that this was only done when device IOTLB is not enabled. We
could use similar method to optimize it in the future.

Tests shows at most about 22% improvement on TX PPS when using
virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:

        SMAP on | SMAP off
Before: 5.0Mpps | 6.6Mpps
After:  6.1Mpps | 7.4Mpps

Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/vhost/vhost.c | 288 +++++++++++++++++++++++++++++++++++++++++-
 drivers/vhost/vhost.h |  13 ++
 mm/shmem.c            |   1 +
 3 files changed, 300 insertions(+), 2 deletions(-)

Comments

Michael S. Tsirkin Jan. 23, 2019, 2:08 p.m. UTC | #1
On Wed, Jan 23, 2019 at 05:55:57PM +0800, Jason Wang wrote:
> It was noticed that the copy_user() friends that was used to access
> virtqueue metdata tends to be very expensive for dataplane
> implementation like vhost since it involves lots of software checks,
> speculation barrier, hardware feature toggling (e.g SMAP). The
> extra cost will be more obvious when transferring small packets since
> the time spent on metadata accessing become more significant.
> 
> This patch tries to eliminate those overheads by accessing them
> through kernel virtual address by vmap(). To make the pages can be
> migrated, instead of pinning them through GUP, we use MMU notifiers to
> invalidate vmaps and re-establish vmaps during each round of metadata
> prefetching if necessary. For devices that doesn't use metadata
> prefetching, the memory accessors fallback to normal copy_user()
> implementation gracefully. The invalidation was synchronized with
> datapath through vq mutex, and in order to avoid hold vq mutex during
> range checking, MMU notifier was teared down when trying to modify vq
> metadata.
> 
> Another thing is kernel lacks efficient solution for tracking dirty
> pages by vmap(), this will lead issues if vhost is using file backed
> memory which needs care of writeback. This patch solves this issue by
> just skipping the vma that is file backed and fallback to normal
> copy_user() friends. This might introduce some overheads for file
> backed users but consider this use case is rare we could do
> optimizations on top.
> 
> Note that this was only done when device IOTLB is not enabled. We
> could use similar method to optimize it in the future.
> 
> Tests shows at most about 22% improvement on TX PPS when using
> virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:
> 
>         SMAP on | SMAP off
> Before: 5.0Mpps | 6.6Mpps
> After:  6.1Mpps | 7.4Mpps
> 
> Signed-off-by: Jason Wang <jasowang@redhat.com>


So this is the bulk of the change.
Threee things that I need to look into
- Are there any security issues with bypassing the speculation barrier
  that is normally present after access_ok?
- How hard does the special handling for
  file backed storage make testing?
  On the one hand we could add a module parameter to
  force copy to/from user. on the other that's
  another configuration we need to support.
  But iotlb is not using vmap, so maybe that's enough
  for testing.
- How hard is it to figure out which mode uses which code.



Meanwhile, could you pls post data comparing this last patch with the
below?  This removes the speculation barrier replacing it with a
(useless but at least more lightweight) data dependency.

Thanks!


diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index bac939af8dbb..352ee7e14476 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -739,7 +739,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 	int ret;
 
 	if (!vq->iotlb)
-		return __copy_to_user(to, from, size);
+		return copy_to_user(to, from, size);
 	else {
 		/* This function should be called after iotlb
 		 * prefetch, which means we're sure that all vq
@@ -752,7 +752,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 				     VHOST_ADDR_USED);
 
 		if (uaddr)
-			return __copy_to_user(uaddr, from, size);
+			return copy_to_user(uaddr, from, size);
 
 		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 				     ARRAY_SIZE(vq->iotlb_iov),
@@ -774,7 +774,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 	int ret;
 
 	if (!vq->iotlb)
-		return __copy_from_user(to, from, size);
+		return copy_from_user(to, from, size);
 	else {
 		/* This function should be called after iotlb
 		 * prefetch, which means we're sure that vq
@@ -787,7 +787,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 		struct iov_iter f;
 
 		if (uaddr)
-			return __copy_from_user(to, uaddr, size);
+			return copy_from_user(to, uaddr, size);
 
 		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 				     ARRAY_SIZE(vq->iotlb_iov),
@@ -855,13 +855,13 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 ({ \
 	int ret = -EFAULT; \
 	if (!vq->iotlb) { \
-		ret = __put_user(x, ptr); \
+		ret = put_user(x, ptr); \
 	} else { \
 		__typeof__(ptr) to = \
 			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
 					  sizeof(*ptr), VHOST_ADDR_USED); \
 		if (to != NULL) \
-			ret = __put_user(x, to); \
+			ret = put_user(x, to); \
 		else \
 			ret = -EFAULT;	\
 	} \
@@ -872,14 +872,14 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 ({ \
 	int ret; \
 	if (!vq->iotlb) { \
-		ret = __get_user(x, ptr); \
+		ret = get_user(x, ptr); \
 	} else { \
 		__typeof__(ptr) from = \
 			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 							   sizeof(*ptr), \
 							   type); \
 		if (from != NULL) \
-			ret = __get_user(x, from); \
+			ret = get_user(x, from); \
 		else \
 			ret = -EFAULT; \
 	} \
Jason Wang Jan. 24, 2019, 4:07 a.m. UTC | #2
On 2019/1/23 下午10:08, Michael S. Tsirkin wrote:
> On Wed, Jan 23, 2019 at 05:55:57PM +0800, Jason Wang wrote:
>> It was noticed that the copy_user() friends that was used to access
>> virtqueue metdata tends to be very expensive for dataplane
>> implementation like vhost since it involves lots of software checks,
>> speculation barrier, hardware feature toggling (e.g SMAP). The
>> extra cost will be more obvious when transferring small packets since
>> the time spent on metadata accessing become more significant.
>>
>> This patch tries to eliminate those overheads by accessing them
>> through kernel virtual address by vmap(). To make the pages can be
>> migrated, instead of pinning them through GUP, we use MMU notifiers to
>> invalidate vmaps and re-establish vmaps during each round of metadata
>> prefetching if necessary. For devices that doesn't use metadata
>> prefetching, the memory accessors fallback to normal copy_user()
>> implementation gracefully. The invalidation was synchronized with
>> datapath through vq mutex, and in order to avoid hold vq mutex during
>> range checking, MMU notifier was teared down when trying to modify vq
>> metadata.
>>
>> Another thing is kernel lacks efficient solution for tracking dirty
>> pages by vmap(), this will lead issues if vhost is using file backed
>> memory which needs care of writeback. This patch solves this issue by
>> just skipping the vma that is file backed and fallback to normal
>> copy_user() friends. This might introduce some overheads for file
>> backed users but consider this use case is rare we could do
>> optimizations on top.
>>
>> Note that this was only done when device IOTLB is not enabled. We
>> could use similar method to optimize it in the future.
>>
>> Tests shows at most about 22% improvement on TX PPS when using
>> virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:
>>
>>          SMAP on | SMAP off
>> Before: 5.0Mpps | 6.6Mpps
>> After:  6.1Mpps | 7.4Mpps
>>
>> Signed-off-by: Jason Wang <jasowang@redhat.com>
>
> So this is the bulk of the change.
> Threee things that I need to look into
> - Are there any security issues with bypassing the speculation barrier
>    that is normally present after access_ok?


If we can make sure the bypassing was only used in a kthread (vhost), it 
should be fine I think.


> - How hard does the special handling for
>    file backed storage make testing?


It's as simple as un-commenting vhost_can_vmap()? Or I can try to hack 
qemu or dpdk to test this.


>    On the one hand we could add a module parameter to
>    force copy to/from user. on the other that's
>    another configuration we need to support.


That sounds sub-optimal since it leave the choice to users.


>    But iotlb is not using vmap, so maybe that's enough
>    for testing.
> - How hard is it to figure out which mode uses which code.
>
>
>
> Meanwhile, could you pls post data comparing this last patch with the
> below?  This removes the speculation barrier replacing it with a
> (useless but at least more lightweight) data dependency.


SMAP off

Your patch: 7.2MPPs

vmap: 7.4Mpps

I don't test SMAP on, since it will be much slow for sure.

Thanks


>
> Thanks!
>
>
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index bac939af8dbb..352ee7e14476 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -739,7 +739,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
>   	int ret;
>   
>   	if (!vq->iotlb)
> -		return __copy_to_user(to, from, size);
> +		return copy_to_user(to, from, size);
>   	else {
>   		/* This function should be called after iotlb
>   		 * prefetch, which means we're sure that all vq
> @@ -752,7 +752,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
>   				     VHOST_ADDR_USED);
>   
>   		if (uaddr)
> -			return __copy_to_user(uaddr, from, size);
> +			return copy_to_user(uaddr, from, size);
>   
>   		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
>   				     ARRAY_SIZE(vq->iotlb_iov),
> @@ -774,7 +774,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
>   	int ret;
>   
>   	if (!vq->iotlb)
> -		return __copy_from_user(to, from, size);
> +		return copy_from_user(to, from, size);
>   	else {
>   		/* This function should be called after iotlb
>   		 * prefetch, which means we're sure that vq
> @@ -787,7 +787,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
>   		struct iov_iter f;
>   
>   		if (uaddr)
> -			return __copy_from_user(to, uaddr, size);
> +			return copy_from_user(to, uaddr, size);
>   
>   		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
>   				     ARRAY_SIZE(vq->iotlb_iov),
> @@ -855,13 +855,13 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
>   ({ \
>   	int ret = -EFAULT; \
>   	if (!vq->iotlb) { \
> -		ret = __put_user(x, ptr); \
> +		ret = put_user(x, ptr); \
>   	} else { \
>   		__typeof__(ptr) to = \
>   			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
>   					  sizeof(*ptr), VHOST_ADDR_USED); \
>   		if (to != NULL) \
> -			ret = __put_user(x, to); \
> +			ret = put_user(x, to); \
>   		else \
>   			ret = -EFAULT;	\
>   	} \
> @@ -872,14 +872,14 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
>   ({ \
>   	int ret; \
>   	if (!vq->iotlb) { \
> -		ret = __get_user(x, ptr); \
> +		ret = get_user(x, ptr); \
>   	} else { \
>   		__typeof__(ptr) from = \
>   			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
>   							   sizeof(*ptr), \
>   							   type); \
>   		if (from != NULL) \
> -			ret = __get_user(x, from); \
> +			ret = get_user(x, from); \
>   		else \
>   			ret = -EFAULT; \
>   	} \
Jason Wang Jan. 24, 2019, 4:11 a.m. UTC | #3
On 2019/1/24 下午12:07, Jason Wang wrote:
>
> On 2019/1/23 下午10:08, Michael S. Tsirkin wrote:
>> On Wed, Jan 23, 2019 at 05:55:57PM +0800, Jason Wang wrote:
>>> It was noticed that the copy_user() friends that was used to access
>>> virtqueue metdata tends to be very expensive for dataplane
>>> implementation like vhost since it involves lots of software checks,
>>> speculation barrier, hardware feature toggling (e.g SMAP). The
>>> extra cost will be more obvious when transferring small packets since
>>> the time spent on metadata accessing become more significant.
>>>
>>> This patch tries to eliminate those overheads by accessing them
>>> through kernel virtual address by vmap(). To make the pages can be
>>> migrated, instead of pinning them through GUP, we use MMU notifiers to
>>> invalidate vmaps and re-establish vmaps during each round of metadata
>>> prefetching if necessary. For devices that doesn't use metadata
>>> prefetching, the memory accessors fallback to normal copy_user()
>>> implementation gracefully. The invalidation was synchronized with
>>> datapath through vq mutex, and in order to avoid hold vq mutex during
>>> range checking, MMU notifier was teared down when trying to modify vq
>>> metadata.
>>>
>>> Another thing is kernel lacks efficient solution for tracking dirty
>>> pages by vmap(), this will lead issues if vhost is using file backed
>>> memory which needs care of writeback. This patch solves this issue by
>>> just skipping the vma that is file backed and fallback to normal
>>> copy_user() friends. This might introduce some overheads for file
>>> backed users but consider this use case is rare we could do
>>> optimizations on top.
>>>
>>> Note that this was only done when device IOTLB is not enabled. We
>>> could use similar method to optimize it in the future.
>>>
>>> Tests shows at most about 22% improvement on TX PPS when using
>>> virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:
>>>
>>>          SMAP on | SMAP off
>>> Before: 5.0Mpps | 6.6Mpps
>>> After:  6.1Mpps | 7.4Mpps
>>>
>>> Signed-off-by: Jason Wang <jasowang@redhat.com>
>>
>> So this is the bulk of the change.
>> Threee things that I need to look into
>> - Are there any security issues with bypassing the speculation barrier
>>    that is normally present after access_ok?
>
>
> If we can make sure the bypassing was only used in a kthread (vhost), 
> it should be fine I think.
>
>
>> - How hard does the special handling for
>>    file backed storage make testing?
>
>
> It's as simple as un-commenting vhost_can_vmap()? Or I can try to hack 
> qemu or dpdk to test this.
>
>
>>    On the one hand we could add a module parameter to
>>    force copy to/from user. on the other that's
>>    another configuration we need to support.
>
>
> That sounds sub-optimal since it leave the choice to users.
>
>
>>    But iotlb is not using vmap, so maybe that's enough
>>    for testing.
>> - How hard is it to figure out which mode uses which code.


It's as simple as tracing __get_user() usage in vhost process?

Thanks


>>
>>
>>
>> Meanwhile, could you pls post data comparing this last patch with the
>> below?  This removes the speculation barrier replacing it with a
>> (useless but at least more lightweight) data dependency.
>
>
> SMAP off
>
> Your patch: 7.2MPPs
>
> vmap: 7.4Mpps
>
> I don't test SMAP on, since it will be much slow for sure.
>
> Thanks
>
>
>>
>> Thanks!
>>
>>
>> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
>> index bac939af8dbb..352ee7e14476 100644
>> --- a/drivers/vhost/vhost.c
>> +++ b/drivers/vhost/vhost.c
>> @@ -739,7 +739,7 @@ static int vhost_copy_to_user(struct 
>> vhost_virtqueue *vq, void __user *to,
>>       int ret;
>>         if (!vq->iotlb)
>> -        return __copy_to_user(to, from, size);
>> +        return copy_to_user(to, from, size);
>>       else {
>>           /* This function should be called after iotlb
>>            * prefetch, which means we're sure that all vq
>> @@ -752,7 +752,7 @@ static int vhost_copy_to_user(struct 
>> vhost_virtqueue *vq, void __user *to,
>>                        VHOST_ADDR_USED);
>>             if (uaddr)
>> -            return __copy_to_user(uaddr, from, size);
>> +            return copy_to_user(uaddr, from, size);
>>             ret = translate_desc(vq, (u64)(uintptr_t)to, size, 
>> vq->iotlb_iov,
>>                        ARRAY_SIZE(vq->iotlb_iov),
>> @@ -774,7 +774,7 @@ static int vhost_copy_from_user(struct 
>> vhost_virtqueue *vq, void *to,
>>       int ret;
>>         if (!vq->iotlb)
>> -        return __copy_from_user(to, from, size);
>> +        return copy_from_user(to, from, size);
>>       else {
>>           /* This function should be called after iotlb
>>            * prefetch, which means we're sure that vq
>> @@ -787,7 +787,7 @@ static int vhost_copy_from_user(struct 
>> vhost_virtqueue *vq, void *to,
>>           struct iov_iter f;
>>             if (uaddr)
>> -            return __copy_from_user(to, uaddr, size);
>> +            return copy_from_user(to, uaddr, size);
>>             ret = translate_desc(vq, (u64)(uintptr_t)from, size, 
>> vq->iotlb_iov,
>>                        ARRAY_SIZE(vq->iotlb_iov),
>> @@ -855,13 +855,13 @@ static inline void __user 
>> *__vhost_get_user(struct vhost_virtqueue *vq,
>>   ({ \
>>       int ret = -EFAULT; \
>>       if (!vq->iotlb) { \
>> -        ret = __put_user(x, ptr); \
>> +        ret = put_user(x, ptr); \
>>       } else { \
>>           __typeof__(ptr) to = \
>>               (__typeof__(ptr)) __vhost_get_user(vq, ptr,    \
>>                         sizeof(*ptr), VHOST_ADDR_USED); \
>>           if (to != NULL) \
>> -            ret = __put_user(x, to); \
>> +            ret = put_user(x, to); \
>>           else \
>>               ret = -EFAULT;    \
>>       } \
>> @@ -872,14 +872,14 @@ static inline void __user 
>> *__vhost_get_user(struct vhost_virtqueue *vq,
>>   ({ \
>>       int ret; \
>>       if (!vq->iotlb) { \
>> -        ret = __get_user(x, ptr); \
>> +        ret = get_user(x, ptr); \
>>       } else { \
>>           __typeof__(ptr) from = \
>>               (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
>>                                  sizeof(*ptr), \
>>                                  type); \
>>           if (from != NULL) \
>> -            ret = __get_user(x, from); \
>> +            ret = get_user(x, from); \
>>           else \
>>               ret = -EFAULT; \
>>       } \
Michael S. Tsirkin Jan. 24, 2019, 4:51 a.m. UTC | #4
On Thu, Jan 24, 2019 at 12:07:54PM +0800, Jason Wang wrote:
> 
> On 2019/1/23 下午10:08, Michael S. Tsirkin wrote:
> > On Wed, Jan 23, 2019 at 05:55:57PM +0800, Jason Wang wrote:
> > > It was noticed that the copy_user() friends that was used to access
> > > virtqueue metdata tends to be very expensive for dataplane
> > > implementation like vhost since it involves lots of software checks,
> > > speculation barrier, hardware feature toggling (e.g SMAP). The
> > > extra cost will be more obvious when transferring small packets since
> > > the time spent on metadata accessing become more significant.
> > > 
> > > This patch tries to eliminate those overheads by accessing them
> > > through kernel virtual address by vmap(). To make the pages can be
> > > migrated, instead of pinning them through GUP, we use MMU notifiers to
> > > invalidate vmaps and re-establish vmaps during each round of metadata
> > > prefetching if necessary. For devices that doesn't use metadata
> > > prefetching, the memory accessors fallback to normal copy_user()
> > > implementation gracefully. The invalidation was synchronized with
> > > datapath through vq mutex, and in order to avoid hold vq mutex during
> > > range checking, MMU notifier was teared down when trying to modify vq
> > > metadata.
> > > 
> > > Another thing is kernel lacks efficient solution for tracking dirty
> > > pages by vmap(), this will lead issues if vhost is using file backed
> > > memory which needs care of writeback. This patch solves this issue by
> > > just skipping the vma that is file backed and fallback to normal
> > > copy_user() friends. This might introduce some overheads for file
> > > backed users but consider this use case is rare we could do
> > > optimizations on top.
> > > 
> > > Note that this was only done when device IOTLB is not enabled. We
> > > could use similar method to optimize it in the future.
> > > 
> > > Tests shows at most about 22% improvement on TX PPS when using
> > > virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:
> > > 
> > >          SMAP on | SMAP off
> > > Before: 5.0Mpps | 6.6Mpps
> > > After:  6.1Mpps | 7.4Mpps
> > > 
> > > Signed-off-by: Jason Wang <jasowang@redhat.com>
> > 
> > So this is the bulk of the change.
> > Threee things that I need to look into
> > - Are there any security issues with bypassing the speculation barrier
> >    that is normally present after access_ok?
> 
> 
> If we can make sure the bypassing was only used in a kthread (vhost), it
> should be fine I think.
> 
> 
> > - How hard does the special handling for
> >    file backed storage make testing?
> 
> 
> It's as simple as un-commenting vhost_can_vmap()? Or I can try to hack qemu
> or dpdk to test this.
> 
> 
> >    On the one hand we could add a module parameter to
> >    force copy to/from user. on the other that's
> >    another configuration we need to support.
> 
> 
> That sounds sub-optimal since it leave the choice to users.
> 
> 
> >    But iotlb is not using vmap, so maybe that's enough
> >    for testing.
> > - How hard is it to figure out which mode uses which code.
> > 
> > 
> > 
> > Meanwhile, could you pls post data comparing this last patch with the
> > below?  This removes the speculation barrier replacing it with a
> > (useless but at least more lightweight) data dependency.
> 
> 
> SMAP off
> 
> Your patch: 7.2MPPs
> 
> vmap: 7.4Mpps
> 

Sounds more or less as expected. Up to 3% gain with vmap - I think
that's a bit higher than what we saw previously when we switched from
get_user to __get_user and that's probably because of all the
array_index_nospec trickery.

> I don't test SMAP on, since it will be much slow for sure.

Right. So bypassing SMAP remains the main reason to do vmap tricks.

> Thanks

> 
> > 
> > Thanks!
> > 
> > 
> > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > index bac939af8dbb..352ee7e14476 100644
> > --- a/drivers/vhost/vhost.c
> > +++ b/drivers/vhost/vhost.c
> > @@ -739,7 +739,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
> >   	int ret;
> >   	if (!vq->iotlb)
> > -		return __copy_to_user(to, from, size);
> > +		return copy_to_user(to, from, size);
> >   	else {
> >   		/* This function should be called after iotlb
> >   		 * prefetch, which means we're sure that all vq
> > @@ -752,7 +752,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
> >   				     VHOST_ADDR_USED);
> >   		if (uaddr)
> > -			return __copy_to_user(uaddr, from, size);
> > +			return copy_to_user(uaddr, from, size);
> >   		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
> >   				     ARRAY_SIZE(vq->iotlb_iov),
> > @@ -774,7 +774,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
> >   	int ret;
> >   	if (!vq->iotlb)
> > -		return __copy_from_user(to, from, size);
> > +		return copy_from_user(to, from, size);
> >   	else {
> >   		/* This function should be called after iotlb
> >   		 * prefetch, which means we're sure that vq
> > @@ -787,7 +787,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
> >   		struct iov_iter f;
> >   		if (uaddr)
> > -			return __copy_from_user(to, uaddr, size);
> > +			return copy_from_user(to, uaddr, size);
> >   		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
> >   				     ARRAY_SIZE(vq->iotlb_iov),
> > @@ -855,13 +855,13 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
> >   ({ \
> >   	int ret = -EFAULT; \
> >   	if (!vq->iotlb) { \
> > -		ret = __put_user(x, ptr); \
> > +		ret = put_user(x, ptr); \
> >   	} else { \
> >   		__typeof__(ptr) to = \
> >   			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
> >   					  sizeof(*ptr), VHOST_ADDR_USED); \
> >   		if (to != NULL) \
> > -			ret = __put_user(x, to); \
> > +			ret = put_user(x, to); \
> >   		else \
> >   			ret = -EFAULT;	\
> >   	} \
> > @@ -872,14 +872,14 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
> >   ({ \
> >   	int ret; \
> >   	if (!vq->iotlb) { \
> > -		ret = __get_user(x, ptr); \
> > +		ret = get_user(x, ptr); \
> >   	} else { \
> >   		__typeof__(ptr) from = \
> >   			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
> >   							   sizeof(*ptr), \
> >   							   type); \
> >   		if (from != NULL) \
> > -			ret = __get_user(x, from); \
> > +			ret = get_user(x, from); \
> >   		else \
> >   			ret = -EFAULT; \
> >   	} \
Michael S. Tsirkin Jan. 24, 2019, 4:53 a.m. UTC | #5
On Thu, Jan 24, 2019 at 12:11:28PM +0800, Jason Wang wrote:
> 
> On 2019/1/24 下午12:07, Jason Wang wrote:
> > 
> > On 2019/1/23 下午10:08, Michael S. Tsirkin wrote:
> > > On Wed, Jan 23, 2019 at 05:55:57PM +0800, Jason Wang wrote:
> > > > It was noticed that the copy_user() friends that was used to access
> > > > virtqueue metdata tends to be very expensive for dataplane
> > > > implementation like vhost since it involves lots of software checks,
> > > > speculation barrier, hardware feature toggling (e.g SMAP). The
> > > > extra cost will be more obvious when transferring small packets since
> > > > the time spent on metadata accessing become more significant.
> > > > 
> > > > This patch tries to eliminate those overheads by accessing them
> > > > through kernel virtual address by vmap(). To make the pages can be
> > > > migrated, instead of pinning them through GUP, we use MMU notifiers to
> > > > invalidate vmaps and re-establish vmaps during each round of metadata
> > > > prefetching if necessary. For devices that doesn't use metadata
> > > > prefetching, the memory accessors fallback to normal copy_user()
> > > > implementation gracefully. The invalidation was synchronized with
> > > > datapath through vq mutex, and in order to avoid hold vq mutex during
> > > > range checking, MMU notifier was teared down when trying to modify vq
> > > > metadata.
> > > > 
> > > > Another thing is kernel lacks efficient solution for tracking dirty
> > > > pages by vmap(), this will lead issues if vhost is using file backed
> > > > memory which needs care of writeback. This patch solves this issue by
> > > > just skipping the vma that is file backed and fallback to normal
> > > > copy_user() friends. This might introduce some overheads for file
> > > > backed users but consider this use case is rare we could do
> > > > optimizations on top.
> > > > 
> > > > Note that this was only done when device IOTLB is not enabled. We
> > > > could use similar method to optimize it in the future.
> > > > 
> > > > Tests shows at most about 22% improvement on TX PPS when using
> > > > virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:
> > > > 
> > > >          SMAP on | SMAP off
> > > > Before: 5.0Mpps | 6.6Mpps
> > > > After:  6.1Mpps | 7.4Mpps
> > > > 
> > > > Signed-off-by: Jason Wang <jasowang@redhat.com>
> > > 
> > > So this is the bulk of the change.
> > > Threee things that I need to look into
> > > - Are there any security issues with bypassing the speculation barrier
> > >    that is normally present after access_ok?
> > 
> > 
> > If we can make sure the bypassing was only used in a kthread (vhost), it
> > should be fine I think.
> > 
> > 
> > > - How hard does the special handling for
> > >    file backed storage make testing?
> > 
> > 
> > It's as simple as un-commenting vhost_can_vmap()? Or I can try to hack
> > qemu or dpdk to test this.
> > 
> > 
> > >    On the one hand we could add a module parameter to
> > >    force copy to/from user. on the other that's
> > >    another configuration we need to support.
> > 
> > 
> > That sounds sub-optimal since it leave the choice to users.
> > 
> > 
> > >    But iotlb is not using vmap, so maybe that's enough
> > >    for testing.
> > > - How hard is it to figure out which mode uses which code.
> 
> 
> It's as simple as tracing __get_user() usage in vhost process?
> 
> Thanks

Well there are now mtu notifiers etc etc. It's hardly as well
contained as that.


> 
> > > 
> > > 
> > > 
> > > Meanwhile, could you pls post data comparing this last patch with the
> > > below?  This removes the speculation barrier replacing it with a
> > > (useless but at least more lightweight) data dependency.
> > 
> > 
> > SMAP off
> > 
> > Your patch: 7.2MPPs
> > 
> > vmap: 7.4Mpps
> > 
> > I don't test SMAP on, since it will be much slow for sure.
> > 
> > Thanks
> > 
> > 
> > > 
> > > Thanks!
> > > 
> > > 
> > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > > index bac939af8dbb..352ee7e14476 100644
> > > --- a/drivers/vhost/vhost.c
> > > +++ b/drivers/vhost/vhost.c
> > > @@ -739,7 +739,7 @@ static int vhost_copy_to_user(struct
> > > vhost_virtqueue *vq, void __user *to,
> > >       int ret;
> > >         if (!vq->iotlb)
> > > -        return __copy_to_user(to, from, size);
> > > +        return copy_to_user(to, from, size);
> > >       else {
> > >           /* This function should be called after iotlb
> > >            * prefetch, which means we're sure that all vq
> > > @@ -752,7 +752,7 @@ static int vhost_copy_to_user(struct
> > > vhost_virtqueue *vq, void __user *to,
> > >                        VHOST_ADDR_USED);
> > >             if (uaddr)
> > > -            return __copy_to_user(uaddr, from, size);
> > > +            return copy_to_user(uaddr, from, size);
> > >             ret = translate_desc(vq, (u64)(uintptr_t)to, size,
> > > vq->iotlb_iov,
> > >                        ARRAY_SIZE(vq->iotlb_iov),
> > > @@ -774,7 +774,7 @@ static int vhost_copy_from_user(struct
> > > vhost_virtqueue *vq, void *to,
> > >       int ret;
> > >         if (!vq->iotlb)
> > > -        return __copy_from_user(to, from, size);
> > > +        return copy_from_user(to, from, size);
> > >       else {
> > >           /* This function should be called after iotlb
> > >            * prefetch, which means we're sure that vq
> > > @@ -787,7 +787,7 @@ static int vhost_copy_from_user(struct
> > > vhost_virtqueue *vq, void *to,
> > >           struct iov_iter f;
> > >             if (uaddr)
> > > -            return __copy_from_user(to, uaddr, size);
> > > +            return copy_from_user(to, uaddr, size);
> > >             ret = translate_desc(vq, (u64)(uintptr_t)from, size,
> > > vq->iotlb_iov,
> > >                        ARRAY_SIZE(vq->iotlb_iov),
> > > @@ -855,13 +855,13 @@ static inline void __user
> > > *__vhost_get_user(struct vhost_virtqueue *vq,
> > >   ({ \
> > >       int ret = -EFAULT; \
> > >       if (!vq->iotlb) { \
> > > -        ret = __put_user(x, ptr); \
> > > +        ret = put_user(x, ptr); \
> > >       } else { \
> > >           __typeof__(ptr) to = \
> > >               (__typeof__(ptr)) __vhost_get_user(vq, ptr,    \
> > >                         sizeof(*ptr), VHOST_ADDR_USED); \
> > >           if (to != NULL) \
> > > -            ret = __put_user(x, to); \
> > > +            ret = put_user(x, to); \
> > >           else \
> > >               ret = -EFAULT;    \
> > >       } \
> > > @@ -872,14 +872,14 @@ static inline void __user
> > > *__vhost_get_user(struct vhost_virtqueue *vq,
> > >   ({ \
> > >       int ret; \
> > >       if (!vq->iotlb) { \
> > > -        ret = __get_user(x, ptr); \
> > > +        ret = get_user(x, ptr); \
> > >       } else { \
> > >           __typeof__(ptr) from = \
> > >               (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
> > >                                  sizeof(*ptr), \
> > >                                  type); \
> > >           if (from != NULL) \
> > > -            ret = __get_user(x, from); \
> > > +            ret = get_user(x, from); \
> > >           else \
> > >               ret = -EFAULT; \
> > >       } \
Jason Wang Jan. 25, 2019, 2:33 a.m. UTC | #6
On 2019/1/24 下午12:53, Michael S. Tsirkin wrote:
>>>> - How hard is it to figure out which mode uses which code.
>> It's as simple as tracing __get_user() usage in vhost process?
>>
>> Thanks
> Well there are now mtu notifiers etc etc. It's hardly as well
> contained as that.
>
>

We can setup filter out exactly what sets of function that we wan to 
trace. E.g we can only trace the usage of __get_user() and 
invalidate_range_start(). This should be sufficient.

In the long run, we may want to have some tracepoints for vhost_net.

Thanks
Michael S. Tsirkin Jan. 25, 2019, 3 a.m. UTC | #7
On Thu, Jan 24, 2019 at 12:07:54PM +0800, Jason Wang wrote:
> > Meanwhile, could you pls post data comparing this last patch with the
> > below?  This removes the speculation barrier replacing it with a
> > (useless but at least more lightweight) data dependency.
> 
> 
> SMAP off
> 
> Your patch: 7.2MPPs
> 
> vmap: 7.4Mpps

OK so while we keep looking into vmap, why don't we merge something like
the below?  Seems quite straight forward ...


> I don't test SMAP on, since it will be much slow for sure.
> 
> Thanks
> 
> 
> > 
> > Thanks!
> > 
> > 
> > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > index bac939af8dbb..352ee7e14476 100644
> > --- a/drivers/vhost/vhost.c
> > +++ b/drivers/vhost/vhost.c
> > @@ -739,7 +739,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
> >   	int ret;
> >   	if (!vq->iotlb)
> > -		return __copy_to_user(to, from, size);
> > +		return copy_to_user(to, from, size);
> >   	else {
> >   		/* This function should be called after iotlb
> >   		 * prefetch, which means we're sure that all vq
> > @@ -752,7 +752,7 @@ static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
> >   				     VHOST_ADDR_USED);
> >   		if (uaddr)
> > -			return __copy_to_user(uaddr, from, size);
> > +			return copy_to_user(uaddr, from, size);
> >   		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
> >   				     ARRAY_SIZE(vq->iotlb_iov),
> > @@ -774,7 +774,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
> >   	int ret;
> >   	if (!vq->iotlb)
> > -		return __copy_from_user(to, from, size);
> > +		return copy_from_user(to, from, size);
> >   	else {
> >   		/* This function should be called after iotlb
> >   		 * prefetch, which means we're sure that vq
> > @@ -787,7 +787,7 @@ static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
> >   		struct iov_iter f;
> >   		if (uaddr)
> > -			return __copy_from_user(to, uaddr, size);
> > +			return copy_from_user(to, uaddr, size);
> >   		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
> >   				     ARRAY_SIZE(vq->iotlb_iov),
> > @@ -855,13 +855,13 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
> >   ({ \
> >   	int ret = -EFAULT; \
> >   	if (!vq->iotlb) { \
> > -		ret = __put_user(x, ptr); \
> > +		ret = put_user(x, ptr); \
> >   	} else { \
> >   		__typeof__(ptr) to = \
> >   			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
> >   					  sizeof(*ptr), VHOST_ADDR_USED); \
> >   		if (to != NULL) \
> > -			ret = __put_user(x, to); \
> > +			ret = put_user(x, to); \
> >   		else \
> >   			ret = -EFAULT;	\
> >   	} \
> > @@ -872,14 +872,14 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
> >   ({ \
> >   	int ret; \
> >   	if (!vq->iotlb) { \
> > -		ret = __get_user(x, ptr); \
> > +		ret = get_user(x, ptr); \
> >   	} else { \
> >   		__typeof__(ptr) from = \
> >   			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
> >   							   sizeof(*ptr), \
> >   							   type); \
> >   		if (from != NULL) \
> > -			ret = __get_user(x, from); \
> > +			ret = get_user(x, from); \
> >   		else \
> >   			ret = -EFAULT; \
> >   	} \
Michael S. Tsirkin Jan. 25, 2019, 3:03 a.m. UTC | #8
On Wed, Jan 23, 2019 at 05:55:57PM +0800, Jason Wang wrote:
> It was noticed that the copy_user() friends that was used to access
> virtqueue metdata tends to be very expensive for dataplane
> implementation like vhost since it involves lots of software checks,
> speculation barrier, hardware feature toggling (e.g SMAP). The
> extra cost will be more obvious when transferring small packets since
> the time spent on metadata accessing become more significant.
> 
> This patch tries to eliminate those overheads by accessing them
> through kernel virtual address by vmap(). To make the pages can be
> migrated, instead of pinning them through GUP, we use MMU notifiers to
> invalidate vmaps and re-establish vmaps during each round of metadata
> prefetching if necessary. For devices that doesn't use metadata
> prefetching, the memory accessors fallback to normal copy_user()
> implementation gracefully. The invalidation was synchronized with
> datapath through vq mutex, and in order to avoid hold vq mutex during
> range checking, MMU notifier was teared down when trying to modify vq
> metadata.
> 
> Another thing is kernel lacks efficient solution for tracking dirty
> pages by vmap(), this will lead issues if vhost is using file backed
> memory which needs care of writeback. This patch solves this issue by
> just skipping the vma that is file backed and fallback to normal
> copy_user() friends. This might introduce some overheads for file
> backed users but consider this use case is rare we could do
> optimizations on top.
> 
> Note that this was only done when device IOTLB is not enabled. We
> could use similar method to optimize it in the future.
> 
> Tests shows at most about 22% improvement on TX PPS when using
> virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:
> 
>         SMAP on | SMAP off
> Before: 5.0Mpps | 6.6Mpps
> After:  6.1Mpps | 7.4Mpps
> 
> Signed-off-by: Jason Wang <jasowang@redhat.com>
> ---
>  drivers/vhost/vhost.c | 288 +++++++++++++++++++++++++++++++++++++++++-
>  drivers/vhost/vhost.h |  13 ++
>  mm/shmem.c            |   1 +
>  3 files changed, 300 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index 37e2cac8e8b0..096ae3298d62 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -440,6 +440,9 @@ void vhost_dev_init(struct vhost_dev *dev,
>  		vq->indirect = NULL;
>  		vq->heads = NULL;
>  		vq->dev = dev;
> +		memset(&vq->avail_ring, 0, sizeof(vq->avail_ring));
> +		memset(&vq->used_ring, 0, sizeof(vq->used_ring));
> +		memset(&vq->desc_ring, 0, sizeof(vq->desc_ring));
>  		mutex_init(&vq->mutex);
>  		vhost_vq_reset(dev, vq);
>  		if (vq->handle_kick)
> @@ -510,6 +513,73 @@ static size_t vhost_get_desc_size(struct vhost_virtqueue *vq, int num)
>  	return sizeof(*vq->desc) * num;
>  }
>  
> +static void vhost_uninit_vmap(struct vhost_vmap *map)
> +{
> +	if (map->addr)
> +		vunmap(map->unmap_addr);
> +
> +	map->addr = NULL;
> +	map->unmap_addr = NULL;
> +}
> +
> +static int vhost_invalidate_vmap(struct vhost_virtqueue *vq,
> +				 struct vhost_vmap *map,
> +				 unsigned long ustart,
> +				 size_t size,
> +				 unsigned long start,
> +				 unsigned long end,
> +				 bool blockable)
> +{
> +	if (end < ustart || start > ustart - 1 + size)
> +		return 0;
> +
> +	if (!blockable)
> +		return -EAGAIN;
> +
> +	mutex_lock(&vq->mutex);
> +	vhost_uninit_vmap(map);
> +	mutex_unlock(&vq->mutex);
> +
> +	return 0;
> +}
> +
> +static int vhost_invalidate_range_start(struct mmu_notifier *mn,
> +					const struct mmu_notifier_range *range)
> +{
> +	struct vhost_dev *dev = container_of(mn, struct vhost_dev,
> +					     mmu_notifier);
> +	int i;
> +
> +	for (i = 0; i < dev->nvqs; i++) {
> +		struct vhost_virtqueue *vq = dev->vqs[i];
> +
> +		if (vhost_invalidate_vmap(vq, &vq->avail_ring,
> +					  (unsigned long)vq->avail,
> +					  vhost_get_avail_size(vq, vq->num),
> +					  range->start, range->end,
> +					  range->blockable))
> +			return -EAGAIN;
> +		if (vhost_invalidate_vmap(vq, &vq->desc_ring,
> +					  (unsigned long)vq->desc,
> +					  vhost_get_desc_size(vq, vq->num),
> +					  range->start, range->end,
> +					  range->blockable))
> +			return -EAGAIN;
> +		if (vhost_invalidate_vmap(vq, &vq->used_ring,
> +					  (unsigned long)vq->used,
> +					  vhost_get_used_size(vq, vq->num),
> +					  range->start, range->end,
> +					  range->blockable))
> +			return -EAGAIN;
> +	}
> +
> +	return 0;
> +}
> +
> +static const struct mmu_notifier_ops vhost_mmu_notifier_ops = {
> +	.invalidate_range_start = vhost_invalidate_range_start,
> +};
> +
>  /* Caller should have device mutex */
>  long vhost_dev_set_owner(struct vhost_dev *dev)
>  {

It seems questionable to merely track .invalidate_range_start.
Don't we care about keeping pages young/accessed?
MMU will think they aren't and will penalize vhost by pushing
them out.

I note that MMU documentation says
        * invalidate_range_start() and invalidate_range_end() must be
         * paired
and it seems questionable that they are not paired here.


I also wonder about things like write-protecting the pages.
It does not look like a range is invalidated when page
is write-protected, even though I might have missed that.
If not we can be corrupting memory in a variety of ways
e.g. when using KSM, or with COW.




> @@ -541,7 +611,14 @@ long vhost_dev_set_owner(struct vhost_dev *dev)
>  	if (err)
>  		goto err_cgroup;
>  
> +	dev->mmu_notifier.ops = &vhost_mmu_notifier_ops;
> +	err = mmu_notifier_register(&dev->mmu_notifier, dev->mm);
> +	if (err)
> +		goto err_mmu_notifier;
> +
>  	return 0;
> +err_mmu_notifier:
> +	vhost_dev_free_iovecs(dev);
>  err_cgroup:
>  	kthread_stop(worker);
>  	dev->worker = NULL;
> @@ -632,6 +709,97 @@ static void vhost_clear_msg(struct vhost_dev *dev)
>  	spin_unlock(&dev->iotlb_lock);
>  }
>  
> +/* Suppress the vma that needs writeback since we can not track dirty
> + * pages now.
> + */
> +static bool vma_can_vmap(struct vm_area_struct *vma)
> +{
> +	return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
> +	       vma_is_shmem(vma);
> +}
> +

IIUC a second but anonymous memory needs writeback too, just to swap.
I'm not an MM person so I might be off.


> +static int vhost_init_vmap(struct vhost_dev *dev,
> +			   struct vhost_vmap *map, unsigned long uaddr,
> +			   size_t size, int write)
> +{
> +	struct mm_struct *mm = dev->mm;
> +	struct vm_area_struct *vma;
> +	struct page **pages;
> +	int npages = DIV_ROUND_UP(size, PAGE_SIZE);
> +	int npinned;
> +	void *vaddr;
> +	int err = 0;
> +
> +	down_read(&mm->mmap_sem);
> +	vma = find_vma(mm, uaddr);
> +	if (!vma || !vma_can_vmap(vma) ||
> +	    vma->vm_end < uaddr - 1 + size) {
> +		err = -EINVAL;
> +		goto err_vma;
> +	}
> +
> +	pages = kmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
> +	if (!pages) {
> +		err = -ENOMEM;
> +		goto err_alloc;
> +	}
> +
> +	npinned = get_user_pages_fast(uaddr, npages, write, pages);
> +	if (npinned != npages) {
> +		err = -EFAULT;
> +		goto err_gup;
> +	}
> +
> +	vaddr = vmap(pages, npages, VM_MAP, PAGE_KERNEL);
> +	if (!vaddr) {
> +		err = EFAULT;
> +		goto err_gup;
> +	}
> +
> +	map->addr = vaddr + (uaddr & (PAGE_SIZE - 1));
> +	map->unmap_addr = vaddr;
> +
> +err_gup:
> +	/* Don't pin pages, mmu notifier will notify us about page
> +	 * migration.
> +	 */
> +	if (npinned > 0)
> +		release_pages(pages, npinned);
> +err_alloc:
> +	kfree(pages);
> +err_vma:
> +	up_read(&mm->mmap_sem);
> +	return err;
> +}
> +
> +static void vhost_clean_vmaps(struct vhost_virtqueue *vq)
> +{
> +	vhost_uninit_vmap(&vq->avail_ring);
> +	vhost_uninit_vmap(&vq->desc_ring);
> +	vhost_uninit_vmap(&vq->used_ring);
> +}
> +
> +static int vhost_setup_avail_vmap(struct vhost_virtqueue *vq,
> +				  unsigned long avail)
> +{
> +	return vhost_init_vmap(vq->dev, &vq->avail_ring, avail,
> +			       vhost_get_avail_size(vq, vq->num), false);
> +}
> +
> +static int vhost_setup_desc_vmap(struct vhost_virtqueue *vq,
> +				 unsigned long desc)
> +{
> +	return vhost_init_vmap(vq->dev, &vq->desc_ring, desc,
> +			       vhost_get_desc_size(vq, vq->num), false);
> +}
> +
> +static int vhost_setup_used_vmap(struct vhost_virtqueue *vq,
> +				 unsigned long used)
> +{
> +	return vhost_init_vmap(vq->dev, &vq->used_ring, used,
> +			       vhost_get_used_size(vq, vq->num), true);
> +}
> +
>  void vhost_dev_cleanup(struct vhost_dev *dev)
>  {
>  	int i;
> @@ -661,8 +829,12 @@ void vhost_dev_cleanup(struct vhost_dev *dev)
>  		kthread_stop(dev->worker);
>  		dev->worker = NULL;
>  	}
> -	if (dev->mm)
> +	if (dev->mm) {
> +		mmu_notifier_unregister(&dev->mmu_notifier, dev->mm);
>  		mmput(dev->mm);
> +	}
> +	for (i = 0; i < dev->nvqs; i++)
> +		vhost_clean_vmaps(dev->vqs[i]);
>  	dev->mm = NULL;
>  }
>  EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
> @@ -891,6 +1063,16 @@ static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
>  
>  static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_used *used = vq->used_ring.addr;
> +
> +		if (likely(used)) {
> +			*((__virtio16 *)&used->ring[vq->num]) =
> +				cpu_to_vhost16(vq, vq->avail_idx);
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
>  			      vhost_avail_event(vq));
>  }
> @@ -899,6 +1081,16 @@ static inline int vhost_put_used(struct vhost_virtqueue *vq,
>  				 struct vring_used_elem *head, int idx,
>  				 int count)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_used *used = vq->used_ring.addr;
> +
> +		if (likely(used)) {
> +			memcpy(used->ring + idx, head,
> +			       count * sizeof(*head));
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_copy_to_user(vq, vq->used->ring + idx, head,
>  				  count * sizeof(*head));
>  }
> @@ -906,6 +1098,15 @@ static inline int vhost_put_used(struct vhost_virtqueue *vq,
>  static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
>  
>  {
> +	if (!vq->iotlb) {
> +		struct vring_used *used = vq->used_ring.addr;
> +
> +		if (likely(used)) {
> +			used->flags = cpu_to_vhost16(vq, vq->used_flags);
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
>  			      &vq->used->flags);
>  }
> @@ -913,6 +1114,15 @@ static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
>  static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
>  
>  {
> +	if (!vq->iotlb) {
> +		struct vring_used *used = vq->used_ring.addr;
> +
> +		if (likely(used)) {
> +			used->idx = cpu_to_vhost16(vq, vq->last_used_idx);
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
>  			      &vq->used->idx);
>  }
> @@ -958,12 +1168,30 @@ static void vhost_dev_unlock_vqs(struct vhost_dev *d)
>  static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
>  				      __virtio16 *idx)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_avail *avail = vq->avail_ring.addr;
> +
> +		if (likely(avail)) {
> +			*idx = avail->idx;
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_get_avail(vq, *idx, &vq->avail->idx);
>  }
>  
>  static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
>  				       __virtio16 *head, int idx)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_avail *avail = vq->avail_ring.addr;
> +
> +		if (likely(avail)) {
> +			*head = avail->ring[idx & (vq->num - 1)];
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_get_avail(vq, *head,
>  			       &vq->avail->ring[idx & (vq->num - 1)]);
>  }
> @@ -971,24 +1199,60 @@ static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
>  static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
>  					__virtio16 *flags)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_avail *avail = vq->avail_ring.addr;
> +
> +		if (likely(avail)) {
> +			*flags = avail->flags;
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_get_avail(vq, *flags, &vq->avail->flags);
>  }
>  
>  static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
>  				       __virtio16 *event)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_avail *avail = vq->avail_ring.addr;
> +
> +		if (likely(avail)) {
> +			*event = (__virtio16)avail->ring[vq->num];
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_get_avail(vq, *event, vhost_used_event(vq));
>  }
>  
>  static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
>  				     __virtio16 *idx)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_used *used = vq->used_ring.addr;
> +
> +		if (likely(used)) {
> +			*idx = used->idx;
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_get_used(vq, *idx, &vq->used->idx);
>  }
>  
>  static inline int vhost_get_desc(struct vhost_virtqueue *vq,
>  				 struct vring_desc *desc, int idx)
>  {
> +	if (!vq->iotlb) {
> +		struct vring_desc *d = vq->desc_ring.addr;
> +
> +		if (likely(d)) {
> +			*desc = *(d + idx);
> +			return 0;
> +		}
> +	}
> +
>  	return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
>  }
>  
> @@ -1329,8 +1593,16 @@ int vq_meta_prefetch(struct vhost_virtqueue *vq)
>  {
>  	unsigned int num = vq->num;
>  
> -	if (!vq->iotlb)
> +	if (!vq->iotlb) {
> +		if (unlikely(!vq->avail_ring.addr))
> +			vhost_setup_avail_vmap(vq, (unsigned long)vq->avail);
> +		if (unlikely(!vq->desc_ring.addr))
> +			vhost_setup_desc_vmap(vq, (unsigned long)vq->desc);
> +		if (unlikely(!vq->used_ring.addr))
> +			vhost_setup_used_vmap(vq, (unsigned long)vq->used);
> +
>  		return 1;
> +	}
>  
>  	return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
>  			       vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
> @@ -1482,6 +1754,13 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
>  
>  	mutex_lock(&vq->mutex);
>  
> +	/* Unregister MMU notifer to allow invalidation callback
> +	 * can access vq->avail, vq->desc , vq->used and vq->num
> +	 * without holding vq->mutex.
> +	 */
> +	if (d->mm)
> +		mmu_notifier_unregister(&d->mmu_notifier, d->mm);
> +
>  	switch (ioctl) {
>  	case VHOST_SET_VRING_NUM:
>  		/* Resizing ring with an active backend?
> @@ -1498,6 +1777,7 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
>  			r = -EINVAL;
>  			break;
>  		}
> +		vhost_clean_vmaps(vq);
>  		vq->num = s.num;
>  		break;
>  	case VHOST_SET_VRING_BASE:
> @@ -1575,6 +1855,8 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
>  			}
>  		}
>  
> +		vhost_clean_vmaps(vq);
> +
>  		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
>  		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
>  		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
> @@ -1655,6 +1937,8 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
>  	if (pollstart && vq->handle_kick)
>  		r = vhost_poll_start(&vq->poll, vq->kick);
>  
> +	if (d->mm)
> +		mmu_notifier_register(&d->mmu_notifier, d->mm);
>  	mutex_unlock(&vq->mutex);
>  
>  	if (pollstop && vq->handle_kick)
> diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h
> index 4e21011b6628..c04bc327db9f 100644
> --- a/drivers/vhost/vhost.h
> +++ b/drivers/vhost/vhost.h
> @@ -12,6 +12,8 @@
>  #include <linux/virtio_config.h>
>  #include <linux/virtio_ring.h>
>  #include <linux/atomic.h>
> +#include <linux/pagemap.h>
> +#include <linux/mmu_notifier.h>
>  
>  struct vhost_work;
>  typedef void (*vhost_work_fn_t)(struct vhost_work *work);
> @@ -80,6 +82,11 @@ enum vhost_uaddr_type {
>  	VHOST_NUM_ADDRS = 3,
>  };
>  
> +struct vhost_vmap {
> +	void *addr;
> +	void *unmap_addr;
> +};
> +
>  /* The virtqueue structure describes a queue attached to a device. */
>  struct vhost_virtqueue {
>  	struct vhost_dev *dev;
> @@ -90,6 +97,11 @@ struct vhost_virtqueue {
>  	struct vring_desc __user *desc;
>  	struct vring_avail __user *avail;
>  	struct vring_used __user *used;
> +
> +	struct vhost_vmap avail_ring;
> +	struct vhost_vmap desc_ring;
> +	struct vhost_vmap used_ring;
> +
>  	const struct vhost_umem_node *meta_iotlb[VHOST_NUM_ADDRS];
>  	struct file *kick;
>  	struct eventfd_ctx *call_ctx;
> @@ -158,6 +170,7 @@ struct vhost_msg_node {
>  
>  struct vhost_dev {
>  	struct mm_struct *mm;
> +	struct mmu_notifier mmu_notifier;
>  	struct mutex mutex;
>  	struct vhost_virtqueue **vqs;
>  	int nvqs;
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 6ece1e2fe76e..745e7c7f7a6c 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -237,6 +237,7 @@ bool vma_is_shmem(struct vm_area_struct *vma)
>  {
>  	return vma->vm_ops == &shmem_vm_ops;
>  }
> +EXPORT_SYMBOL_GPL(vma_is_shmem);
>  
>  static LIST_HEAD(shmem_swaplist);
>  static DEFINE_MUTEX(shmem_swaplist_mutex);
> -- 
> 2.17.1
Jason Wang Jan. 25, 2019, 9:16 a.m. UTC | #9
On 2019/1/25 上午11:00, Michael S. Tsirkin wrote:
> On Thu, Jan 24, 2019 at 12:07:54PM +0800, Jason Wang wrote:
>>> Meanwhile, could you pls post data comparing this last patch with the
>>> below?  This removes the speculation barrier replacing it with a
>>> (useless but at least more lightweight) data dependency.
>> SMAP off
>>
>> Your patch: 7.2MPPs
>>
>> vmap: 7.4Mpps
> OK so while we keep looking into vmap, why don't we merge something like
> the below?  Seems quite straight forward ...
>

The problem is it gives ~8% regression on PPS when SMAP is on. This is 
probably because the latency of lfence is hided by previous stac as 
mentioned in the commit b3bbfb3fb5d25776b8e3f361d2eedaabb0b496cd.

Thanks
Jason Wang Jan. 25, 2019, 9:21 a.m. UTC | #10
On 2019/1/25 上午11:03, Michael S. Tsirkin wrote:
> On Wed, Jan 23, 2019 at 05:55:57PM +0800, Jason Wang wrote:
>> It was noticed that the copy_user() friends that was used to access
>> virtqueue metdata tends to be very expensive for dataplane
>> implementation like vhost since it involves lots of software checks,
>> speculation barrier, hardware feature toggling (e.g SMAP). The
>> extra cost will be more obvious when transferring small packets since
>> the time spent on metadata accessing become more significant.
>>
>> This patch tries to eliminate those overheads by accessing them
>> through kernel virtual address by vmap(). To make the pages can be
>> migrated, instead of pinning them through GUP, we use MMU notifiers to
>> invalidate vmaps and re-establish vmaps during each round of metadata
>> prefetching if necessary. For devices that doesn't use metadata
>> prefetching, the memory accessors fallback to normal copy_user()
>> implementation gracefully. The invalidation was synchronized with
>> datapath through vq mutex, and in order to avoid hold vq mutex during
>> range checking, MMU notifier was teared down when trying to modify vq
>> metadata.
>>
>> Another thing is kernel lacks efficient solution for tracking dirty
>> pages by vmap(), this will lead issues if vhost is using file backed
>> memory which needs care of writeback. This patch solves this issue by
>> just skipping the vma that is file backed and fallback to normal
>> copy_user() friends. This might introduce some overheads for file
>> backed users but consider this use case is rare we could do
>> optimizations on top.
>>
>> Note that this was only done when device IOTLB is not enabled. We
>> could use similar method to optimize it in the future.
>>
>> Tests shows at most about 22% improvement on TX PPS when using
>> virtio-user + vhost_net + xdp1 + TAP on 2.6GHz Broadwell:
>>
>>          SMAP on | SMAP off
>> Before: 5.0Mpps | 6.6Mpps
>> After:  6.1Mpps | 7.4Mpps
>>
>> Signed-off-by: Jason Wang <jasowang@redhat.com>
>> ---
>>   drivers/vhost/vhost.c | 288 +++++++++++++++++++++++++++++++++++++++++-
>>   drivers/vhost/vhost.h |  13 ++
>>   mm/shmem.c            |   1 +
>>   3 files changed, 300 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
>> index 37e2cac8e8b0..096ae3298d62 100644
>> --- a/drivers/vhost/vhost.c
>> +++ b/drivers/vhost/vhost.c
>> @@ -440,6 +440,9 @@ void vhost_dev_init(struct vhost_dev *dev,
>>   		vq->indirect = NULL;
>>   		vq->heads = NULL;
>>   		vq->dev = dev;
>> +		memset(&vq->avail_ring, 0, sizeof(vq->avail_ring));
>> +		memset(&vq->used_ring, 0, sizeof(vq->used_ring));
>> +		memset(&vq->desc_ring, 0, sizeof(vq->desc_ring));
>>   		mutex_init(&vq->mutex);
>>   		vhost_vq_reset(dev, vq);
>>   		if (vq->handle_kick)
>> @@ -510,6 +513,73 @@ static size_t vhost_get_desc_size(struct vhost_virtqueue *vq, int num)
>>   	return sizeof(*vq->desc) * num;
>>   }
>>   
>> +static void vhost_uninit_vmap(struct vhost_vmap *map)
>> +{
>> +	if (map->addr)
>> +		vunmap(map->unmap_addr);
>> +
>> +	map->addr = NULL;
>> +	map->unmap_addr = NULL;
>> +}
>> +
>> +static int vhost_invalidate_vmap(struct vhost_virtqueue *vq,
>> +				 struct vhost_vmap *map,
>> +				 unsigned long ustart,
>> +				 size_t size,
>> +				 unsigned long start,
>> +				 unsigned long end,
>> +				 bool blockable)
>> +{
>> +	if (end < ustart || start > ustart - 1 + size)
>> +		return 0;
>> +
>> +	if (!blockable)
>> +		return -EAGAIN;
>> +
>> +	mutex_lock(&vq->mutex);
>> +	vhost_uninit_vmap(map);
>> +	mutex_unlock(&vq->mutex);
>> +
>> +	return 0;
>> +}
>> +
>> +static int vhost_invalidate_range_start(struct mmu_notifier *mn,
>> +					const struct mmu_notifier_range *range)
>> +{
>> +	struct vhost_dev *dev = container_of(mn, struct vhost_dev,
>> +					     mmu_notifier);
>> +	int i;
>> +
>> +	for (i = 0; i < dev->nvqs; i++) {
>> +		struct vhost_virtqueue *vq = dev->vqs[i];
>> +
>> +		if (vhost_invalidate_vmap(vq, &vq->avail_ring,
>> +					  (unsigned long)vq->avail,
>> +					  vhost_get_avail_size(vq, vq->num),
>> +					  range->start, range->end,
>> +					  range->blockable))
>> +			return -EAGAIN;
>> +		if (vhost_invalidate_vmap(vq, &vq->desc_ring,
>> +					  (unsigned long)vq->desc,
>> +					  vhost_get_desc_size(vq, vq->num),
>> +					  range->start, range->end,
>> +					  range->blockable))
>> +			return -EAGAIN;
>> +		if (vhost_invalidate_vmap(vq, &vq->used_ring,
>> +					  (unsigned long)vq->used,
>> +					  vhost_get_used_size(vq, vq->num),
>> +					  range->start, range->end,
>> +					  range->blockable))
>> +			return -EAGAIN;
>> +	}
>> +
>> +	return 0;
>> +}
>> +
>> +static const struct mmu_notifier_ops vhost_mmu_notifier_ops = {
>> +	.invalidate_range_start = vhost_invalidate_range_start,
>> +};
>> +
>>   /* Caller should have device mutex */
>>   long vhost_dev_set_owner(struct vhost_dev *dev)
>>   {
> It seems questionable to merely track .invalidate_range_start.
> Don't we care about keeping pages young/accessed?


My understanding is the young stuffs were only needed for secondary MMU 
where the hva is not used. This is not the case of vhost since anyway 
guest will access those pages through userspace address.


> MMU will think they aren't and will penalize vhost by pushing
> them out.
>
> I note that MMU documentation says
>          * invalidate_range_start() and invalidate_range_end() must be
>           * paired
> and it seems questionable that they are not paired here.


I can see some users with the unpaired invalidate_range_start(). Maybe I 
miss something but I can not find anything that we need to do after the 
page is unmaped.


>
>
> I also wonder about things like write-protecting the pages.
> It does not look like a range is invalidated when page
> is write-protected, even though I might have missed that.
> If not we can be corrupting memory in a variety of ways
> e.g. when using KSM, or with COW.


Yes, we probably need to implement change_pte() method which will do 
vunmap().

Thanks
Jason Wang Jan. 25, 2019, 9:24 a.m. UTC | #11
On 2019/1/25 上午11:03, Michael S. Tsirkin wrote:
>> +/* Suppress the vma that needs writeback since we can not track dirty
>> + * pages now.
>> + */
>> +static bool vma_can_vmap(struct vm_area_struct *vma)
>> +{
>> +	return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
>> +	       vma_is_shmem(vma);
>> +}
>> +
> IIUC a second but anonymous memory needs writeback too, just to swap.
> I'm not an MM person so I might be off.


Right, my fault, I mean the vma that needs dirty page tracking.

Thanks

Patch
diff mbox series

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index 37e2cac8e8b0..096ae3298d62 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -440,6 +440,9 @@  void vhost_dev_init(struct vhost_dev *dev,
 		vq->indirect = NULL;
 		vq->heads = NULL;
 		vq->dev = dev;
+		memset(&vq->avail_ring, 0, sizeof(vq->avail_ring));
+		memset(&vq->used_ring, 0, sizeof(vq->used_ring));
+		memset(&vq->desc_ring, 0, sizeof(vq->desc_ring));
 		mutex_init(&vq->mutex);
 		vhost_vq_reset(dev, vq);
 		if (vq->handle_kick)
@@ -510,6 +513,73 @@  static size_t vhost_get_desc_size(struct vhost_virtqueue *vq, int num)
 	return sizeof(*vq->desc) * num;
 }
 
+static void vhost_uninit_vmap(struct vhost_vmap *map)
+{
+	if (map->addr)
+		vunmap(map->unmap_addr);
+
+	map->addr = NULL;
+	map->unmap_addr = NULL;
+}
+
+static int vhost_invalidate_vmap(struct vhost_virtqueue *vq,
+				 struct vhost_vmap *map,
+				 unsigned long ustart,
+				 size_t size,
+				 unsigned long start,
+				 unsigned long end,
+				 bool blockable)
+{
+	if (end < ustart || start > ustart - 1 + size)
+		return 0;
+
+	if (!blockable)
+		return -EAGAIN;
+
+	mutex_lock(&vq->mutex);
+	vhost_uninit_vmap(map);
+	mutex_unlock(&vq->mutex);
+
+	return 0;
+}
+
+static int vhost_invalidate_range_start(struct mmu_notifier *mn,
+					const struct mmu_notifier_range *range)
+{
+	struct vhost_dev *dev = container_of(mn, struct vhost_dev,
+					     mmu_notifier);
+	int i;
+
+	for (i = 0; i < dev->nvqs; i++) {
+		struct vhost_virtqueue *vq = dev->vqs[i];
+
+		if (vhost_invalidate_vmap(vq, &vq->avail_ring,
+					  (unsigned long)vq->avail,
+					  vhost_get_avail_size(vq, vq->num),
+					  range->start, range->end,
+					  range->blockable))
+			return -EAGAIN;
+		if (vhost_invalidate_vmap(vq, &vq->desc_ring,
+					  (unsigned long)vq->desc,
+					  vhost_get_desc_size(vq, vq->num),
+					  range->start, range->end,
+					  range->blockable))
+			return -EAGAIN;
+		if (vhost_invalidate_vmap(vq, &vq->used_ring,
+					  (unsigned long)vq->used,
+					  vhost_get_used_size(vq, vq->num),
+					  range->start, range->end,
+					  range->blockable))
+			return -EAGAIN;
+	}
+
+	return 0;
+}
+
+static const struct mmu_notifier_ops vhost_mmu_notifier_ops = {
+	.invalidate_range_start = vhost_invalidate_range_start,
+};
+
 /* Caller should have device mutex */
 long vhost_dev_set_owner(struct vhost_dev *dev)
 {
@@ -541,7 +611,14 @@  long vhost_dev_set_owner(struct vhost_dev *dev)
 	if (err)
 		goto err_cgroup;
 
+	dev->mmu_notifier.ops = &vhost_mmu_notifier_ops;
+	err = mmu_notifier_register(&dev->mmu_notifier, dev->mm);
+	if (err)
+		goto err_mmu_notifier;
+
 	return 0;
+err_mmu_notifier:
+	vhost_dev_free_iovecs(dev);
 err_cgroup:
 	kthread_stop(worker);
 	dev->worker = NULL;
@@ -632,6 +709,97 @@  static void vhost_clear_msg(struct vhost_dev *dev)
 	spin_unlock(&dev->iotlb_lock);
 }
 
+/* Suppress the vma that needs writeback since we can not track dirty
+ * pages now.
+ */
+static bool vma_can_vmap(struct vm_area_struct *vma)
+{
+	return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) ||
+	       vma_is_shmem(vma);
+}
+
+static int vhost_init_vmap(struct vhost_dev *dev,
+			   struct vhost_vmap *map, unsigned long uaddr,
+			   size_t size, int write)
+{
+	struct mm_struct *mm = dev->mm;
+	struct vm_area_struct *vma;
+	struct page **pages;
+	int npages = DIV_ROUND_UP(size, PAGE_SIZE);
+	int npinned;
+	void *vaddr;
+	int err = 0;
+
+	down_read(&mm->mmap_sem);
+	vma = find_vma(mm, uaddr);
+	if (!vma || !vma_can_vmap(vma) ||
+	    vma->vm_end < uaddr - 1 + size) {
+		err = -EINVAL;
+		goto err_vma;
+	}
+
+	pages = kmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
+	if (!pages) {
+		err = -ENOMEM;
+		goto err_alloc;
+	}
+
+	npinned = get_user_pages_fast(uaddr, npages, write, pages);
+	if (npinned != npages) {
+		err = -EFAULT;
+		goto err_gup;
+	}
+
+	vaddr = vmap(pages, npages, VM_MAP, PAGE_KERNEL);
+	if (!vaddr) {
+		err = EFAULT;
+		goto err_gup;
+	}
+
+	map->addr = vaddr + (uaddr & (PAGE_SIZE - 1));
+	map->unmap_addr = vaddr;
+
+err_gup:
+	/* Don't pin pages, mmu notifier will notify us about page
+	 * migration.
+	 */
+	if (npinned > 0)
+		release_pages(pages, npinned);
+err_alloc:
+	kfree(pages);
+err_vma:
+	up_read(&mm->mmap_sem);
+	return err;
+}
+
+static void vhost_clean_vmaps(struct vhost_virtqueue *vq)
+{
+	vhost_uninit_vmap(&vq->avail_ring);
+	vhost_uninit_vmap(&vq->desc_ring);
+	vhost_uninit_vmap(&vq->used_ring);
+}
+
+static int vhost_setup_avail_vmap(struct vhost_virtqueue *vq,
+				  unsigned long avail)
+{
+	return vhost_init_vmap(vq->dev, &vq->avail_ring, avail,
+			       vhost_get_avail_size(vq, vq->num), false);
+}
+
+static int vhost_setup_desc_vmap(struct vhost_virtqueue *vq,
+				 unsigned long desc)
+{
+	return vhost_init_vmap(vq->dev, &vq->desc_ring, desc,
+			       vhost_get_desc_size(vq, vq->num), false);
+}
+
+static int vhost_setup_used_vmap(struct vhost_virtqueue *vq,
+				 unsigned long used)
+{
+	return vhost_init_vmap(vq->dev, &vq->used_ring, used,
+			       vhost_get_used_size(vq, vq->num), true);
+}
+
 void vhost_dev_cleanup(struct vhost_dev *dev)
 {
 	int i;
@@ -661,8 +829,12 @@  void vhost_dev_cleanup(struct vhost_dev *dev)
 		kthread_stop(dev->worker);
 		dev->worker = NULL;
 	}
-	if (dev->mm)
+	if (dev->mm) {
+		mmu_notifier_unregister(&dev->mmu_notifier, dev->mm);
 		mmput(dev->mm);
+	}
+	for (i = 0; i < dev->nvqs; i++)
+		vhost_clean_vmaps(dev->vqs[i]);
 	dev->mm = NULL;
 }
 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
@@ -891,6 +1063,16 @@  static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 
 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 {
+	if (!vq->iotlb) {
+		struct vring_used *used = vq->used_ring.addr;
+
+		if (likely(used)) {
+			*((__virtio16 *)&used->ring[vq->num]) =
+				cpu_to_vhost16(vq, vq->avail_idx);
+			return 0;
+		}
+	}
+
 	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 			      vhost_avail_event(vq));
 }
@@ -899,6 +1081,16 @@  static inline int vhost_put_used(struct vhost_virtqueue *vq,
 				 struct vring_used_elem *head, int idx,
 				 int count)
 {
+	if (!vq->iotlb) {
+		struct vring_used *used = vq->used_ring.addr;
+
+		if (likely(used)) {
+			memcpy(used->ring + idx, head,
+			       count * sizeof(*head));
+			return 0;
+		}
+	}
+
 	return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 				  count * sizeof(*head));
 }
@@ -906,6 +1098,15 @@  static inline int vhost_put_used(struct vhost_virtqueue *vq,
 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 
 {
+	if (!vq->iotlb) {
+		struct vring_used *used = vq->used_ring.addr;
+
+		if (likely(used)) {
+			used->flags = cpu_to_vhost16(vq, vq->used_flags);
+			return 0;
+		}
+	}
+
 	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 			      &vq->used->flags);
 }
@@ -913,6 +1114,15 @@  static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 
 {
+	if (!vq->iotlb) {
+		struct vring_used *used = vq->used_ring.addr;
+
+		if (likely(used)) {
+			used->idx = cpu_to_vhost16(vq, vq->last_used_idx);
+			return 0;
+		}
+	}
+
 	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 			      &vq->used->idx);
 }
@@ -958,12 +1168,30 @@  static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
 				      __virtio16 *idx)
 {
+	if (!vq->iotlb) {
+		struct vring_avail *avail = vq->avail_ring.addr;
+
+		if (likely(avail)) {
+			*idx = avail->idx;
+			return 0;
+		}
+	}
+
 	return vhost_get_avail(vq, *idx, &vq->avail->idx);
 }
 
 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
 				       __virtio16 *head, int idx)
 {
+	if (!vq->iotlb) {
+		struct vring_avail *avail = vq->avail_ring.addr;
+
+		if (likely(avail)) {
+			*head = avail->ring[idx & (vq->num - 1)];
+			return 0;
+		}
+	}
+
 	return vhost_get_avail(vq, *head,
 			       &vq->avail->ring[idx & (vq->num - 1)]);
 }
@@ -971,24 +1199,60 @@  static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
 					__virtio16 *flags)
 {
+	if (!vq->iotlb) {
+		struct vring_avail *avail = vq->avail_ring.addr;
+
+		if (likely(avail)) {
+			*flags = avail->flags;
+			return 0;
+		}
+	}
+
 	return vhost_get_avail(vq, *flags, &vq->avail->flags);
 }
 
 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
 				       __virtio16 *event)
 {
+	if (!vq->iotlb) {
+		struct vring_avail *avail = vq->avail_ring.addr;
+
+		if (likely(avail)) {
+			*event = (__virtio16)avail->ring[vq->num];
+			return 0;
+		}
+	}
+
 	return vhost_get_avail(vq, *event, vhost_used_event(vq));
 }
 
 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
 				     __virtio16 *idx)
 {
+	if (!vq->iotlb) {
+		struct vring_used *used = vq->used_ring.addr;
+
+		if (likely(used)) {
+			*idx = used->idx;
+			return 0;
+		}
+	}
+
 	return vhost_get_used(vq, *idx, &vq->used->idx);
 }
 
 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
 				 struct vring_desc *desc, int idx)
 {
+	if (!vq->iotlb) {
+		struct vring_desc *d = vq->desc_ring.addr;
+
+		if (likely(d)) {
+			*desc = *(d + idx);
+			return 0;
+		}
+	}
+
 	return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
 }
 
@@ -1329,8 +1593,16 @@  int vq_meta_prefetch(struct vhost_virtqueue *vq)
 {
 	unsigned int num = vq->num;
 
-	if (!vq->iotlb)
+	if (!vq->iotlb) {
+		if (unlikely(!vq->avail_ring.addr))
+			vhost_setup_avail_vmap(vq, (unsigned long)vq->avail);
+		if (unlikely(!vq->desc_ring.addr))
+			vhost_setup_desc_vmap(vq, (unsigned long)vq->desc);
+		if (unlikely(!vq->used_ring.addr))
+			vhost_setup_used_vmap(vq, (unsigned long)vq->used);
+
 		return 1;
+	}
 
 	return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
 			       vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
@@ -1482,6 +1754,13 @@  long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
 
 	mutex_lock(&vq->mutex);
 
+	/* Unregister MMU notifer to allow invalidation callback
+	 * can access vq->avail, vq->desc , vq->used and vq->num
+	 * without holding vq->mutex.
+	 */
+	if (d->mm)
+		mmu_notifier_unregister(&d->mmu_notifier, d->mm);
+
 	switch (ioctl) {
 	case VHOST_SET_VRING_NUM:
 		/* Resizing ring with an active backend?
@@ -1498,6 +1777,7 @@  long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
 			r = -EINVAL;
 			break;
 		}
+		vhost_clean_vmaps(vq);
 		vq->num = s.num;
 		break;
 	case VHOST_SET_VRING_BASE:
@@ -1575,6 +1855,8 @@  long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
 			}
 		}
 
+		vhost_clean_vmaps(vq);
+
 		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
@@ -1655,6 +1937,8 @@  long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
 	if (pollstart && vq->handle_kick)
 		r = vhost_poll_start(&vq->poll, vq->kick);
 
+	if (d->mm)
+		mmu_notifier_register(&d->mmu_notifier, d->mm);
 	mutex_unlock(&vq->mutex);
 
 	if (pollstop && vq->handle_kick)
diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h
index 4e21011b6628..c04bc327db9f 100644
--- a/drivers/vhost/vhost.h
+++ b/drivers/vhost/vhost.h
@@ -12,6 +12,8 @@ 
 #include <linux/virtio_config.h>
 #include <linux/virtio_ring.h>
 #include <linux/atomic.h>
+#include <linux/pagemap.h>
+#include <linux/mmu_notifier.h>
 
 struct vhost_work;
 typedef void (*vhost_work_fn_t)(struct vhost_work *work);
@@ -80,6 +82,11 @@  enum vhost_uaddr_type {
 	VHOST_NUM_ADDRS = 3,
 };
 
+struct vhost_vmap {
+	void *addr;
+	void *unmap_addr;
+};
+
 /* The virtqueue structure describes a queue attached to a device. */
 struct vhost_virtqueue {
 	struct vhost_dev *dev;
@@ -90,6 +97,11 @@  struct vhost_virtqueue {
 	struct vring_desc __user *desc;
 	struct vring_avail __user *avail;
 	struct vring_used __user *used;
+
+	struct vhost_vmap avail_ring;
+	struct vhost_vmap desc_ring;
+	struct vhost_vmap used_ring;
+
 	const struct vhost_umem_node *meta_iotlb[VHOST_NUM_ADDRS];
 	struct file *kick;
 	struct eventfd_ctx *call_ctx;
@@ -158,6 +170,7 @@  struct vhost_msg_node {
 
 struct vhost_dev {
 	struct mm_struct *mm;
+	struct mmu_notifier mmu_notifier;
 	struct mutex mutex;
 	struct vhost_virtqueue **vqs;
 	int nvqs;
diff --git a/mm/shmem.c b/mm/shmem.c
index 6ece1e2fe76e..745e7c7f7a6c 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -237,6 +237,7 @@  bool vma_is_shmem(struct vm_area_struct *vma)
 {
 	return vma->vm_ops == &shmem_vm_ops;
 }
+EXPORT_SYMBOL_GPL(vma_is_shmem);
 
 static LIST_HEAD(shmem_swaplist);
 static DEFINE_MUTEX(shmem_swaplist_mutex);