[PATCH 0/2] use vfio_iova_rw() to read/write IOVAs from CPU side

[PATCH 0/2] use vfio_iova_rw() to read/write IOVAs from CPU side

[Yan Zhao]

When device models read/write memory pointed by a range of IOVAs, it is
better to use vfio interfaces. And because this read/write is from CPUs,
there's no need to call vfio_pin_pages() to pin those memory.

patch 1 introduces interface vfio_iova_rw() in vfio to read/write
userspace IOVAs without pinning userspace pages.

patch 2 let gvt switch from kvm side rw interface to vfio_iova_rw().


Yan Zhao (2):
  vfio: introduce vfio_iova_rw to read/write a range of IOVAs
  drm/i915/gvt: subsitute kvm_read/write_guest with vfio_iova_rw

 drivers/gpu/drm/i915/gvt/kvmgt.c | 26 +++-------
 drivers/vfio/vfio.c              | 45 ++++++++++++++++++
 drivers/vfio/vfio_iommu_type1.c  | 81 ++++++++++++++++++++++++++++++++
 include/linux/vfio.h             |  5 ++
 4 files changed, 138 insertions(+), 19 deletions(-)

-- 
2.17.1

[PATCH 1/2] vfio: introduce vfio_iova_rw to read/write a range of IOVAs

[Yan Zhao]

vfio_iova_rw will read/write a range of userspace memory (starting form
device iova to iova + len -1) into a kenrel buffer without pinning the
userspace memory.

TODO: vfio needs to mark the iova dirty if vfio_iova_rw(write) is
called.

Cc: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
---
 drivers/vfio/vfio.c             | 45 ++++++++++++++++++
 drivers/vfio/vfio_iommu_type1.c | 81 +++++++++++++++++++++++++++++++++
 include/linux/vfio.h            |  5 ++
 3 files changed, 131 insertions(+)

diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c
index c8482624ca34..36e91e647ed5 100644
--- a/drivers/vfio/vfio.c
+++ b/drivers/vfio/vfio.c
@@ -1961,6 +1961,51 @@ int vfio_unpin_pages(struct device *dev, unsigned long *user_pfn, int npage)
 }
 EXPORT_SYMBOL(vfio_unpin_pages);
 
+/*
+ * Read/Write a range of userspace IOVAs for a device into/from a kernel
+ * buffer without pinning the userspace memory
+ * @dev [in]  : device
+ * @iova [in] : base IOVA of a userspace buffer
+ * @data [in] : pointer to kernel buffer
+ * @len [in]  : kernel buffer length
+ * @write     : indicate read or write
+ * Return error on failure or 0 on success.
+ */
+int vfio_iova_rw(struct device *dev, unsigned long iova, void *data,
+		   unsigned long len, bool write)
+{
+	struct vfio_container *container;
+	struct vfio_group *group;
+	struct vfio_iommu_driver *driver;
+	int ret = 0;
+
+	if (!dev || !data || len <= 0)
+		return -EINVAL;
+
+	group = vfio_group_get_from_dev(dev);
+	if (!group)
+		return -ENODEV;
+
+	ret = vfio_group_add_container_user(group);
+	if (ret)
+		goto out;
+
+	container = group->container;
+	driver = container->iommu_driver;
+
+	if (likely(driver && driver->ops->iova_rw))
+		ret = driver->ops->iova_rw(container->iommu_data,
+					   iova, data, len, write);
+	else
+		ret = -ENOTTY;
+
+	vfio_group_try_dissolve_container(group);
+out:
+	vfio_group_put(group);
+	return ret;
+}
+EXPORT_SYMBOL(vfio_iova_rw);
+
 static int vfio_register_iommu_notifier(struct vfio_group *group,
 					unsigned long *events,
 					struct notifier_block *nb)
diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index 2ada8e6cdb88..aee191077235 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -27,6 +27,7 @@
 #include <linux/iommu.h>
 #include <linux/module.h>
 #include <linux/mm.h>
+#include <linux/mmu_context.h>
 #include <linux/rbtree.h>
 #include <linux/sched/signal.h>
 #include <linux/sched/mm.h>
@@ -2326,6 +2327,85 @@ static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
 	return blocking_notifier_chain_unregister(&iommu->notifier, nb);
 }
 
+static int next_segment(unsigned long len, int offset)
+{
+	if (len > PAGE_SIZE - offset)
+		return PAGE_SIZE - offset;
+	else
+		return len;
+}
+
+static int vfio_iommu_type1_rw_iova_seg(struct vfio_iommu *iommu,
+					  unsigned long iova, void *data,
+					  unsigned long seg_len,
+					  unsigned long offset,
+					  bool write)
+{
+	struct mm_struct *mm;
+	unsigned long vaddr;
+	struct vfio_dma *dma;
+	bool kthread = current->mm == NULL;
+	int ret = 0;
+
+	dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
+	if (!dma)
+		return -EINVAL;
+
+	mm = get_task_mm(dma->task);
+
+	if (!mm)
+		return -ENODEV;
+
+	if (kthread)
+		use_mm(mm);
+	else if (current->mm != mm) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	vaddr = dma->vaddr + iova - dma->iova + offset;
+
+	ret = write ? __copy_to_user((void __user *)vaddr,
+			data, seg_len) :
+		__copy_from_user(data, (void __user *)vaddr,
+				seg_len);
+	if (ret)
+		ret = -EFAULT;
+
+	if (kthread)
+		unuse_mm(mm);
+out:
+	mmput(mm);
+	return ret;
+}
+
+static int vfio_iommu_type1_iova_rw(void *iommu_data, unsigned long iova,
+				    void *data, unsigned long len, bool write)
+{
+	struct vfio_iommu *iommu = iommu_data;
+	int offset = iova & ~PAGE_MASK;
+	int seg_len;
+	int ret = 0;
+
+	iova = iova & PAGE_MASK;
+
+	mutex_lock(&iommu->lock);
+	while ((seg_len = next_segment(len, offset)) > 0) {
+		ret = vfio_iommu_type1_rw_iova_seg(iommu, iova, data,
+						   seg_len, offset, write);
+		if (ret)
+			break;
+
+		offset = 0;
+		len -= seg_len;
+		data += seg_len;
+		iova += PAGE_SIZE;
+	}
+
+	mutex_unlock(&iommu->lock);
+	return ret;
+}
+
 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
 	.name			= "vfio-iommu-type1",
 	.owner			= THIS_MODULE,
@@ -2338,6 +2418,7 @@ static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
 	.unpin_pages		= vfio_iommu_type1_unpin_pages,
 	.register_notifier	= vfio_iommu_type1_register_notifier,
 	.unregister_notifier	= vfio_iommu_type1_unregister_notifier,
+	.iova_rw		= vfio_iommu_type1_iova_rw,
 };
 
 static int __init vfio_iommu_type1_init(void)
diff --git a/include/linux/vfio.h b/include/linux/vfio.h
index e42a711a2800..7bf18a31bbcf 100644
--- a/include/linux/vfio.h
+++ b/include/linux/vfio.h
@@ -82,6 +82,8 @@ struct vfio_iommu_driver_ops {
 					     struct notifier_block *nb);
 	int		(*unregister_notifier)(void *iommu_data,
 					       struct notifier_block *nb);
+	int		(*iova_rw)(void *iommu_data, unsigned long iova,
+				   void *data, unsigned long len, bool write);
 };
 
 extern int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops);
@@ -107,6 +109,9 @@ extern int vfio_pin_pages(struct device *dev, unsigned long *user_pfn,
 extern int vfio_unpin_pages(struct device *dev, unsigned long *user_pfn,
 			    int npage);
 
+extern int vfio_iova_rw(struct device *dev, unsigned long iova, void *data,
+			unsigned long len, bool write);
+
 /* each type has independent events */
 enum vfio_notify_type {
 	VFIO_IOMMU_NOTIFY = 0,
-- 
2.17.1

[PATCH 2/2] drm/i915/gvt: subsitute kvm_read/write_guest with vfio_iova_rw

[Yan Zhao]

As a device model, it is better to read/write guest memory using vfio
interface, so that vfio is able to maintain dirty info of device IOVAs.

Compared to CPU side interfaces kvm_read/write_guest(), vfio_iova_rw()
has ~600 cycles more overhead on average.
-------------------------------------
|    interface     | avg cpu cycles |
|-----------------------------------|
| kvm_write_guest  |     1546       |
| ----------------------------------|
| kvm_read_guest   |     686        |
|-----------------------------------|
| vfio_iova_rw(w)  |     2233       |
|-----------------------------------|
| vfio_iova_rw(r)  |     1262       |
-------------------------------------

Comparison of benchmarks scores are as blow:
---------------------------------------------------------
|  avg score  | kvm_read/write_guest   | vfio_iova_rw   |
---------------------------------------------------------
|   Glmark2   |         1132           |      1138.2    |
---------------------------------------------------------
|  Lightsmark |        61.558          |      61.538    |
|--------------------------------------------------------
|  OpenArena  |        142.77          |      136.6     |
---------------------------------------------------------
|   Heaven    |         698            |      686.8     |
--------------------------------------------------------
No obvious performance downgrade found.

Cc: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
---
 drivers/gpu/drm/i915/gvt/kvmgt.c | 26 +++++++-------------------
 1 file changed, 7 insertions(+), 19 deletions(-)

diff --git a/drivers/gpu/drm/i915/gvt/kvmgt.c b/drivers/gpu/drm/i915/gvt/kvmgt.c
index bd79a9718cc7..576b05db3998 100644
--- a/drivers/gpu/drm/i915/gvt/kvmgt.c
+++ b/drivers/gpu/drm/i915/gvt/kvmgt.c
@@ -1966,31 +1966,19 @@ static int kvmgt_rw_gpa(unsigned long handle, unsigned long gpa,
 			void *buf, unsigned long len, bool write)
 {
 	struct kvmgt_guest_info *info;
-	struct kvm *kvm;
-	int idx, ret;
-	bool kthread = current->mm == NULL;
+	int ret;
+	struct intel_vgpu *vgpu;
+	struct device *dev;
 
 	if (!handle_valid(handle))
 		return -ESRCH;
 
 	info = (struct kvmgt_guest_info *)handle;
-	kvm = info->kvm;
-
-	if (kthread) {
-		if (!mmget_not_zero(kvm->mm))
-			return -EFAULT;
-		use_mm(kvm->mm);
-	}
-
-	idx = srcu_read_lock(&kvm->srcu);
-	ret = write ? kvm_write_guest(kvm, gpa, buf, len) :
-		      kvm_read_guest(kvm, gpa, buf, len);
-	srcu_read_unlock(&kvm->srcu, idx);
+	vgpu = info->vgpu;
+	dev = mdev_dev(vgpu->vdev.mdev);
 
-	if (kthread) {
-		unuse_mm(kvm->mm);
-		mmput(kvm->mm);
-	}
+	ret = write ? vfio_iova_rw(dev, gpa, buf, len, true) :
+			vfio_iova_rw(dev, gpa, buf, len, false);
 
 	return ret;
 }
-- 
2.17.1

Re: [PATCH 2/2] drm/i915/gvt: subsitute kvm_read/write_guest with vfio_iova_rw

[Paolo Bonzini]

On 03/01/20 02:03, Yan Zhao wrote:
> +	ret = write ? vfio_iova_rw(dev, gpa, buf, len, true) :
> +			vfio_iova_rw(dev, gpa, buf, len, false);
>  
>  	return ret;

"Write" can be just the final argument to vfio_iova_rw, that is

   return vfio_iova_rw(dev, gpa, buf, len, write):

Thanks,

Paolo

Re: [PATCH 1/2] vfio: introduce vfio_iova_rw to read/write a range of IOVAs

[Alex Williamson]

On Thu,  2 Jan 2020 20:02:17 -0500
Yan Zhao <yan.y.zhao@intel.com> wrote:

> vfio_iova_rw will read/write a range of userspace memory (starting form
> device iova to iova + len -1) into a kenrel buffer without pinning the
> userspace memory.
> 
> TODO: vfio needs to mark the iova dirty if vfio_iova_rw(write) is
> called.
> 
> Cc: Kevin Tian <kevin.tian@intel.com>
> Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
> ---
>  drivers/vfio/vfio.c             | 45 ++++++++++++++++++
>  drivers/vfio/vfio_iommu_type1.c | 81 +++++++++++++++++++++++++++++++++
>  include/linux/vfio.h            |  5 ++
>  3 files changed, 131 insertions(+)
> 
> diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c
> index c8482624ca34..36e91e647ed5 100644
> --- a/drivers/vfio/vfio.c
> +++ b/drivers/vfio/vfio.c
> @@ -1961,6 +1961,51 @@ int vfio_unpin_pages(struct device *dev, unsigned long *user_pfn, int npage)
>  }
>  EXPORT_SYMBOL(vfio_unpin_pages);
>  
> +/*
> + * Read/Write a range of userspace IOVAs for a device into/from a kernel
> + * buffer without pinning the userspace memory
> + * @dev [in]  : device
> + * @iova [in] : base IOVA of a userspace buffer
> + * @data [in] : pointer to kernel buffer
> + * @len [in]  : kernel buffer length
> + * @write     : indicate read or write
> + * Return error on failure or 0 on success.
> + */
> +int vfio_iova_rw(struct device *dev, unsigned long iova, void *data,
> +		   unsigned long len, bool write)

Shouldn't iova be a dma_addr_t and len be a size_t?  AIUI this function
performs the equivalent behavior of the device itself performing a DMA.
Hmm, should the interface be named vfio_dma_rw()?

> +{
> +	struct vfio_container *container;
> +	struct vfio_group *group;
> +	struct vfio_iommu_driver *driver;
> +	int ret = 0;
> +
> +	if (!dev || !data || len <= 0)
> +		return -EINVAL;
> +
> +	group = vfio_group_get_from_dev(dev);
> +	if (!group)
> +		return -ENODEV;
> +
> +	ret = vfio_group_add_container_user(group);
> +	if (ret)
> +		goto out;
> +
> +	container = group->container;
> +	driver = container->iommu_driver;
> +
> +	if (likely(driver && driver->ops->iova_rw))
> +		ret = driver->ops->iova_rw(container->iommu_data,
> +					   iova, data, len, write);
> +	else
> +		ret = -ENOTTY;
> +
> +	vfio_group_try_dissolve_container(group);
> +out:
> +	vfio_group_put(group);
> +	return ret;
> +}
> +EXPORT_SYMBOL(vfio_iova_rw);
> +
>  static int vfio_register_iommu_notifier(struct vfio_group *group,
>  					unsigned long *events,
>  					struct notifier_block *nb)
> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> index 2ada8e6cdb88..aee191077235 100644
> --- a/drivers/vfio/vfio_iommu_type1.c
> +++ b/drivers/vfio/vfio_iommu_type1.c
> @@ -27,6 +27,7 @@
>  #include <linux/iommu.h>
>  #include <linux/module.h>
>  #include <linux/mm.h>
> +#include <linux/mmu_context.h>
>  #include <linux/rbtree.h>
>  #include <linux/sched/signal.h>
>  #include <linux/sched/mm.h>
> @@ -2326,6 +2327,85 @@ static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
>  	return blocking_notifier_chain_unregister(&iommu->notifier, nb);
>  }
>  
> +static int next_segment(unsigned long len, int offset)
> +{
> +	if (len > PAGE_SIZE - offset)
> +		return PAGE_SIZE - offset;
> +	else
> +		return len;
> +}
> +
> +static int vfio_iommu_type1_rw_iova_seg(struct vfio_iommu *iommu,
> +					  unsigned long iova, void *data,
> +					  unsigned long seg_len,
> +					  unsigned long offset,
> +					  bool write)
> +{
> +	struct mm_struct *mm;
> +	unsigned long vaddr;
> +	struct vfio_dma *dma;
> +	bool kthread = current->mm == NULL;
> +	int ret = 0;
> +
> +	dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
> +	if (!dma)
> +		return -EINVAL;
> +
> +	mm = get_task_mm(dma->task);
> +
> +	if (!mm)
> +		return -ENODEV;
> +
> +	if (kthread)
> +		use_mm(mm);
> +	else if (current->mm != mm) {
> +		ret = -EINVAL;
> +		goto out;
> +	}
> +
> +	vaddr = dma->vaddr + iova - dma->iova + offset;

Parenthesis here would be useful and might prevent overflow, ie:

  dma->vaddr + (iova - dma->iova) + offset

> +
> +	ret = write ? __copy_to_user((void __user *)vaddr,
> +			data, seg_len) :
> +		__copy_from_user(data, (void __user *)vaddr,
> +				seg_len);
> +	if (ret)
> +		ret = -EFAULT;
> +
> +	if (kthread)
> +		unuse_mm(mm);
> +out:
> +	mmput(mm);
> +	return ret;
> +}
> +
> +static int vfio_iommu_type1_iova_rw(void *iommu_data, unsigned long iova,
> +				    void *data, unsigned long len, bool write)
> +{
> +	struct vfio_iommu *iommu = iommu_data;
> +	int offset = iova & ~PAGE_MASK;
> +	int seg_len;
> +	int ret = 0;
> +
> +	iova = iova & PAGE_MASK;
> +
> +	mutex_lock(&iommu->lock);
> +	while ((seg_len = next_segment(len, offset)) > 0) {
> +		ret = vfio_iommu_type1_rw_iova_seg(iommu, iova, data,
> +						   seg_len, offset, write);

Why do we need to split operations at page boundaries?  It seems really
inefficient that at each page crossing we need to lookup the vfio_dma
again (probably the same one), switch to the mm (probably the same one),
and perform another copy_{to,from}_user() when potentially have
everything we need to perform a larger copy.  Thanks,

Alex

> +		if (ret)
> +			break;
> +
> +		offset = 0;
> +		len -= seg_len;
> +		data += seg_len;
> +		iova += PAGE_SIZE;
> +	}
> +
> +	mutex_unlock(&iommu->lock);
> +	return ret;
> +}
> +
>  static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
>  	.name			= "vfio-iommu-type1",
>  	.owner			= THIS_MODULE,
> @@ -2338,6 +2418,7 @@ static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
>  	.unpin_pages		= vfio_iommu_type1_unpin_pages,
>  	.register_notifier	= vfio_iommu_type1_register_notifier,
>  	.unregister_notifier	= vfio_iommu_type1_unregister_notifier,
> +	.iova_rw		= vfio_iommu_type1_iova_rw,
>  };
>  
>  static int __init vfio_iommu_type1_init(void)
> diff --git a/include/linux/vfio.h b/include/linux/vfio.h
> index e42a711a2800..7bf18a31bbcf 100644
> --- a/include/linux/vfio.h
> +++ b/include/linux/vfio.h
> @@ -82,6 +82,8 @@ struct vfio_iommu_driver_ops {
>  					     struct notifier_block *nb);
>  	int		(*unregister_notifier)(void *iommu_data,
>  					       struct notifier_block *nb);
> +	int		(*iova_rw)(void *iommu_data, unsigned long iova,
> +				   void *data, unsigned long len, bool write);
>  };
>  
>  extern int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops);
> @@ -107,6 +109,9 @@ extern int vfio_pin_pages(struct device *dev, unsigned long *user_pfn,
>  extern int vfio_unpin_pages(struct device *dev, unsigned long *user_pfn,
>  			    int npage);
>  
> +extern int vfio_iova_rw(struct device *dev, unsigned long iova, void *data,
> +			unsigned long len, bool write);
> +
>  /* each type has independent events */
>  enum vfio_notify_type {
>  	VFIO_IOMMU_NOTIFY = 0,

Re: [PATCH 1/2] vfio: introduce vfio_iova_rw to read/write a range of IOVAs

[Yan Zhao]

On Fri, Jan 10, 2020 at 02:16:36AM +0800, Alex Williamson wrote:
> On Thu,  2 Jan 2020 20:02:17 -0500
> Yan Zhao <yan.y.zhao@intel.com> wrote:
> 
> > vfio_iova_rw will read/write a range of userspace memory (starting form
> > device iova to iova + len -1) into a kenrel buffer without pinning the
> > userspace memory.
> > 
> > TODO: vfio needs to mark the iova dirty if vfio_iova_rw(write) is
> > called.
> > 
> > Cc: Kevin Tian <kevin.tian@intel.com>
> > Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
> > ---
> >  drivers/vfio/vfio.c             | 45 ++++++++++++++++++
> >  drivers/vfio/vfio_iommu_type1.c | 81 +++++++++++++++++++++++++++++++++
> >  include/linux/vfio.h            |  5 ++
> >  3 files changed, 131 insertions(+)
> > 
> > diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c
> > index c8482624ca34..36e91e647ed5 100644
> > --- a/drivers/vfio/vfio.c
> > +++ b/drivers/vfio/vfio.c
> > @@ -1961,6 +1961,51 @@ int vfio_unpin_pages(struct device *dev, unsigned long *user_pfn, int npage)
> >  }
> >  EXPORT_SYMBOL(vfio_unpin_pages);
> >  
> > +/*
> > + * Read/Write a range of userspace IOVAs for a device into/from a kernel
> > + * buffer without pinning the userspace memory
> > + * @dev [in]  : device
> > + * @iova [in] : base IOVA of a userspace buffer
> > + * @data [in] : pointer to kernel buffer
> > + * @len [in]  : kernel buffer length
> > + * @write     : indicate read or write
> > + * Return error on failure or 0 on success.
> > + */
> > +int vfio_iova_rw(struct device *dev, unsigned long iova, void *data,
> > +		   unsigned long len, bool write)
> 
> Shouldn't iova be a dma_addr_t and len be a size_t?  AIUI this function
> performs the equivalent behavior of the device itself performing a DMA.
> Hmm, should the interface be named vfio_dma_rw()?
>
ok. will rename the interface to vfio_dma_rw(). thanks :)

> > +{
> > +	struct vfio_container *container;
> > +	struct vfio_group *group;
> > +	struct vfio_iommu_driver *driver;
> > +	int ret = 0;
> > +
> > +	if (!dev || !data || len <= 0)
> > +		return -EINVAL;
> > +
> > +	group = vfio_group_get_from_dev(dev);
> > +	if (!group)
> > +		return -ENODEV;
> > +
> > +	ret = vfio_group_add_container_user(group);
> > +	if (ret)
> > +		goto out;
> > +
> > +	container = group->container;
> > +	driver = container->iommu_driver;
> > +
> > +	if (likely(driver && driver->ops->iova_rw))
> > +		ret = driver->ops->iova_rw(container->iommu_data,
> > +					   iova, data, len, write);
> > +	else
> > +		ret = -ENOTTY;
> > +
> > +	vfio_group_try_dissolve_container(group);
> > +out:
> > +	vfio_group_put(group);
> > +	return ret;
> > +}
> > +EXPORT_SYMBOL(vfio_iova_rw);
> > +
> >  static int vfio_register_iommu_notifier(struct vfio_group *group,
> >  					unsigned long *events,
> >  					struct notifier_block *nb)
> > diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> > index 2ada8e6cdb88..aee191077235 100644
> > --- a/drivers/vfio/vfio_iommu_type1.c
> > +++ b/drivers/vfio/vfio_iommu_type1.c
> > @@ -27,6 +27,7 @@
> >  #include <linux/iommu.h>
> >  #include <linux/module.h>
> >  #include <linux/mm.h>
> > +#include <linux/mmu_context.h>
> >  #include <linux/rbtree.h>
> >  #include <linux/sched/signal.h>
> >  #include <linux/sched/mm.h>
> > @@ -2326,6 +2327,85 @@ static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
> >  	return blocking_notifier_chain_unregister(&iommu->notifier, nb);
> >  }
> >  
> > +static int next_segment(unsigned long len, int offset)
> > +{
> > +	if (len > PAGE_SIZE - offset)
> > +		return PAGE_SIZE - offset;
> > +	else
> > +		return len;
> > +}
> > +
> > +static int vfio_iommu_type1_rw_iova_seg(struct vfio_iommu *iommu,
> > +					  unsigned long iova, void *data,
> > +					  unsigned long seg_len,
> > +					  unsigned long offset,
> > +					  bool write)
> > +{
> > +	struct mm_struct *mm;
> > +	unsigned long vaddr;
> > +	struct vfio_dma *dma;
> > +	bool kthread = current->mm == NULL;
> > +	int ret = 0;
> > +
> > +	dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
> > +	if (!dma)
> > +		return -EINVAL;
> > +
> > +	mm = get_task_mm(dma->task);
> > +
> > +	if (!mm)
> > +		return -ENODEV;
> > +
> > +	if (kthread)
> > +		use_mm(mm);
> > +	else if (current->mm != mm) {
> > +		ret = -EINVAL;
> > +		goto out;
> > +	}
> > +
> > +	vaddr = dma->vaddr + iova - dma->iova + offset;
> 
> Parenthesis here would be useful and might prevent overflow, ie:
> 
>   dma->vaddr + (iova - dma->iova) + offset
>
Yes, thanks for pointing it out!

> > +
> > +	ret = write ? __copy_to_user((void __user *)vaddr,
> > +			data, seg_len) :
> > +		__copy_from_user(data, (void __user *)vaddr,
> > +				seg_len);
> > +	if (ret)
> > +		ret = -EFAULT;
> > +
> > +	if (kthread)
> > +		unuse_mm(mm);
> > +out:
> > +	mmput(mm);
> > +	return ret;
> > +}
> > +
> > +static int vfio_iommu_type1_iova_rw(void *iommu_data, unsigned long iova,
> > +				    void *data, unsigned long len, bool write)
> > +{
> > +	struct vfio_iommu *iommu = iommu_data;
> > +	int offset = iova & ~PAGE_MASK;
> > +	int seg_len;
> > +	int ret = 0;
> > +
> > +	iova = iova & PAGE_MASK;
> > +
> > +	mutex_lock(&iommu->lock);
> > +	while ((seg_len = next_segment(len, offset)) > 0) {
> > +		ret = vfio_iommu_type1_rw_iova_seg(iommu, iova, data,
> > +						   seg_len, offset, write);
> 
> Why do we need to split operations at page boundaries?  It seems really
> inefficient that at each page crossing we need to lookup the vfio_dma
> again (probably the same one), switch to the mm (probably the same one),
> and perform another copy_{to,from}_user() when potentially have
> everything we need to perform a larger copy.  Thanks,
>
you are right. Maybe I can first search dma with size of 1 and then
check the size of the found dma to perform a larger copy.

Thanks
Yan

> 
> > +		if (ret)
> > +			break;
> > +
> > +		offset = 0;
> > +		len -= seg_len;
> > +		data += seg_len;
> > +		iova += PAGE_SIZE;
> > +	}
> > +
> > +	mutex_unlock(&iommu->lock);
> > +	return ret;
> > +}
> > +
> >  static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
> >  	.name			= "vfio-iommu-type1",
> >  	.owner			= THIS_MODULE,
> > @@ -2338,6 +2418,7 @@ static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
> >  	.unpin_pages		= vfio_iommu_type1_unpin_pages,
> >  	.register_notifier	= vfio_iommu_type1_register_notifier,
> >  	.unregister_notifier	= vfio_iommu_type1_unregister_notifier,
> > +	.iova_rw		= vfio_iommu_type1_iova_rw,
> >  };
> >  
> >  static int __init vfio_iommu_type1_init(void)
> > diff --git a/include/linux/vfio.h b/include/linux/vfio.h
> > index e42a711a2800..7bf18a31bbcf 100644
> > --- a/include/linux/vfio.h
> > +++ b/include/linux/vfio.h
> > @@ -82,6 +82,8 @@ struct vfio_iommu_driver_ops {
> >  					     struct notifier_block *nb);
> >  	int		(*unregister_notifier)(void *iommu_data,
> >  					       struct notifier_block *nb);
> > +	int		(*iova_rw)(void *iommu_data, unsigned long iova,
> > +				   void *data, unsigned long len, bool write);
> >  };
> >  
> >  extern int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops);
> > @@ -107,6 +109,9 @@ extern int vfio_pin_pages(struct device *dev, unsigned long *user_pfn,
> >  extern int vfio_unpin_pages(struct device *dev, unsigned long *user_pfn,
> >  			    int npage);
> >  
> > +extern int vfio_iova_rw(struct device *dev, unsigned long iova, void *data,
> > +			unsigned long len, bool write);
> > +
> >  /* each type has independent events */
> >  enum vfio_notify_type {
> >  	VFIO_IOMMU_NOTIFY = 0,
>