[PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[zanghongyong]

From: Hongyong Zang <zanghongyong@huawei.com>

Vhost-net uses its own vhost_memory, which results from user space (qemu) info,
to translate GPA to HVA. Since kernel's kvm structure already maintains the 
address relationship in its member *kvm_memslots*, these patches use kernel's 
kvm_memslots directly without the need of initialization and maintenance of 
vhost_memory.

Hongyong Zang (2):
  kvm: Introduce get_kvm_from_task
  vhost-net: Use kvm_memslots for address translation

 drivers/vhost/vhost.c    |   53 +++++++++++++++++----------------------------
 include/linux/kvm_host.h |    2 +-
 virt/kvm/kvm_main.c      |   13 +++++++++++
 3 files changed, 34 insertions(+), 34 deletions(-)

[PATCH 1/2] kvm: Introduce get_kvm_from_task

[zanghongyong]

From: Hongyong Zang <zanghongyong@huawei.com>

This function finds the kvm structure from its corresponding user
space process, such as qemu process.

Signed-off-by: Hongyong Zang <zanghongyong@huawei.com>
---
 include/linux/kvm_host.h |    2 +-
 virt/kvm/kvm_main.c      |   13 +++++++++++++
 2 files changed, 14 insertions(+), 1 deletions(-)

diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 8c5c303..1b2f027 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -801,4 +801,4 @@ static inline bool kvm_check_request(int req, struct kvm_vcpu *vcpu)
 }
 
 #endif
-
+struct kvm *get_kvm_from_task(struct task_struct *task);
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index e289486..458fe29 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -2664,6 +2664,19 @@ static const struct file_operations *stat_fops[] = {
 	[KVM_STAT_VM]   = &vm_stat_fops,
 };
 
+struct kvm* get_kvm_from_task(struct task_struct *task)
+{
+        struct kvm *kvm;
+
+        list_for_each_entry(kvm, &vm_list, vm_list) {
+                if(kvm->mm == task->mm)
+                    return kvm;
+        }
+
+        return NULL;
+}
+EXPORT_SYMBOL_GPL(get_kvm_from_task);
+
 static void kvm_init_debug(void)
 {
 	struct kvm_stats_debugfs_item *p;
-- 
1.7.1

[PATCH 2/2] vhost-net: Use kvm_memslots for address translation

[zanghongyong]

From: Hongyong Zang <zanghongyong@huawei.com>

Use kvm's memslots instead of vhost_memory to traslate address
from GPA to HVA.

Signed-off-by: Hongyong Zang <zanghongyong@huawei.com>
---
 drivers/vhost/vhost.c |   53 ++++++++++++++++++------------------------------
 1 files changed, 20 insertions(+), 33 deletions(-)

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index c14c42b..63e4322 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -11,6 +11,7 @@
  * Generic code for virtio server in host kernel.
  */
 
+#include <linux/kvm_host.h>
 #include <linux/eventfd.h>
 #include <linux/vhost.h>
 #include <linux/virtio_net.h>
@@ -904,23 +905,6 @@ done:
 	return r;
 }
 
-static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
-						     __u64 addr, __u32 len)
-{
-	struct vhost_memory_region *reg;
-	int i;
-
-	/* linear search is not brilliant, but we really have on the order of 6
-	 * regions in practice */
-	for (i = 0; i < mem->nregions; ++i) {
-		reg = mem->regions + i;
-		if (reg->guest_phys_addr <= addr &&
-		    reg->guest_phys_addr + reg->memory_size - 1 >= addr)
-			return reg;
-	}
-	return NULL;
-}
-
 /* TODO: This is really inefficient.  We need something like get_user()
  * (instruction directly accesses the data, with an exception table entry
  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
@@ -1046,40 +1030,36 @@ int vhost_init_used(struct vhost_virtqueue *vq)
 	return get_user(vq->last_used_idx, &vq->used->idx);
 }
 
-static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
+static int translate_desc(struct kvm *kvm, u64 addr, u32 len,
 			  struct iovec iov[], int iov_size)
 {
-	const struct vhost_memory_region *reg;
-	struct vhost_memory *mem;
+	const struct kvm_memory_slot *slot;
+	gfn_t gfn = addr >> PAGE_SHIFT;
 	struct iovec *_iov;
 	u64 s = 0;
 	int ret = 0;
 
-	rcu_read_lock();
-
-	mem = rcu_dereference(dev->memory);
 	while ((u64)len > s) {
 		u64 size;
 		if (unlikely(ret >= iov_size)) {
 			ret = -ENOBUFS;
 			break;
 		}
-		reg = find_region(mem, addr, len);
-		if (unlikely(!reg)) {
+		slot = gfn_to_memslot(kvm, gfn);
+		if (unlikely(!slot)) {
 			ret = -EFAULT;
 			break;
 		}
 		_iov = iov + ret;
-		size = reg->memory_size - addr + reg->guest_phys_addr;
+		size = slot->npages*VHOST_PAGE_SIZE - addr + (slot->base_gfn<<PAGE_SHIFT);
 		_iov->iov_len = min((u64)len, size);
 		_iov->iov_base = (void __user *)(unsigned long)
-			(reg->userspace_addr + addr - reg->guest_phys_addr);
+			(slot->userspace_addr + addr - (slot->base_gfn<<PAGE_SHIFT));
 		s += size;
 		addr += size;
 		++ret;
 	}
 
-	rcu_read_unlock();
 	return ret;
 }
 
@@ -1104,7 +1084,7 @@ static unsigned next_desc(struct vring_desc *desc)
 	return next;
 }
 
-static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
+static int get_indirect(struct kvm *kvm, struct vhost_virtqueue *vq,
 			struct iovec iov[], unsigned int iov_size,
 			unsigned int *out_num, unsigned int *in_num,
 			struct vhost_log *log, unsigned int *log_num,
@@ -1123,7 +1103,7 @@ static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
 		return -EINVAL;
 	}
 
-	ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
+	ret = translate_desc(kvm, indirect->addr, indirect->len, vq->indirect,
 			     UIO_MAXIOV);
 	if (unlikely(ret < 0)) {
 		vq_err(vq, "Translation failure %d in indirect.\n", ret);
@@ -1163,7 +1143,7 @@ static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
 			return -EINVAL;
 		}
 
-		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
+		ret = translate_desc(kvm, desc.addr, desc.len, iov + iov_count,
 				     iov_size - iov_count);
 		if (unlikely(ret < 0)) {
 			vq_err(vq, "Translation failure %d indirect idx %d\n",
@@ -1209,6 +1189,13 @@ int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
 	unsigned int i, head, found = 0;
 	u16 last_avail_idx;
 	int ret;
+	struct kvm *kvm = get_kvm_from_task(current);
+
+	if(unlikely(kvm == NULL)){
+		vq_err(vq, "Failed to get corresponding kvm struct of vhost-%d\n",
+			current->pid);
+		return -EFAULT;
+	}
 
 	/* Check it isn't doing very strange things with descriptor numbers. */
 	last_avail_idx = vq->last_avail_idx;
@@ -1274,7 +1261,7 @@ int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
 			return -EFAULT;
 		}
 		if (desc.flags & VRING_DESC_F_INDIRECT) {
-			ret = get_indirect(dev, vq, iov, iov_size,
+			ret = get_indirect(kvm, vq, iov, iov_size,
 					   out_num, in_num,
 					   log, log_num, &desc);
 			if (unlikely(ret < 0)) {
@@ -1285,7 +1272,7 @@ int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
 			continue;
 		}
 
-		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
+		ret = translate_desc(kvm, desc.addr, desc.len, iov + iov_count,
 				     iov_size - iov_count);
 		if (unlikely(ret < 0)) {
 			vq_err(vq, "Translation failure %d descriptor idx %d\n",
-- 
1.7.1

Re: [PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[Takuya Yoshikawa]

(2011/12/16 14:32), zanghongyong@huawei.com wrote:
> From: Hongyong Zang<zanghongyong@huawei.com>
> 
> Vhost-net uses its own vhost_memory, which results from user space (qemu) info,
> to translate GPA to HVA. Since kernel's kvm structure already maintains the
> address relationship in its member *kvm_memslots*, these patches use kernel's
> kvm_memslots directly without the need of initialization and maintenance of
> vhost_memory.

Isn't vhost independent from KVM?

	Takuya

Re: [PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[Sasha Levin]

On Fri, 2011-12-16 at 13:32 +0800, zanghongyong@huawei.com wrote:
> From: Hongyong Zang <zanghongyong@huawei.com>
> 
> Vhost-net uses its own vhost_memory, which results from user space (qemu) info,
> to translate GPA to HVA. Since kernel's kvm structure already maintains the 
> address relationship in its member *kvm_memslots*, these patches use kernel's 
> kvm_memslots directly without the need of initialization and maintenance of 
> vhost_memory.

Conceptually, vhost isn't aware of KVM - it's just a driver which moves
data from vq to a tap device and back. You can't simply add KVM specific
code into vhost.

Whats the performance benefit?

-- 

Sasha.

Re: [PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[Zang Hongyong]

于 2011/12/16,星期五 15:05, Sasha Levin 写道:
> On Fri, 2011-12-16 at 13:32 +0800, zanghongyong@huawei.com wrote:
>> From: Hongyong Zang<zanghongyong@huawei.com>
>>
>> Vhost-net uses its own vhost_memory, which results from user space (qemu) info,
>> to translate GPA to HVA. Since kernel's kvm structure already maintains the
>> address relationship in its member *kvm_memslots*, these patches use kernel's
>> kvm_memslots directly without the need of initialization and maintenance of
>> vhost_memory.
> Conceptually, vhost isn't aware of KVM - it's just a driver which moves
> data from vq to a tap device and back. You can't simply add KVM specific
> code into vhost.
>
> Whats the performance benefit?
>
But vhost-net is only used in virtualization situation. vhost_memory is 
maintained
by user space qemu.
In this way, the memory relationship can be accquired from kernel 
without the
need of maintainence of vhost_memory from qemu.

Re: [PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[Sasha Levin]

On Fri, 2011-12-16 at 15:40 +0800, Zang Hongyong wrote:
> 于 2011/12/16,星期五 15:05, Sasha Levin 写道:
> > On Fri, 2011-12-16 at 13:32 +0800, zanghongyong@huawei.com wrote:
> >> From: Hongyong Zang<zanghongyong@huawei.com>
> >>
> >> Vhost-net uses its own vhost_memory, which results from user space (qemu) info,
> >> to translate GPA to HVA. Since kernel's kvm structure already maintains the
> >> address relationship in its member *kvm_memslots*, these patches use kernel's
> >> kvm_memslots directly without the need of initialization and maintenance of
> >> vhost_memory.
> > Conceptually, vhost isn't aware of KVM - it's just a driver which moves
> > data from vq to a tap device and back. You can't simply add KVM specific
> > code into vhost.
> >
> > Whats the performance benefit?
> >
> But vhost-net is only used in virtualization situation. vhost_memory is 
> maintained
> by user space qemu.
> In this way, the memory relationship can be accquired from kernel 
> without the
> need of maintainence of vhost_memory from qemu.

You can't assume that vhost-* is used only along with qemu/kvm. Just as
virtio has more uses than just virtualization (heres one:
https://lkml.org/lkml/2011/10/25/139 ), there are more uses for vhost as
well.

There has been a great deal of effort to keep vhost and kvm untangled.
One example is the memory translation it has to do, another one is the
eventfd/irqfd thing it does just so it could signal an IRQ in the guest
instead of accessing the guest directly.

If you do see a great performance increase when tying vhost and KVM
together, it may be worth it to create some sort of an in-kernel
vhost-kvm bridging thing, but if the performance isn't noticeable we're
better off just leaving it as is and keeping the vhost code general.

-- 

Sasha.

Re: [PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[Sasha Levin]

On Fri, 2011-12-16 at 09:59 +0200, Sasha Levin wrote:
> There has been a great deal of effort to keep vhost and kvm untangled.
> One example is the memory translation it has to do, another one is the
> eventfd/irqfd thing it does just so it could signal an IRQ in the guest
> instead of accessing the guest directly.

Actually, CONFIG_VHOST_NET doesn't even depend on CONFIG_KVM, so your
patch will break build when (CONFIG_VHOST_NET=y && CONFIG_KVM=n).

-- 

Sasha.

Re: [PATCH 0/2] vhot-net: Use kvm_memslots instead of vhost_memory to translate GPA to HVA

[Zang Hongyong]

于 2011/12/16,星期五 15:59, Sasha Levin 写道:
> On Fri, 2011-12-16 at 15:40 +0800, Zang Hongyong wrote:
>> 于 2011/12/16,星期五 15:05, Sasha Levin 写道:
>>> On Fri, 2011-12-16 at 13:32 +0800, zanghongyong@huawei.com wrote:
>>>> From: Hongyong Zang<zanghongyong@huawei.com>
>>>>
>>>> Vhost-net uses its own vhost_memory, which results from user space (qemu) info,
>>>> to translate GPA to HVA. Since kernel's kvm structure already maintains the
>>>> address relationship in its member *kvm_memslots*, these patches use kernel's
>>>> kvm_memslots directly without the need of initialization and maintenance of
>>>> vhost_memory.
>>> Conceptually, vhost isn't aware of KVM - it's just a driver which moves
>>> data from vq to a tap device and back. You can't simply add KVM specific
>>> code into vhost.
>>>
>>> Whats the performance benefit?
>>>
>> But vhost-net is only used in virtualization situation. vhost_memory is 
>> maintained
>> by user space qemu.
>> In this way, the memory relationship can be accquired from kernel 
>> without the
>> need of maintainence of vhost_memory from qemu.
> You can't assume that vhost-* is used only along with qemu/kvm. Just as
> virtio has more uses than just virtualization (heres one:
> https://lkml.org/lkml/2011/10/25/139 ), there are more uses for vhost as
> well.
>
> There has been a great deal of effort to keep vhost and kvm untangled.
> One example is the memory translation it has to do, another one is the
> eventfd/irqfd thing it does just so it could signal an IRQ in the guest
> instead of accessing the guest directly.
>
> If you do see a great performance increase when tying vhost and KVM
> together, it may be worth it to create some sort of an in-kernel
> vhost-kvm bridging thing, but if the performance isn't noticeable we're
> better off just leaving it as is and keeping the vhost code general.
>
Thanks for your explanation.
Since memory layout is seldom changed after guest boots, the situation
manily occurs during initialization. There's no need for vhost-kvm
bridge now.