Linux Virtualization - Apr 2018 - [RFC] vhost: introduce mdev based hardware vhost backend

On Tue, Apr 10, 2018 at 10:52:52AM +0800, Jason Wang
wrote:
> On 2018?04?02? 23:23, Tiwei Bie wrote:
> > This patch introduces a mdev (mediated device) based hardware
> > vhost backend. This backend is an abstraction of the various
> > hardware vhost accelerators (potentially any device that uses
> > virtio ring can be used as a vhost accelerator). Some generic
> > mdev parent ops are provided for accelerator drivers to support
> > generating mdev instances.
> > 
> > What's this
> > ==========> > 
> > The idea is that we can setup a virtio ring compatible device
> > with the messages available at the vhost-backend. Originally,
> > these messages are used to implement a software vhost backend,
> > but now we will use these messages to setup a virtio ring
> > compatible hardware device. Then the hardware device will be
> > able to work with the guest virtio driver in the VM just like
> > what the software backend does. That is to say, we can implement
> > a hardware based vhost backend in QEMU, and any virtio ring
> > compatible devices potentially can be used with this backend.
> > (We also call it vDPA -- vhost Data Path Acceleration).
> > 
> > One problem is that, different virtio ring compatible devices
> > may have different device interfaces. That is to say, we will
> > need different drivers in QEMU. It could be troublesome. And
> > that's what this patch trying to fix. The idea behind this
> > patch is very simple: mdev is a standard way to emulate device
> > in kernel.
> 
> So you just move the abstraction layer from qemu to kernel, and you still
> need different drivers in kernel for different device interfaces of
> accelerators. This looks even more complex than leaving it in qemu. As you
> said, another idea is to implement userspace vhost backend for accelerators
> which seems easier and could co-work with other parts of qemu without
> inventing new type of messages.
I'm not quite sure. Do you think it's acceptable to
add various vendor specific hardware drivers in QEMU?
> 
> Need careful thought here to seek a best solution here.
Yeah, definitely! :)
And your opinions would be very helpful!
> 
> >   So we defined a standard device based on mdev, which
> > is able to accept vhost messages. When the mdev emulation code
> > (i.e. the generic mdev parent ops provided by this patch) gets
> > vhost messages, it will parse and deliver them to accelerator
> > drivers. Drivers can use these messages to setup accelerators.
> > 
> > That is to say, the generic mdev parent ops (e.g. read()/write()/
> > ioctl()/...) will be provided for accelerator drivers to register
> > accelerators as mdev parent devices. And each accelerator device
> > will support generating standard mdev instance(s).
> > 
> > With this standard device interface, we will be able to just
> > develop one userspace driver to implement the hardware based
> > vhost backend in QEMU.
> > 
> > Difference between vDPA and PCI passthru
> > =======================================> > 
> > The key difference between vDPA and PCI passthru is that, in
> > vDPA only the data path of the device (e.g. DMA ring, notify
> > region and queue interrupt) is pass-throughed to the VM, the
> > device control path (e.g. PCI configuration space and MMIO
> > regions) is still defined and emulated by QEMU.
> > 
> > The benefits of keeping virtio device emulation in QEMU compared
> > with virtio device PCI passthru include (but not limit to):
> > 
> > - consistent device interface for guest OS in the VM;
> > - max flexibility on the hardware design, especially the
> >    accelerator for each vhost backend doesn't have to be a
> >    full PCI device;
> > - leveraging the existing virtio live-migration framework;
> > 
> > The interface of this mdev based device
> > ======================================> > 
> > 1. BAR0
> > 
> > The MMIO region described by BAR0 is the main control
> > interface. Messages will be written to or read from
> > this region.
> > 
> > The message type is determined by the `request` field
> > in message header. The message size is encoded in the
> > message header too. The message format looks like this:
> > 
> > struct vhost_vfio_op {
> > 	__u64 request;
> > 	__u32 flags;
> > 	/* Flag values: */
> > #define VHOST_VFIO_NEED_REPLY 0x1 /* Whether need reply */
> > 	__u32 size;
> > 	union {
> > 		__u64 u64;
> > 		struct vhost_vring_state state;
> > 		struct vhost_vring_addr addr;
> > 		struct vhost_memory memory;
> > 	} payload;
> > };
> > 
> > The existing vhost-kernel ioctl cmds are reused as
> > the message requests in above structure.
> > 
> > Each message will be written to or read from this
> > region at offset 0:
> > 
> > int vhost_vfio_write(struct vhost_dev *dev, struct vhost_vfio_op *op)
> > {
> > 	int count = VHOST_VFIO_OP_HDR_SIZE + op->size;
> > 	struct vhost_vfio *vfio = dev->opaque;
> > 	int ret;
> > 
> > 	ret = pwrite64(vfio->device_fd, op, count, vfio->bar0_offset);
> > 	if (ret != count)
> > 		return -1;
> > 
> > 	return 0;
> > }
> > 
> > int vhost_vfio_read(struct vhost_dev *dev, struct vhost_vfio_op *op)
> > {
> > 	int count = VHOST_VFIO_OP_HDR_SIZE + op->size;
> > 	struct vhost_vfio *vfio = dev->opaque;
> > 	uint64_t request = op->request;
> > 	int ret;
> > 
> > 	ret = pread64(vfio->device_fd, op, count, vfio->bar0_offset);
> > 	if (ret != count || request != op->request)
> > 		return -1;
> > 
> > 	return 0;
> > }
> > 
> > It's quite straightforward to set things to the device.
> > Just need to write the message to device directly:
> > 
> > int vhost_vfio_set_features(struct vhost_dev *dev, uint64_t features)
> > {
> > 	struct vhost_vfio_op op;
> > 
> > 	op.request = VHOST_SET_FEATURES;
> > 	op.flags = 0;
> > 	op.size = sizeof(features);
> > 	op.payload.u64 = features;
> > 
> > 	return vhost_vfio_write(dev, &op);
> > }
> > 
> > To get things from the device, two steps are needed.
> > Take VHOST_GET_FEATURE as an example:
> > 
> > int vhost_vfio_get_features(struct vhost_dev *dev, uint64_t *features)
> > {
> > 	struct vhost_vfio_op op;
> > 	int ret;
> > 
> > 	op.request = VHOST_GET_FEATURES;
> > 	op.flags = VHOST_VFIO_NEED_REPLY;
> > 	op.size = 0;
> > 
> > 	/* Just need to write the header */
> > 	ret = vhost_vfio_write(dev, &op);
> > 	if (ret != 0)
> > 		goto out;
> > 
> > 	/* `op` wasn't changed during write */
> > 	op.flags = 0;
> > 	op.size = sizeof(*features);
> > 
> > 	ret = vhost_vfio_read(dev, &op);
> > 	if (ret != 0)
> > 		goto out;
> > 
> > 	*features = op.payload.u64;
> > out:
> > 	return ret;
> > }
> > 
> > 2. BAR1 (mmap-able)
> > 
> > The MMIO region described by BAR1 will be used to notify the
> > device.
> > 
> > Each queue will has a page for notification, and it can be
> > mapped to VM (if hardware also supports), and the virtio
> > driver in the VM will be able to notify the device directly.
> > 
> > The MMIO region described by BAR1 is also write-able. If the
> > accelerator's notification register(s) cannot be mapped to the
> > VM, write() can also be used to notify the device. Something
> > like this:
> > 
> > void notify_relay(void *opaque)
> > {
> > 	......
> > 	offset = 0x1000 * queue_idx; /* XXX assume page size is 4K here. */
> > 
> > 	ret = pwrite64(vfio->device_fd, &queue_idx, sizeof(queue_idx),
> > 			vfio->bar1_offset + offset);
> > 	......
> > }
> > 
> > Other BARs are reserved.
> > 
> > 3. VFIO interrupt ioctl API
> > 
> > VFIO interrupt ioctl API is used to setup device interrupts.
> > IRQ-bypass will also be supported.
> > 
> > Currently, only VFIO_PCI_MSIX_IRQ_INDEX is supported.
> > 
> > The API for drivers to provide mdev instances
> > ============================================> > 
> > The read()/write()/ioctl()/mmap()/open()/release() mdev
> > parent ops have been provided for accelerators' drivers
> > to provide mdev instances.
> > 
> > ssize_t vdpa_read(struct mdev_device *mdev, char __user *buf,
> > 		  size_t count, loff_t *ppos);
> > ssize_t vdpa_write(struct mdev_device *mdev, const char __user *buf,
> > 		   size_t count, loff_t *ppos);
> > long vdpa_ioctl(struct mdev_device *mdev, unsigned int cmd, unsigned
long arg);
> > int vdpa_mmap(struct mdev_device *mdev, struct vm_area_struct *vma);
> > int vdpa_open(struct mdev_device *mdev);
> > void vdpa_close(struct mdev_device *mdev);
> > 
> > Each accelerator driver just needs to implement its own
> > create()/remove() ops, and provide a vdpa device ops
> > which will be called by the generic mdev emulation code.
> > 
> > Currently, the vdpa device ops are defined as:
> > 
> > typedef int (*vdpa_start_device_t)(struct vdpa_dev *vdpa);
> > typedef int (*vdpa_stop_device_t)(struct vdpa_dev *vdpa);
> > typedef int (*vdpa_dma_map_t)(struct vdpa_dev *vdpa);
> > typedef int (*vdpa_dma_unmap_t)(struct vdpa_dev *vdpa);
> > typedef int (*vdpa_set_eventfd_t)(struct vdpa_dev *vdpa, int vector,
int fd);
> > typedef u64 (*vdpa_supported_features_t)(struct vdpa_dev *vdpa);
> > typedef void (*vdpa_notify_device_t)(struct vdpa_dev *vdpa, int qid);
> > typedef u64 (*vdpa_get_notify_addr_t)(struct vdpa_dev *vdpa, int qid);
> > 
> > struct vdpa_device_ops {
> > 	vdpa_start_device_t		start;
> > 	vdpa_stop_device_t		stop;
> > 	vdpa_dma_map_t			dma_map;
> > 	vdpa_dma_unmap_t		dma_unmap;
> > 	vdpa_set_eventfd_t		set_eventfd;
> > 	vdpa_supported_features_t	supported_features;
> > 	vdpa_notify_device_t		notify;
> > 	vdpa_get_notify_addr_t		get_notify_addr;
> > };
> > 
> > struct vdpa_dev {
> > 	struct mdev_device *mdev;
> > 	struct mutex ops_lock;
> > 	u8 vconfig[VDPA_CONFIG_SIZE];
> > 	int nr_vring;
> > 	u64 features;
> > 	u64 state;
> > 	struct vhost_memory *mem_table;
> > 	bool pending_reply;
> > 	struct vhost_vfio_op pending;
> > 	const struct vdpa_device_ops *ops;
> > 	void *private;
> > 	int max_vrings;
> > 	struct vdpa_vring_info vring_info[0];
> > };
> > 
> > struct vdpa_dev *vdpa_alloc(struct mdev_device *mdev, void *private,
> > 			    int max_vrings);
> > void vdpa_free(struct vdpa_dev *vdpa);
> > 
> > A simple example
> > ===============> > 
> > # Query the number of available mdev instances
> > $ cat
/sys/class/mdev_bus/0000:06:00.2/mdev_supported_types/ifcvf_vdpa-vdpa_virtio/available_instances
> > 
> > # Create a mdev instance
> > $ echo $UUID >
/sys/class/mdev_bus/0000:06:00.2/mdev_supported_types/ifcvf_vdpa-vdpa_virtio/create
> > 
> > # Launch QEMU with a virtio-net device
> > $ qemu \
> > 	...... \
> > 	-netdev type=vhost-vfio,sysfsdev=/sys/bus/mdev/devices/$UUID,id=$ID \
> > 	-device virtio-net-pci,netdev=$ID
> > 
> > -------- END --------
> > 
> > Most of above words will be refined and moved to a doc in
> > the formal patch. In this RFC, all introductions and code
> > are gathered in this patch, the idea is to make it easier
> > to find all the relevant information. Anyone who wants to
> > comment could use inline comment and just keep the relevant
> > parts. Sorry for the big RFC patch..
> > 
> > This patch is just a RFC for now, and something is still
> > missing or needs to be refined. But it's never too early
> > to hear the thoughts from the community. So any comments
> > would be appreciated! Thanks! :-)
> 
> I don't see vhost_vfio_write() and other above functions in the patch.
Looks
> like some part of the patch is missed, it would be better to post a
complete
> series with an example driver (vDPA) to get a full picture.
No problem. We will send out the QEMU changes soon!

Thanks!
> 
> Thanks
> 
[...]

On 2018?04?10? 12:57, Tiwei Bie wrote:
> On Tue, Apr 10, 2018 at 10:52:52AM +0800, Jason Wang wrote:
>> On 2018?04?02? 23:23, Tiwei Bie wrote:
>>> This patch introduces a mdev (mediated device) based hardware
>>> vhost backend. This backend is an abstraction of the various
>>> hardware vhost accelerators (potentially any device that uses
>>> virtio ring can be used as a vhost accelerator). Some generic
>>> mdev parent ops are provided for accelerator drivers to support
>>> generating mdev instances.
>>>
>>> What's this
>>> ==========>>>
>>> The idea is that we can setup a virtio ring compatible device
>>> with the messages available at the vhost-backend. Originally,
>>> these messages are used to implement a software vhost backend,
>>> but now we will use these messages to setup a virtio ring
>>> compatible hardware device. Then the hardware device will be
>>> able to work with the guest virtio driver in the VM just like
>>> what the software backend does. That is to say, we can implement
>>> a hardware based vhost backend in QEMU, and any virtio ring
>>> compatible devices potentially can be used with this backend.
>>> (We also call it vDPA -- vhost Data Path Acceleration).
>>>
>>> One problem is that, different virtio ring compatible devices
>>> may have different device interfaces. That is to say, we will
>>> need different drivers in QEMU. It could be troublesome. And
>>> that's what this patch trying to fix. The idea behind this
>>> patch is very simple: mdev is a standard way to emulate device
>>> in kernel.
>> So you just move the abstraction layer from qemu to kernel, and you
still
>> need different drivers in kernel for different device interfaces of
>> accelerators. This looks even more complex than leaving it in qemu. As
you
>> said, another idea is to implement userspace vhost backend for
accelerators
>> which seems easier and could co-work with other parts of qemu without
>> inventing new type of messages.
> I'm not quite sure. Do you think it's acceptable to
> add various vendor specific hardware drivers in QEMU?
>
I don't object but we need to figure out the advantages of doing it in 
qemu too.

Thanks

On 10/04/2018 06:57, Tiwei Bie wrote:
>> So you just move the abstraction layer from qemu to kernel, and you
still
>> need different drivers in kernel for different device interfaces of
>> accelerators. This looks even more complex than leaving it in qemu. As
you
>> said, another idea is to implement userspace vhost backend for
accelerators
>> which seems easier and could co-work with other parts of qemu without
>> inventing new type of messages.
> 
> I'm not quite sure. Do you think it's acceptable to
> add various vendor specific hardware drivers in QEMU?
I think so.  We have vendor-specific quirks, and at some point there was
an idea of using quirks to implement (vendor-specific) live migration
support for assigned devices.

Paolo

> -----Original Message-----
> From: Paolo Bonzini [mailto:pbonzini at redhat.com]
> Sent: Tuesday, April 10, 2018 3:52 PM
> To: Bie, Tiwei <tiwei.bie at intel.com>; Jason Wang <jasowang at
redhat.com>
> Cc: mst at redhat.com; alex.williamson at redhat.com; ddutile at
redhat.com;
> Duyck, Alexander H <alexander.h.duyck at intel.com>; virtio-dev at
lists.oasis-
> open.org; linux-kernel at vger.kernel.org; kvm at vger.kernel.org;
> virtualization at lists.linux-foundation.org; netdev at vger.kernel.org;
Daly, Dan
> <dan.daly at intel.com>; Liang, Cunming <cunming.liang at
intel.com>; Wang,
> Zhihong <zhihong.wang at intel.com>; Tan, Jianfeng <jianfeng.tan
at intel.com>;
> Wang, Xiao W <xiao.w.wang at intel.com>
> Subject: Re: [virtio-dev] Re: [RFC] vhost: introduce mdev based hardware
> vhost backend
> 
> On 10/04/2018 06:57, Tiwei Bie wrote:
> >> So you just move the abstraction layer from qemu to kernel, and
you
> >> still need different drivers in kernel for different device
> >> interfaces of accelerators. This looks even more complex than
leaving
> >> it in qemu. As you said, another idea is to implement userspace
vhost
> >> backend for accelerators which seems easier and could co-work with
> >> other parts of qemu without inventing new type of messages.
> >
> > I'm not quite sure. Do you think it's acceptable to add
various vendor
> > specific hardware drivers in QEMU?
> 
> I think so.  We have vendor-specific quirks, and at some point there was an
> idea of using quirks to implement (vendor-specific) live migration support
for
> assigned devices.
Vendor-specific quirks of accessing VGA is a small portion. Other major portions
are still handled by guest driver.

While in this case, when saying various vendor specific drivers in QEMU, it says
QEMU takes over and provides the entire user space device drivers. Some parts
are even not relevant to vhost, they're basic device function enabling.
Moreover, it could be different kinds of devices(network/block/...) under vhost.
No matter # of vendors or # of types, total LOC is not small.

The idea is to avoid introducing these extra complexity out of QEMU, keeping
vhost adapter simple. As vhost protocol is de factor standard, it leverages
kernel device driver to provide the diversity. Changing once in QEMU, then it
supports multi-vendor devices whose drivers naturally providing kernel driver
there.

If QEMU is going to build a user space driver framework there, we're open
mind on that, even leveraging DPDK as the underlay library. Looking forward to
more others' comments from community.

Steve
> 
> Paolo

On Tue, Apr 10, 2018 at 03:25:45PM +0800, Jason Wang
wrote:
> > > > One problem is that, different virtio ring compatible
devices
> > > > may have different device interfaces. That is to say, we
will
> > > > need different drivers in QEMU. It could be troublesome. And
> > > > that's what this patch trying to fix. The idea behind
this
> > > > patch is very simple: mdev is a standard way to emulate
device
> > > > in kernel.
> > > So you just move the abstraction layer from qemu to kernel, and
you still
> > > need different drivers in kernel for different device interfaces
of
> > > accelerators. This looks even more complex than leaving it in
qemu. As you
> > > said, another idea is to implement userspace vhost backend for
accelerators
> > > which seems easier and could co-work with other parts of qemu
without
> > > inventing new type of messages.
> > I'm not quite sure. Do you think it's acceptable to
> > add various vendor specific hardware drivers in QEMU?
> > 
> 
> I don't object but we need to figure out the advantages of doing it in
qemu
> too.
> 
> Thanks
To be frank kernel is exactly where device drivers belong.  DPDK did
move them to userspace but that's merely a requirement for data path.
*If* you can have them in kernel that is best:
- update kernel and there's no need to rebuild userspace
- apps can be written in any language no need to maintain multiple
  libraries or add wrappers
- security concerns are much smaller (ok people are trying to
  raise the bar with IOMMUs and such, but it's already pretty
  good even without)

The biggest issue is that you let userspace poke at the
device which is also allowed by the IOMMU to poke at
kernel memory (needed for kernel driver to work).

Yes, maybe if device is not buggy it's all fine, but
it's better if we do not have to trust the device
otherwise the security picture becomes more murky.

I suggested attaching a PASID to (some) queues - see my old post "using
PASIDs to enable a safe variant of direct ring access".

Then using IOMMU with VFIO to limit access through queue to corrent
ranges of memory.


-- 
MST