On 2020/5/30 上午2:30, Rob Miller wrote:
Given the need for 4K doorbell such that QEMU can easily map, ect, and
assuming that I have a HW device which exposes 2 VQ's, with a
notification area off of BAR3, offset=whatever, notifier_multiplier=4,
we don't need to have 2 x 4K pages mapped into the VM for both
doorbells do we? The guest driver would ring DB0 at BAR4+offset, and
DB1 at BAR4+offset+(4*1).
This requires qemu to advertise notifier_multiplier = 4 to guest, it has
several implications:
- guest may think control virtqueue sit in the same page then we can't
trap access to control virtqueue, qemu may lose the state of the device
- it requires the destination to have exact the same queue_notify_off
and notifier_multiplier in order to let migration to work
So for doorbell mapping itself, we can support the mapping of doorbells
at once which may occupy one or more pages, but it may bring troubles
for other features.
The 4K per DB is useful how? This allows for QEMU trapping of
individual DBs, that can then be used to do what, just forward the DBs
via some other scheme - this makes sense for non-HW related Virtio
devices I guess. Is this why there is a qemu option?
The page per vq option provides extra flexibility. Note that it's the
layout seen by guest, so actually for hardware vendor, the most easy way
is to:
- use a single doorbell register for all virtqueues except for the
control virtqueue (driver can differ queue index by the value wrote by
the driver)
- do not share the page with other registers
In this way, it doesn't require much BAR space and we can still:
- map doorbell separately to guest: guest may see a page per vq doorbell
layout but in fact all those pages are mapped into the same page
- trap the control virtqueue, (don't do mmap for control vq)
Thanks
Rob Miller
rob.miller@xxxxxxxxxxxx <mailto:rob.miller@xxxxxxxxxxxx>
(919)721-3339
On Fri, May 29, 2020 at 4:03 AM Jason Wang <jasowang@xxxxxxxxxx
<mailto:jasowang@xxxxxxxxxx>> wrote:
Currently the doorbell is relayed via eventfd which may have
significant overhead because of the cost of vmexits or syscall. This
patch introduces mmap() based doorbell mapping which can eliminate the
overhead caused by vmexit or syscall.
To ease the userspace modeling of the doorbell layout (usually
virtio-pci), this patch starts from a doorbell per page
model. Vhost-vdpa only support the hardware doorbell that sit at the
boundary of a page and does not share the page with other registers.
Doorbell of each virtqueue must be mapped separately, pgoff is the
index of the virtqueue. This allows userspace to map a subset of the
doorbell which may be useful for the implementation of software
assisted virtqueue (control vq) in the future.
Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx
<mailto:jasowang@xxxxxxxxxx>>
---
drivers/vhost/vdpa.c | 59
++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 59 insertions(+)
diff --git a/drivers/vhost/vdpa.c b/drivers/vhost/vdpa.c
index 6ff72289f488..bbe23cea139a 100644
--- a/drivers/vhost/vdpa.c
+++ b/drivers/vhost/vdpa.c
@@ -15,6 +15,7 @@
#include <linux/module.h>
#include <linux/cdev.h>
#include <linux/device.h>
+#include <linux/mm.h>
#include <linux/iommu.h>
#include <linux/uuid.h>
#include <linux/vdpa.h>
@@ -741,12 +742,70 @@ static int vhost_vdpa_release(struct inode
*inode, struct file *filep)
return 0;
}
+static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
+{
+ struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
+ struct vdpa_device *vdpa = v->vdpa;
+ const struct vdpa_config_ops *ops = vdpa->config;
+ struct vdpa_notification_area notify;
+ struct vm_area_struct *vma = vmf->vma;
+ u16 index = vma->vm_pgoff;
+
+ notify = ops->get_vq_notification(vdpa, index);
+
+ vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+ if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
+ notify.addr >> PAGE_SHIFT, PAGE_SIZE,
+ vma->vm_page_prot))
+ return VM_FAULT_SIGBUS;
+
+ return VM_FAULT_NOPAGE;
+}
+
+static const struct vm_operations_struct vhost_vdpa_vm_ops = {
+ .fault = vhost_vdpa_fault,
+};
+
+static int vhost_vdpa_mmap(struct file *file, struct
vm_area_struct *vma)
+{
+ struct vhost_vdpa *v = vma->vm_file->private_data;
+ struct vdpa_device *vdpa = v->vdpa;
+ const struct vdpa_config_ops *ops = vdpa->config;
+ struct vdpa_notification_area notify;
+ int index = vma->vm_pgoff;
+
+ if (vma->vm_end - vma->vm_start != PAGE_SIZE)
+ return -EINVAL;
+ if ((vma->vm_flags & VM_SHARED) == 0)
+ return -EINVAL;
+ if (vma->vm_flags & VM_READ)
+ return -EINVAL;
+ if (index > 65535)
+ return -EINVAL;
+ if (!ops->get_vq_notification)
+ return -ENOTSUPP;
+
+ /* To be safe and easily modelled by userspace, We only
+ * support the doorbell which sits on the page boundary and
+ * does not share the page with other registers.
+ */
+ notify = ops->get_vq_notification(vdpa, index);
+ if (notify.addr & (PAGE_SIZE - 1))
+ return -EINVAL;
+ if (vma->vm_end - vma->vm_start != notify.size)
+ return -ENOTSUPP;
+
+ vma->vm_ops = &vhost_vdpa_vm_ops;
+ return 0;
+}
+
static const struct file_operations vhost_vdpa_fops = {
.owner = THIS_MODULE,
.open = vhost_vdpa_open,
.release = vhost_vdpa_release,
.write_iter = vhost_vdpa_chr_write_iter,
.unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
+ .mmap = vhost_vdpa_mmap,
.compat_ioctl = compat_ptr_ioctl,
};
--
2.20.1
