在 2021/6/3 上午1:18, Eugenio Perez Martin 写道:
On Wed, Jun 2, 2021 at 11:51 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
在 2021/5/20 上午12:28, Eugenio Pérez 写道:
Initial version of shadow virtqueue that actually forward buffers. The
exposed addresses are the qemu's virtual address, so devices with IOMMU
that does not allow full mapping of qemu's address space does not work
at the moment.
Also for simplicity it only supports modern devices, that expects vring
in little endian, with split ring and no event idx or indirect
descriptors.
It reuses the VirtQueue code for the device part. The driver part is
based on Linux's virtio_ring driver, but with stripped functionality
and optimizations so it's easier to review.
Later commits will solve some of these concerns.
It would be more more easier to review if you squash those
"enhancements" into this patch.
Ok, they will be in the same commit for the next version.
Code also need to map used ring (device part) as RW in, and only in,
vhost-net. To map (or call vhost_device_iotlb_miss) inconditionally
would print an error in case of vhost devices with its own mapping
(vdpa).
I think we should not depend on the IOTLB miss. Instead, we should
program the device IOTLB before starting the svq. Or is there anything
that prevent you from doing this?
Sorry for being unclear, that is what I meant in the message: No other
device than kernel vhost needs the map (as "sent iotlb miss ahead"),
so we must make it conditional. Doing it unconditionally would make
nothing but an error appear on the screen, but it is incorrect anyway.
Is it clearer this way?
So what I'm worrying is the following code:
int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write)
{
...
if (dev->shadow_vqs_enabled) {
/* Shadow virtqueue translations in its Virtual Address Space */
const VhostDMAMap *result;
const VhostDMAMap needle = {
.iova = iova,
};
result = vhost_iova_tree_find_taddr(&dev->iova_map, &needle);
...
}
I wonder the reason for doing that (sorry if I've asked this in the
previous version).
I think the correct way is to use map those in the device IOTLB before,
instead of using the miss.
Then we can have a unified code for vhost-vDPA and vhost-kernel.
To know if this call is needed, vhost_sw_live_migration_start_vq and
vhost_sw_live_migration_stop copy the test performed in
vhost_dev_start. Testing for the actual backend type could be cleaner,
or checking for non-NULL vhost_force_iommu, enable_custom_iommu, or
another vhostOp. We could extract this test in its own function too,
so its name could give a better hint. Just copy the vhost_dev_start
check at the moment.
Signed-off-by: Eugenio Pérez <eperezma@xxxxxxxxxx>
---
hw/virtio/vhost-shadow-virtqueue.c | 205 +++++++++++++++++++++++++++--
hw/virtio/vhost.c | 134 ++++++++++++++++++-
2 files changed, 325 insertions(+), 14 deletions(-)
diff --git a/hw/virtio/vhost-shadow-virtqueue.c b/hw/virtio/vhost-shadow-virtqueue.c
index ff50f12410..6d767fe248 100644
--- a/hw/virtio/vhost-shadow-virtqueue.c
+++ b/hw/virtio/vhost-shadow-virtqueue.c
@@ -9,6 +9,7 @@
#include "hw/virtio/vhost-shadow-virtqueue.h"
#include "hw/virtio/vhost.h"
+#include "hw/virtio/virtio-access.h"
#include "standard-headers/linux/vhost_types.h"
@@ -48,9 +49,93 @@ typedef struct VhostShadowVirtqueue {
/* Virtio device */
VirtIODevice *vdev;
+
+ /* Map for returning guest's descriptors */
+ VirtQueueElement **ring_id_maps;
+
+ /* Next head to expose to device */
+ uint16_t avail_idx_shadow;
+
+ /* Next free descriptor */
+ uint16_t free_head;
+
+ /* Last seen used idx */
+ uint16_t shadow_used_idx;
+
+ /* Next head to consume from device */
+ uint16_t used_idx;
} VhostShadowVirtqueue;
-/* Forward guest notifications */
+static void vhost_vring_write_descs(VhostShadowVirtqueue *svq,
+ const struct iovec *iovec,
+ size_t num, bool more_descs, bool write)
+{
+ uint16_t i = svq->free_head, last = svq->free_head;
+ unsigned n;
+ uint16_t flags = write ? cpu_to_le16(VRING_DESC_F_WRITE) : 0;
+ vring_desc_t *descs = svq->vring.desc;
+
+ if (num == 0) {
+ return;
+ }
+
+ for (n = 0; n < num; n++) {
+ if (more_descs || (n + 1 < num)) {
+ descs[i].flags = flags | cpu_to_le16(VRING_DESC_F_NEXT);
+ } else {
+ descs[i].flags = flags;
+ }
+ descs[i].addr = cpu_to_le64((hwaddr)iovec[n].iov_base);
+ descs[i].len = cpu_to_le32(iovec[n].iov_len);
+
+ last = i;
+ i = cpu_to_le16(descs[i].next);
+ }
+
+ svq->free_head = le16_to_cpu(descs[last].next);
+}
+
+static unsigned vhost_shadow_vq_add_split(VhostShadowVirtqueue *svq,
+ VirtQueueElement *elem)
+{
+ int head;
+ unsigned avail_idx;
+ vring_avail_t *avail = svq->vring.avail;
+
+ head = svq->free_head;
+
+ /* We need some descriptors here */
+ assert(elem->out_num || elem->in_num);
+
+ vhost_vring_write_descs(svq, elem->out_sg, elem->out_num,
+ elem->in_num > 0, false);
+ vhost_vring_write_descs(svq, elem->in_sg, elem->in_num, false, true);
+
+ /*
+ * Put entry in available array (but don't update avail->idx until they
+ * do sync).
+ */
+ avail_idx = svq->avail_idx_shadow & (svq->vring.num - 1);
+ avail->ring[avail_idx] = cpu_to_le16(head);
+ svq->avail_idx_shadow++;
+
+ /* Expose descriptors to device */
It's better to describe the detail order.
E.g "Update avail index after the descriptor is wrote"
Agree, I will replace it with your wording.
+ smp_wmb();
+ avail->idx = cpu_to_le16(svq->avail_idx_shadow);
+
+ return head;
+
+}
+
+static void vhost_shadow_vq_add(VhostShadowVirtqueue *svq,
+ VirtQueueElement *elem)
+{
+ unsigned qemu_head = vhost_shadow_vq_add_split(svq, elem);
+
+ svq->ring_id_maps[qemu_head] = elem;
+}
+
+/* Handle guest->device notifications */
static void vhost_handle_guest_kick(EventNotifier *n)
{
VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue,
@@ -60,7 +145,67 @@ static void vhost_handle_guest_kick(EventNotifier *n)
return;
}
- event_notifier_set(&svq->kick_notifier);
+ /* Make available as many buffers as possible */
+ do {
+ if (virtio_queue_get_notification(svq->vq)) {
+ /* No more notifications until process all available */
+ virtio_queue_set_notification(svq->vq, false);
+ }
+
+ while (true) {
+ VirtQueueElement *elem = virtqueue_pop(svq->vq, sizeof(*elem));
+ if (!elem) {
+ break;
+ }
+
+ vhost_shadow_vq_add(svq, elem);
+ event_notifier_set(&svq->kick_notifier);
+ }
+
+ virtio_queue_set_notification(svq->vq, true);
+ } while (!virtio_queue_empty(svq->vq));
+}
+
+static bool vhost_shadow_vq_more_used(VhostShadowVirtqueue *svq)
+{
+ if (svq->used_idx != svq->shadow_used_idx) {
+ return true;
+ }
+
+ /* Get used idx must not be reordered */
+ smp_rmb();
+ svq->shadow_used_idx = cpu_to_le16(svq->vring.used->idx);
+
+ return svq->used_idx != svq->shadow_used_idx;
+}
+
+static VirtQueueElement *vhost_shadow_vq_get_buf(VhostShadowVirtqueue *svq)
+{
+ vring_desc_t *descs = svq->vring.desc;
+ const vring_used_t *used = svq->vring.used;
+ vring_used_elem_t used_elem;
+ uint16_t last_used;
+
+ if (!vhost_shadow_vq_more_used(svq)) {
+ return NULL;
+ }
+
+ last_used = svq->used_idx & (svq->vring.num - 1);
+ used_elem.id = le32_to_cpu(used->ring[last_used].id);
+ used_elem.len = le32_to_cpu(used->ring[last_used].len);
+
+ if (unlikely(used_elem.id >= svq->vring.num)) {
+ error_report("Device %s says index %u is available", svq->vdev->name,
+ used_elem.id);
+ return NULL;
+ }
+
+ descs[used_elem.id].next = svq->free_head;
+ svq->free_head = used_elem.id;
+
+ svq->used_idx++;
+ svq->ring_id_maps[used_elem.id]->len = used_elem.len;
+ return g_steal_pointer(&svq->ring_id_maps[used_elem.id]);
}
/* Forward vhost notifications */
@@ -69,17 +214,33 @@ static void vhost_shadow_vq_handle_call_no_test(EventNotifier *n)
VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue,
call_notifier);
EventNotifier *masked_notifier;
+ VirtQueue *vq = svq->vq;
masked_notifier = svq->masked_notifier.n;
- if (!masked_notifier) {
- unsigned n = virtio_get_queue_index(svq->vq);
- virtio_queue_invalidate_signalled_used(svq->vdev, n);
- virtio_notify_irqfd(svq->vdev, svq->vq);
- } else if (!svq->masked_notifier.signaled) {
- svq->masked_notifier.signaled = true;
- event_notifier_set(svq->masked_notifier.n);
- }
+ /* Make as many buffers as possible used. */
+ do {
+ unsigned i = 0;
+
+ /* TODO: Use VRING_AVAIL_F_NO_INTERRUPT */
+ while (true) {
+ g_autofree VirtQueueElement *elem = vhost_shadow_vq_get_buf(svq);
+ if (!elem) {
+ break;
+ }
+
+ assert(i < svq->vring.num);
+ virtqueue_fill(vq, elem, elem->len, i++);
+ }
+
+ virtqueue_flush(vq, i);
+ if (!masked_notifier) {
+ virtio_notify_irqfd(svq->vdev, svq->vq);
Any reason that you don't use virtio_notify() here?
No reason but to make sure guest_notifier is used. I'm not sure if
this is an implementation detail though.
The difference is that virtio_notify() will go through the memory API
which will finally go to irqfd in this case.
I can test to switch to virtio_notify, what would be the advantage here?
Probably no.
+ } else if (!svq->masked_notifier.signaled) {
+ svq->masked_notifier.signaled = true;
+ event_notifier_set(svq->masked_notifier.n);
+ }
This is an example of the extra complexity if you do shadow virtqueue at
.> virtio level.
If you do everything at e.g vhost-vdpa, you don't need to care about
masked_notifer at all.
Correct me if I'm wrong, what you mean is to use the backend
vhost_set_vring_call to set the guest notifier (from SVQ point of
view), and then set it unconditionally. The function
vhost_virtqueue_mask would set the masked one by itself, no
modification is needed here.
Something like this, from the point of vhost, it doesn't even need to
know whether or not the notifier is masked or not. All it needs is to
write to the eventfd set via vq call.
Backend would still need a conditional checking if SVQ is enabled, so
it either sends call_fd to backend or to SVQ.
Yes.
The call to
virtqueue_fill, would still be needed if we don't want to duplicate
all the device virtio's logic in the vhost-vdpa backend.
Yes, you can make the buffer forwarding a common library then it could
be used other vhost backend in the future.
The point is to avoid touching vhost protocols to reduce the changeset
and have someting minimal for our requirements (vhost-vDPA mainly).
Another possibility would be to just store guest_notifier in SVQ, and
replace it with masked notifier and back. I think this is more aligned
with what you have in mind, but it still needs changes to
vhost_virtqueue_mask. Note that the boolean store
masked_notifier.signaled is just a (maybe premature) optimization to
skip the unneeded write syscall, but it could be omitted for brevity.
Or maybe a cleaner solution is to use io_uring for this write? :).
Looks like not what I meant :)
To clarify, it works like:
1) record the vq call fd1 during vhost_vdpa_set_vring_call
2) when svq is not enabled, set this fd1 to vhost-VDPA via
VHOST_SET_VRING_CALL
3) when svq is enabled, initialize and set fd2 to vhost-vDPA, poll and
handle guest kick via fd1 and rely fd1 to fd2
So we don't need to care much about the masking, in the svq codes, we
just stick to use the fd that is set via recent vhost_vdpa_set_vring_call().
That means, if virtqueue is masked, we're using mased_notifier actually,
but it's totally transparent to us.
So the idea is behave like a normal vhost-vDPA backend, and hide the
shadowing from the virtio codes.
Thanks
Thanks!
Thanks
_______________________________________________
Virtualization mailing list
Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
