On 9/4/19 6:06 AM, Jason Wang wrote:
>
> On 2019/9/3 下午4:49, Tiwei Bie wrote:
>> On Tue, Sep 03, 2019 at 09:40:25AM +0200, Maxime Coquelin wrote:
>>> On 9/3/19 7:30 AM, Tiwei Bie wrote:
>>>> On Thu, Aug 29, 2019 at 09:59:55AM +0200, Maxime Coquelin wrote:
>>>>> In order to support multi-queue, we need to implement the control
>>>>> path. The problem is that both the Vhost-user master and slave use
>>>>> VAs in their processes address spaces as IOVAs, which creates
>>>>> collusions between the data rings IOVAs managed the master, and
>>>>> the Control ring IOVAs. The trick here is to remmap the Control
>>>>> ring memory to another range, after the slave is aware of master's
>>>>> ranges.
>>>>>
>>>>> Signed-off-by: Maxime Coquelin <maxime.coque...@redhat.com>
>>>>> ---
>>>>> drivers/net/virtio/virtio_vdpa.c | 255
>>>>> +++++++++++++++++++++++++++++++
>>>>> 1 file changed, 255 insertions(+)
>>>>>
>>>>> diff --git a/drivers/net/virtio/virtio_vdpa.c
>>>>> b/drivers/net/virtio/virtio_vdpa.c
>>>>> index fc52a8e92..13b4dd07d 100644
>>>>> --- a/drivers/net/virtio/virtio_vdpa.c
>>>>> +++ b/drivers/net/virtio/virtio_vdpa.c
>>>>> @@ -106,6 +106,127 @@ find_internal_resource_by_dev(struct
>>>>> rte_pci_device *pdev)
>>>>> return list;
>>>>> }
>>>>> +static int
>>>>> +virtio_vdpa_dma_map_ctrl_queue(struct virtio_vdpa_device *dev, int
>>>>> do_map,
>>>>> + uint64_t iova)
>>>>> +{
>>>>> + const struct rte_memzone *mz;
>>>>> + int ret;
>>>>> +
>>>>> + /*
>>>>> + * IOVAs are processes VAs. We cannot use them as the Data and
>>>>> Control
>>>>> + * paths are run in different processes, which may (does) lead to
>>>>> + * collusions. The trick here is to fixup Ctrl path IOVAs so
>>>>> that they
>>>>> + * start after the Data path ranges.
>>>>> + */
>>>>> + if (do_map) {
>>>>> + mz = dev->cvq->cq.mz;
>>>>> + ret = rte_vfio_container_dma_map(dev->vfio_container_fd,
>>>>> + (uint64_t)(uintptr_t)mz->addr,
>>>>> + iova, mz->len);
>>>>> + if (ret < 0) {
>>>>> + DRV_LOG(ERR, "Failed to map ctrl ring (%d)", ret);
>>>>> + return ret;
>>>>> + }
>>>>> +
>>>>> + dev->cvq->vq_ring_mem = iova;
>>>>> + iova += mz->len;
>>>>> +
>>>>> + mz = dev->cvq->cq.virtio_net_hdr_mz;
>>>>> + ret = rte_vfio_container_dma_map(dev->vfio_container_fd,
>>>>> + (uint64_t)(uintptr_t)mz->addr,
>>>>> + iova, mz->len);
>>>>> + if (ret < 0) {
>>>>> + DRV_LOG(ERR, "Failed to map ctrl headers (%d)", ret);
>>>>> + return ret;
>>>>> + }
>>>> This will allow guest to access the cq.mz and cq.virtio_net_hdr_mz
>>>> via the device which may have potential risks.
>>> I get what you mean, but I'm not sure to see how we could avoid that.
>>> AFAIU, we need to map the control queue in the device IOMMU, otherwise
>>> how could the host (in case of virtual device) or the NIC (in case of
>>> Virtio offload), could access the ring?
>>> Any thoughts?
>> I also don't see a way to avoid that. That's why I said in below
>> thread that I think the control queue based interface seems not a
>> quite good interface for a backend device:
>>
>> https://lkml.org/lkml/2019/9/2/934
>>
>> In IFCVF NIC, we added a MMIO based interface to replace control
>> queue for the multiqueue setup in vDPA mode.
>>
>> Jason is proposing some changes to make virtio device suitable
>> for backend device. I'm not sure whether it's possible to cover
>> this case as well..
>
>
> A silly question, can we do dynamic mapping like what kernel driver did
> here?
Not silly at all, it is of course possible.
I will implement that in my v2.
Thanks!
Maxime
> Thanks
>
>
>>
>> Regards,
>> Tiwei
>>
>>> Thanks,
>>> Maxime
>>>> Regards,
>>>> Tiwei
>>>>
>>>>> +
>>>>> + dev->cvq->cq.virtio_net_hdr_mem = iova;
>>>>> + } else {
>>>>> + mz = dev->cvq->cq.mz;
>>>>> + ret = rte_vfio_container_dma_unmap(dev->vfio_container_fd,
>>>>> + (uint64_t)(uintptr_t)mz->addr,
>>>>> + iova, mz->len);
>>>>> + if (ret < 0) {
>>>>> + DRV_LOG(ERR, "Failed to unmap ctrl ring (%d)", ret);
>>>>> + return ret;
>>>>> + }
>>>>> +
>>>>> + dev->cvq->vq_ring_mem = 0;
>>>>> + iova += mz->len;
>>>>> +
>>>>> + mz = dev->cvq->cq.virtio_net_hdr_mz;
>>>>> + ret = rte_vfio_container_dma_unmap(dev->vfio_container_fd,
>>>>> + (uint64_t)(uintptr_t)mz->addr,
>>>>> + iova, mz->len);
>>>>> + if (ret < 0) {
>>>>> + DRV_LOG(ERR, "Failed to unmap ctrl headers (%d)", ret);
>>>>> + return ret;
>>>>> + }
>>>>> +
>>>>> + dev->cvq->cq.virtio_net_hdr_mem = 0;
>>>>> + }
>>>>> +
>>>>> + return 0;
>>>>> +}
>>>>> +
>>>>> +static int
>>>>> +virtio_vdpa_dma_map(struct virtio_vdpa_device *dev, int do_map)
>>>>> +{
>>>>> + uint32_t i;
>>>>> + int ret;
>>>>> + struct rte_vhost_memory *mem = NULL;
>>>>> + int vfio_container_fd;
>>>>> + uint64_t avail_iova = 0;
>>>>> +
>>>>> + ret = rte_vhost_get_mem_table(dev->vid, &mem);
>>>>> + if (ret < 0 || !mem) {
>>>>> + DRV_LOG(ERR, "failed to get VM memory layout.");
>>>>> + return ret;
>>>>> + }
>>>>> +
>>>>> + vfio_container_fd = dev->vfio_container_fd;
>>>>> +
>>>>> + for (i = 0; i < mem->nregions; i++) {
>>>>> + struct rte_vhost_mem_region *reg;
>>>>> +
>>>>> + reg = &mem->regions[i];
>>>>> + DRV_LOG(INFO, "%s, region %u: HVA 0x%" PRIx64 ", "
>>>>> + "GPA 0x%" PRIx64 ", size 0x%" PRIx64 ".",
>>>>> + do_map ? "DMA map" : "DMA unmap", i,
>>>>> + reg->host_user_addr, reg->guest_phys_addr, reg->size);
>>>>> +
>>>>> + if (reg->guest_phys_addr + reg->size > avail_iova)
>>>>> + avail_iova = reg->guest_phys_addr + reg->size;
>>>>> +
>>>>> + if (do_map) {
>>>>> + ret = rte_vfio_container_dma_map(vfio_container_fd,
>>>>> + reg->host_user_addr, reg->guest_phys_addr,
>>>>> + reg->size);
>>>>> + if (ret < 0) {
>>>>> + DRV_LOG(ERR, "DMA map failed.");
>>>>> + goto exit;
>>>>> + }
>>>>> + } else {
>>>>> + ret = rte_vfio_container_dma_unmap(vfio_container_fd,
>>>>> + reg->host_user_addr, reg->guest_phys_addr,
>>>>> + reg->size);
>>>>> + if (ret < 0) {
>>>>> + DRV_LOG(ERR, "DMA unmap failed.");
>>>>> + goto exit;
>>>>> + }
>>>>> + }
>>>>> + }
>>>>> +
>>>>> + if (dev->cvq)
>>>>> + ret = virtio_vdpa_dma_map_ctrl_queue(dev, do_map,
>>>>> avail_iova);
>>>>> +
>>>>> +exit:
>>>>> + free(mem);
>>>>> +
>>>>> + return ret;
>>>>> +}
>>>>> +
>>>>> static int
>>>>> virtio_vdpa_vfio_setup(struct virtio_vdpa_device *dev)
>>>>> {
>>>>> @@ -216,10 +337,144 @@ virtio_vdpa_get_protocol_features(int did
>>>>> __rte_unused, uint64_t *features)
>>>>> return 0;
>>>>> }
>>>>> +static uint64_t
>>>>> +hva_to_gpa(int vid, uint64_t hva)
>>>>> +{
>>>>> + struct rte_vhost_memory *mem = NULL;
>>>>> + struct rte_vhost_mem_region *reg;
>>>>> + uint32_t i;
>>>>> + uint64_t gpa = 0;
>>>>> +
>>>>> + if (rte_vhost_get_mem_table(vid, &mem) < 0)
>>>>> + goto exit;
>>>>> +
>>>>> + for (i = 0; i < mem->nregions; i++) {
>>>>> + reg = &mem->regions[i];
>>>>> +
>>>>> + if (hva >= reg->host_user_addr &&
>>>>> + hva < reg->host_user_addr + reg->size) {
>>>>> + gpa = hva - reg->host_user_addr + reg->guest_phys_addr;
>>>>> + break;
>>>>> + }
>>>>> + }
>>>>> +
>>>>> +exit:
>>>>> + if (mem)
>>>>> + free(mem);
>>>>> + return gpa;
>>>>> +}
>>>>> +
>>>>> +static int
>>>>> +virtio_vdpa_start(struct virtio_vdpa_device *dev)
>>>>> +{
>>>>> + struct virtio_hw *hw = &dev->hw;
>>>>> + int i, vid, nr_vring, ret;
>>>>> + struct rte_vhost_vring vr;
>>>>> + struct virtio_pmd_ctrl ctrl;
>>>>> + int dlen[1];
>>>>> +
>>>>> + vid = dev->vid;
>>>>> + nr_vring = rte_vhost_get_vring_num(vid);
>>>>> +
>>>>> + if (dev->vqs)
>>>>> + rte_free(dev->vqs);
>>>>> +
>>>>> + dev->vqs = rte_zmalloc("virtio_vdpa", sizeof(*dev->vqs) *
>>>>> nr_vring, 0);
>>>>> +
>>>>> + for (i = 0; i < nr_vring; i++) {
>>>>> + struct virtqueue *vq = &dev->vqs[i];
>>>>> +
>>>>> + rte_vhost_get_vhost_vring(vid, i, &vr);
>>>>> +
>>>>> + vq->vq_queue_index = i;
>>>>> + vq->vq_nentries = vr.size;
>>>>> + vq->vq_ring_mem = hva_to_gpa(vid,
>>>>> (uint64_t)(uintptr_t)vr.desc);
>>>>> + if (vq->vq_ring_mem == 0) {
>>>>> + DRV_LOG(ERR, "Fail to get GPA for descriptor ring.");
>>>>> + ret = -1;
>>>>> + goto out_free_vqs;
>>>>> + }
>>>>> +
>>>>> + ret = VTPCI_OPS(hw)->setup_queue(hw, vq);
>>>>> + if (ret) {
>>>>> + DRV_LOG(ERR, "Fail to setup queue.");
>>>>> + goto out_free_vqs;
>>>>> + }
>>>>> + }
>>>>> +
>>>>> + if (dev->cvq) {
>>>>> + ret = VTPCI_OPS(hw)->setup_queue(hw, dev->cvq);
>>>>> + if (ret) {
>>>>> + DRV_LOG(ERR, "Fail to setup ctrl queue.");
>>>>> + goto out_free_vqs;
>>>>> + }
>>>>> + }
>>>>> +
>>>>> + vtpci_set_status(hw, VIRTIO_CONFIG_STATUS_DRIVER_OK);
>>>>> +
>>>>> + if (!dev->cvq)
>>>>> + return 0;
>>>>> +
>>>>> + ctrl.hdr.class = VIRTIO_NET_CTRL_MQ;
>>>>> + ctrl.hdr.cmd = VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET;
>>>>> + memcpy(ctrl.data, &dev->max_queue_pairs, sizeof(uint16_t));
>>>>> +
>>>>> + dlen[0] = sizeof(uint16_t);
>>>>> +
>>>>> + ret = virtio_send_command(hw->cvq, &ctrl, dlen, 1);
>>>>> + if (ret) {
>>>>> + DRV_LOG(ERR, "Multiqueue configured but send command "
>>>>> + "failed, this is too late now...");
>>>>> + ret = -EINVAL;
>>>>> + goto out_free_vqs;
>>>>> + }
>>>>> +
>>>>> + return 0;
>>>>> +out_free_vqs:
>>>>> + rte_free(dev->vqs);
>>>>> +
>>>>> + return ret;
>>>>> +}
>>>>> +
>>>>> +static int
>>>>> +virtio_vdpa_dev_config(int vid)
>>>>> +{
>>>>> + int did, ret;
>>>>> + struct internal_list *list;
>>>>> + struct virtio_vdpa_device *dev;
>>>>> +
>>>>> + did = rte_vhost_get_vdpa_device_id(vid);
>>>>> + list = find_internal_resource_by_did(did);
>>>>> + if (list == NULL) {
>>>>> + DRV_LOG(ERR, "Invalid device id: %d", did);
>>>>> + return -1;
>>>>> + }
>>>>> +
>>>>> + dev = list->dev;
>>>>> + dev->vid = vid;
>>>>> +
>>>>> + rte_spinlock_lock(&dev->lock);
>>>>> +
>>>>> + ret = virtio_vdpa_dma_map(dev, 1);
>>>>> + if (ret)
>>>>> + goto out_unlock;
>>>>> +
>>>>> + ret = virtio_vdpa_start(dev);
>>>>> +
>>>>> + if (rte_vhost_host_notifier_ctrl(vid, true) != 0)
>>>>> + DRV_LOG(NOTICE, "vDPA (%d): software relay is used.", did);
>>>>> +
>>>>> +out_unlock:
>>>>> + rte_spinlock_unlock(&dev->lock);
>>>>> +
>>>>> + return ret;
>>>>> +}
>>>>> +
>>>>> static struct rte_vdpa_dev_ops virtio_vdpa_ops = {
>>>>> .get_queue_num = virtio_vdpa_get_queue_num,
>>>>> .get_features = virtio_vdpa_get_features,
>>>>> .get_protocol_features = virtio_vdpa_get_protocol_features,
>>>>> + .dev_conf = virtio_vdpa_dev_config,
>>>>> };
>>>>> static inline int
>>>>> --
>>>>> 2.21.0
>>>>>
