net/ifc: support SW assisted VDPA live migration
authorXiao Wang <xiao.w.wang@intel.com>
Tue, 18 Dec 2018 08:02:06 +0000 (16:02 +0800)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 21 Dec 2018 15:22:40 +0000 (16:22 +0100)
In SW assisted live migration mode, driver will stop the device and
setup a mediated virtio ring to relay the communication between the
virtio driver and the VDPA device.

This data path intervention will allow SW to help on guest dirty page
logging for live migration.

This SW fallback is event driven relay thread, so when the network
throughput is low, this SW fallback will take little CPU resource, but
when the throughput goes up, the relay thread's CPU usage will goes up
accordingly.

User needs to take all the factors including CPU usage, guest perf
degradation, etc. into consideration when selecting the live migration
support mode.

Signed-off-by: Xiao Wang <xiao.w.wang@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
drivers/net/ifc/base/ifcvf.h
drivers/net/ifc/ifcvf_vdpa.c

index c15c691..9be2770 100644 (file)
@@ -50,6 +50,7 @@
 #define IFCVF_LM_ENABLE_VF             0x1
 #define IFCVF_LM_ENABLE_PF             0x3
 #define IFCVF_LOG_BASE                 0x100000000000
+#define IFCVF_MEDIATED_VRING           0x200000000000
 
 #define IFCVF_32_BIT_MASK              0xffffffff
 
index 4edba3e..972033b 100644 (file)
@@ -63,6 +63,9 @@ struct ifcvf_internal {
        rte_atomic32_t running;
        rte_spinlock_t lock;
        bool sw_lm;
+       bool sw_fallback_running;
+       /* mediated vring for sw fallback */
+       struct vring m_vring[IFCVF_MAX_QUEUES * 2];
 };
 
 struct internal_list {
@@ -308,6 +311,9 @@ vdpa_ifcvf_stop(struct ifcvf_internal *internal)
                rte_vhost_set_vring_base(vid, i, hw->vring[i].last_avail_idx,
                                hw->vring[i].last_used_idx);
 
+       if (internal->sw_lm)
+               return;
+
        rte_vhost_get_negotiated_features(vid, &features);
        if (RTE_VHOST_NEED_LOG(features)) {
                ifcvf_disable_logging(hw);
@@ -539,6 +545,318 @@ err:
        return ret;
 }
 
+static int
+m_ifcvf_start(struct ifcvf_internal *internal)
+{
+       struct ifcvf_hw *hw = &internal->hw;
+       uint32_t i, nr_vring;
+       int vid, ret;
+       struct rte_vhost_vring vq;
+       void *vring_buf;
+       uint64_t m_vring_iova = IFCVF_MEDIATED_VRING;
+       uint64_t size;
+       uint64_t gpa;
+
+       vid = internal->vid;
+       nr_vring = rte_vhost_get_vring_num(vid);
+       rte_vhost_get_negotiated_features(vid, &hw->req_features);
+
+       for (i = 0; i < nr_vring; i++) {
+               rte_vhost_get_vhost_vring(vid, i, &vq);
+
+               size = RTE_ALIGN_CEIL(vring_size(vq.size, PAGE_SIZE),
+                               PAGE_SIZE);
+               vring_buf = rte_zmalloc("ifcvf", size, PAGE_SIZE);
+               vring_init(&internal->m_vring[i], vq.size, vring_buf,
+                               PAGE_SIZE);
+
+               ret = rte_vfio_container_dma_map(internal->vfio_container_fd,
+                       (uint64_t)(uintptr_t)vring_buf, m_vring_iova, size);
+               if (ret < 0) {
+                       DRV_LOG(ERR, "mediated vring DMA map failed.");
+                       goto error;
+               }
+
+               gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.desc);
+               if (gpa == 0) {
+                       DRV_LOG(ERR, "Fail to get GPA for descriptor ring.");
+                       return -1;
+               }
+               hw->vring[i].desc = gpa;
+
+               hw->vring[i].avail = m_vring_iova +
+                       (char *)internal->m_vring[i].avail -
+                       (char *)internal->m_vring[i].desc;
+
+               hw->vring[i].used = m_vring_iova +
+                       (char *)internal->m_vring[i].used -
+                       (char *)internal->m_vring[i].desc;
+
+               hw->vring[i].size = vq.size;
+
+               rte_vhost_get_vring_base(vid, i, &hw->vring[i].last_avail_idx,
+                               &hw->vring[i].last_used_idx);
+
+               m_vring_iova += size;
+       }
+       hw->nr_vring = nr_vring;
+
+       return ifcvf_start_hw(&internal->hw);
+
+error:
+       for (i = 0; i < nr_vring; i++)
+               if (internal->m_vring[i].desc)
+                       rte_free(internal->m_vring[i].desc);
+
+       return -1;
+}
+
+static int
+m_ifcvf_stop(struct ifcvf_internal *internal)
+{
+       int vid;
+       uint32_t i;
+       struct rte_vhost_vring vq;
+       struct ifcvf_hw *hw = &internal->hw;
+       uint64_t m_vring_iova = IFCVF_MEDIATED_VRING;
+       uint64_t size, len;
+
+       vid = internal->vid;
+       ifcvf_stop_hw(hw);
+
+       for (i = 0; i < hw->nr_vring; i++) {
+               rte_vhost_get_vhost_vring(vid, i, &vq);
+               len = IFCVF_USED_RING_LEN(vq.size);
+               rte_vhost_log_used_vring(vid, i, 0, len);
+
+               size = RTE_ALIGN_CEIL(vring_size(vq.size, PAGE_SIZE),
+                               PAGE_SIZE);
+               rte_vfio_container_dma_unmap(internal->vfio_container_fd,
+                       (uint64_t)(uintptr_t)internal->m_vring[i].desc,
+                       m_vring_iova, size);
+
+               rte_vhost_set_vring_base(vid, i, hw->vring[i].last_avail_idx,
+                               hw->vring[i].last_used_idx);
+               rte_free(internal->m_vring[i].desc);
+               m_vring_iova += size;
+       }
+
+       return 0;
+}
+
+static int
+m_enable_vfio_intr(struct ifcvf_internal *internal)
+{
+       uint32_t nr_vring;
+       struct rte_intr_handle *intr_handle = &internal->pdev->intr_handle;
+       int ret;
+
+       nr_vring = rte_vhost_get_vring_num(internal->vid);
+
+       ret = rte_intr_efd_enable(intr_handle, nr_vring);
+       if (ret)
+               return -1;
+
+       ret = rte_intr_enable(intr_handle);
+       if (ret)
+               return -1;
+
+       return 0;
+}
+
+static void
+m_disable_vfio_intr(struct ifcvf_internal *internal)
+{
+       struct rte_intr_handle *intr_handle = &internal->pdev->intr_handle;
+
+       rte_intr_efd_disable(intr_handle);
+       rte_intr_disable(intr_handle);
+}
+
+static void
+update_avail_ring(struct ifcvf_internal *internal, uint16_t qid)
+{
+       rte_vdpa_relay_vring_avail(internal->vid, qid, &internal->m_vring[qid]);
+       ifcvf_notify_queue(&internal->hw, qid);
+}
+
+static void
+update_used_ring(struct ifcvf_internal *internal, uint16_t qid)
+{
+       rte_vdpa_relay_vring_used(internal->vid, qid, &internal->m_vring[qid]);
+       rte_vhost_vring_call(internal->vid, qid);
+}
+
+static void *
+vring_relay(void *arg)
+{
+       int i, vid, epfd, fd, nfds;
+       struct ifcvf_internal *internal = (struct ifcvf_internal *)arg;
+       struct rte_vhost_vring vring;
+       struct rte_intr_handle *intr_handle;
+       uint16_t qid, q_num;
+       struct epoll_event events[IFCVF_MAX_QUEUES * 4];
+       struct epoll_event ev;
+       int nbytes;
+       uint64_t buf;
+
+       vid = internal->vid;
+       q_num = rte_vhost_get_vring_num(vid);
+       /* prepare the mediated vring */
+       for (qid = 0; qid < q_num; qid++) {
+               rte_vhost_get_vring_base(vid, qid,
+                               &internal->m_vring[qid].avail->idx,
+                               &internal->m_vring[qid].used->idx);
+               rte_vdpa_relay_vring_avail(vid, qid, &internal->m_vring[qid]);
+       }
+
+       /* add notify fd and interrupt fd to epoll */
+       epfd = epoll_create(IFCVF_MAX_QUEUES * 2);
+       if (epfd < 0) {
+               DRV_LOG(ERR, "failed to create epoll instance.");
+               return NULL;
+       }
+       internal->epfd = epfd;
+
+       for (qid = 0; qid < q_num; qid++) {
+               ev.events = EPOLLIN | EPOLLPRI;
+               rte_vhost_get_vhost_vring(vid, qid, &vring);
+               ev.data.u64 = qid << 1 | (uint64_t)vring.kickfd << 32;
+               if (epoll_ctl(epfd, EPOLL_CTL_ADD, vring.kickfd, &ev) < 0) {
+                       DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
+                       return NULL;
+               }
+       }
+
+       intr_handle = &internal->pdev->intr_handle;
+       for (qid = 0; qid < q_num; qid++) {
+               ev.events = EPOLLIN | EPOLLPRI;
+               ev.data.u64 = 1 | qid << 1 |
+                       (uint64_t)intr_handle->efds[qid] << 32;
+               if (epoll_ctl(epfd, EPOLL_CTL_ADD, intr_handle->efds[qid], &ev)
+                               < 0) {
+                       DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
+                       return NULL;
+               }
+       }
+
+       /* start relay with a first kick */
+       for (qid = 0; qid < q_num; qid++)
+               ifcvf_notify_queue(&internal->hw, qid);
+
+       /* listen to the events and react accordingly */
+       for (;;) {
+               nfds = epoll_wait(epfd, events, q_num * 2, -1);
+               if (nfds < 0) {
+                       if (errno == EINTR)
+                               continue;
+                       DRV_LOG(ERR, "epoll_wait return fail\n");
+                       return NULL;
+               }
+
+               for (i = 0; i < nfds; i++) {
+                       fd = (uint32_t)(events[i].data.u64 >> 32);
+                       do {
+                               nbytes = read(fd, &buf, 8);
+                               if (nbytes < 0) {
+                                       if (errno == EINTR ||
+                                           errno == EWOULDBLOCK ||
+                                           errno == EAGAIN)
+                                               continue;
+                                       DRV_LOG(INFO, "Error reading "
+                                               "kickfd: %s",
+                                               strerror(errno));
+                               }
+                               break;
+                       } while (1);
+
+                       qid = events[i].data.u32 >> 1;
+
+                       if (events[i].data.u32 & 1)
+                               update_used_ring(internal, qid);
+                       else
+                               update_avail_ring(internal, qid);
+               }
+       }
+
+       return NULL;
+}
+
+static int
+setup_vring_relay(struct ifcvf_internal *internal)
+{
+       int ret;
+
+       ret = pthread_create(&internal->tid, NULL, vring_relay,
+                       (void *)internal);
+       if (ret) {
+               DRV_LOG(ERR, "failed to create ring relay pthread.");
+               return -1;
+       }
+       return 0;
+}
+
+static int
+unset_vring_relay(struct ifcvf_internal *internal)
+{
+       void *status;
+
+       if (internal->tid) {
+               pthread_cancel(internal->tid);
+               pthread_join(internal->tid, &status);
+       }
+       internal->tid = 0;
+
+       if (internal->epfd >= 0)
+               close(internal->epfd);
+       internal->epfd = -1;
+
+       return 0;
+}
+
+static int
+ifcvf_sw_fallback_switchover(struct ifcvf_internal *internal)
+{
+       int ret;
+
+       /* stop the direct IO data path */
+       unset_notify_relay(internal);
+       vdpa_ifcvf_stop(internal);
+       vdpa_disable_vfio_intr(internal);
+
+       ret = rte_vhost_host_notifier_ctrl(internal->vid, false);
+       if (ret && ret != -ENOTSUP)
+               goto error;
+
+       /* set up interrupt for interrupt relay */
+       ret = m_enable_vfio_intr(internal);
+       if (ret)
+               goto unmap;
+
+       /* config the VF */
+       ret = m_ifcvf_start(internal);
+       if (ret)
+               goto unset_intr;
+
+       /* set up vring relay thread */
+       ret = setup_vring_relay(internal);
+       if (ret)
+               goto stop_vf;
+
+       internal->sw_fallback_running = true;
+
+       return 0;
+
+stop_vf:
+       m_ifcvf_stop(internal);
+unset_intr:
+       m_disable_vfio_intr(internal);
+unmap:
+       ifcvf_dma_map(internal, 0);
+error:
+       return -1;
+}
+
 static int
 ifcvf_dev_config(int vid)
 {
@@ -579,8 +897,25 @@ ifcvf_dev_close(int vid)
        }
 
        internal = list->internal;
-       rte_atomic32_set(&internal->dev_attached, 0);
-       update_datapath(internal);
+
+       if (internal->sw_fallback_running) {
+               /* unset ring relay */
+               unset_vring_relay(internal);
+
+               /* reset VF */
+               m_ifcvf_stop(internal);
+
+               /* remove interrupt setting */
+               m_disable_vfio_intr(internal);
+
+               /* unset DMA map for guest memory */
+               ifcvf_dma_map(internal, 0);
+
+               internal->sw_fallback_running = false;
+       } else {
+               rte_atomic32_set(&internal->dev_attached, 0);
+               update_datapath(internal);
+       }
 
        return 0;
 }
@@ -604,7 +939,12 @@ ifcvf_set_features(int vid)
        internal = list->internal;
        rte_vhost_get_negotiated_features(vid, &features);
 
-       if (RTE_VHOST_NEED_LOG(features)) {
+       if (!RTE_VHOST_NEED_LOG(features))
+               return 0;
+
+       if (internal->sw_lm) {
+               ifcvf_sw_fallback_switchover(internal);
+       } else {
                rte_vhost_get_log_base(vid, &log_base, &log_size);
                rte_vfio_container_dma_map(internal->vfio_container_fd,
                                log_base, IFCVF_LOG_BASE, log_size);