net/virtio: rationalize queue flushing
[dpdk.git] / lib / librte_vhost / vhost_user.c
index 6bf5626..1dd1a61 100644 (file)
@@ -154,7 +154,22 @@ vhost_user_set_features(struct virtio_net *dev, uint64_t features)
                return -1;
        }
 
-       if ((dev->flags & VIRTIO_DEV_RUNNING) && dev->features != features) {
+       if (dev->flags & VIRTIO_DEV_RUNNING) {
+               if (dev->features == features)
+                       return 0;
+
+               /*
+                * Error out if master tries to change features while device is
+                * in running state. The exception being VHOST_F_LOG_ALL, which
+                * is enabled when the live-migration starts.
+                */
+               if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
+                       RTE_LOG(ERR, VHOST_CONFIG,
+                               "(%d) features changed while device is running.\n",
+                               dev->vid);
+                       return -1;
+               }
+
                if (dev->notify_ops->features_changed)
                        dev->notify_ops->features_changed(dev->vid, features);
        }
@@ -172,6 +187,25 @@ vhost_user_set_features(struct virtio_net *dev, uint64_t features)
                (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
                (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 
+       if (!(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
+               /*
+                * Remove all but first queue pair if MQ hasn't been
+                * negotiated. This is safe because the device is not
+                * running at this stage.
+                */
+               while (dev->nr_vring > 2) {
+                       struct vhost_virtqueue *vq;
+
+                       vq = dev->virtqueue[--dev->nr_vring];
+                       if (!vq)
+                               continue;
+
+                       dev->virtqueue[dev->nr_vring] = NULL;
+                       cleanup_vq(vq, 1);
+                       free_vq(vq);
+               }
+       }
+
        return 0;
 }
 
@@ -198,6 +232,7 @@ vhost_user_set_vring_num(struct virtio_net *dev,
                                "zero copy is force disabled\n");
                        dev->dequeue_zero_copy = 0;
                }
+               TAILQ_INIT(&vq->zmbuf_list);
        }
 
        vq->shadow_used_ring = rte_malloc(NULL,
@@ -232,6 +267,9 @@ numa_realloc(struct virtio_net *dev, int index)
        int oldnode, newnode;
        struct virtio_net *old_dev;
        struct vhost_virtqueue *old_vq, *vq;
+       struct zcopy_mbuf *new_zmbuf;
+       struct vring_used_elem *new_shadow_used_ring;
+       struct batch_copy_elem *new_batch_copy_elems;
        int ret;
 
        old_dev = dev;
@@ -256,6 +294,33 @@ numa_realloc(struct virtio_net *dev, int index)
                        return dev;
 
                memcpy(vq, old_vq, sizeof(*vq));
+               TAILQ_INIT(&vq->zmbuf_list);
+
+               new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
+                       sizeof(struct zcopy_mbuf), 0, newnode);
+               if (new_zmbuf) {
+                       rte_free(vq->zmbufs);
+                       vq->zmbufs = new_zmbuf;
+               }
+
+               new_shadow_used_ring = rte_malloc_socket(NULL,
+                       vq->size * sizeof(struct vring_used_elem),
+                       RTE_CACHE_LINE_SIZE,
+                       newnode);
+               if (new_shadow_used_ring) {
+                       rte_free(vq->shadow_used_ring);
+                       vq->shadow_used_ring = new_shadow_used_ring;
+               }
+
+               new_batch_copy_elems = rte_malloc_socket(NULL,
+                       vq->size * sizeof(struct batch_copy_elem),
+                       RTE_CACHE_LINE_SIZE,
+                       newnode);
+               if (new_batch_copy_elems) {
+                       rte_free(vq->batch_copy_elems);
+                       vq->batch_copy_elems = new_batch_copy_elems;
+               }
+
                rte_free(old_vq);
        }
 
@@ -544,6 +609,30 @@ dump_guest_pages(struct virtio_net *dev)
 #define dump_guest_pages(dev)
 #endif
 
+static bool
+vhost_memory_changed(struct VhostUserMemory *new,
+                    struct rte_vhost_memory *old)
+{
+       uint32_t i;
+
+       if (new->nregions != old->nregions)
+               return true;
+
+       for (i = 0; i < new->nregions; ++i) {
+               VhostUserMemoryRegion *new_r = &new->regions[i];
+               struct rte_vhost_mem_region *old_r = &old->regions[i];
+
+               if (new_r->guest_phys_addr != old_r->guest_phys_addr)
+                       return true;
+               if (new_r->memory_size != old_r->size)
+                       return true;
+               if (new_r->userspace_addr != old_r->guest_user_addr)
+                       return true;
+       }
+
+       return false;
+}
+
 static int
 vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
 {
@@ -556,6 +645,16 @@ vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
        uint32_t i;
        int fd;
 
+       if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
+               RTE_LOG(INFO, VHOST_CONFIG,
+                       "(%d) memory regions not changed\n", dev->vid);
+
+               for (i = 0; i < memory.nregions; i++)
+                       close(pmsg->fds[i]);
+
+               return 0;
+       }
+
        if (dev->mem) {
                free_mem_region(dev);
                rte_free(dev->mem);
@@ -1161,12 +1260,47 @@ vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev, VhostUserMsg *msg)
        return alloc_vring_queue(dev, vring_idx);
 }
 
+static void
+vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
+{
+       unsigned int i = 0;
+       unsigned int vq_num = 0;
+
+       while (vq_num < dev->nr_vring) {
+               struct vhost_virtqueue *vq = dev->virtqueue[i];
+
+               if (vq) {
+                       rte_spinlock_lock(&vq->access_lock);
+                       vq_num++;
+               }
+               i++;
+       }
+}
+
+static void
+vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
+{
+       unsigned int i = 0;
+       unsigned int vq_num = 0;
+
+       while (vq_num < dev->nr_vring) {
+               struct vhost_virtqueue *vq = dev->virtqueue[i];
+
+               if (vq) {
+                       rte_spinlock_unlock(&vq->access_lock);
+                       vq_num++;
+               }
+               i++;
+       }
+}
+
 int
 vhost_user_msg_handler(int vid, int fd)
 {
        struct virtio_net *dev;
        struct VhostUserMsg msg;
        int ret;
+       int unlock_required = 0;
 
        dev = get_device(vid);
        if (dev == NULL)
@@ -1212,6 +1346,38 @@ vhost_user_msg_handler(int vid, int fd)
                return -1;
        }
 
+       /*
+        * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE,
+        * since it is sent when virtio stops and device is destroyed.
+        * destroy_device waits for queues to be inactive, so it is safe.
+        * Otherwise taking the access_lock would cause a dead lock.
+        */
+       switch (msg.request.master) {
+       case VHOST_USER_SET_FEATURES:
+       case VHOST_USER_SET_PROTOCOL_FEATURES:
+       case VHOST_USER_SET_OWNER:
+       case VHOST_USER_RESET_OWNER:
+       case VHOST_USER_SET_MEM_TABLE:
+       case VHOST_USER_SET_LOG_BASE:
+       case VHOST_USER_SET_LOG_FD:
+       case VHOST_USER_SET_VRING_NUM:
+       case VHOST_USER_SET_VRING_ADDR:
+       case VHOST_USER_SET_VRING_BASE:
+       case VHOST_USER_SET_VRING_KICK:
+       case VHOST_USER_SET_VRING_CALL:
+       case VHOST_USER_SET_VRING_ERR:
+       case VHOST_USER_SET_VRING_ENABLE:
+       case VHOST_USER_SEND_RARP:
+       case VHOST_USER_NET_SET_MTU:
+       case VHOST_USER_SET_SLAVE_REQ_FD:
+               vhost_user_lock_all_queue_pairs(dev);
+               unlock_required = 1;
+               break;
+       default:
+               break;
+
+       }
+
        switch (msg.request.master) {
        case VHOST_USER_GET_FEATURES:
                msg.payload.u64 = vhost_user_get_features(dev);
@@ -1219,7 +1385,9 @@ vhost_user_msg_handler(int vid, int fd)
                send_vhost_reply(fd, &msg);
                break;
        case VHOST_USER_SET_FEATURES:
-               vhost_user_set_features(dev, msg.payload.u64);
+               ret = vhost_user_set_features(dev, msg.payload.u64);
+               if (ret)
+                       return -1;
                break;
 
        case VHOST_USER_GET_PROTOCOL_FEATURES:
@@ -1313,6 +1481,9 @@ vhost_user_msg_handler(int vid, int fd)
 
        }
 
+       if (unlock_required)
+               vhost_user_unlock_all_queue_pairs(dev);
+
        if (msg.flags & VHOST_USER_NEED_REPLY) {
                msg.payload.u64 = !!ret;
                msg.size = sizeof(msg.payload.u64);