vhost: replace SMP with thread fence for control path
[dpdk.git] / lib / librte_vhost / vhost.h
index a29c663..23e11ff 100644 (file)
 
 #define MAX_PKT_BURST 32
 
-#define ASYNC_MAX_POLL_SEG 255
-
 #define VHOST_MAX_ASYNC_IT (MAX_PKT_BURST * 2)
-#define VHOST_MAX_ASYNC_VEC (BUF_VECTOR_MAX * 2)
+#define VHOST_MAX_ASYNC_VEC (BUF_VECTOR_MAX * 4)
 
 #define PACKED_DESC_ENQUEUE_USED_FLAG(w)       \
        ((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED | VRING_DESC_F_WRITE) : \
@@ -94,20 +92,6 @@ struct buf_vector {
        uint32_t desc_idx;
 };
 
-/*
- * A structure to hold some fields needed in zero copy code path,
- * mainly for associating an mbuf with the right desc_idx.
- */
-struct zcopy_mbuf {
-       struct rte_mbuf *mbuf;
-       uint32_t desc_idx;
-       uint16_t desc_count;
-       uint16_t in_use;
-
-       TAILQ_ENTRY(zcopy_mbuf) next;
-};
-TAILQ_HEAD(zcopy_mbuf_list, zcopy_mbuf);
-
 /*
  * Structure contains the info for each batched memory copy.
  */
@@ -185,12 +169,6 @@ struct vhost_virtqueue {
        struct rte_vhost_resubmit_info *resubmit_inflight;
        uint64_t                global_counter;
 
-       uint16_t                nr_zmbuf;
-       uint16_t                zmbuf_size;
-       uint16_t                last_zmbuf_idx;
-       struct zcopy_mbuf       *zmbufs;
-       struct zcopy_mbuf_list  zmbuf_list;
-
        union {
                struct vring_used_elem  *shadow_used_split;
                struct vring_used_elem_packed *shadow_used_packed;
@@ -220,17 +198,15 @@ struct vhost_virtqueue {
        /* operation callbacks for async dma */
        struct rte_vhost_async_channel_ops      async_ops;
 
-       struct rte_vhost_iov_iter it_pool[VHOST_MAX_ASYNC_IT];
-       struct iovec vec_pool[VHOST_MAX_ASYNC_VEC];
+       struct rte_vhost_iov_iter *it_pool;
+       struct iovec *vec_pool;
 
        /* async data transfer status */
        uintptr_t       **async_pkts_pending;
-       #define         ASYNC_PENDING_INFO_N_MSK 0xFFFF
-       #define         ASYNC_PENDING_INFO_N_SFT 16
-       uint64_t        *async_pending_info;
+       struct async_inflight_info *async_pkts_info;
        uint16_t        async_pkts_idx;
        uint16_t        async_pkts_inflight_n;
-       uint16_t        async_last_seg_n;
+       uint16_t        async_last_pkts_n;
 
        /* vq async features */
        bool            async_inorder;
@@ -239,6 +215,7 @@ struct vhost_virtqueue {
 } __rte_cache_aligned;
 
 /* Virtio device status as per Virtio specification */
+#define VIRTIO_DEVICE_STATUS_RESET             0x00
 #define VIRTIO_DEVICE_STATUS_ACK               0x01
 #define VIRTIO_DEVICE_STATUS_DRIVER            0x02
 #define VIRTIO_DEVICE_STATUS_DRIVER_OK         0x04
@@ -379,7 +356,6 @@ struct virtio_net {
        /* to tell if we need broadcast rarp packet */
        int16_t                 broadcast_rarp;
        uint32_t                nr_vring;
-       int                     dequeue_zero_copy;
        int                     async_copy;
        int                     extbuf;
        int                     linearbuf;
@@ -587,38 +563,67 @@ static __rte_always_inline int guest_page_addrcmp(const void *p1,
        return 0;
 }
 
-/* Convert guest physical address to host physical address */
 static __rte_always_inline rte_iova_t
-gpa_to_hpa(struct virtio_net *dev, uint64_t gpa, uint64_t size)
+gpa_to_first_hpa(struct virtio_net *dev, uint64_t gpa,
+       uint64_t gpa_size, uint64_t *hpa_size)
 {
        uint32_t i;
        struct guest_page *page;
        struct guest_page key;
 
+       *hpa_size = gpa_size;
        if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
-               key.guest_phys_addr = gpa;
+               key.guest_phys_addr = gpa & ~(dev->guest_pages[0].size - 1);
                page = bsearch(&key, dev->guest_pages, dev->nr_guest_pages,
                               sizeof(struct guest_page), guest_page_addrcmp);
                if (page) {
-                       if (gpa + size < page->guest_phys_addr + page->size)
+                       if (gpa + gpa_size <=
+                                       page->guest_phys_addr + page->size) {
+                               return gpa - page->guest_phys_addr +
+                                       page->host_phys_addr;
+                       } else if (gpa < page->guest_phys_addr +
+                                               page->size) {
+                               *hpa_size = page->guest_phys_addr +
+                                       page->size - gpa;
                                return gpa - page->guest_phys_addr +
                                        page->host_phys_addr;
+                       }
                }
        } else {
                for (i = 0; i < dev->nr_guest_pages; i++) {
                        page = &dev->guest_pages[i];
 
-                       if (gpa >= page->guest_phys_addr &&
-                           gpa + size < page->guest_phys_addr +
-                           page->size)
-                               return gpa - page->guest_phys_addr +
-                                      page->host_phys_addr;
+                       if (gpa >= page->guest_phys_addr) {
+                               if (gpa + gpa_size <=
+                                       page->guest_phys_addr + page->size) {
+                                       return gpa - page->guest_phys_addr +
+                                               page->host_phys_addr;
+                               } else if (gpa < page->guest_phys_addr +
+                                                       page->size) {
+                                       *hpa_size = page->guest_phys_addr +
+                                               page->size - gpa;
+                                       return gpa - page->guest_phys_addr +
+                                               page->host_phys_addr;
+                               }
+                       }
                }
        }
 
+       *hpa_size = 0;
        return 0;
 }
 
+/* Convert guest physical address to host physical address */
+static __rte_always_inline rte_iova_t
+gpa_to_hpa(struct virtio_net *dev, uint64_t gpa, uint64_t size)
+{
+       rte_iova_t hpa;
+       uint64_t hpa_size;
+
+       hpa = gpa_to_first_hpa(dev, gpa, size, &hpa_size);
+       return hpa_size == size ? hpa : 0;
+}
+
 static __rte_always_inline uint64_t
 hva_to_gpa(struct virtio_net *dev, uint64_t vva, uint64_t len)
 {
@@ -667,7 +672,6 @@ int alloc_vring_queue(struct virtio_net *dev, uint32_t vring_idx);
 void vhost_attach_vdpa_device(int vid, struct rte_vdpa_device *dev);
 
 void vhost_set_ifname(int, const char *if_name, unsigned int if_len);
-void vhost_enable_dequeue_zero_copy(int vid);
 void vhost_set_builtin_virtio_net(int vid, bool enable);
 void vhost_enable_extbuf(int vid);
 void vhost_enable_linearbuf(int vid);
@@ -724,7 +728,7 @@ static __rte_always_inline void
 vhost_vring_call_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 {
        /* Flush used->idx update before we read avail->flags. */
-       rte_smp_mb();
+       rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 
        /* Don't kick guest if we don't reach index specified by guest. */
        if (dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX)) {
@@ -766,7 +770,7 @@ vhost_vring_call_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
        bool signalled_used_valid, kick = false;
 
        /* Flush used desc update. */
-       rte_smp_mb();
+       rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
 
        if (!(dev->features & (1ULL << VIRTIO_RING_F_EVENT_IDX))) {
                if (vq->driver_event->flags !=
@@ -792,7 +796,7 @@ vhost_vring_call_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
                goto kick;
        }
 
-       rte_smp_rmb();
+       rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
 
        off_wrap = vq->driver_event->off_wrap;
        off = off_wrap & ~(1 << 15);
@@ -847,10 +851,4 @@ mbuf_is_consumed(struct rte_mbuf *m)
        return true;
 }
 
-static __rte_always_inline void
-put_zmbuf(struct zcopy_mbuf *zmbuf)
-{
-       zmbuf->in_use = 0;
-}
-
 #endif /* _VHOST_NET_CDEV_H_ */