vhost: add packed ring batch dequeue
[dpdk.git] / lib / librte_vhost / virtio_net.c
index 241d467..7643520 100644 (file)
@@ -138,11 +138,7 @@ flush_shadow_used_ring_packed(struct virtio_net *dev,
                        head_flags = flags;
                }
 
-               vq->last_used_idx += vq->shadow_used_packed[i].count;
-               if (vq->last_used_idx >= vq->size) {
-                       vq->used_wrap_counter ^= 1;
-                       vq->last_used_idx -= vq->size;
-               }
+               vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
        }
 
        __atomic_store_n(&vq->desc_packed[head_idx].flags, head_flags,
@@ -177,7 +173,8 @@ do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 
        for (i = 0; i < count; i++) {
                rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
-               vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
+               vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
+                                          elem[i].len);
                PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
        }
 
@@ -631,7 +628,7 @@ copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 
                PRINT_PACKET(dev, (uintptr_t)dst,
                                (uint32_t)len, 0);
-               vhost_log_cache_write(dev, vq,
+               vhost_log_cache_write_iova(dev, vq,
                                iova, len);
 
                remain -= len;
@@ -731,7 +728,7 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        } else {
                                PRINT_PACKET(dev, (uintptr_t)hdr_addr,
                                                dev->vhost_hlen, 0);
-                               vhost_log_cache_write(dev, vq,
+                               vhost_log_cache_write_iova(dev, vq,
                                                buf_vec[0].buf_iova,
                                                dev->vhost_hlen);
                        }
@@ -746,8 +743,9 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
                                rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
                                cpy_len);
-                       vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
-                                       cpy_len);
+                       vhost_log_cache_write_iova(dev, vq,
+                                                  buf_iova + buf_offset,
+                                                  cpy_len);
                        PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
                                cpy_len, 0);
                } else {
@@ -772,6 +770,60 @@ out:
        return error;
 }
 
+static __rte_always_inline int
+vhost_enqueue_single_packed(struct virtio_net *dev,
+                           struct vhost_virtqueue *vq,
+                           struct rte_mbuf *pkt,
+                           struct buf_vector *buf_vec,
+                           uint16_t *nr_descs)
+{
+       uint16_t nr_vec = 0;
+       uint16_t avail_idx = vq->last_avail_idx;
+       uint16_t max_tries, tries = 0;
+       uint16_t buf_id = 0;
+       uint32_t len = 0;
+       uint16_t desc_count;
+       uint32_t size = pkt->pkt_len + dev->vhost_hlen;
+       uint16_t num_buffers = 0;
+
+       if (rxvq_is_mergeable(dev))
+               max_tries = vq->size - 1;
+       else
+               max_tries = 1;
+
+       while (size > 0) {
+               /*
+                * if we tried all available ring items, and still
+                * can't get enough buf, it means something abnormal
+                * happened.
+                */
+               if (unlikely(++tries > max_tries))
+                       return -1;
+
+               if (unlikely(fill_vec_buf_packed(dev, vq,
+                                               avail_idx, &desc_count,
+                                               buf_vec, &nr_vec,
+                                               &buf_id, &len,
+                                               VHOST_ACCESS_RW) < 0))
+                       return -1;
+
+               len = RTE_MIN(len, size);
+               size -= len;
+
+               num_buffers += 1;
+
+               *nr_descs += desc_count;
+               avail_idx += desc_count;
+               if (avail_idx >= vq->size)
+                       avail_idx -= vq->size;
+       }
+
+       if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
+               return -1;
+
+       return 0;
+}
+
 static __rte_noinline uint32_t
 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mbuf **pkts, uint32_t count)
@@ -829,6 +881,102 @@ virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        return pkt_idx;
 }
 
+static __rte_unused int
+virtio_dev_rx_batch_packed(struct virtio_net *dev,
+                          struct vhost_virtqueue *vq,
+                          struct rte_mbuf **pkts)
+{
+       bool wrap_counter = vq->avail_wrap_counter;
+       struct vring_packed_desc *descs = vq->desc_packed;
+       uint16_t avail_idx = vq->last_avail_idx;
+       uint64_t desc_addrs[PACKED_BATCH_SIZE];
+       struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
+       uint32_t buf_offset = dev->vhost_hlen;
+       uint64_t lens[PACKED_BATCH_SIZE];
+       uint16_t i;
+
+       if (unlikely(avail_idx & PACKED_BATCH_MASK))
+               return -1;
+
+       if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
+               return -1;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(pkts[i]->next != NULL))
+                       return -1;
+               if (unlikely(!desc_is_avail(&descs[avail_idx + i],
+                                           wrap_counter)))
+                       return -1;
+       }
+
+       rte_smp_rmb();
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               lens[i] = descs[avail_idx + i].len;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
+                       return -1;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               desc_addrs[i] = vhost_iova_to_vva(dev, vq,
+                                                 descs[avail_idx + i].addr,
+                                                 &lens[i],
+                                                 VHOST_ACCESS_RW);
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(lens[i] != descs[avail_idx + i].len))
+                       return -1;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
+               hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
+                                       (uintptr_t)desc_addrs[i];
+               lens[i] = pkts[i]->pkt_len + dev->vhost_hlen;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
+
+       vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
+                          rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
+                          pkts[i]->pkt_len);
+       }
+
+       return 0;
+}
+
+static __rte_unused int16_t
+virtio_dev_rx_single_packed(struct virtio_net *dev,
+                           struct vhost_virtqueue *vq,
+                           struct rte_mbuf *pkt)
+{
+       struct buf_vector buf_vec[BUF_VECTOR_MAX];
+       uint16_t nr_descs = 0;
+
+       rte_smp_rmb();
+       if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
+                                                &nr_descs) < 0)) {
+               VHOST_LOG_DEBUG(VHOST_DATA,
+                               "(%d) failed to get enough desc from vring\n",
+                               dev->vid);
+               return -1;
+       }
+
+       VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
+                       dev->vid, vq->last_avail_idx,
+                       vq->last_avail_idx + nr_descs);
+
+       vq_inc_last_avail_packed(vq, nr_descs);
+
+       return 0;
+}
+
 static __rte_noinline uint32_t
 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mbuf **pkts, uint32_t count)
@@ -863,11 +1011,7 @@ virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        break;
                }
 
-               vq->last_avail_idx += nr_descs;
-               if (vq->last_avail_idx >= vq->size) {
-                       vq->last_avail_idx -= vq->size;
-                       vq->avail_wrap_counter ^= 1;
-               }
+               vq_inc_last_avail_packed(vq, nr_descs);
        }
 
        do_data_copy_enqueue(dev, vq);
@@ -1287,6 +1431,93 @@ again:
        return NULL;
 }
 
+static void
+virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
+{
+       rte_free(opaque);
+}
+
+static int
+virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
+{
+       struct rte_mbuf_ext_shared_info *shinfo = NULL;
+       uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
+       uint16_t buf_len;
+       rte_iova_t iova;
+       void *buf;
+
+       /* Try to use pkt buffer to store shinfo to reduce the amount of memory
+        * required, otherwise store shinfo in the new buffer.
+        */
+       if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo))
+               shinfo = rte_pktmbuf_mtod(pkt,
+                                         struct rte_mbuf_ext_shared_info *);
+       else {
+               total_len += sizeof(*shinfo) + sizeof(uintptr_t);
+               total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
+       }
+
+       if (unlikely(total_len > UINT16_MAX))
+               return -ENOSPC;
+
+       buf_len = total_len;
+       buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
+       if (unlikely(buf == NULL))
+               return -ENOMEM;
+
+       /* Initialize shinfo */
+       if (shinfo) {
+               shinfo->free_cb = virtio_dev_extbuf_free;
+               shinfo->fcb_opaque = buf;
+               rte_mbuf_ext_refcnt_set(shinfo, 1);
+       } else {
+               shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
+                                             virtio_dev_extbuf_free, buf);
+               if (unlikely(shinfo == NULL)) {
+                       rte_free(buf);
+                       RTE_LOG(ERR, VHOST_DATA, "Failed to init shinfo\n");
+                       return -1;
+               }
+       }
+
+       iova = rte_malloc_virt2iova(buf);
+       rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
+       rte_pktmbuf_reset_headroom(pkt);
+
+       return 0;
+}
+
+/*
+ * Allocate a host supported pktmbuf.
+ */
+static __rte_always_inline struct rte_mbuf *
+virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
+                        uint32_t data_len)
+{
+       struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
+
+       if (unlikely(pkt == NULL))
+               return NULL;
+
+       if (rte_pktmbuf_tailroom(pkt) >= data_len)
+               return pkt;
+
+       /* attach an external buffer if supported */
+       if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
+               return pkt;
+
+       /* check if chained buffers are allowed */
+       if (!dev->linearbuf)
+               return pkt;
+
+       /* Data doesn't fit into the buffer and the host supports
+        * only linear buffers
+        */
+       rte_pktmbuf_free(pkt);
+
+       return NULL;
+}
+
 static __rte_noinline uint16_t
 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
@@ -1341,26 +1572,23 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        for (i = 0; i < count; i++) {
                struct buf_vector buf_vec[BUF_VECTOR_MAX];
                uint16_t head_idx;
-               uint32_t dummy_len;
+               uint32_t buf_len;
                uint16_t nr_vec = 0;
                int err;
 
                if (unlikely(fill_vec_buf_split(dev, vq,
                                                vq->last_avail_idx + i,
                                                &nr_vec, buf_vec,
-                                               &head_idx, &dummy_len,
+                                               &head_idx, &buf_len,
                                                VHOST_ACCESS_RO) < 0))
                        break;
 
                if (likely(dev->dequeue_zero_copy == 0))
                        update_shadow_used_ring_split(vq, head_idx, 0);
 
-               pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
-               if (unlikely(pkts[i] == NULL)) {
-                       RTE_LOG(ERR, VHOST_DATA,
-                               "Failed to allocate memory for mbuf.\n");
+               pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
+               if (unlikely(pkts[i] == NULL))
                        break;
-               }
 
                err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
                                mbuf_pool);
@@ -1407,6 +1635,172 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
        return i;
 }
 
+static __rte_always_inline int
+vhost_reserve_avail_batch_packed(struct virtio_net *dev,
+                                struct vhost_virtqueue *vq,
+                                struct rte_mempool *mbuf_pool,
+                                struct rte_mbuf **pkts,
+                                uint16_t avail_idx,
+                                uintptr_t *desc_addrs,
+                                uint16_t *ids)
+{
+       bool wrap = vq->avail_wrap_counter;
+       struct vring_packed_desc *descs = vq->desc_packed;
+       struct virtio_net_hdr *hdr;
+       uint64_t lens[PACKED_BATCH_SIZE];
+       uint64_t buf_lens[PACKED_BATCH_SIZE];
+       uint32_t buf_offset = dev->vhost_hlen;
+       uint16_t flags, i;
+
+       if (unlikely(avail_idx & PACKED_BATCH_MASK))
+               return -1;
+       if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
+               return -1;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               flags = descs[avail_idx + i].flags;
+               if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
+                            (wrap == !!(flags & VRING_DESC_F_USED))  ||
+                            (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
+                       return -1;
+       }
+
+       rte_smp_rmb();
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               lens[i] = descs[avail_idx + i].len;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               desc_addrs[i] = vhost_iova_to_vva(dev, vq,
+                                                 descs[avail_idx + i].addr,
+                                                 &lens[i], VHOST_ACCESS_RW);
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely((lens[i] != descs[avail_idx + i].len)))
+                       return -1;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
+               if (!pkts[i])
+                       goto free_buf;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
+                       goto free_buf;
+       }
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+               pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
+               pkts[i]->data_len = pkts[i]->pkt_len;
+               ids[i] = descs[avail_idx + i].id;
+       }
+
+       if (virtio_net_with_host_offload(dev)) {
+               vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+                       hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
+                       vhost_dequeue_offload(hdr, pkts[i]);
+               }
+       }
+
+       return 0;
+
+free_buf:
+       for (i = 0; i < PACKED_BATCH_SIZE; i++)
+               rte_pktmbuf_free(pkts[i]);
+
+       return -1;
+}
+
+static __rte_unused int
+virtio_dev_tx_batch_packed(struct virtio_net *dev,
+                          struct vhost_virtqueue *vq,
+                          struct rte_mempool *mbuf_pool,
+                          struct rte_mbuf **pkts)
+{
+       uint16_t avail_idx = vq->last_avail_idx;
+       uint32_t buf_offset = dev->vhost_hlen;
+       uintptr_t desc_addrs[PACKED_BATCH_SIZE];
+       uint16_t ids[PACKED_BATCH_SIZE];
+       uint16_t i;
+
+       if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
+                                            avail_idx, desc_addrs, ids))
+               return -1;
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
+
+       vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+               rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
+                          (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
+                          pkts[i]->pkt_len);
+
+       vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
+
+       return 0;
+}
+
+static __rte_always_inline int
+vhost_dequeue_single_packed(struct virtio_net *dev,
+                           struct vhost_virtqueue *vq,
+                           struct rte_mempool *mbuf_pool,
+                           struct rte_mbuf **pkts,
+                           uint16_t *buf_id,
+                           uint16_t *desc_count)
+{
+       struct buf_vector buf_vec[BUF_VECTOR_MAX];
+       uint32_t buf_len;
+       uint16_t nr_vec = 0;
+       int err;
+
+       if (unlikely(fill_vec_buf_packed(dev, vq,
+                                        vq->last_avail_idx, desc_count,
+                                        buf_vec, &nr_vec,
+                                        buf_id, &buf_len,
+                                        VHOST_ACCESS_RO) < 0))
+               return -1;
+
+       *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
+       if (unlikely(*pkts == NULL)) {
+               RTE_LOG(ERR, VHOST_DATA,
+                       "Failed to allocate memory for mbuf.\n");
+               return -1;
+       }
+
+       err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
+                               mbuf_pool);
+       if (unlikely(err)) {
+               rte_pktmbuf_free(*pkts);
+               return -1;
+       }
+
+       return 0;
+}
+
+static __rte_unused int
+virtio_dev_tx_single_packed(struct virtio_net *dev,
+                           struct vhost_virtqueue *vq,
+                           struct rte_mempool *mbuf_pool,
+                           struct rte_mbuf **pkts)
+{
+
+       uint16_t buf_id, desc_count;
+
+       if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
+                                       &desc_count))
+               return -1;
+
+       vq_inc_last_avail_packed(vq, desc_count);
+
+       return 0;
+}
+
 static __rte_noinline uint16_t
 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
        struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
@@ -1449,14 +1843,14 @@ virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
        for (i = 0; i < count; i++) {
                struct buf_vector buf_vec[BUF_VECTOR_MAX];
                uint16_t buf_id;
-               uint32_t dummy_len;
+               uint32_t buf_len;
                uint16_t desc_count, nr_vec = 0;
                int err;
 
                if (unlikely(fill_vec_buf_packed(dev, vq,
                                                vq->last_avail_idx, &desc_count,
                                                buf_vec, &nr_vec,
-                                               &buf_id, &dummy_len,
+                                               &buf_id, &buf_len,
                                                VHOST_ACCESS_RO) < 0))
                        break;
 
@@ -1464,12 +1858,9 @@ virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        update_shadow_used_ring_packed(vq, buf_id, 0,
                                        desc_count);
 
-               pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
-               if (unlikely(pkts[i] == NULL)) {
-                       RTE_LOG(ERR, VHOST_DATA,
-                               "Failed to allocate memory for mbuf.\n");
+               pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
+               if (unlikely(pkts[i] == NULL))
                        break;
-               }
 
                err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
                                mbuf_pool);
@@ -1502,11 +1893,7 @@ virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
                }
 
-               vq->last_avail_idx += desc_count;
-               if (vq->last_avail_idx >= vq->size) {
-                       vq->last_avail_idx -= vq->size;
-                       vq->avail_wrap_counter ^= 1;
-               }
+               vq_inc_last_avail_packed(vq, desc_count);
        }
 
        if (likely(dev->dequeue_zero_copy == 0)) {