vhost: unify Rx mergeable and non-mergeable paths
authorMaxime Coquelin <maxime.coquelin@redhat.com>
Tue, 29 May 2018 09:45:13 +0000 (11:45 +0200)
committerFerruh Yigit <ferruh.yigit@intel.com>
Thu, 14 Jun 2018 17:27:50 +0000 (19:27 +0200)
This patch reworks the vhost enqueue path so that a single
code path is used for both Rx mergeable or non-mergeable cases.

Acked-by: Zhihong Wang <zhihong.wang@intel.com>
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
lib/librte_vhost/virtio_net.c

index 76ec5f0..e66b8f4 100644 (file)
 
 #define MAX_BATCH_LEN 256
 
+static  __rte_always_inline bool
+rxvq_is_mergeable(struct virtio_net *dev)
+{
+       return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
+}
+
 static bool
 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
 {
@@ -154,7 +160,7 @@ do_data_copy_dequeue(struct vhost_virtqueue *vq)
                (var) = (val);                  \
 } while (0)
 
-static void
+static __rte_always_inline void
 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 {
        uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
@@ -215,317 +221,6 @@ virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
        }
 }
 
-static __rte_always_inline int
-copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
-                 struct vring_desc *descs, struct rte_mbuf *m,
-                 uint16_t desc_idx, uint32_t size)
-{
-       uint32_t desc_avail, desc_offset;
-       uint32_t mbuf_avail, mbuf_offset;
-       uint32_t cpy_len;
-       uint64_t desc_chunck_len;
-       struct vring_desc *desc;
-       uint64_t desc_addr, desc_gaddr;
-       /* A counter to avoid desc dead loop chain */
-       uint16_t nr_desc = 1;
-       struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
-       uint16_t copy_nb = vq->batch_copy_nb_elems;
-       int error = 0;
-
-       desc = &descs[desc_idx];
-       desc_chunck_len = desc->len;
-       desc_gaddr = desc->addr;
-       desc_addr = vhost_iova_to_vva(dev, vq, desc_gaddr,
-                                       &desc_chunck_len, VHOST_ACCESS_RW);
-       /*
-        * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
-        * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
-        * otherwise stores offset on the stack instead of in a register.
-        */
-       if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
-               error = -1;
-               goto out;
-       }
-
-       rte_prefetch0((void *)(uintptr_t)desc_addr);
-
-       if (likely(desc_chunck_len >= dev->vhost_hlen)) {
-               virtio_enqueue_offload(m,
-                               (struct virtio_net_hdr *)(uintptr_t)desc_addr);
-               PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
-               vhost_log_cache_write(dev, vq, desc_gaddr, dev->vhost_hlen);
-       } else {
-               struct virtio_net_hdr vnet_hdr;
-               uint64_t remain = dev->vhost_hlen;
-               uint64_t len;
-               uint64_t src = (uint64_t)(uintptr_t)&vnet_hdr, dst;
-               uint64_t guest_addr = desc_gaddr;
-
-               virtio_enqueue_offload(m, &vnet_hdr);
-
-               while (remain) {
-                       len = remain;
-                       dst = vhost_iova_to_vva(dev, vq, guest_addr,
-                                       &len, VHOST_ACCESS_RW);
-                       if (unlikely(!dst || !len)) {
-                               error = -1;
-                               goto out;
-                       }
-
-                       rte_memcpy((void *)(uintptr_t)dst,
-                                       (void *)(uintptr_t)src, len);
-
-                       PRINT_PACKET(dev, (uintptr_t)dst, (uint32_t)len, 0);
-                       vhost_log_cache_write(dev, vq, guest_addr, len);
-                       remain -= len;
-                       guest_addr += len;
-                       src += len;
-               }
-       }
-
-       desc_avail  = desc->len - dev->vhost_hlen;
-       if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
-               desc_chunck_len = desc_avail;
-               desc_gaddr = desc->addr + dev->vhost_hlen;
-               desc_addr = vhost_iova_to_vva(dev,
-                               vq, desc_gaddr,
-                               &desc_chunck_len,
-                               VHOST_ACCESS_RW);
-               if (unlikely(!desc_addr)) {
-                       error = -1;
-                       goto out;
-               }
-
-               desc_offset = 0;
-       } else {
-               desc_offset = dev->vhost_hlen;
-               desc_chunck_len -= dev->vhost_hlen;
-       }
-
-       mbuf_avail  = rte_pktmbuf_data_len(m);
-       mbuf_offset = 0;
-       while (mbuf_avail != 0 || m->next != NULL) {
-               /* done with current mbuf, fetch next */
-               if (mbuf_avail == 0) {
-                       m = m->next;
-
-                       mbuf_offset = 0;
-                       mbuf_avail  = rte_pktmbuf_data_len(m);
-               }
-
-               /* done with current desc buf, fetch next */
-               if (desc_avail == 0) {
-                       if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
-                               /* Room in vring buffer is not enough */
-                               error = -1;
-                               goto out;
-                       }
-                       if (unlikely(desc->next >= size || ++nr_desc > size)) {
-                               error = -1;
-                               goto out;
-                       }
-
-                       desc = &descs[desc->next];
-                       desc_chunck_len = desc->len;
-                       desc_gaddr = desc->addr;
-                       desc_addr = vhost_iova_to_vva(dev, vq, desc_gaddr,
-                                                       &desc_chunck_len,
-                                                       VHOST_ACCESS_RW);
-                       if (unlikely(!desc_addr)) {
-                               error = -1;
-                               goto out;
-                       }
-
-                       desc_offset = 0;
-                       desc_avail  = desc->len;
-               } else if (unlikely(desc_chunck_len == 0)) {
-                       desc_chunck_len = desc_avail;
-                       desc_gaddr += desc_offset;
-                       desc_addr = vhost_iova_to_vva(dev,
-                                       vq, desc_gaddr,
-                                       &desc_chunck_len, VHOST_ACCESS_RW);
-                       if (unlikely(!desc_addr)) {
-                               error = -1;
-                               goto out;
-                       }
-                       desc_offset = 0;
-               }
-
-               cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
-               if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
-                       rte_memcpy((void *)((uintptr_t)(desc_addr +
-                                                       desc_offset)),
-                               rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
-                               cpy_len);
-                       vhost_log_cache_write(dev, vq, desc_gaddr + desc_offset,
-                                       cpy_len);
-                       PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
-                                    cpy_len, 0);
-               } else {
-                       batch_copy[copy_nb].dst =
-                               (void *)((uintptr_t)(desc_addr + desc_offset));
-                       batch_copy[copy_nb].src =
-                               rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
-                       batch_copy[copy_nb].log_addr = desc_gaddr + desc_offset;
-                       batch_copy[copy_nb].len = cpy_len;
-                       copy_nb++;
-               }
-
-               mbuf_avail  -= cpy_len;
-               mbuf_offset += cpy_len;
-               desc_avail  -= cpy_len;
-               desc_offset += cpy_len;
-               desc_chunck_len -= cpy_len;
-       }
-
-out:
-       vq->batch_copy_nb_elems = copy_nb;
-
-       return error;
-}
-
-/**
- * This function adds buffers to the virtio devices RX virtqueue. Buffers can
- * be received from the physical port or from another virtio device. A packet
- * count is returned to indicate the number of packets that are successfully
- * added to the RX queue. This function works when the mbuf is scattered, but
- * it doesn't support the mergeable feature.
- */
-static __rte_always_inline uint32_t
-virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
-             struct rte_mbuf **pkts, uint32_t count)
-{
-       struct vhost_virtqueue *vq;
-       uint16_t avail_idx, free_entries, start_idx;
-       uint16_t desc_indexes[MAX_PKT_BURST];
-       struct vring_desc *descs;
-       uint16_t used_idx;
-       uint32_t i, sz;
-
-       VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
-       if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
-               RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
-                       dev->vid, __func__, queue_id);
-               return 0;
-       }
-
-       vq = dev->virtqueue[queue_id];
-
-       rte_spinlock_lock(&vq->access_lock);
-
-       if (unlikely(vq->enabled == 0))
-               goto out_access_unlock;
-
-       if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
-               vhost_user_iotlb_rd_lock(vq);
-
-       if (unlikely(vq->access_ok == 0)) {
-               if (unlikely(vring_translate(dev, vq) < 0)) {
-                       count = 0;
-                       goto out;
-               }
-       }
-
-       avail_idx = *((volatile uint16_t *)&vq->avail->idx);
-       start_idx = vq->last_used_idx;
-       free_entries = avail_idx - start_idx;
-       count = RTE_MIN(count, free_entries);
-       count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
-       if (count == 0)
-               goto out;
-
-       VHOST_LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
-               dev->vid, start_idx, start_idx + count);
-
-       vq->batch_copy_nb_elems = 0;
-
-       /* Retrieve all of the desc indexes first to avoid caching issues. */
-       rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
-       for (i = 0; i < count; i++) {
-               used_idx = (start_idx + i) & (vq->size - 1);
-               desc_indexes[i] = vq->avail->ring[used_idx];
-               vq->used->ring[used_idx].id = desc_indexes[i];
-               vq->used->ring[used_idx].len = pkts[i]->pkt_len +
-                                              dev->vhost_hlen;
-               vhost_log_cache_used_vring(dev, vq,
-                       offsetof(struct vring_used, ring[used_idx]),
-                       sizeof(vq->used->ring[used_idx]));
-       }
-
-       rte_prefetch0(&vq->desc[desc_indexes[0]]);
-       for (i = 0; i < count; i++) {
-               struct vring_desc *idesc = NULL;
-               uint16_t desc_idx = desc_indexes[i];
-               int err;
-
-               if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
-                       uint64_t dlen = vq->desc[desc_idx].len;
-                       descs = (struct vring_desc *)(uintptr_t)
-                               vhost_iova_to_vva(dev,
-                                               vq, vq->desc[desc_idx].addr,
-                                               &dlen, VHOST_ACCESS_RO);
-                       if (unlikely(!descs)) {
-                               count = i;
-                               break;
-                       }
-
-                       if (unlikely(dlen < vq->desc[desc_idx].len)) {
-                               /*
-                                * The indirect desc table is not contiguous
-                                * in process VA space, we have to copy it.
-                                */
-                               idesc = alloc_copy_ind_table(dev, vq,
-                                                       &vq->desc[desc_idx]);
-                               if (unlikely(!idesc))
-                                       break;
-
-                               descs = idesc;
-                       }
-
-                       desc_idx = 0;
-                       sz = vq->desc[desc_idx].len / sizeof(*descs);
-               } else {
-                       descs = vq->desc;
-                       sz = vq->size;
-               }
-
-               err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
-               if (unlikely(err)) {
-                       count = i;
-                       free_ind_table(idesc);
-                       break;
-               }
-
-               if (i + 1 < count)
-                       rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
-
-               if (unlikely(!!idesc))
-                       free_ind_table(idesc);
-       }
-
-       do_data_copy_enqueue(dev, vq);
-
-       rte_smp_wmb();
-
-       vhost_log_cache_sync(dev, vq);
-
-       *(volatile uint16_t *)&vq->used->idx += count;
-       vq->last_used_idx += count;
-       vhost_log_used_vring(dev, vq,
-               offsetof(struct vring_used, idx),
-               sizeof(vq->used->idx));
-
-       vhost_vring_call(dev, vq);
-out:
-       if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
-               vhost_user_iotlb_rd_unlock(vq);
-
-out_access_unlock:
-       rte_spinlock_unlock(&vq->access_lock);
-
-       return count;
-}
-
 static __rte_always_inline int
 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                         uint32_t avail_idx, uint32_t *vec_idx,
@@ -596,13 +291,13 @@ fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
  * Returns -1 on fail, 0 on success
  */
 static inline int
-reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
+reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                                uint32_t size, struct buf_vector *buf_vec,
                                uint16_t *num_buffers, uint16_t avail_head)
 {
        uint16_t cur_idx;
        uint32_t vec_idx = 0;
-       uint16_t tries = 0;
+       uint16_t max_tries, tries = 0;
 
        uint16_t head_idx = 0;
        uint16_t len = 0;
@@ -610,6 +305,11 @@ reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
        *num_buffers = 0;
        cur_idx  = vq->last_avail_idx;
 
+       if (rxvq_is_mergeable(dev))
+               max_tries = vq->size;
+       else
+               max_tries = 1;
+
        while (size > 0) {
                if (unlikely(cur_idx == avail_head))
                        return -1;
@@ -630,7 +330,7 @@ reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
                 * can't get enough buf, it means something abnormal
                 * happened.
                 */
-               if (unlikely(tries >= vq->size))
+               if (unlikely(tries > max_tries))
                        return -1;
        }
 
@@ -638,7 +338,7 @@ reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 }
 
 static __rte_always_inline int
-copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
+copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                            struct rte_mbuf *m, struct buf_vector *buf_vec,
                            uint16_t num_buffers)
 {
@@ -748,7 +448,9 @@ copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 
                if (hdr_addr) {
                        virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
-                       ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
+                       if (rxvq_is_mergeable(dev))
+                               ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
+                                               num_buffers);
 
                        if (unlikely(hdr == &tmp_hdr)) {
                                uint64_t len;
@@ -824,7 +526,7 @@ out:
 }
 
 static __rte_always_inline uint32_t
-virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
+virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
        struct rte_mbuf **pkts, uint32_t count)
 {
        struct vhost_virtqueue *vq;
@@ -867,7 +569,7 @@ virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
        for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
                uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 
-               if (unlikely(reserve_avail_buf_mergeable(dev, vq,
+               if (unlikely(reserve_avail_buf(dev, vq,
                                                pkt_len, buf_vec, &num_buffers,
                                                avail_head) < 0)) {
                        VHOST_LOG_DEBUG(VHOST_DATA,
@@ -881,7 +583,7 @@ virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
                        dev->vid, vq->last_avail_idx,
                        vq->last_avail_idx + num_buffers);
 
-               if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
+               if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
                                                buf_vec, num_buffers) < 0) {
                        vq->shadow_used_idx -= num_buffers;
                        break;
@@ -923,10 +625,7 @@ rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
                return 0;
        }
 
-       if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
-               return virtio_dev_merge_rx(dev, queue_id, pkts, count);
-       else
-               return virtio_dev_rx(dev, queue_id, pkts, count);
+       return virtio_dev_rx(dev, queue_id, pkts, count);
 }
 
 static inline bool