> -----Original Message----- > From: Maxime Coquelin [mailto:maxime.coquelin at redhat.com] > Sent: Monday, September 12, 2016 11:46 PM > To: Wang, Zhihong <zhihong.wang at intel.com>; dev at dpdk.org > Cc: yuanhan.liu at linux.intel.com; thomas.monjalon at 6wind.com > Subject: Re: [PATCH v5 5/6] vhost: batch update used ring > > > > On 09/09/2016 05:39 AM, Zhihong Wang wrote: > > This patch enables batch update of the used ring for better efficiency. > > > > Signed-off-by: Zhihong Wang <zhihong.wang at intel.com> > > --- > > Changes in v4: > > > > 1. Free shadow used ring in the right place. > > > > 2. Add failure check for shadow used ring malloc. > > > > lib/librte_vhost/vhost.c | 20 ++++++++++++-- > > lib/librte_vhost/vhost.h | 4 +++ > > lib/librte_vhost/vhost_user.c | 31 +++++++++++++++++---- > > lib/librte_vhost/virtio_net.c | 64 > +++++++++++++++++++++++++++++++++++-------- > > 4 files changed, 101 insertions(+), 18 deletions(-) > > > > diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c > > index 46095c3..cb31cdd 100644 > > --- a/lib/librte_vhost/vhost.c > > +++ b/lib/librte_vhost/vhost.c > > @@ -119,10 +119,26 @@ cleanup_device(struct virtio_net *dev, int > destroy) > > static void > > free_device(struct virtio_net *dev) > > { > > + struct vhost_virtqueue *vq_0; > > + struct vhost_virtqueue *vq_1; > > uint32_t i; > > > > - for (i = 0; i < dev->virt_qp_nb; i++) > > - rte_free(dev->virtqueue[i * VIRTIO_QNUM]); > > + for (i = 0; i < dev->virt_qp_nb; i++) { > > + vq_0 = dev->virtqueue[i * VIRTIO_QNUM]; > > + if (vq_0->shadow_used_ring) { > > + rte_free(vq_0->shadow_used_ring); > > + vq_0->shadow_used_ring = NULL; > > + } > > + > > + vq_1 = dev->virtqueue[i * VIRTIO_QNUM + 1]; > > + if (vq_1->shadow_used_ring) { > > + rte_free(vq_1->shadow_used_ring); > > + vq_1->shadow_used_ring = NULL; > > + } > > + > > + /* malloc together, free together */ > > + rte_free(vq_0); > > + } > > > > rte_free(dev); > > } > > diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h > > index 9707dfc..381dc27 100644 > > --- a/lib/librte_vhost/vhost.h > > +++ b/lib/librte_vhost/vhost.h > > @@ -85,6 +85,10 @@ struct vhost_virtqueue { > > > > /* Physical address of used ring, for logging */ > > uint64_t log_guest_addr; > > + > > + /* Shadow used ring for performance */ > > + struct vring_used_elem *shadow_used_ring; > > + uint32_t shadow_used_idx; > > } __rte_cache_aligned; > > > > /* Old kernels have no such macro defined */ > > diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c > > index eee99e9..d7cf1ed 100644 > > --- a/lib/librte_vhost/vhost_user.c > > +++ b/lib/librte_vhost/vhost_user.c > > @@ -193,7 +193,21 @@ static int > > vhost_user_set_vring_num(struct virtio_net *dev, > > struct vhost_vring_state *state) > > { > > - dev->virtqueue[state->index]->size = state->num; > > + struct vhost_virtqueue *vq; > > + > > + vq = dev->virtqueue[state->index]; > > + vq->size = state->num; > > + if (!vq->shadow_used_ring) { > > + vq->shadow_used_ring = rte_malloc(NULL, > > + vq->size * sizeof(struct vring_used_elem), > > + RTE_CACHE_LINE_SIZE); > > + if (!vq->shadow_used_ring) { > > + RTE_LOG(ERR, VHOST_CONFIG, > > + "Failed to allocate memory" > > + " for shadow used ring.\n"); > > + return -1; > > + } > > + } > > > > return 0; > > } > > @@ -611,14 +625,21 @@ static int > > vhost_user_get_vring_base(struct virtio_net *dev, > > struct vhost_vring_state *state) > > { > > + struct vhost_virtqueue *vq; > > + > > /* We have to stop the queue (virtio) if it is running. */ > > if (dev->flags & VIRTIO_DEV_RUNNING) { > > dev->flags &= ~VIRTIO_DEV_RUNNING; > > notify_ops->destroy_device(dev->vid); > > } > > > > + vq = dev->virtqueue[state->index]; > > /* Here we are safe to get the last used index */ > > - state->num = dev->virtqueue[state->index]->last_used_idx; > > + state->num = vq->last_used_idx; > > + if (vq->shadow_used_ring) { > > + rte_free(vq->shadow_used_ring); > > + vq->shadow_used_ring = NULL; > > + } > > > > RTE_LOG(INFO, VHOST_CONFIG, > > "vring base idx:%d file:%d\n", state->index, state->num); > > @@ -627,10 +648,10 @@ vhost_user_get_vring_base(struct virtio_net > *dev, > > * sent and only sent in vhost_vring_stop. > > * TODO: cleanup the vring, it isn't usable since here. > > */ > > - if (dev->virtqueue[state->index]->kickfd >= 0) > > - close(dev->virtqueue[state->index]->kickfd); > > + if (vq->kickfd >= 0) > > + close(vq->kickfd); > > > > - dev->virtqueue[state->index]->kickfd = > VIRTIO_UNINITIALIZED_EVENTFD; > > + vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD; > > > > return 0; > > } > > diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c > > index b38f18f..e9f6353 100644 > > --- a/lib/librte_vhost/virtio_net.c > > +++ b/lib/librte_vhost/virtio_net.c > > @@ -134,17 +134,52 @@ virtio_enqueue_offload(struct rte_mbuf > *m_buf, struct virtio_net_hdr *net_hdr) > > } > > > > static inline void __attribute__((always_inline)) > > -update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq, > > - uint32_t desc_chain_head, uint32_t desc_chain_len) > > +update_used_ring(struct vhost_virtqueue *vq, uint32_t > desc_chain_head, > > + uint32_t desc_chain_len) > > { > > - uint32_t used_idx = vq->last_used_idx & (vq->size - 1); > > - > > - vq->used->ring[used_idx].id = desc_chain_head; > > - vq->used->ring[used_idx].len = desc_chain_len; > > + vq->shadow_used_ring[vq->shadow_used_idx].id = > desc_chain_head; > > + vq->shadow_used_ring[vq->shadow_used_idx].len = > desc_chain_len; > > + vq->shadow_used_idx++; > > vq->last_used_idx++; > > - vhost_log_used_vring(dev, vq, offsetof(struct vring_used, > > - ring[used_idx]), > > - sizeof(vq->used->ring[used_idx])); > > +} > > + > > +static inline void __attribute__((always_inline)) > > +flush_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq, > > + uint32_t used_idx_start) > > +{ > > + if (used_idx_start + vq->shadow_used_idx < vq->size) { > > + rte_memcpy(&vq->used->ring[used_idx_start], > > + &vq->shadow_used_ring[0], > > + vq->shadow_used_idx * > > + sizeof(struct vring_used_elem)); > > + vhost_log_used_vring(dev, vq, > > + offsetof(struct vring_used, > > + ring[used_idx_start]), > > + vq->shadow_used_idx * > > + sizeof(struct vring_used_elem)); > > + } else { > > + uint32_t part_1 = vq->size - used_idx_start; > > + uint32_t part_2 = vq->shadow_used_idx - part_1; > > + > > + rte_memcpy(&vq->used->ring[used_idx_start], > > + &vq->shadow_used_ring[0], > > + part_1 * > > + sizeof(struct vring_used_elem)); > > + vhost_log_used_vring(dev, vq, > > + offsetof(struct vring_used, > > + ring[used_idx_start]), > > + part_1 * > > + sizeof(struct vring_used_elem)); > > + rte_memcpy(&vq->used->ring[0], > > + &vq->shadow_used_ring[part_1], > > + part_2 * > > + sizeof(struct vring_used_elem)); > > + vhost_log_used_vring(dev, vq, > > + offsetof(struct vring_used, > > + ring[0]), > > + part_2 * > > + sizeof(struct vring_used_elem)); > > + } > > } > Is expanding the code done for performance purpose?
Hi Maxime, Yes theoretically this has the least branch number. And I think the logic is simpler this way. Thanks Zhihong > Or maybe we could have a loop to do that? > Something like this (not compiled, not tested): > > static inline void __attribute__((always_inline)) > flush_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq, > { > uint32_t to = used_idx_start; > uint32_t from = 0; > uint32_t count; > > if (used_idx_start + vq->shadow_used_idx < vq->size) > count = vq->shadow_used_idx; > else > count = vq->size - used_idx_start; > > do { > rte_memcpy(&vq->used->ring[to], > &vq->shadow_used_ring[from], > count * sizeof(struct vring_used_elem)); > vhost_log_used_vring(dev, vq, > offsetof(struct vring_used, ring[to]), > count * sizeof(struct vring_used_elem)); > > to = (to + count) & (vq->size - 1); > from += count; > count = vq->shadow_used_idx - count; > } while (count); > } > > Regards, > Maxime