net/virtio: fix avail descriptor ID
[dpdk.git] / drivers / net / virtio / virtqueue.h
index e3d38b5..4f3dd31 100644 (file)
 #include <rte_atomic.h>
 #include <rte_memory.h>
 #include <rte_mempool.h>
+#include <rte_net.h>
 
-#include "virtio_pci.h"
+#include "virtio.h"
 #include "virtio_ring.h"
 #include "virtio_logs.h"
 #include "virtio_rxtx.h"
 
 struct rte_mbuf;
 
+#define DEFAULT_TX_FREE_THRESH 32
+#define DEFAULT_RX_FREE_THRESH 32
+
+#define VIRTIO_MBUF_BURST_SZ 64
 /*
  * Per virtio_ring.h in Linux.
  *     For virtio_pci on SMP, we don't need to order with respect to MMIO
- *     accesses through relaxed memory I/O windows, so smp_mb() et al are
+ *     accesses through relaxed memory I/O windows, so thread_fence is
  *     sufficient.
  *
  *     For using virtio to talk to real devices (eg. vDPA) we do need real
@@ -31,7 +36,7 @@ static inline void
 virtio_mb(uint8_t weak_barriers)
 {
        if (weak_barriers)
-               rte_smp_mb();
+               rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
        else
                rte_mb();
 }
@@ -40,18 +45,18 @@ static inline void
 virtio_rmb(uint8_t weak_barriers)
 {
        if (weak_barriers)
-               rte_smp_rmb();
+               rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
        else
-               rte_cio_rmb();
+               rte_io_rmb();
 }
 
 static inline void
 virtio_wmb(uint8_t weak_barriers)
 {
        if (weak_barriers)
-               rte_smp_wmb();
+               rte_atomic_thread_fence(__ATOMIC_RELEASE);
        else
-               rte_cio_wmb();
+               rte_io_wmb();
 }
 
 static inline uint16_t
@@ -61,20 +66,19 @@ virtqueue_fetch_flags_packed(struct vring_packed_desc *dp,
        uint16_t flags;
 
        if (weak_barriers) {
-/* x86 prefers to using rte_smp_rmb over __atomic_load_n as it reports
+/* x86 prefers to using rte_io_rmb over __atomic_load_n as it reports
  * a better perf(~1.5%), which comes from the saved branch by the compiler.
- * The if and else branch are identical with the smp and cio barriers both
- * defined as compiler barriers on x86.
+ * The if and else branch are identical  on the platforms except Arm.
  */
-#ifdef RTE_ARCH_X86_64
-               flags = dp->flags;
-               rte_smp_rmb();
-#else
+#ifdef RTE_ARCH_ARM
                flags = __atomic_load_n(&dp->flags, __ATOMIC_ACQUIRE);
+#else
+               flags = dp->flags;
+               rte_io_rmb();
 #endif
        } else {
                flags = dp->flags;
-               rte_cio_rmb();
+               rte_io_rmb();
        }
 
        return flags;
@@ -85,22 +89,22 @@ virtqueue_store_flags_packed(struct vring_packed_desc *dp,
                              uint16_t flags, uint8_t weak_barriers)
 {
        if (weak_barriers) {
-/* x86 prefers to using rte_smp_wmb over __atomic_store_n as it reports
+/* x86 prefers to using rte_io_wmb over __atomic_store_n as it reports
  * a better perf(~1.5%), which comes from the saved branch by the compiler.
- * The if and else branch are identical with the smp and cio barriers both
- * defined as compiler barriers on x86.
+ * The if and else branch are identical on the platforms except Arm.
  */
-#ifdef RTE_ARCH_X86_64
-               rte_smp_wmb();
-               dp->flags = flags;
-#else
+#ifdef RTE_ARCH_ARM
                __atomic_store_n(&dp->flags, flags, __ATOMIC_RELEASE);
+#else
+               rte_io_wmb();
+               dp->flags = flags;
 #endif
        } else {
-               rte_cio_wmb();
+               rte_io_wmb();
                dp->flags = flags;
        }
 }
+
 #ifdef RTE_PMD_PACKET_PREFETCH
 #define rte_packet_prefetch(p)  rte_prefetch1(p)
 #else
@@ -109,10 +113,9 @@ virtqueue_store_flags_packed(struct vring_packed_desc *dp,
 
 #define VIRTQUEUE_MAX_NAME_SZ 32
 
-#ifdef RTE_VIRTIO_USER
 /**
- * Return the physical address (or virtual address in case of
- * virtio-user) of mbuf data buffer.
+ * Return the IOVA (or virtual address in case of virtio-user) of mbuf
+ * data buffer.
  *
  * The address is firstly casted to the word size (sizeof(uintptr_t))
  * before casting it to uint64_t. This is to make it work with different
@@ -120,10 +123,7 @@ virtqueue_store_flags_packed(struct vring_packed_desc *dp,
  * (virtio-pci and virtio-user).
  */
 #define VIRTIO_MBUF_ADDR(mb, vq) \
-       ((uint64_t)(*(uintptr_t *)((uintptr_t)(mb) + (vq)->offset)))
-#else
-#define VIRTIO_MBUF_ADDR(mb, vq) ((mb)->buf_iova)
-#endif
+       ((uint64_t)(*(uintptr_t *)((uintptr_t)(mb) + (vq)->mbuf_addr_offset)))
 
 /**
  * Return the physical address (or virtual address in case of
@@ -236,6 +236,10 @@ struct vq_desc_extra {
        uint16_t next;
 };
 
+#define virtnet_rxq_to_vq(rxvq) container_of(rxvq, struct virtqueue, rxq)
+#define virtnet_txq_to_vq(txvq) container_of(txvq, struct virtqueue, txq)
+#define virtnet_cq_to_vq(cvq) container_of(cvq, struct virtqueue, cq)
+
 struct virtqueue {
        struct virtio_hw  *hw; /**< virtio_hw structure pointer. */
        union {
@@ -259,8 +263,18 @@ struct virtqueue {
        uint16_t vq_avail_idx; /**< sync until needed */
        uint16_t vq_free_thresh; /**< free threshold */
 
+       /**
+        * Head of the free chain in the descriptor table. If
+        * there are no free descriptors, this will be set to
+        * VQ_RING_DESC_CHAIN_END.
+        */
+       uint16_t  vq_desc_head_idx;
+       uint16_t  vq_desc_tail_idx;
+       uint16_t  vq_queue_index;   /**< PCI queue index */
+
        void *vq_ring_virt_mem;  /**< linear address of vring*/
        unsigned int vq_ring_size;
+       uint16_t mbuf_addr_offset;
 
        union {
                struct virtnet_rx rxq;
@@ -271,15 +285,6 @@ struct virtqueue {
        rte_iova_t vq_ring_mem; /**< physical address of vring,
                                 * or virtual address for virtio_user. */
 
-       /**
-        * Head of the free chain in the descriptor table. If
-        * there are no free descriptors, this will be set to
-        * VQ_RING_DESC_CHAIN_END.
-        */
-       uint16_t  vq_desc_head_idx;
-       uint16_t  vq_desc_tail_idx;
-       uint16_t  vq_queue_index;   /**< PCI queue index */
-       uint16_t offset; /**< relative offset to obtain addr in mbuf */
        uint16_t  *notify_addr;
        struct rte_mbuf **sw_ring;  /**< RX software ring. */
        struct vq_desc_extra vq_descx[0];
@@ -324,8 +329,11 @@ struct virtio_net_hdr_mrg_rxbuf {
 #define VIRTIO_MAX_TX_INDIRECT 8
 struct virtio_tx_region {
        struct virtio_net_hdr_mrg_rxbuf tx_hdr;
-       struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT]
-               __rte_aligned(16);
+       union {
+               struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT];
+               struct vring_packed_desc
+                       tx_packed_indir[VIRTIO_MAX_TX_INDIRECT];
+       } __rte_aligned(16);
 };
 
 static inline int
@@ -363,6 +371,16 @@ vring_desc_init_split(struct vring_desc *dp, uint16_t n)
        dp[i].next = VQ_RING_DESC_CHAIN_END;
 }
 
+static inline void
+vring_desc_init_indirect_packed(struct vring_packed_desc *dp, int n)
+{
+       int i;
+       for (i = 0; i < n; i++) {
+               dp[i].id = (uint16_t)i;
+               dp[i].flags = VRING_DESC_F_WRITE;
+       }
+}
+
 /**
  * Tell the backend not to interrupt us. Implementation for packed virtqueues.
  */
@@ -391,7 +409,7 @@ virtqueue_disable_intr_split(struct virtqueue *vq)
 static inline void
 virtqueue_disable_intr(struct virtqueue *vq)
 {
-       if (vtpci_packed_queue(vq->hw))
+       if (virtio_with_packed_queue(vq->hw))
                virtqueue_disable_intr_packed(vq);
        else
                virtqueue_disable_intr_split(vq);
@@ -425,7 +443,7 @@ virtqueue_enable_intr_split(struct virtqueue *vq)
 static inline void
 virtqueue_enable_intr(struct virtqueue *vq)
 {
-       if (vtpci_packed_queue(vq->hw))
+       if (virtio_with_packed_queue(vq->hw))
                virtqueue_enable_intr_packed(vq);
        else
                virtqueue_enable_intr_split(vq);
@@ -454,18 +472,43 @@ virtqueue_full(const struct virtqueue *vq)
 }
 
 static inline int
-virtio_get_queue_type(struct virtio_hw *hw, uint16_t vtpci_queue_idx)
+virtio_get_queue_type(struct virtio_hw *hw, uint16_t vq_idx)
 {
-       if (vtpci_queue_idx == hw->max_queue_pairs * 2)
+       if (vq_idx == hw->max_queue_pairs * 2)
                return VTNET_CQ;
-       else if (vtpci_queue_idx % 2 == 0)
+       else if (vq_idx % 2 == 0)
                return VTNET_RQ;
        else
                return VTNET_TQ;
 }
 
-#define VIRTQUEUE_NUSED(vq) ((uint16_t)((vq)->vq_split.ring.used->idx - \
-                                       (vq)->vq_used_cons_idx))
+/* virtqueue_nused has load-acquire or rte_io_rmb insed */
+static inline uint16_t
+virtqueue_nused(const struct virtqueue *vq)
+{
+       uint16_t idx;
+
+       if (vq->hw->weak_barriers) {
+       /**
+        * x86 prefers to using rte_smp_rmb over __atomic_load_n as it
+        * reports a slightly better perf, which comes from the saved
+        * branch by the compiler.
+        * The if and else branches are identical with the smp and io
+        * barriers both defined as compiler barriers on x86.
+        */
+#ifdef RTE_ARCH_X86_64
+               idx = vq->vq_split.ring.used->idx;
+               rte_smp_rmb();
+#else
+               idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx,
+                               __ATOMIC_ACQUIRE);
+#endif
+       } else {
+               idx = vq->vq_split.ring.used->idx;
+               rte_io_rmb();
+       }
+       return idx - vq->vq_used_cons_idx;
+}
 
 void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx);
 void vq_ring_free_chain_packed(struct virtqueue *vq, uint16_t used_idx);
@@ -475,8 +518,24 @@ void vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx,
 static inline void
 vq_update_avail_idx(struct virtqueue *vq)
 {
-       virtio_wmb(vq->hw->weak_barriers);
-       vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
+       if (vq->hw->weak_barriers) {
+       /* x86 prefers to using rte_smp_wmb over __atomic_store_n as
+        * it reports a slightly better perf, which comes from the
+        * saved branch by the compiler.
+        * The if and else branches are identical with the smp and
+        * io barriers both defined as compiler barriers on x86.
+        */
+#ifdef RTE_ARCH_X86_64
+               rte_smp_wmb();
+               vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
+#else
+               __atomic_store_n(&vq->vq_split.ring.avail->idx,
+                                vq->vq_avail_idx, __ATOMIC_RELEASE);
+#endif
+       } else {
+               rte_io_wmb();
+               vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
+       }
 }
 
 static inline void
@@ -528,15 +587,16 @@ virtqueue_kick_prepare_packed(struct virtqueue *vq)
 static inline void
 virtqueue_notify(struct virtqueue *vq)
 {
-       VTPCI_OPS(vq->hw)->notify_queue(vq->hw, vq);
+       VIRTIO_OPS(vq->hw)->notify_queue(vq->hw, vq);
 }
 
 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
 #define VIRTQUEUE_DUMP(vq) do { \
        uint16_t used_idx, nused; \
-       used_idx = (vq)->vq_split.ring.used->idx; \
+       used_idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx, \
+                                  __ATOMIC_RELAXED); \
        nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \
-       if (vtpci_packed_queue((vq)->hw)) { \
+       if (virtio_with_packed_queue((vq)->hw)) { \
                PMD_INIT_LOG(DEBUG, \
                "VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \
                " cached_flags=0x%x; used_wrap_counter=%d", \
@@ -549,13 +609,344 @@ virtqueue_notify(struct virtqueue *vq)
          "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \
          " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \
          " avail.flags=0x%x; used.flags=0x%x", \
-         (vq)->vq_nentries, (vq)->vq_free_cnt, nused, \
-         (vq)->vq_desc_head_idx, (vq)->vq_split.ring.avail->idx, \
-         (vq)->vq_used_cons_idx, (vq)->vq_split.ring.used->idx, \
+         (vq)->vq_nentries, (vq)->vq_free_cnt, nused, (vq)->vq_desc_head_idx, \
+         (vq)->vq_split.ring.avail->idx, (vq)->vq_used_cons_idx, \
+         __atomic_load_n(&(vq)->vq_split.ring.used->idx, __ATOMIC_RELAXED), \
          (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \
 } while (0)
 #else
 #define VIRTQUEUE_DUMP(vq) do { } while (0)
 #endif
 
+/* avoid write operation when necessary, to lessen cache issues */
+#define ASSIGN_UNLESS_EQUAL(var, val) do {     \
+       typeof(var) *const var_ = &(var);       \
+       typeof(val)  const val_ = (val);        \
+       if (*var_ != val_)                      \
+               *var_ = val_;                   \
+} while (0)
+
+#define virtqueue_clear_net_hdr(hdr) do {              \
+       typeof(hdr) hdr_ = (hdr);                       \
+       ASSIGN_UNLESS_EQUAL((hdr_)->csum_start, 0);     \
+       ASSIGN_UNLESS_EQUAL((hdr_)->csum_offset, 0);    \
+       ASSIGN_UNLESS_EQUAL((hdr_)->flags, 0);          \
+       ASSIGN_UNLESS_EQUAL((hdr_)->gso_type, 0);       \
+       ASSIGN_UNLESS_EQUAL((hdr_)->gso_size, 0);       \
+       ASSIGN_UNLESS_EQUAL((hdr_)->hdr_len, 0);        \
+} while (0)
+
+static inline void
+virtqueue_xmit_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *cookie)
+{
+       uint64_t csum_l4 = cookie->ol_flags & PKT_TX_L4_MASK;
+       uint16_t o_l23_len = (cookie->ol_flags & PKT_TX_TUNNEL_MASK) ?
+                            cookie->outer_l2_len + cookie->outer_l3_len : 0;
+
+       if (cookie->ol_flags & PKT_TX_TCP_SEG)
+               csum_l4 |= PKT_TX_TCP_CKSUM;
+
+       switch (csum_l4) {
+       case PKT_TX_UDP_CKSUM:
+               hdr->csum_start = o_l23_len + cookie->l2_len + cookie->l3_len;
+               hdr->csum_offset = offsetof(struct rte_udp_hdr, dgram_cksum);
+               hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
+               break;
+
+       case PKT_TX_TCP_CKSUM:
+               hdr->csum_start = o_l23_len + cookie->l2_len + cookie->l3_len;
+               hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
+               hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
+               break;
+
+       default:
+               ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
+               ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
+               ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
+               break;
+       }
+
+       /* TCP Segmentation Offload */
+       if (cookie->ol_flags & PKT_TX_TCP_SEG) {
+               hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
+                       VIRTIO_NET_HDR_GSO_TCPV6 :
+                       VIRTIO_NET_HDR_GSO_TCPV4;
+               hdr->gso_size = cookie->tso_segsz;
+               hdr->hdr_len = o_l23_len + cookie->l2_len + cookie->l3_len +
+                              cookie->l4_len;
+       } else {
+               ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
+               ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
+               ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
+       }
+}
+
+static inline void
+virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
+                             uint16_t needed, int use_indirect, int can_push,
+                             int in_order)
+{
+       struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
+       struct vq_desc_extra *dxp;
+       struct virtqueue *vq = virtnet_txq_to_vq(txvq);
+       struct vring_packed_desc *start_dp, *head_dp;
+       uint16_t idx, id, head_idx, head_flags;
+       int16_t head_size = vq->hw->vtnet_hdr_size;
+       struct virtio_net_hdr *hdr;
+       uint16_t prev;
+       bool prepend_header = false;
+       uint16_t seg_num = cookie->nb_segs;
+
+       id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
+
+       dxp = &vq->vq_descx[id];
+       dxp->ndescs = needed;
+       dxp->cookie = cookie;
+
+       head_idx = vq->vq_avail_idx;
+       idx = head_idx;
+       prev = head_idx;
+       start_dp = vq->vq_packed.ring.desc;
+
+       head_dp = &vq->vq_packed.ring.desc[idx];
+       head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
+       head_flags |= vq->vq_packed.cached_flags;
+
+       if (can_push) {
+               /* prepend cannot fail, checked by caller */
+               hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
+                                             -head_size);
+               prepend_header = true;
+
+               /* if offload disabled, it is not zeroed below, do it now */
+               if (!vq->hw->has_tx_offload)
+                       virtqueue_clear_net_hdr(hdr);
+       } else if (use_indirect) {
+               /* setup tx ring slot to point to indirect
+                * descriptor list stored in reserved region.
+                *
+                * the first slot in indirect ring is already preset
+                * to point to the header in reserved region
+                */
+               start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
+                       RTE_PTR_DIFF(&txr[idx].tx_packed_indir, txr);
+               start_dp[idx].len   = (seg_num + 1) *
+                       sizeof(struct vring_packed_desc);
+               /* Packed descriptor id needs to be restored when inorder. */
+               if (in_order)
+                       start_dp[idx].id = idx;
+               /* reset flags for indirect desc */
+               head_flags = VRING_DESC_F_INDIRECT;
+               head_flags |= vq->vq_packed.cached_flags;
+               hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
+
+               /* loop below will fill in rest of the indirect elements */
+               start_dp = txr[idx].tx_packed_indir;
+               idx = 1;
+       } else {
+               /* setup first tx ring slot to point to header
+                * stored in reserved region.
+                */
+               start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
+                       RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
+               start_dp[idx].len   = vq->hw->vtnet_hdr_size;
+               hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
+               idx++;
+               if (idx >= vq->vq_nentries) {
+                       idx -= vq->vq_nentries;
+                       vq->vq_packed.cached_flags ^=
+                               VRING_PACKED_DESC_F_AVAIL_USED;
+               }
+       }
+
+       if (vq->hw->has_tx_offload)
+               virtqueue_xmit_offload(hdr, cookie);
+
+       do {
+               uint16_t flags;
+
+               start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
+               start_dp[idx].len  = cookie->data_len;
+               if (prepend_header) {
+                       start_dp[idx].addr -= head_size;
+                       start_dp[idx].len += head_size;
+                       prepend_header = false;
+               }
+
+               if (likely(idx != head_idx)) {
+                       flags = cookie->next ? VRING_DESC_F_NEXT : 0;
+                       flags |= vq->vq_packed.cached_flags;
+                       start_dp[idx].flags = flags;
+               }
+               prev = idx;
+               idx++;
+               if (idx >= vq->vq_nentries) {
+                       idx -= vq->vq_nentries;
+                       vq->vq_packed.cached_flags ^=
+                               VRING_PACKED_DESC_F_AVAIL_USED;
+               }
+       } while ((cookie = cookie->next) != NULL);
+
+       start_dp[prev].id = id;
+
+       if (use_indirect) {
+               idx = head_idx;
+               if (++idx >= vq->vq_nentries) {
+                       idx -= vq->vq_nentries;
+                       vq->vq_packed.cached_flags ^=
+                               VRING_PACKED_DESC_F_AVAIL_USED;
+               }
+       }
+
+       vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
+       vq->vq_avail_idx = idx;
+
+       if (!in_order) {
+               vq->vq_desc_head_idx = dxp->next;
+               if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
+                       vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
+       }
+
+       virtqueue_store_flags_packed(head_dp, head_flags,
+                                    vq->hw->weak_barriers);
+}
+
+static void
+vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
+{
+       struct vq_desc_extra *dxp;
+
+       dxp = &vq->vq_descx[id];
+       vq->vq_free_cnt += dxp->ndescs;
+
+       if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
+               vq->vq_desc_head_idx = id;
+       else
+               vq->vq_descx[vq->vq_desc_tail_idx].next = id;
+
+       vq->vq_desc_tail_idx = id;
+       dxp->next = VQ_RING_DESC_CHAIN_END;
+}
+
+static void
+virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, uint16_t num)
+{
+       uint16_t used_idx, id, curr_id, free_cnt = 0;
+       uint16_t size = vq->vq_nentries;
+       struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
+       struct vq_desc_extra *dxp;
+       int nb = num;
+
+       used_idx = vq->vq_used_cons_idx;
+       /* desc_is_used has a load-acquire or rte_io_rmb inside
+        * and wait for used desc in virtqueue.
+        */
+       while (nb > 0 && desc_is_used(&desc[used_idx], vq)) {
+               id = desc[used_idx].id;
+               do {
+                       curr_id = used_idx;
+                       dxp = &vq->vq_descx[used_idx];
+                       used_idx += dxp->ndescs;
+                       free_cnt += dxp->ndescs;
+                       nb -= dxp->ndescs;
+                       if (used_idx >= size) {
+                               used_idx -= size;
+                               vq->vq_packed.used_wrap_counter ^= 1;
+                       }
+                       if (dxp->cookie != NULL) {
+                               rte_pktmbuf_free(dxp->cookie);
+                               dxp->cookie = NULL;
+                       }
+               } while (curr_id != id);
+       }
+       vq->vq_used_cons_idx = used_idx;
+       vq->vq_free_cnt += free_cnt;
+}
+
+static void
+virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, uint16_t num)
+{
+       uint16_t used_idx, id;
+       uint16_t size = vq->vq_nentries;
+       struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
+       struct vq_desc_extra *dxp;
+
+       used_idx = vq->vq_used_cons_idx;
+       /* desc_is_used has a load-acquire or rte_io_rmb inside
+        * and wait for used desc in virtqueue.
+        */
+       while (num-- && desc_is_used(&desc[used_idx], vq)) {
+               id = desc[used_idx].id;
+               dxp = &vq->vq_descx[id];
+               vq->vq_used_cons_idx += dxp->ndescs;
+               if (vq->vq_used_cons_idx >= size) {
+                       vq->vq_used_cons_idx -= size;
+                       vq->vq_packed.used_wrap_counter ^= 1;
+               }
+               vq_ring_free_id_packed(vq, id);
+               if (dxp->cookie != NULL) {
+                       rte_pktmbuf_free(dxp->cookie);
+                       dxp->cookie = NULL;
+               }
+               used_idx = vq->vq_used_cons_idx;
+       }
+}
+
+/* Cleanup from completed transmits. */
+static inline void
+virtio_xmit_cleanup_packed(struct virtqueue *vq, uint16_t num, int in_order)
+{
+       if (in_order)
+               virtio_xmit_cleanup_inorder_packed(vq, num);
+       else
+               virtio_xmit_cleanup_normal_packed(vq, num);
+}
+
+static inline void
+virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
+{
+       uint16_t i, used_idx, desc_idx;
+       for (i = 0; i < num; i++) {
+               struct vring_used_elem *uep;
+               struct vq_desc_extra *dxp;
+
+               used_idx = (uint16_t)(vq->vq_used_cons_idx &
+                               (vq->vq_nentries - 1));
+               uep = &vq->vq_split.ring.used->ring[used_idx];
+
+               desc_idx = (uint16_t)uep->id;
+               dxp = &vq->vq_descx[desc_idx];
+               vq->vq_used_cons_idx++;
+               vq_ring_free_chain(vq, desc_idx);
+
+               if (dxp->cookie != NULL) {
+                       rte_pktmbuf_free(dxp->cookie);
+                       dxp->cookie = NULL;
+               }
+       }
+}
+
+/* Cleanup from completed inorder transmits. */
+static __rte_always_inline void
+virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
+{
+       uint16_t i, idx = vq->vq_used_cons_idx;
+       int16_t free_cnt = 0;
+       struct vq_desc_extra *dxp = NULL;
+
+       if (unlikely(num == 0))
+               return;
+
+       for (i = 0; i < num; i++) {
+               dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
+               free_cnt += dxp->ndescs;
+               if (dxp->cookie != NULL) {
+                       rte_pktmbuf_free(dxp->cookie);
+                       dxp->cookie = NULL;
+               }
+       }
+
+       vq->vq_free_cnt += free_cnt;
+       vq->vq_used_cons_idx = idx;
+}
 #endif /* _VIRTQUEUE_H_ */