static ucs_status_t
uct_ud_mlx5_ep_am_short(uct_ep_h tl_ep, uint8_t id, uint64_t hdr,
const void *buffer, unsigned length)
{
uct_ud_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_ud_mlx5_ep_t);
uct_ud_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface,
uct_ud_mlx5_iface_t);
struct mlx5_wqe_ctrl_seg *ctrl;
struct mlx5_wqe_inl_data_seg *inl;
uct_ud_am_short_hdr_t *am;
uct_ud_neth_t *neth;
unsigned wqe_size;
uct_ud_send_skb_t *skb;
/* data a written directly into tx wqe, so it is impossible to use
* common ud am code
*/
UCT_CHECK_AM_ID(id);
UCT_CHECK_LENGTH(sizeof(uct_ud_neth_t) + sizeof(hdr) + length,
iface->super.config.max_inline, "am_short");
uct_ud_enter(&iface->super);
uct_ud_iface_progress_pending_tx(&iface->super);
skb = uct_ud_ep_get_tx_skb(&iface->super, &ep->super);
if (!skb) {
uct_ud_leave(&iface->super);
return UCS_ERR_NO_RESOURCE;
}
ctrl = iface->tx.wq.curr;
/* Set inline segment which has AM id, AM header, and AM payload */
inl = uct_ib_mlx5_get_next_seg(&iface->tx.wq, ctrl, UCT_UD_MLX5_WQE_SIZE);
wqe_size = length + sizeof(*am) + sizeof(*neth);
inl->byte_count = htonl(wqe_size | MLX5_INLINE_SEG);
/* assume that neth and am header fit into one bb */
ucs_assert(sizeof(*am) + sizeof(*neth) < MLX5_SEND_WQE_BB);
neth = (void*)(inl + 1);
uct_ud_am_set_neth(neth, &ep->super, id);
am = (void*)(neth + 1);
am->hdr = hdr;
uct_ib_mlx5_inline_copy(am + 1, buffer, length, &iface->tx.wq);
wqe_size += UCT_UD_MLX5_WQE_SIZE + sizeof(*inl);
UCT_CHECK_LENGTH(wqe_size, UCT_IB_MLX5_MAX_BB * MLX5_SEND_WQE_BB,
"am_short");
UCT_UD_EP_HOOK_CALL_TX(&ep->super, neth);
uct_ud_mlx5_post_send(iface, ep, ctrl, wqe_size);
skb->len = sizeof(*neth) + sizeof(*am);
memcpy(skb->neth, neth, skb->len);
uct_ud_iface_complete_tx_inl(&iface->super, &ep->super, skb,
(char *)skb->neth + skb->len, buffer, length);
UCT_TL_EP_STAT_OP(&ep->super.super, AM, SHORT, sizeof(hdr) + length);
uct_ud_leave(&iface->super);
return UCS_OK;
}
static ucs_status_t
uct_ud_mlx5_ep_put_short(uct_ep_h tl_ep, const void *buffer, unsigned length,
uint64_t remote_addr, uct_rkey_t rkey)
{
uct_ud_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_ud_mlx5_ep_t);
uct_ud_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface,
uct_ud_mlx5_iface_t);
struct mlx5_wqe_ctrl_seg *ctrl;
struct mlx5_wqe_inl_data_seg *inl;
unsigned wqe_size;
uct_ud_put_hdr_t *put_hdr;
uct_ud_neth_t *neth;
uct_ud_send_skb_t *skb;
uct_ud_enter(&iface->super);
uct_ud_iface_progress_pending_tx(&iface->super);
skb = uct_ud_ep_get_tx_skb(&iface->super, &ep->super);
if (!skb) {
uct_ud_leave(&iface->super);
return UCS_ERR_NO_RESOURCE;
}
ctrl = iface->tx.wq.curr;
/* Set inline segment which has AM id, AM header, and AM payload */
inl = uct_ib_mlx5_get_next_seg(&iface->tx.wq, ctrl, UCT_UD_MLX5_WQE_SIZE);
wqe_size = length + sizeof(*put_hdr) + sizeof(*neth);
inl->byte_count = htonl(wqe_size | MLX5_INLINE_SEG);
/* assume that neth and am header fit into one bb */
ucs_assert(sizeof(*put_hdr) + sizeof(*neth) < MLX5_SEND_WQE_BB);
neth = (void*)(inl + 1);
uct_ud_neth_init_data(&ep->super, neth);
uct_ud_neth_set_type_put(&ep->super, neth);
uct_ud_neth_ack_req(&ep->super, neth);
put_hdr = (uct_ud_put_hdr_t *)(neth+1);
put_hdr->rva = remote_addr;
uct_ib_mlx5_inline_copy(put_hdr + 1, buffer, length, &iface->tx.wq);
wqe_size += UCT_UD_MLX5_WQE_SIZE + sizeof(*inl);
UCT_CHECK_LENGTH(wqe_size, UCT_IB_MLX5_MAX_BB * MLX5_SEND_WQE_BB,
"put_short");
UCT_UD_EP_HOOK_CALL_TX(&ep->super, neth);
uct_ud_mlx5_post_send(iface, ep, ctrl, wqe_size);
skb->len = sizeof(*neth) + sizeof(*put_hdr);
memcpy(skb->neth, neth, skb->len);
uct_ud_iface_complete_tx_inl(&iface->super, &ep->super, skb,
(char *)skb->neth + skb->len, buffer, length);
UCT_TL_EP_STAT_OP(&ep->super.super, PUT, SHORT, length);
uct_ud_leave(&iface->super);
return UCS_OK;
}
static UCS_F_ALWAYS_INLINE void
uct_ud_mlx5_ep_tx_skb(uct_ud_mlx5_iface_t *iface, uct_ud_mlx5_ep_t *ep,
uct_ud_send_skb_t *skb)
{
struct mlx5_wqe_ctrl_seg *ctrl;
struct mlx5_wqe_data_seg *dptr;
ctrl = iface->tx.wq.curr;
dptr = uct_ib_mlx5_get_next_seg(&iface->tx.wq, ctrl, UCT_UD_MLX5_WQE_SIZE);
uct_ib_mlx5_set_data_seg(dptr, skb->neth, skb->len, skb->lkey);
UCT_UD_EP_HOOK_CALL_TX(&ep->super, skb->neth);
uct_ud_mlx5_post_send(iface, ucs_derived_of(ep, uct_ud_mlx5_ep_t),
ctrl, UCT_UD_MLX5_WQE_SIZE + sizeof(*dptr));
}
static inline void
uct_ud_mlx5_ep_tx_inl(uct_ud_mlx5_iface_t *iface, uct_ud_mlx5_ep_t *ep,
const void *buf, unsigned length)
{
struct mlx5_wqe_ctrl_seg *ctrl;
struct mlx5_wqe_inl_data_seg *inl;
ctrl = iface->tx.wq.curr;
inl = uct_ib_mlx5_get_next_seg(&iface->tx.wq, ctrl, UCT_UD_MLX5_WQE_SIZE);
inl->byte_count = htonl(length | MLX5_INLINE_SEG);
uct_ib_mlx5_inline_copy(inl + 1, buf, length, &iface->tx.wq);
UCT_UD_EP_HOOK_CALL_TX(&ep->super, (uct_ud_neth_t *)buf);
uct_ud_mlx5_post_send(iface, ep, ctrl,
UCT_UD_MLX5_WQE_SIZE + sizeof(*inl) + length);
}