ucs_status_t uct_rc_verbs_ep_flush(uct_ep_h tl_ep, unsigned flags,
uct_completion_t *comp)
{
uct_rc_verbs_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_rc_verbs_iface_t);
uct_rc_verbs_ep_t *ep = ucs_derived_of(tl_ep, uct_rc_verbs_ep_t);
ucs_status_t status;
if (!uct_rc_iface_has_tx_resources(&iface->super)) {
return UCS_ERR_NO_RESOURCE;
}
if (uct_rc_txqp_available(&ep->super.txqp) == iface->config.tx_max_wr) {
UCT_TL_EP_STAT_FLUSH(&ep->super.super);
return UCS_OK;
}
if (uct_rc_txqp_unsignaled(&ep->super.txqp) != 0) {
if (IBV_DEVICE_HAS_NOP(&uct_ib_iface_device(&iface->super.super)->dev_attr)) {
status = uct_rc_verbs_ep_nop(ep);
} else {
status = uct_rc_verbs_ep_put_short(tl_ep, NULL, 0, 0, 0);
}
if (status != UCS_OK) {
return status;
}
} else if (!uct_rc_ep_has_tx_resources(&ep->super)) {
return UCS_ERR_NO_RESOURCE;
}
uct_rc_txqp_add_send_comp(&iface->super, &ep->super.txqp, comp, ep->txcnt.pi);
UCT_TL_EP_STAT_FLUSH_WAIT(&ep->super.super);
return UCS_INPROGRESS;
}
static UCS_F_ALWAYS_INLINE void
uct_dc_mlx5_iface_zcopy_post(uct_dc_mlx5_iface_t *iface, uct_dc_mlx5_ep_t *ep,
unsigned opcode, const uct_iov_t *iov, size_t iovcnt,
/* SEND */ uint8_t am_id, const void *am_hdr, unsigned am_hdr_len,
/* RDMA */ uint64_t rdma_raddr, uct_rkey_t rdma_rkey,
/* TAG */ uct_tag_t tag, uint32_t app_ctx, uint32_t ib_imm_be,
uct_completion_t *comp, uint8_t send_flags)
{
uint16_t sn;
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
UCT_DC_MLX5_IFACE_TXQP_GET(iface, ep, txqp, txwq);
sn = txwq->sw_pi;
uct_rc_mlx5_txqp_dptr_post_iov(&iface->super, UCT_IB_QPT_DCI, txqp,
txwq, opcode, iov, iovcnt,
am_id, am_hdr, am_hdr_len,
rdma_raddr, uct_ib_md_direct_rkey(rdma_rkey),
tag, app_ctx, ib_imm_be,
&ep->av, uct_dc_mlx5_ep_get_grh(ep),
uct_ib_mlx5_wqe_av_size(&ep->av),
MLX5_WQE_CTRL_CQ_UPDATE | send_flags,
UCT_IB_MAX_ZCOPY_LOG_SGE(&iface->super.super.super));
uct_rc_txqp_add_send_comp(&iface->super.super, txqp, comp, sn,
UCT_RC_IFACE_SEND_OP_FLAG_ZCOPY);
}
/*
* Helper function for zero-copy post.
* Adds user completion to the callback queue.
*/
static UCS_F_ALWAYS_INLINE ucs_status_t
uct_rc_mlx5_ep_zcopy_post(uct_rc_mlx5_ep_t *ep,
unsigned opcode, const uct_iov_t *iov, size_t iovcnt,
/* SEND */ uint8_t am_id, const void *am_hdr, unsigned am_hdr_len,
/* RDMA */ uint64_t rdma_raddr, uct_rkey_t rdma_rkey,
int force_sig, uct_completion_t *comp)
{
uct_rc_iface_t *iface = ucs_derived_of(ep->super.super.super.iface,
uct_rc_iface_t);
uint16_t sn;
UCT_RC_CHECK_RES(iface, &ep->super);
sn = ep->tx.wq.sw_pi;
uct_rc_mlx5_txqp_dptr_post_iov(iface, IBV_QPT_RC,
&ep->super.txqp, &ep->tx.wq,
opcode, iov, iovcnt,
am_id, am_hdr, am_hdr_len,
rdma_raddr, uct_ib_md_direct_rkey(rdma_rkey),
NULL, 0,
(comp == NULL) ? force_sig : MLX5_WQE_CTRL_CQ_UPDATE);
uct_rc_txqp_add_send_comp(iface, &ep->super.txqp, comp, sn);
return UCS_INPROGRESS;
}
ucs_status_t uct_rc_mlx5_ep_flush(uct_ep_h tl_ep, unsigned flags,
uct_completion_t *comp)
{
uct_rc_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_rc_mlx5_ep_t);
uct_rc_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_rc_mlx5_iface_t);
uint16_t sn;
if (!uct_rc_iface_has_tx_resources(&iface->super)) {
return UCS_ERR_NO_RESOURCE;
}
if (uct_rc_txqp_available(&ep->super.txqp) == ep->tx.wq.bb_max) {
UCT_TL_EP_STAT_FLUSH(&ep->super.super);
return UCS_OK;
}
if (uct_rc_txqp_unsignaled(&ep->super.txqp) != 0) {
sn = ep->tx.wq.sw_pi;
UCT_RC_CHECK_RES(&iface->super, &ep->super);
uct_rc_mlx5_txqp_inline_post(&iface->super, IBV_QPT_RC,
&ep->super.txqp, &ep->tx.wq,
MLX5_OPCODE_NOP, NULL, 0,
0, 0,
0, 0,
NULL, 0);
} else if (!uct_rc_ep_has_tx_resources(&ep->super)) {
return UCS_ERR_NO_RESOURCE;
} else {
sn = ep->tx.wq.sig_pi;
}
uct_rc_txqp_add_send_comp(&iface->super, &ep->super.txqp, comp, sn);
UCT_TL_EP_STAT_FLUSH_WAIT(&ep->super.super);
return UCS_INPROGRESS;
}
static inline void uct_dc_mlx5_iface_add_send_comp(uct_dc_mlx5_iface_t *iface,
uct_dc_mlx5_ep_t *ep,
uct_completion_t *comp)
{
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
UCT_DC_MLX5_IFACE_TXQP_GET(iface, ep, txqp, txwq);
uct_rc_txqp_add_send_comp(&iface->super.super, txqp, comp, txwq->sig_pi);
}
static inline ucs_status_t
uct_rc_verbs_ep_rdma_zcopy(uct_rc_verbs_ep_t *ep, const uct_iov_t *iov,
size_t iovcnt, uint64_t remote_addr, uct_rkey_t rkey,
uct_completion_t *comp, int opcode)
{
uct_rc_verbs_iface_t *iface = ucs_derived_of(ep->super.super.super.iface,
uct_rc_verbs_iface_t);
struct ibv_sge sge[UCT_IB_MAX_IOV];
struct ibv_send_wr wr;
size_t sge_cnt;
UCT_RC_CHECK_RES(&iface->super, &ep->super);
sge_cnt = uct_ib_verbs_sge_fill_iov(sge, iov, iovcnt);
UCT_SKIP_ZERO_LENGTH(sge_cnt);
UCT_RC_VERBS_FILL_RDMA_WR_IOV(wr, wr.opcode, opcode, sge, sge_cnt, remote_addr, rkey);
wr.next = NULL;
uct_rc_verbs_ep_post_send(iface, ep, &wr, IBV_SEND_SIGNALED);
uct_rc_txqp_add_send_comp(&iface->super, &ep->super.txqp, comp, ep->txcnt.pi);
return UCS_INPROGRESS;
}
static UCS_F_ALWAYS_INLINE void
uct_dc_mlx5_iface_zcopy_post(uct_dc_mlx5_iface_t *iface, uct_dc_mlx5_ep_t *ep,
unsigned opcode, const uct_iov_t *iov, size_t iovcnt,
/* SEND */ uint8_t am_id, const void *am_hdr, unsigned am_hdr_len,
/* RDMA */ uint64_t rdma_raddr, uct_rkey_t rdma_rkey,
uct_completion_t *comp)
{
uint16_t sn;
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
UCT_DC_MLX5_IFACE_TXQP_GET(iface, ep, txqp, txwq);
sn = txwq->sw_pi;
uct_rc_mlx5_txqp_dptr_post_iov(&iface->super.super, IBV_EXP_QPT_DC_INI, txqp,
txwq, opcode, iov, iovcnt,
am_id, am_hdr, am_hdr_len,
rdma_raddr, uct_ib_md_direct_rkey(rdma_rkey),
&ep->av, uct_ib_mlx5_wqe_av_size(&ep->av),
MLX5_WQE_CTRL_CQ_UPDATE);
uct_rc_txqp_add_send_comp(&iface->super.super, txqp, comp, sn);
}
