static inline uint64_t
select_ddp_flags(struct socket *so, int flags, int db_idx)
{
uint64_t ddp_flags = V_TF_DDP_INDICATE_OUT(0);
int waitall = flags & MSG_WAITALL;
int nb = so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO);
KASSERT(db_idx == 0 || db_idx == 1,
("%s: bad DDP buffer index %d", __func__, db_idx));
if (db_idx == 0) {
ddp_flags |= V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_ACTIVE_BUF(0);
if (waitall)
ddp_flags |= V_TF_DDP_PUSH_DISABLE_0(1);
else if (nb)
ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
else
ddp_flags |= V_TF_DDP_BUF0_FLUSH(0);
} else {
ddp_flags |= V_TF_DDP_BUF1_VALID(1) | V_TF_DDP_ACTIVE_BUF(1);
if (waitall)
ddp_flags |= V_TF_DDP_PUSH_DISABLE_1(1);
else if (nb)
ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
else
ddp_flags |= V_TF_DDP_BUF1_FLUSH(0);
}
return (ddp_flags);
}
void
enable_ddp(struct adapter *sc, struct toepcb *toep)
{
KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
("%s: toep %p has bad ddp_flags 0x%x",
__func__, toep, toep->ddp_flags));
CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
__func__, toep->tid, time_uptime);
toep->ddp_flags |= DDP_SC_REQ;
t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS,
V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1));
t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS,
V_TF_RCV_COALESCE_ENABLE(1), 0);
}
static struct wrqe *
mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
int offset, uint64_t ddp_flags)
{
struct ddp_buffer *db = toep->db[db_idx];
struct wrqe *wr;
struct work_request_hdr *wrh;
struct ulp_txpkt *ulpmc;
int len;
KASSERT(db_idx == 0 || db_idx == 1,
("%s: bad DDP buffer index %d", __func__, db_idx));
/*
* We'll send a compound work request that has 3 SET_TCB_FIELDs and an
* RX_DATA_ACK (with RX_MODULATE to speed up delivery).
*
* The work request header is 16B and always ends at a 16B boundary.
* The ULPTX master commands that follow must all end at 16B boundaries
* too so we round up the size to 16.
*/
len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
roundup2(LEN__RX_DATA_ACK_ULP, 16);
wr = alloc_wrqe(len, toep->ctrlq);
if (wr == NULL)
return (NULL);
wrh = wrtod(wr);
INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */
ulpmc = (struct ulp_txpkt *)(wrh + 1);
/* Write the buffer's tag */
ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
W_TCB_RX_DDP_BUF0_TAG + db_idx,
V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
V_TCB_RX_DDP_BUF0_TAG(db->tag));
/* Update the current offset in the DDP buffer and its total length */
if (db_idx == 0)
ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
W_TCB_RX_DDP_BUF0_OFFSET,
V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
V_TCB_RX_DDP_BUF0_OFFSET(offset) |
V_TCB_RX_DDP_BUF0_LEN(db->len));
else
ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
W_TCB_RX_DDP_BUF1_OFFSET,
V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
V_TCB_RX_DDP_BUF1_OFFSET(offset) |
V_TCB_RX_DDP_BUF1_LEN((u64)db->len << 32));
/* Update DDP flags */
ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF1_FLUSH(1) |
V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PUSH_DISABLE_1(1) |
V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1) |
V_TF_DDP_ACTIVE_BUF(1) | V_TF_DDP_INDICATE_OUT(1), ddp_flags);
/* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
return (wr);
}