void ci_sock_cmn_timestamp_q_enqueue(ci_netif* ni, ci_sock_cmn* s,
ci_ip_pkt_fmt* pkt)
{
ci_ip_pkt_queue* qu = &s->timestamp_q;
oo_pkt_p prev_head = qu->head;
/* This part is effectively ci_ip_queue_enqueue(ni, &s->timestamp_q, p);
* but inlined to allow using tsq_next field
*/
pkt->tsq_next = OO_PP_NULL;
if( ci_ip_queue_is_empty(qu) ) {
ci_assert(OO_PP_IS_NULL(qu->head));
qu->head = OO_PKT_P(pkt);
}
else {
ci_assert(OO_PP_NOT_NULL(qu->head));
/* This assumes the netif lock is held, so use
ci_ip_queue_enqueue_nnl() if it's not */
PKT(ni, qu->tail)->tsq_next = OO_PKT_P(pkt);
}
qu->tail = OO_PKT_P(pkt);
qu->num++;
if( OO_PP_IS_NULL(prev_head) ) {
ci_assert(OO_PP_IS_NULL(s->timestamp_q_extract));
s->timestamp_q_extract = qu->head;
}
else {
ci_sock_cmn_timestamp_q_reap(ni, s);
}
/* Tells post-poll loop to put socket on the [reap_list]. */
s->b.sb_flags |= CI_SB_FLAG_RX_DELIVERED;
}
static void ci_udp_pkt_to_zc_msg(ci_netif* ni, ci_ip_pkt_fmt* pkt,
struct onload_zc_msg* zc_msg)
{
int i, bytes_left = pkt->pf.udp.pay_len;
ci_ip_pkt_fmt* frag;
ci_ip_pkt_fmt* handle_frag;
handle_frag = frag = pkt;
i = 0;
ci_assert_nequal(zc_msg->iov, NULL);
/* Ignore first frag if zero length and there is another frag, but
* still pass the zero-length buffer as the onload_zc_handle so it
* will get freed correctly
*/
if( oo_offbuf_left(&frag->buf) == 0 && OO_PP_NOT_NULL(frag->frag_next) )
frag = PKT_CHK_NNL(ni, frag->frag_next);
do {
zc_msg->iov[i].iov_len = CI_MIN(oo_offbuf_left(&frag->buf),
bytes_left);
zc_msg->iov[i].iov_base = oo_offbuf_ptr(&frag->buf);
zc_msg->iov[i].buf = (onload_zc_handle)handle_frag;
zc_msg->iov[i].iov_flags = 0;
bytes_left -= zc_msg->iov[i].iov_len;
++i;
if( OO_PP_IS_NULL(frag->frag_next) ||
(i == CI_UDP_ZC_IOVEC_MAX) ||
(bytes_left == 0) )
break;
frag = PKT_CHK_NNL(ni, frag->frag_next);
handle_frag = frag;
} while( 1 );
zc_msg->msghdr.msg_iovlen = i;
}
void ci_sock_cmn_timestamp_q_drop(ci_netif* netif, ci_sock_cmn* s)
{
ci_ip_pkt_queue* qu = &s->timestamp_q;
ci_ip_pkt_fmt* p;
CI_DEBUG(int i = qu->num);
ci_assert(netif);
ci_assert(qu);
while( OO_PP_NOT_NULL(qu->head) CI_DEBUG( && i-- > 0) ) {
p = PKT_CHK(netif, qu->head);
qu->head = p->tsq_next;
ci_netif_pkt_release(netif, p);
}
ci_assert_equal(i, 0);
ci_assert(OO_PP_IS_NULL(qu->head));
qu->num = 0;
}
static int ci_udp_recvmsg_socklocked_slowpath(ci_udp_iomsg_args* a,
ci_msghdr* msg,
ci_iovec_ptr *piov, int flags)
{
int rc = 0;
ci_netif* ni = a->ni;
ci_udp_state* us = a->us;
if(CI_UNLIKELY( ni->state->rxq_low ))
ci_netif_rxq_low_on_recv(ni, &us->s,
1 /* assume at least one pkt freed */);
/* In the kernel recv() with flags is not called.
* only read(). So flags may only contain MSG_DONTWAIT */
#ifdef __KERNEL__
ci_assert_equal(flags, 0);
#endif
#ifndef __KERNEL__
if( flags & MSG_ERRQUEUE_CHK ) {
if( OO_PP_NOT_NULL(us->timestamp_q.extract) ) {
ci_ip_pkt_fmt* pkt;
struct timespec ts[3];
struct cmsg_state cmsg_state;
ci_udp_hdr* udp;
int paylen;
/* TODO is this necessary? - mirroring ci_udp_recvmsg_get() */
ci_rmb();
pkt = PKT_CHK_NNL(ni, us->timestamp_q.extract);
if( pkt->tx_hw_stamp.tv_sec == CI_PKT_TX_HW_STAMP_CONSUMED ) {
if( OO_PP_IS_NULL(pkt->tsq_next) )
goto errqueue_empty;
us->timestamp_q.extract = pkt->tsq_next;
pkt = PKT_CHK_NNL(ni, us->timestamp_q.extract);
ci_assert(pkt->tx_hw_stamp.tv_sec != CI_PKT_TX_HW_STAMP_CONSUMED);
}
udp = oo_ip_data(pkt);
paylen = CI_BSWAP_BE16(oo_ip_hdr(pkt)->ip_tot_len_be16) -
sizeof(ci_ip4_hdr) - sizeof(udp);
msg->msg_flags = 0;
cmsg_state.msg = msg;
cmsg_state.cm = msg->msg_control;
cmsg_state.cmsg_bytes_used = 0;
ci_iovec_ptr_init_nz(piov, msg->msg_iov, msg->msg_iovlen);
memset(ts, 0, sizeof(ts));
if( us->s.timestamping_flags & ONLOAD_SOF_TIMESTAMPING_RAW_HARDWARE ) {
ts[2].tv_sec = pkt->tx_hw_stamp.tv_sec;
ts[2].tv_nsec = pkt->tx_hw_stamp.tv_nsec;
}
if( (us->s.timestamping_flags & ONLOAD_SOF_TIMESTAMPING_SYS_HARDWARE) &&
(pkt->tx_hw_stamp.tv_nsec & CI_IP_PKT_HW_STAMP_FLAG_IN_SYNC) ) {
ts[1].tv_sec = pkt->tx_hw_stamp.tv_sec;
ts[1].tv_nsec = pkt->tx_hw_stamp.tv_nsec;
}
ci_put_cmsg(&cmsg_state, SOL_SOCKET, ONLOAD_SCM_TIMESTAMPING,
sizeof(ts), &ts);
oo_offbuf_set_start(&pkt->buf, udp + 1);
oo_offbuf_set_len(&pkt->buf, paylen);
rc = oo_copy_pkt_to_iovec_no_adv(ni, pkt, piov, paylen);
/* Mark this packet/timestamp as consumed */
pkt->tx_hw_stamp.tv_sec = CI_PKT_TX_HW_STAMP_CONSUMED;
ci_ip_cmsg_finish(&cmsg_state);
msg->msg_flags |= MSG_ERRQUEUE_CHK;
return rc;
}
errqueue_empty:
/* ICMP is handled via OS, so get OS error */
rc = oo_os_sock_recvmsg(ni, SC_SP(&us->s), msg, flags);
if( rc < 0 ) {
ci_assert(-rc == errno);
return -1;
}
else
return rc;
}
#endif
if( (rc = ci_get_so_error(&us->s)) != 0 ) {
CI_SET_ERROR(rc, rc);
return rc;
}
if( msg->msg_iovlen > 0 && msg->msg_iov == NULL ) {
CI_SET_ERROR(rc, EFAULT);
return rc;
}
#if MSG_OOB_CHK
if( flags & MSG_OOB_CHK ) {
CI_SET_ERROR(rc, EOPNOTSUPP);
return rc;
}
#endif
#if CI_CFG_POSIX_RECV
if( ! udp_lport_be16(us)) {
LOG_UV(log("%s: -1 (ENOTCONN)", __FUNCTION__));
CI_SET_ERROR(rc, ENOTCONN);
return rc;
}
#endif
if( msg->msg_iovlen == 0 ) {
/* We have a difference in behaviour from the Linux stack here. When
** msg_iovlen is 0 Linux 2.4.21-15.EL does not set MSG_TRUNC when a
** datagram has non-zero length. We do. */
CI_IOVEC_LEN(&piov->io) = piov->iovlen = 0;
return IOVLEN_WORKAROUND_RC_VALUE;
}
return 0;
}
