static void sdp_poll_tx_timeout(unsigned long data)
{
struct sdp_sock *ssk = (struct sdp_sock *)data;
struct sock *sk = sk_ssk(ssk);
u32 inflight, wc_processed;
sdp_prf1(sk_ssk(ssk), NULL, "TX timeout: inflight=%d, head=%d tail=%d",
(u32) tx_ring_posted(ssk),
ring_head(ssk->tx_ring), ring_tail(ssk->tx_ring));
/* Only process if the socket is not in use */
bh_lock_sock(sk);
if (sock_owned_by_user(sk)) {
sdp_prf(sk_ssk(ssk), NULL, "TX comp: socket is busy");
if (sdp_tx_handler_select(ssk) && sk->sk_state != TCP_CLOSE &&
likely(ssk->qp_active)) {
sdp_prf1(sk, NULL, "schedule a timer");
mod_timer(&ssk->tx_ring.timer, jiffies + SDP_TX_POLL_TIMEOUT);
}
SDPSTATS_COUNTER_INC(tx_poll_busy);
goto out;
}
if (unlikely(!ssk->qp || sk->sk_state == TCP_CLOSE)) {
SDPSTATS_COUNTER_INC(tx_poll_no_op);
goto out;
}
wc_processed = sdp_process_tx_cq(ssk);
if (!wc_processed)
SDPSTATS_COUNTER_INC(tx_poll_miss);
else {
sdp_post_sends(ssk, GFP_ATOMIC);
SDPSTATS_COUNTER_INC(tx_poll_hit);
}
inflight = (u32) tx_ring_posted(ssk);
sdp_prf1(sk_ssk(ssk), NULL, "finished tx proccessing. inflight = %d",
tx_ring_posted(ssk));
/* If there are still packets in flight and the timer has not already
* been scheduled by the Tx routine then schedule it here to guarantee
* completion processing of these packets */
if (inflight && likely(ssk->qp_active))
mod_timer(&ssk->tx_ring.timer, jiffies + SDP_TX_POLL_TIMEOUT);
out:
if (ssk->tx_ring.rdma_inflight && ssk->tx_ring.rdma_inflight->busy) {
sdp_prf1(sk, NULL, "RDMA is inflight - arming irq");
sdp_arm_tx_cq(sk);
}
bh_unlock_sock(sk);
}
int
sdp_tx_ring_create(struct sdp_sock *ssk, struct ib_device *device)
{
struct ib_cq *tx_cq;
int rc = 0;
sdp_dbg(ssk->socket, "tx ring create\n");
callout_init_rw(&ssk->tx_ring.timer, &ssk->lock, 0);
callout_init_rw(&ssk->nagle_timer, &ssk->lock, 0);
atomic_set(&ssk->tx_ring.head, 1);
atomic_set(&ssk->tx_ring.tail, 1);
ssk->tx_ring.buffer = kzalloc(
sizeof *ssk->tx_ring.buffer * SDP_TX_SIZE, GFP_KERNEL);
if (!ssk->tx_ring.buffer) {
rc = -ENOMEM;
sdp_warn(ssk->socket, "Can't allocate TX Ring size %zd.\n",
sizeof(*ssk->tx_ring.buffer) * SDP_TX_SIZE);
goto out;
}
tx_cq = ib_create_cq(device, sdp_tx_irq, sdp_tx_cq_event_handler,
ssk, SDP_TX_SIZE, IB_CQ_VECTOR_LEAST_ATTACHED);
if (IS_ERR(tx_cq)) {
rc = PTR_ERR(tx_cq);
sdp_warn(ssk->socket, "Unable to allocate TX CQ: %d.\n", rc);
goto err_cq;
}
ssk->tx_ring.cq = tx_cq;
ssk->tx_ring.poll_cnt = 0;
sdp_arm_tx_cq(ssk);
return 0;
err_cq:
kfree(ssk->tx_ring.buffer);
ssk->tx_ring.buffer = NULL;
out:
return rc;
}
static int sdp_process_tx_cq(struct sdp_sock *ssk)
{
struct ib_wc ibwc[SDP_NUM_WC];
int n, i;
int wc_processed = 0;
if (!ssk->tx_ring.cq) {
sdp_dbg(sk_ssk(ssk), "tx irq on destroyed tx_cq\n");
return 0;
}
do {
n = ib_poll_cq(ssk->tx_ring.cq, SDP_NUM_WC, ibwc);
for (i = 0; i < n; ++i) {
sdp_process_tx_wc(ssk, ibwc + i);
wc_processed++;
}
} while (n == SDP_NUM_WC);
if (wc_processed) {
struct sock *sk = sk_ssk(ssk);
sdp_prf1(sk, NULL, "Waking sendmsg. inflight=%d",
(u32) tx_ring_posted(ssk));
sk_mem_reclaim(sk);
sk_stream_write_space(sk_ssk(ssk));
if (sk->sk_write_pending &&
test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
tx_ring_posted(ssk)) {
/* a write is pending and still no room in tx queue,
* arm tx cq
*/
sdp_prf(sk_ssk(ssk), NULL, "pending tx - rearming");
sdp_arm_tx_cq(sk);
}
}
return wc_processed;
}
static void
sdp_poll_tx(struct sdp_sock *ssk)
{
struct socket *sk = ssk->socket;
u32 inflight, wc_processed;
sdp_prf1(ssk->socket, NULL, "TX timeout: inflight=%d, head=%d tail=%d",
(u32) tx_ring_posted(ssk),
ring_head(ssk->tx_ring), ring_tail(ssk->tx_ring));
if (unlikely(ssk->state == TCPS_CLOSED)) {
sdp_warn(sk, "Socket is closed\n");
goto out;
}
wc_processed = sdp_process_tx_cq(ssk);
if (!wc_processed)
SDPSTATS_COUNTER_INC(tx_poll_miss);
else
SDPSTATS_COUNTER_INC(tx_poll_hit);
inflight = (u32) tx_ring_posted(ssk);
sdp_prf1(ssk->socket, NULL, "finished tx proccessing. inflight = %d",
inflight);
/* If there are still packets in flight and the timer has not already
* been scheduled by the Tx routine then schedule it here to guarantee
* completion processing of these packets */
if (inflight)
callout_reset(&ssk->tx_ring.timer, SDP_TX_POLL_TIMEOUT,
sdp_poll_tx_timeout, ssk);
out:
#ifdef SDP_ZCOPY
if (ssk->tx_ring.rdma_inflight && ssk->tx_ring.rdma_inflight->busy) {
sdp_prf1(sk, NULL, "RDMA is inflight - arming irq");
sdp_arm_tx_cq(ssk);
}
#endif
return;
}
void
sdp_post_sends(struct sdp_sock *ssk, int wait)
{
struct mbuf *mb;
int post_count = 0;
struct socket *sk;
int low;
sk = ssk->socket;
if (unlikely(!ssk->id)) {
if (sk->so_snd.sb_sndptr) {
sdp_dbg(ssk->socket,
"Send on socket without cmid ECONNRESET.\n");
sdp_notify(ssk, ECONNRESET);
}
return;
}
again:
if (sdp_tx_ring_slots_left(ssk) < SDP_TX_SIZE / 2)
sdp_xmit_poll(ssk, 1);
if (ssk->recv_request &&
ring_tail(ssk->rx_ring) >= ssk->recv_request_head &&
tx_credits(ssk) >= SDP_MIN_TX_CREDITS &&
sdp_tx_ring_slots_left(ssk)) {
mb = sdp_alloc_mb_chrcvbuf_ack(sk,
ssk->recv_bytes - SDP_HEAD_SIZE, wait);
if (mb == NULL)
goto allocfail;
ssk->recv_request = 0;
sdp_post_send(ssk, mb);
post_count++;
}
if (tx_credits(ssk) <= SDP_MIN_TX_CREDITS &&
sdp_tx_ring_slots_left(ssk) && sk->so_snd.sb_sndptr &&
sdp_nagle_off(ssk, sk->so_snd.sb_sndptr)) {
SDPSTATS_COUNTER_INC(send_miss_no_credits);
}
while (tx_credits(ssk) > SDP_MIN_TX_CREDITS &&
sdp_tx_ring_slots_left(ssk) && (mb = sk->so_snd.sb_sndptr) &&
sdp_nagle_off(ssk, mb)) {
struct mbuf *n;
SOCKBUF_LOCK(&sk->so_snd);
sk->so_snd.sb_sndptr = mb->m_nextpkt;
sk->so_snd.sb_mb = mb->m_nextpkt;
mb->m_nextpkt = NULL;
SB_EMPTY_FIXUP(&sk->so_snd);
for (n = mb; n != NULL; n = n->m_next)
sbfree(&sk->so_snd, n);
SOCKBUF_UNLOCK(&sk->so_snd);
sdp_post_send(ssk, mb);
post_count++;
}
if (credit_update_needed(ssk) && ssk->state >= TCPS_ESTABLISHED &&
ssk->state < TCPS_FIN_WAIT_2) {
mb = sdp_alloc_mb_data(ssk->socket, wait);
if (mb == NULL)
goto allocfail;
sdp_post_send(ssk, mb);
SDPSTATS_COUNTER_INC(post_send_credits);
post_count++;
}
/* send DisConn if needed
* Do not send DisConn if there is only 1 credit. Compliance with CA4-82
* If one credit is available, an implementation shall only send SDP
* messages that provide additional credits and also do not contain ULP
* payload. */
if ((ssk->flags & SDP_NEEDFIN) && !sk->so_snd.sb_sndptr &&
tx_credits(ssk) > 1) {
mb = sdp_alloc_mb_disconnect(sk, wait);
if (mb == NULL)
goto allocfail;
ssk->flags &= ~SDP_NEEDFIN;
sdp_post_send(ssk, mb);
post_count++;
}
low = (sdp_tx_ring_slots_left(ssk) <= SDP_MIN_TX_CREDITS);
if (post_count || low) {
if (low)
sdp_arm_tx_cq(ssk);
if (sdp_xmit_poll(ssk, low))
goto again;
}
return;
allocfail:
ssk->nagle_last_unacked = -1;
callout_reset(&ssk->nagle_timer, 1, sdp_nagle_timeout, ssk);
return;
}
