/* Like tcp_fin - called when SDP_MID_DISCONNECT is received */
static void
sdp_handle_disconn(struct sdp_sock *ssk)
{
sdp_dbg(ssk->socket, "%s\n", __func__);
SDP_WLOCK_ASSERT(ssk);
if (TCPS_HAVERCVDFIN(ssk->state) == 0)
socantrcvmore(ssk->socket);
switch (ssk->state) {
case TCPS_SYN_RECEIVED:
case TCPS_ESTABLISHED:
ssk->state = TCPS_CLOSE_WAIT;
break;
case TCPS_FIN_WAIT_1:
/* Received a reply FIN - start Infiniband tear down */
sdp_dbg(ssk->socket,
"%s: Starting Infiniband tear down sending DREQ\n",
__func__);
sdp_cancel_dreq_wait_timeout(ssk);
ssk->qp_active = 0;
if (ssk->id) {
struct rdma_cm_id *id;
id = ssk->id;
SDP_WUNLOCK(ssk);
rdma_disconnect(id);
SDP_WLOCK(ssk);
} else {
sdp_warn(ssk->socket,
"%s: ssk->id is NULL\n", __func__);
return;
}
break;
case TCPS_TIME_WAIT:
/* This is a mutual close situation and we've got the DREQ from
the peer before the SDP_MID_DISCONNECT */
break;
case TCPS_CLOSED:
/* FIN arrived after IB teardown started - do nothing */
sdp_dbg(ssk->socket, "%s: fin in state %s\n",
__func__, sdp_state_str(ssk->state));
return;
default:
sdp_warn(ssk->socket,
"%s: FIN in unexpected state. state=%d\n",
__func__, ssk->state);
break;
}
}
static inline void
sdp_process_tx_wc(struct sdp_sock *ssk, struct ib_wc *wc)
{
if (likely(wc->wr_id & SDP_OP_SEND)) {
sdp_handle_send_comp(ssk, wc);
return;
}
#ifdef SDP_ZCOPY
if (wc->wr_id & SDP_OP_RDMA) {
/* TODO: handle failed RDMA read cqe */
sdp_dbg_data(ssk->socket,
"TX comp: RDMA read. status: %d\n", wc->status);
sdp_prf1(sk, NULL, "TX comp: RDMA read");
if (!ssk->tx_ring.rdma_inflight) {
sdp_warn(ssk->socket, "ERROR: unexpected RDMA read\n");
return;
}
if (!ssk->tx_ring.rdma_inflight->busy) {
sdp_warn(ssk->socket,
"ERROR: too many RDMA read completions\n");
return;
}
/* Only last RDMA read WR is signalled. Order is guaranteed -
* therefore if Last RDMA read WR is completed - all other
* have, too */
ssk->tx_ring.rdma_inflight->busy = 0;
sowwakeup(ssk->socket);
sdp_dbg_data(ssk->socket, "woke up sleepers\n");
return;
}
#endif
/* Keepalive probe sent cleanup */
sdp_cnt(sdp_keepalive_probes_sent);
if (likely(!wc->status))
return;
sdp_dbg(ssk->socket, " %s consumes KEEPALIVE status %d\n",
__func__, wc->status);
if (wc->status == IB_WC_WR_FLUSH_ERR)
return;
sdp_notify(ssk, ECONNRESET);
}
/* called only from irq */
static struct mbuf *
sdp_process_rx_wc(struct sdp_sock *ssk, struct ib_wc *wc)
{
struct mbuf *mb;
struct sdp_bsdh *h;
struct socket *sk = ssk->socket;
int mseq;
mb = sdp_recv_completion(ssk, wc->wr_id);
if (unlikely(!mb))
return NULL;
if (unlikely(wc->status)) {
if (ssk->qp_active && sk) {
sdp_dbg(sk, "Recv completion with error. "
"Status %d, vendor: %d\n",
wc->status, wc->vendor_err);
sdp_abort(sk);
ssk->qp_active = 0;
}
m_freem(mb);
return NULL;
}
sdp_dbg_data(sk, "Recv completion. ID %d Length %d\n",
(int)wc->wr_id, wc->byte_len);
if (unlikely(wc->byte_len < sizeof(struct sdp_bsdh))) {
sdp_warn(sk, "SDP BUG! byte_len %d < %zd\n",
wc->byte_len, sizeof(struct sdp_bsdh));
m_freem(mb);
return NULL;
}
/* Use m_adj to trim the tail of data we didn't use. */
m_adj(mb, -(mb->m_pkthdr.len - wc->byte_len));
h = mtod(mb, struct sdp_bsdh *);
SDP_DUMP_PACKET(ssk->socket, "RX", mb, h);
ssk->rx_packets++;
ssk->rx_bytes += mb->m_pkthdr.len;
mseq = ntohl(h->mseq);
atomic_set(&ssk->mseq_ack, mseq);
if (mseq != (int)wc->wr_id)
sdp_warn(sk, "SDP BUG! mseq %d != wrid %d\n",
mseq, (int)wc->wr_id);
return mb;
}
int sdp_tx_ring_create(struct sdp_sock *ssk, struct ib_device *device)
{
struct ib_cq *tx_cq;
int rc = 0, vector = 0;
atomic_set(&ssk->tx_ring.head, 1);
atomic_set(&ssk->tx_ring.tail, 1);
ssk->tx_ring.buffer = kmalloc(
sizeof *ssk->tx_ring.buffer * SDP_TX_SIZE, GFP_KERNEL);
if (!ssk->tx_ring.buffer) {
rc = -ENOMEM;
sdp_warn(sk_ssk(ssk), "Can't allocate TX Ring size %zd.\n",
sizeof(*ssk->tx_ring.buffer) * SDP_TX_SIZE);
goto out;
}
vector = sdp_get_vector_num(device);
tx_cq = ib_create_cq(device, sdp_tx_irq, sdp_tx_cq_event_handler,
sk_ssk(ssk), SDP_TX_SIZE, vector);
if (IS_ERR(tx_cq)) {
rc = PTR_ERR(tx_cq);
sdp_warn(sk_ssk(ssk), "Unable to allocate TX CQ: %d.\n", rc);
goto err_cq;
}
ssk->tx_ring.cq = tx_cq;
setup_timer(&ssk->tx_ring.timer, sdp_poll_tx_timeout,
(unsigned long)ssk);
ssk->tx_ring.poll_cnt = 0;
tasklet_init(&ssk->tx_ring.tasklet, sdp_poll_tx_timeout,
(unsigned long) ssk);
setup_timer(&ssk->nagle_timer, sdp_nagle_timeout, (unsigned long) ssk);
return 0;
err_cq:
kfree(ssk->tx_ring.buffer);
ssk->tx_ring.buffer = NULL;
out:
return rc;
}
void
sdp_tx_ring_destroy(struct sdp_sock *ssk)
{
sdp_dbg(ssk->socket, "tx ring destroy\n");
SDP_WLOCK(ssk);
callout_stop(&ssk->tx_ring.timer);
callout_stop(&ssk->nagle_timer);
SDP_WUNLOCK(ssk);
callout_drain(&ssk->tx_ring.timer);
callout_drain(&ssk->nagle_timer);
if (ssk->tx_ring.buffer) {
sdp_tx_ring_purge(ssk);
kfree(ssk->tx_ring.buffer);
ssk->tx_ring.buffer = NULL;
}
if (ssk->tx_ring.cq) {
if (ib_destroy_cq(ssk->tx_ring.cq)) {
sdp_warn(ssk->socket, "destroy cq(%p) failed\n",
ssk->tx_ring.cq);
} else {
ssk->tx_ring.cq = NULL;
}
}
WARN_ON(ring_head(ssk->tx_ring) != ring_tail(ssk->tx_ring));
}
void sdp_tx_ring_destroy(struct sdp_sock *ssk)
{
del_timer_sync(&ssk->tx_ring.timer);
if (ssk->nagle_timer.function)
del_timer_sync(&ssk->nagle_timer);
if (ssk->tx_ring.buffer) {
sdp_tx_ring_purge(ssk);
kfree(ssk->tx_ring.buffer);
ssk->tx_ring.buffer = NULL;
}
if (ssk->tx_ring.cq) {
if (ib_destroy_cq(ssk->tx_ring.cq)) {
sdp_warn(sk_ssk(ssk), "destroy cq(%p) failed\n",
ssk->tx_ring.cq);
} else {
ssk->tx_ring.cq = NULL;
}
}
tasklet_kill(&ssk->tx_ring.tasklet);
/* tx_cq is destroyed, so no more tx_irq, so no one will schedule this
* tasklet. */
SDP_WARN_ON(ring_head(ssk->tx_ring) != ring_tail(ssk->tx_ring));
}
void
sdp_rx_ring_destroy(struct sdp_sock *ssk)
{
cancel_work_sync(&ssk->rx_comp_work);
rx_ring_destroy_lock(&ssk->rx_ring);
if (ssk->rx_ring.buffer) {
sdp_rx_ring_purge(ssk);
kfree(ssk->rx_ring.buffer);
ssk->rx_ring.buffer = NULL;
}
if (ssk->rx_ring.cq) {
if (ib_destroy_cq(ssk->rx_ring.cq)) {
sdp_warn(ssk->socket, "destroy cq(%p) failed\n",
ssk->rx_ring.cq);
} else {
ssk->rx_ring.cq = NULL;
}
}
WARN_ON(ring_head(ssk->rx_ring) != ring_tail(ssk->rx_ring));
}
static void sdp_qp_event_handler(struct ib_event *event, void *data)
{
if (event->event == IB_EVENT_PATH_MIG) {
sdp_dbg(NULL, "Path migration event\n");
return;
}
sdp_warn(NULL, "unexpected invocation: event: %d, data=%p\n",
event->event, data);
}
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;
}
int
sdp_rx_ring_create(struct sdp_sock *ssk, struct ib_device *device)
{
struct ib_cq *rx_cq;
int rc = 0;
sdp_dbg(ssk->socket, "rx ring created");
INIT_WORK(&ssk->rx_comp_work, sdp_rx_comp_work);
atomic_set(&ssk->rx_ring.head, 1);
atomic_set(&ssk->rx_ring.tail, 1);
ssk->rx_ring.buffer = kmalloc(
sizeof *ssk->rx_ring.buffer * SDP_RX_SIZE, GFP_KERNEL);
if (!ssk->rx_ring.buffer) {
sdp_warn(ssk->socket,
"Unable to allocate RX Ring size %zd.\n",
sizeof(*ssk->rx_ring.buffer) * SDP_RX_SIZE);
return -ENOMEM;
}
rx_cq = ib_create_cq(device, sdp_rx_irq, sdp_rx_cq_event_handler,
ssk->socket, SDP_RX_SIZE, IB_CQ_VECTOR_LEAST_ATTACHED);
if (IS_ERR(rx_cq)) {
rc = PTR_ERR(rx_cq);
sdp_warn(ssk->socket, "Unable to allocate RX CQ: %d.\n", rc);
goto err_cq;
}
sdp_sk(ssk->socket)->rx_ring.cq = rx_cq;
sdp_arm_rx_cq(ssk);
return 0;
err_cq:
kfree(ssk->rx_ring.buffer);
ssk->rx_ring.buffer = NULL;
return rc;
}
static inline void sdp_process_tx_wc(struct sdp_sock *ssk, struct ib_wc *wc)
{
struct sock *sk = sk_ssk(ssk);
if (likely(wc->wr_id & SDP_OP_SEND)) {
struct sk_buff *skb;
skb = sdp_send_completion(ssk, wc->wr_id);
if (likely(skb))
sk_wmem_free_skb(sk, skb);
} else if (wc->wr_id & SDP_OP_RDMA) {
if (ssk->tx_ring.rdma_inflight &&
ssk->tx_ring.rdma_inflight->busy) {
/* Only last RDMA read WR is signalled. Order is guaranteed -
* therefore if Last RDMA read WR is completed - all other
* have, too */
ssk->tx_ring.rdma_inflight->busy = 0;
} else {
sdp_warn(sk, "Unexpected RDMA read completion, "
"probably was canceled already\n");
}
wake_up(sdp_sk_sleep(sk));
} else {
/* Keepalive probe sent cleanup */
sdp_cnt(sdp_keepalive_probes_sent);
}
if (likely(!wc->status) || wc->status == IB_WC_WR_FLUSH_ERR)
return;
sdp_warn(sk, "Send completion with error. wr_id 0x%llx Status %d\n",
wc->wr_id, wc->status);
sdp_set_error(sk, -ECONNRESET);
}
static int sdp_get_max_dev_sge(struct ib_device *dev)
{
struct ib_device_attr attr;
static int max_sges = -1;
int rc;
if (max_sges > 0)
goto out;
rc = ib_query_device(dev, &attr);
if (rc) {
sdp_warn(NULL, "ib_query_device failed: %d\n", rc);
goto out;
}
max_sges = attr.max_sge;
out:
return max_sges;
}
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;
}
static int sdp_init_qp(struct sock *sk, struct rdma_cm_id *id)
{
struct ib_qp_init_attr qp_init_attr = {
.event_handler = sdp_qp_event_handler,
.cap.max_send_wr = SDP_TX_SIZE,
.cap.max_recv_wr = sdp_rx_size,
.cap.max_inline_data = sdp_inline_thresh,
.sq_sig_type = IB_SIGNAL_REQ_WR,
.qp_type = IB_QPT_RC,
};
struct ib_device *device = id->device;
int rc;
sdp_dbg(sk, "%s\n", __func__);
sdp_sk(sk)->max_sge = sdp_get_max_dev_sge(device);
sdp_dbg(sk, "Max sges: %d\n", sdp_sk(sk)->max_sge);
qp_init_attr.cap.max_send_sge = MIN(sdp_sk(sk)->max_sge, SDP_MAX_SEND_SGES);
sdp_dbg(sk, "Setting max send sge to: %d\n", qp_init_attr.cap.max_send_sge);
qp_init_attr.cap.max_recv_sge = MIN(sdp_sk(sk)->max_sge, SDP_MAX_RECV_SGES);
sdp_dbg(sk, "Setting max recv sge to: %d\n", qp_init_attr.cap.max_recv_sge);
sdp_sk(sk)->sdp_dev = ib_get_client_data(device, &sdp_client);
if (!sdp_sk(sk)->sdp_dev) {
sdp_warn(sk, "SDP not available on device %s\n", device->name);
rc = -ENODEV;
goto err_rx;
}
rc = sdp_rx_ring_create(sdp_sk(sk), device);
if (rc)
goto err_rx;
rc = sdp_tx_ring_create(sdp_sk(sk), device);
if (rc)
goto err_tx;
qp_init_attr.recv_cq = sdp_sk(sk)->rx_ring.cq;
qp_init_attr.send_cq = sdp_sk(sk)->tx_ring.cq;
rc = rdma_create_qp(id, sdp_sk(sk)->sdp_dev->pd, &qp_init_attr);
if (rc) {
sdp_warn(sk, "Unable to create QP: %d.\n", rc);
goto err_qp;
}
sdp_sk(sk)->qp = id->qp;
sdp_sk(sk)->ib_device = device;
sdp_sk(sk)->qp_active = 1;
sdp_sk(sk)->context.device = device;
sdp_sk(sk)->inline_thresh = qp_init_attr.cap.max_inline_data;
sdp_dbg(sk, "%s done\n", __func__);
return 0;
err_qp:
sdp_tx_ring_destroy(sdp_sk(sk));
err_tx:
sdp_rx_ring_destroy(sdp_sk(sk));
err_rx:
return rc;
}
static int sdp_get_max_send_frags(u32 buf_size)
{
return MIN(
/* +1 to conpensate on not aligned buffers */
(PAGE_ALIGN(buf_size) >> PAGE_SHIFT) + 1,
SDP_MAX_SEND_SGES - 1);
}
static int sdp_connect_handler(struct sock *sk, struct rdma_cm_id *id,
struct rdma_cm_event *event)
{
struct sockaddr_in *dst_addr;
struct sock *child;
const struct sdp_hh *h;
int rc = 0;
sdp_dbg(sk, "%s %p -> %p\n", __func__, sdp_sk(sk)->id, id);
h = event->param.conn.private_data;
SDP_DUMP_PACKET(sk, "RX", NULL, &h->bsdh);
if (h->ipv_cap & HH_IPV_MASK & ~(HH_IPV4 | HH_IPV6)) {
sdp_warn(sk, "Bad IPV field in SDP Hello header: 0x%x\n",
h->ipv_cap & HH_IPV_MASK);
return -EINVAL;
}
if (!h->max_adverts)
return -EINVAL;
#if (LINUX_VERSION_CODE < KERNEL_VERSION(3, 3, 0))
child = sk_clone(sk, GFP_KERNEL);
#else
child = sk_clone_lock(sk, GFP_KERNEL);
#endif
if (!child)
return -ENOMEM;
sdp_init_sock(child);
dst_addr = (struct sockaddr_in *)&id->route.addr.dst_addr;
sdp_inet_dport(child) = dst_addr->sin_port;
#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
if (inet6_sk(sk)) {
struct ipv6_pinfo *newnp;
newnp = inet_sk(child)->pinet6 = sdp_inet6_sk_generic(child);
memcpy(newnp, inet6_sk(sk), sizeof(struct ipv6_pinfo));
if ((h->ipv_cap & HH_IPV_MASK) == HH_IPV4) {
/* V6 mapped */
sdp_inet_daddr(child) = dst_addr->sin_addr.s_addr;
ipv6_addr_set(&child->sk_v6_daddr, 0, 0, htonl(0x0000FFFF),
h->src_addr.ip4.addr);
ipv6_addr_set(&child->sk_v6_rcv_saddr, 0, 0, htonl(0x0000FFFF),
h->dst_addr.ip4.addr);
ipv6_addr_copy(&child->sk_v6_rcv_saddr, &child->sk_v6_daddr);
} else if ((h->ipv_cap & HH_IPV_MASK) == HH_IPV6) {
struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst_addr;
struct sockaddr_in6 *src_addr6 =
(struct sockaddr_in6 *)&id->route.addr.src_addr;
ipv6_addr_copy(&child->sk_v6_daddr, &dst_addr6->sin6_addr);
ipv6_addr_copy(&child->sk_v6_rcv_saddr, &src_addr6->sin6_addr);
ipv6_addr_copy(&newnp->saddr, &src_addr6->sin6_addr);
} else {
sdp_warn(child, "Bad IPV field: 0x%x\n", h->ipv_cap & HH_IPV_MASK);
}
sdp_inet_daddr(child) = sdp_inet_saddr(child) =
sdp_inet_rcv_saddr(child) = LOOPBACK4_IPV6;
} else
#endif
{
sdp_inet_daddr(child) = dst_addr->sin_addr.s_addr;
}
#ifdef SDP_SOCK_HISTORY
sdp_ssk_hist_rename(sk);
#endif
__sock_put(child, SOCK_REF_CLONE);
down_read(&device_removal_lock);
rc = sdp_init_qp(child, id);
if (rc) {
bh_unlock_sock(child);
up_read(&device_removal_lock);
sdp_sk(child)->destructed_already = 1;
#ifdef SDP_SOCK_HISTORY
sdp_ssk_hist_close(child);
#endif
sk_free(child);
return rc;
}
sdp_sk(child)->max_bufs = ntohs(h->bsdh.bufs);
atomic_set(&sdp_sk(child)->tx_ring.credits, sdp_sk(child)->max_bufs);
sdp_sk(child)->min_bufs = tx_credits(sdp_sk(child)) / 4;
sdp_sk(child)->xmit_size_goal = ntohl(h->localrcvsz) -
sizeof(struct sdp_bsdh);
sdp_sk(child)->send_frags = sdp_get_max_send_frags(sdp_sk(child)->xmit_size_goal);
sdp_init_buffers(sdp_sk(child), rcvbuf_initial_size);
id->context = child;
sdp_sk(child)->id = id;
list_add_tail(&sdp_sk(child)->backlog_queue,
&sdp_sk(sk)->backlog_queue);
sdp_sk(child)->parent = sk;
bh_unlock_sock(child);
sdp_add_sock(sdp_sk(child));
up_read(&device_removal_lock);
sdp_exch_state(child, TCPF_LISTEN | TCPF_CLOSE, TCP_SYN_RECV);
/* child->sk_write_space(child); */
/* child->sk_data_ready(child, 0); */
sk->sk_data_ready(sk);
return 0;
}
static int sdp_response_handler(struct sock *sk, struct rdma_cm_id *id,
struct rdma_cm_event *event)
{
const struct sdp_hah *h;
struct sockaddr_in *dst_addr;
sdp_dbg(sk, "%s\n", __func__);
sdp_exch_state(sk, TCPF_SYN_SENT, TCP_ESTABLISHED);
sdp_set_default_moderation(sdp_sk(sk));
if (sock_flag(sk, SOCK_KEEPOPEN))
sdp_start_keepalive_timer(sk);
if (sock_flag(sk, SOCK_DEAD))
return 0;
h = event->param.conn.private_data;
SDP_DUMP_PACKET(sk, "RX", NULL, &h->bsdh);
sdp_sk(sk)->max_bufs = ntohs(h->bsdh.bufs);
atomic_set(&sdp_sk(sk)->tx_ring.credits, sdp_sk(sk)->max_bufs);
sdp_sk(sk)->min_bufs = tx_credits(sdp_sk(sk)) / 4;
sdp_sk(sk)->xmit_size_goal =
ntohl(h->actrcvsz) - sizeof(struct sdp_bsdh);
sdp_sk(sk)->send_frags = sdp_get_max_send_frags(sdp_sk(sk)->xmit_size_goal);
sdp_sk(sk)->xmit_size_goal = MIN(sdp_sk(sk)->xmit_size_goal,
sdp_sk(sk)->send_frags * PAGE_SIZE);
sdp_sk(sk)->poll_cq = 1;
sk->sk_state_change(sk);
sk_wake_async(sk, 0, POLL_OUT);
dst_addr = (struct sockaddr_in *)&id->route.addr.dst_addr;
sdp_inet_dport(sk) = dst_addr->sin_port;
sdp_inet_daddr(sk) = dst_addr->sin_addr.s_addr;
#ifdef SDP_SOCK_HISTORY
sdp_ssk_hist_rename(sk);
#endif
return 0;
}
static int sdp_connected_handler(struct sock *sk)
{
struct sock *parent;
sdp_dbg(sk, "%s\n", __func__);
parent = sdp_sk(sk)->parent;
BUG_ON(!parent);
sdp_exch_state(sk, TCPF_SYN_RECV, TCP_ESTABLISHED);
#ifdef SDP_SOCK_HISTORY
sdp_ssk_hist_rename(sk);
#endif
sdp_set_default_moderation(sdp_sk(sk));
if (sock_flag(sk, SOCK_KEEPOPEN))
sdp_start_keepalive_timer(sk);
if (sock_flag(sk, SOCK_DEAD))
return 0;
lock_sock(parent);
if (!sdp_sk(parent)->id) { /* TODO: look at SOCK_DEAD? */
sdp_dbg(sk, "parent is going away.\n");
goto done;
}
sk_acceptq_added(parent);
sdp_dbg(parent, "%s child connection established\n", __func__);
list_del_init(&sdp_sk(sk)->backlog_queue);
list_add_tail(&sdp_sk(sk)->accept_queue,
&sdp_sk(parent)->accept_queue);
parent->sk_state_change(parent);
sk_wake_async(parent, 0, POLL_OUT);
done:
release_sock(parent);
return 0;
}
static int sdp_disconnected_handler(struct sock *sk)
{
struct sdp_sock *ssk = sdp_sk(sk);
sdp_dbg(sk, "%s\n", __func__);
if (ssk->tx_ring.cq)
if (sdp_xmit_poll(ssk, 1))
sdp_post_sends(ssk, 0);
if (sk->sk_state == TCP_SYN_RECV) {
sdp_connected_handler(sk);
if (rcv_nxt(ssk))
return 0;
}
return -ECONNRESET;
}
int sdp_cma_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
{
struct rdma_conn_param conn_param;
struct sock *parent = NULL;
struct sock *child = NULL;
struct sock *sk;
struct sdp_hah hah;
struct sdp_hh hh;
int rc = 0, rc2;
sk = id->context;
if (!sk) {
sdp_dbg(NULL, "cm_id is being torn down, event %s\n",
rdma_cm_event_str(event->event));
return event->event == RDMA_CM_EVENT_CONNECT_REQUEST ?
-EINVAL : 0;
}
sdp_add_to_history(sk, rdma_cm_event_str(event->event));
lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
sdp_dbg(sk, "event: %s\n", rdma_cm_event_str(event->event));
if (!sdp_sk(sk)->id) {
sdp_dbg(sk, "socket is being torn down\n");
rc = event->event == RDMA_CM_EVENT_CONNECT_REQUEST ?
-EINVAL : 0;
release_sock(sk);
return rc;
}
switch (event->event) {
case RDMA_CM_EVENT_ADDR_RESOLVED:
if (sdp_link_layer_ib_only &&
rdma_node_get_transport(id->device->node_type) ==
RDMA_TRANSPORT_IB &&
rdma_port_get_link_layer(id->device, id->port_num) !=
IB_LINK_LAYER_INFINIBAND) {
sdp_dbg(sk, "Link layer is: %d. Only IB link layer "
"is allowed\n",
rdma_port_get_link_layer(id->device,
id->port_num));
rc = -ENETUNREACH;
break;
}
rc = rdma_resolve_route(id, SDP_ROUTE_TIMEOUT);
break;
case RDMA_CM_EVENT_ADDR_ERROR:
rc = -ENETUNREACH;
break;
case RDMA_CM_EVENT_ROUTE_RESOLVED:
rc = sdp_init_qp(sk, id);
if (rc)
break;
memset(&hh, 0, sizeof hh);
hh.bsdh.mid = SDP_MID_HELLO;
hh.bsdh.len = htonl(sizeof(struct sdp_hh));
hh.max_adverts = 1;
hh.majv_minv = SDP_MAJV_MINV;
sdp_init_buffers(sdp_sk(sk), rcvbuf_initial_size);
hh.bsdh.bufs = htons(rx_ring_posted(sdp_sk(sk)));
atomic_set(&sdp_sk(sk)->remote_credits,
rx_ring_posted(sdp_sk(sk)));
hh.localrcvsz = hh.desremrcvsz = htonl(sdp_sk(sk)->recv_frags *
PAGE_SIZE + sizeof(struct sdp_bsdh));
#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
if (inet6_sk(sk)) {
struct sockaddr *src_addr = (struct sockaddr *)&id->route.addr.src_addr;
struct sockaddr_in *addr4 = (struct sockaddr_in *)src_addr;
struct sockaddr_in6 *addr6 = (struct sockaddr_in6 *)src_addr;
if (src_addr->sa_family == AF_INET) {
/* IPv4 over IPv6 */
ipv6_addr_set(&sk->sk_v6_rcv_saddr, 0, 0, htonl(0xFFFF),
addr4->sin_addr.s_addr);
} else {
sk->sk_v6_rcv_saddr = addr6->sin6_addr;
}
inet6_sk(sk)->saddr = sk->sk_v6_rcv_saddr;
}
else
#endif
{
sdp_inet_saddr(sk) = sdp_inet_rcv_saddr(sk) =
((struct sockaddr_in *)&id->route.addr.src_addr)->sin_addr.s_addr;
}
memset(&conn_param, 0, sizeof conn_param);
conn_param.private_data_len = sizeof hh;
conn_param.private_data = &hh;
conn_param.responder_resources = 4 /* TODO */;
conn_param.initiator_depth = 4 /* TODO */;
conn_param.retry_count = sdp_retry_count;
SDP_DUMP_PACKET(sk, "TX", NULL, &hh.bsdh);
if (sdp_apm_enable) {
rc = rdma_enable_apm(id, RDMA_ALT_PATH_BEST);
if (rc)
sdp_warn(sk, "APM couldn't be enabled for active side: %d\n", rc);
}
rc = rdma_connect(id, &conn_param);
break;
case RDMA_CM_EVENT_ALT_ROUTE_RESOLVED:
sdp_dbg(sk, "alt route was resolved slid=%d, dlid=%d\n",
id->route.path_rec[1].slid, id->route.path_rec[1].dlid);
break;
case RDMA_CM_EVENT_ALT_PATH_LOADED:
sdp_dbg(sk, "alt route path loaded\n");
break;
case RDMA_CM_EVENT_ALT_ROUTE_ERROR:
sdp_warn(sk, "alt route resolve error\n");
break;
case RDMA_CM_EVENT_ROUTE_ERROR:
rc = -ETIMEDOUT;
break;
case RDMA_CM_EVENT_CONNECT_REQUEST:
rc = sdp_connect_handler(sk, id, event);
if (rc) {
sdp_dbg(sk, "Destroying qp\n");
rdma_reject(id, NULL, 0);
break;
}
child = id->context;
atomic_set(&sdp_sk(child)->remote_credits,
rx_ring_posted(sdp_sk(child)));
memset(&hah, 0, sizeof hah);
hah.bsdh.mid = SDP_MID_HELLO_ACK;
hah.bsdh.bufs = htons(rx_ring_posted(sdp_sk(child)));
hah.bsdh.len = htonl(sizeof(struct sdp_hah));
hah.majv_minv = SDP_MAJV_MINV;
hah.ext_max_adverts = 1; /* Doesn't seem to be mandated by spec,
but just in case */
hah.actrcvsz = htonl(sdp_sk(child)->recv_frags * PAGE_SIZE +
sizeof(struct sdp_bsdh));
memset(&conn_param, 0, sizeof conn_param);
conn_param.private_data_len = sizeof hah;
conn_param.private_data = &hah;
conn_param.responder_resources = 4 /* TODO */;
conn_param.initiator_depth = 4 /* TODO */;
conn_param.retry_count = sdp_retry_count;
SDP_DUMP_PACKET(sk, "TX", NULL, &hah.bsdh);
rc = rdma_accept(id, &conn_param);
if (rc) {
sdp_sk(child)->id = NULL;
id->qp = NULL;
id->context = NULL;
parent = sdp_sk(child)->parent; /* TODO: hold ? */
} else if (sdp_apm_enable) {
rc2 = rdma_enable_apm(id, RDMA_ALT_PATH_BEST);
if (rc2)
sdp_warn(sk, "APM couldn't be enabled for passive side: %d\n", rc2);
}
break;
case RDMA_CM_EVENT_CONNECT_RESPONSE:
rc = sdp_response_handler(sk, id, event);
if (rc) {
sdp_dbg(sk, "Destroying qp\n");
rdma_reject(id, NULL, 0);
} else {
rc = rdma_accept(id, NULL);
if (!rc && sdp_apm_enable) {
rc2 = rdma_enable_apm(id, RDMA_ALT_PATH_BEST);
if (rc2)
sdp_warn(sk, "APM couldn't be enabled for passive side:%d \n", rc2);
}
}
break;
case RDMA_CM_EVENT_CONNECT_ERROR:
rc = -ETIMEDOUT;
break;
case RDMA_CM_EVENT_UNREACHABLE:
rc = -ENETUNREACH;
break;
case RDMA_CM_EVENT_REJECTED:
rc = -ECONNREFUSED;
break;
case RDMA_CM_EVENT_ESTABLISHED:
sdp_inet_saddr(sk) = sdp_inet_rcv_saddr(sk) =
((struct sockaddr_in *)&id->route.addr.src_addr)->sin_addr.s_addr;
rc = sdp_connected_handler(sk);
break;
case RDMA_CM_EVENT_DISCONNECTED: /* This means DREQ/DREP received */
if (sk->sk_state == TCP_LAST_ACK) {
sdp_cancel_dreq_wait_timeout(sdp_sk(sk));
sdp_exch_state(sk, TCPF_LAST_ACK, TCP_TIME_WAIT);
sdp_dbg(sk, "%s: waiting for Infiniband tear down\n",
__func__);
}
sdp_sk(sk)->qp_active = 0;
rdma_disconnect(id);
if (sk->sk_state != TCP_TIME_WAIT) {
if (sk->sk_state == TCP_CLOSE_WAIT) {
sdp_dbg(sk, "IB teardown while in "
"TCP_CLOSE_WAIT taking reference to "
"let close() finish the work\n");
sock_hold(sk, SOCK_REF_CMA);
sdp_start_cma_timewait_timeout(sdp_sk(sk),
SDP_CMA_TIMEWAIT_TIMEOUT);
}
sdp_set_error(sk, -EPIPE);
rc = sdp_disconnected_handler(sk);
}
break;
case RDMA_CM_EVENT_TIMEWAIT_EXIT:
rc = sdp_disconnected_handler(sk);
break;
case RDMA_CM_EVENT_DEVICE_REMOVAL:
rc = -ENETRESET;
break;
case RDMA_CM_EVENT_ADDR_CHANGE:
sdp_dbg(sk, "Got Address change event\n");
rc = 0;
break;
default:
printk(KERN_ERR "SDP: Unexpected CMA event: %d\n",
event->event);
rc = -ECONNABORTED;
break;
}
sdp_dbg(sk, "event: %s handled\n", rdma_cm_event_str(event->event));
if (rc && sdp_sk(sk)->id == id) {
child = sk;
sdp_sk(sk)->id = NULL;
id->qp = NULL;
id->context = NULL;
parent = sdp_sk(sk)->parent;
sdp_reset_sk(sk, rc);
}
release_sock(sk);
sdp_dbg(sk, "event: %s done. status %d\n",
rdma_cm_event_str(event->event), rc);
if (parent) {
lock_sock(parent);
if (!sdp_sk(parent)->id) { /* TODO: look at SOCK_DEAD? */
sdp_dbg(sk, "parent is going away.\n");
child = NULL;
goto done;
}
if (!list_empty(&sdp_sk(child)->backlog_queue))
list_del_init(&sdp_sk(child)->backlog_queue);
else
child = NULL;
done:
release_sock(parent);
if (child)
sdp_common_release(child);
}
return rc;
}
static int
sdp_post_recv(struct sdp_sock *ssk)
{
struct sdp_buf *rx_req;
int i, rc;
u64 addr;
struct ib_device *dev;
struct ib_recv_wr rx_wr = { NULL };
struct ib_sge ibsge[SDP_MAX_RECV_SGES];
struct ib_sge *sge = ibsge;
struct ib_recv_wr *bad_wr;
struct mbuf *mb, *m;
struct sdp_bsdh *h;
int id = ring_head(ssk->rx_ring);
/* Now, allocate and repost recv */
sdp_prf(ssk->socket, mb, "Posting mb");
mb = m_getm2(NULL, ssk->recv_bytes, M_NOWAIT, MT_DATA, M_PKTHDR);
if (mb == NULL) {
/* Retry so we can't stall out with no memory. */
if (!rx_ring_posted(ssk))
queue_work(rx_comp_wq, &ssk->rx_comp_work);
return -1;
}
for (m = mb; m != NULL; m = m->m_next) {
m->m_len = (m->m_flags & M_EXT) ? m->m_ext.ext_size :
((m->m_flags & M_PKTHDR) ? MHLEN : MLEN);
mb->m_pkthdr.len += m->m_len;
}
h = mtod(mb, struct sdp_bsdh *);
rx_req = ssk->rx_ring.buffer + (id & (SDP_RX_SIZE - 1));
rx_req->mb = mb;
dev = ssk->ib_device;
for (i = 0; mb != NULL; i++, mb = mb->m_next, sge++) {
addr = ib_dma_map_single(dev, mb->m_data, mb->m_len,
DMA_TO_DEVICE);
/* TODO: proper error handling */
BUG_ON(ib_dma_mapping_error(dev, addr));
BUG_ON(i >= SDP_MAX_RECV_SGES);
rx_req->mapping[i] = addr;
sge->addr = addr;
sge->length = mb->m_len;
sge->lkey = ssk->sdp_dev->mr->lkey;
}
rx_wr.next = NULL;
rx_wr.wr_id = id | SDP_OP_RECV;
rx_wr.sg_list = ibsge;
rx_wr.num_sge = i;
rc = ib_post_recv(ssk->qp, &rx_wr, &bad_wr);
if (unlikely(rc)) {
sdp_warn(ssk->socket, "ib_post_recv failed. status %d\n", rc);
sdp_cleanup_sdp_buf(ssk, rx_req, DMA_FROM_DEVICE);
m_freem(mb);
sdp_notify(ssk, ECONNRESET);
return -1;
}
atomic_inc(&ssk->rx_ring.head);
SDPSTATS_COUNTER_INC(post_recv);
return 0;
}
/* socket lock should be taken before calling this */
static int
sdp_process_rx_ctl_mb(struct sdp_sock *ssk, struct mbuf *mb)
{
struct sdp_bsdh *h;
struct socket *sk;
SDP_WLOCK_ASSERT(ssk);
sk = ssk->socket;
h = mtod(mb, struct sdp_bsdh *);
switch (h->mid) {
case SDP_MID_DATA:
case SDP_MID_SRCAVAIL:
sdp_dbg(sk, "DATA after socket rcv was shutdown\n");
/* got data in RCV_SHUTDOWN */
if (ssk->state == TCPS_FIN_WAIT_1) {
sdp_dbg(sk, "RX data when state = FIN_WAIT1\n");
sdp_notify(ssk, ECONNRESET);
}
m_freem(mb);
break;
#ifdef SDP_ZCOPY
case SDP_MID_RDMARDCOMPL:
m_freem(mb);
break;
case SDP_MID_SENDSM:
sdp_handle_sendsm(ssk, ntohl(h->mseq_ack));
m_freem(mb);
break;
case SDP_MID_SRCAVAIL_CANCEL:
sdp_dbg_data(sk, "Handling SrcAvailCancel\n");
sdp_prf(sk, NULL, "Handling SrcAvailCancel");
if (ssk->rx_sa) {
ssk->srcavail_cancel_mseq = ntohl(h->mseq);
ssk->rx_sa->flags |= RX_SA_ABORTED;
ssk->rx_sa = NULL; /* TODO: change it into SDP_MID_DATA and get
the dirty logic from recvmsg */
} else {
sdp_dbg(sk, "Got SrcAvailCancel - "
"but no SrcAvail in process\n");
}
m_freem(mb);
break;
case SDP_MID_SINKAVAIL:
sdp_dbg_data(sk, "Got SinkAvail - not supported: ignored\n");
sdp_prf(sk, NULL, "Got SinkAvail - not supported: ignored");
/* FALLTHROUGH */
#endif
case SDP_MID_ABORT:
sdp_dbg_data(sk, "Handling ABORT\n");
sdp_prf(sk, NULL, "Handling ABORT");
sdp_notify(ssk, ECONNRESET);
m_freem(mb);
break;
case SDP_MID_DISCONN:
sdp_dbg_data(sk, "Handling DISCONN\n");
sdp_prf(sk, NULL, "Handling DISCONN");
sdp_handle_disconn(ssk);
break;
case SDP_MID_CHRCVBUF:
sdp_dbg_data(sk, "Handling RX CHRCVBUF\n");
sdp_handle_resize_request(ssk, (struct sdp_chrecvbuf *)(h+1));
m_freem(mb);
break;
case SDP_MID_CHRCVBUF_ACK:
sdp_dbg_data(sk, "Handling RX CHRCVBUF_ACK\n");
sdp_handle_resize_ack(ssk, (struct sdp_chrecvbuf *)(h+1));
m_freem(mb);
break;
default:
/* TODO: Handle other messages */
sdp_warn(sk, "SDP: FIXME MID %d\n", h->mid);
m_freem(mb);
}
return 0;
}