1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
5 /* Greg Joyce <greg@opengridcomputing.com> */
6 /* Copyright (c) 2008-2019, IBM Corporation */
7 /* Copyright (c) 2017, Open Grid Computing, Inc. */
9 #include <linux/errno.h>
10 #include <linux/types.h>
11 #include <linux/net.h>
12 #include <linux/inetdevice.h>
13 #include <net/addrconf.h>
14 #include <linux/workqueue.h>
17 #include <linux/inet.h>
18 #include <linux/tcp.h>
20 #include <rdma/iw_cm.h>
21 #include <rdma/ib_verbs.h>
22 #include <rdma/ib_user_verbs.h>
28 * Set to any combination of
29 * MPA_V2_RDMA_NO_RTR, MPA_V2_RDMA_READ_RTR, MPA_V2_RDMA_WRITE_RTR
31 static __be16 rtr_type = MPA_V2_RDMA_READ_RTR | MPA_V2_RDMA_WRITE_RTR;
32 static const bool relaxed_ird_negotiation = 1;
34 static void siw_cm_llp_state_change(struct sock *s);
35 static void siw_cm_llp_data_ready(struct sock *s);
36 static void siw_cm_llp_write_space(struct sock *s);
37 static void siw_cm_llp_error_report(struct sock *s);
38 static int siw_cm_upcall(struct siw_cep *cep, enum iw_cm_event_type reason,
41 static void siw_sk_assign_cm_upcalls(struct sock *sk)
43 write_lock_bh(&sk->sk_callback_lock);
44 sk->sk_state_change = siw_cm_llp_state_change;
45 sk->sk_data_ready = siw_cm_llp_data_ready;
46 sk->sk_write_space = siw_cm_llp_write_space;
47 sk->sk_error_report = siw_cm_llp_error_report;
48 write_unlock_bh(&sk->sk_callback_lock);
51 static void siw_sk_save_upcalls(struct sock *sk)
53 struct siw_cep *cep = sk_to_cep(sk);
55 write_lock_bh(&sk->sk_callback_lock);
56 cep->sk_state_change = sk->sk_state_change;
57 cep->sk_data_ready = sk->sk_data_ready;
58 cep->sk_write_space = sk->sk_write_space;
59 cep->sk_error_report = sk->sk_error_report;
60 write_unlock_bh(&sk->sk_callback_lock);
63 static void siw_sk_restore_upcalls(struct sock *sk, struct siw_cep *cep)
65 sk->sk_state_change = cep->sk_state_change;
66 sk->sk_data_ready = cep->sk_data_ready;
67 sk->sk_write_space = cep->sk_write_space;
68 sk->sk_error_report = cep->sk_error_report;
69 sk->sk_user_data = NULL;
72 static void siw_qp_socket_assoc(struct siw_cep *cep, struct siw_qp *qp)
74 struct socket *s = cep->sock;
75 struct sock *sk = s->sk;
77 write_lock_bh(&sk->sk_callback_lock);
80 sk->sk_data_ready = siw_qp_llp_data_ready;
81 sk->sk_write_space = siw_qp_llp_write_space;
83 write_unlock_bh(&sk->sk_callback_lock);
86 static void siw_socket_disassoc(struct socket *s)
88 struct sock *sk = s->sk;
92 write_lock_bh(&sk->sk_callback_lock);
95 siw_sk_restore_upcalls(sk, cep);
98 pr_warn("siw: cannot restore sk callbacks: no ep\n");
100 write_unlock_bh(&sk->sk_callback_lock);
102 pr_warn("siw: cannot restore sk callbacks: no sk\n");
106 static void siw_rtr_data_ready(struct sock *sk)
109 struct siw_qp *qp = NULL;
110 read_descriptor_t rd_desc;
112 read_lock(&sk->sk_callback_lock);
116 WARN(1, "No connection endpoint\n");
121 memset(&rd_desc, 0, sizeof(rd_desc));
122 rd_desc.arg.data = qp;
125 tcp_read_sock(sk, &rd_desc, siw_tcp_rx_data);
127 * Check if first frame was successfully processed.
128 * Signal connection full establishment if yes.
129 * Failed data processing would have already scheduled
132 if (!qp->rx_stream.rx_suspend)
133 siw_cm_upcall(cep, IW_CM_EVENT_ESTABLISHED, 0);
135 read_unlock(&sk->sk_callback_lock);
137 siw_qp_socket_assoc(cep, qp);
140 static void siw_sk_assign_rtr_upcalls(struct siw_cep *cep)
142 struct sock *sk = cep->sock->sk;
144 write_lock_bh(&sk->sk_callback_lock);
145 sk->sk_data_ready = siw_rtr_data_ready;
146 sk->sk_write_space = siw_qp_llp_write_space;
147 write_unlock_bh(&sk->sk_callback_lock);
150 static void siw_cep_socket_assoc(struct siw_cep *cep, struct socket *s)
154 s->sk->sk_user_data = cep;
156 siw_sk_save_upcalls(s->sk);
157 siw_sk_assign_cm_upcalls(s->sk);
160 static struct siw_cep *siw_cep_alloc(struct siw_device *sdev)
162 struct siw_cep *cep = kzalloc(sizeof(*cep), GFP_KERNEL);
168 INIT_LIST_HEAD(&cep->listenq);
169 INIT_LIST_HEAD(&cep->devq);
170 INIT_LIST_HEAD(&cep->work_freelist);
172 kref_init(&cep->ref);
173 cep->state = SIW_EPSTATE_IDLE;
174 init_waitqueue_head(&cep->waitq);
175 spin_lock_init(&cep->lock);
177 cep->enhanced_rdma_conn_est = false;
179 spin_lock_irqsave(&sdev->lock, flags);
180 list_add_tail(&cep->devq, &sdev->cep_list);
181 spin_unlock_irqrestore(&sdev->lock, flags);
183 siw_dbg_cep(cep, "new endpoint\n");
187 static void siw_cm_free_work(struct siw_cep *cep)
189 struct list_head *w, *tmp;
190 struct siw_cm_work *work;
192 list_for_each_safe(w, tmp, &cep->work_freelist) {
193 work = list_entry(w, struct siw_cm_work, list);
194 list_del(&work->list);
199 static void siw_cancel_mpatimer(struct siw_cep *cep)
201 spin_lock_bh(&cep->lock);
202 if (cep->mpa_timer) {
203 if (cancel_delayed_work(&cep->mpa_timer->work)) {
205 kfree(cep->mpa_timer); /* not needed again */
207 cep->mpa_timer = NULL;
209 spin_unlock_bh(&cep->lock);
212 static void siw_put_work(struct siw_cm_work *work)
214 INIT_LIST_HEAD(&work->list);
215 spin_lock_bh(&work->cep->lock);
216 list_add(&work->list, &work->cep->work_freelist);
217 spin_unlock_bh(&work->cep->lock);
220 static void siw_cep_set_inuse(struct siw_cep *cep)
224 spin_lock_irqsave(&cep->lock, flags);
227 spin_unlock_irqrestore(&cep->lock, flags);
228 wait_event_interruptible(cep->waitq, !cep->in_use);
229 if (signal_pending(current))
230 flush_signals(current);
234 spin_unlock_irqrestore(&cep->lock, flags);
238 static void siw_cep_set_free(struct siw_cep *cep)
242 spin_lock_irqsave(&cep->lock, flags);
244 spin_unlock_irqrestore(&cep->lock, flags);
246 wake_up(&cep->waitq);
249 static void __siw_cep_dealloc(struct kref *ref)
251 struct siw_cep *cep = container_of(ref, struct siw_cep, ref);
252 struct siw_device *sdev = cep->sdev;
255 WARN_ON(cep->listen_cep);
257 /* kfree(NULL) is safe */
258 kfree(cep->mpa.pdata);
259 spin_lock_bh(&cep->lock);
260 if (!list_empty(&cep->work_freelist))
261 siw_cm_free_work(cep);
262 spin_unlock_bh(&cep->lock);
264 spin_lock_irqsave(&sdev->lock, flags);
265 list_del(&cep->devq);
266 spin_unlock_irqrestore(&sdev->lock, flags);
268 siw_dbg_cep(cep, "free endpoint\n");
272 static struct siw_cm_work *siw_get_work(struct siw_cep *cep)
274 struct siw_cm_work *work = NULL;
276 spin_lock_bh(&cep->lock);
277 if (!list_empty(&cep->work_freelist)) {
278 work = list_entry(cep->work_freelist.next, struct siw_cm_work,
280 list_del_init(&work->list);
282 spin_unlock_bh(&cep->lock);
286 static int siw_cm_alloc_work(struct siw_cep *cep, int num)
288 struct siw_cm_work *work;
291 work = kmalloc(sizeof(*work), GFP_KERNEL);
293 if (!(list_empty(&cep->work_freelist)))
294 siw_cm_free_work(cep);
298 INIT_LIST_HEAD(&work->list);
299 list_add(&work->list, &cep->work_freelist);
307 * Upcall to IWCM to inform about async connection events
309 static int siw_cm_upcall(struct siw_cep *cep, enum iw_cm_event_type reason,
312 struct iw_cm_event event;
315 memset(&event, 0, sizeof(event));
316 event.status = status;
317 event.event = reason;
319 if (reason == IW_CM_EVENT_CONNECT_REQUEST) {
320 event.provider_data = cep;
321 id = cep->listen_cep->cm_id;
325 /* Signal IRD and ORD */
326 if (reason == IW_CM_EVENT_ESTABLISHED ||
327 reason == IW_CM_EVENT_CONNECT_REPLY) {
328 /* Signal negotiated IRD/ORD values we will use */
329 event.ird = cep->ird;
330 event.ord = cep->ord;
331 } else if (reason == IW_CM_EVENT_CONNECT_REQUEST) {
332 event.ird = cep->ord;
333 event.ord = cep->ird;
335 /* Signal private data and address information */
336 if (reason == IW_CM_EVENT_CONNECT_REQUEST ||
337 reason == IW_CM_EVENT_CONNECT_REPLY) {
338 u16 pd_len = be16_to_cpu(cep->mpa.hdr.params.pd_len);
342 * hand over MPA private data
344 event.private_data_len = pd_len;
345 event.private_data = cep->mpa.pdata;
347 /* Hide MPA V2 IRD/ORD control */
348 if (cep->enhanced_rdma_conn_est) {
349 event.private_data_len -=
350 sizeof(struct mpa_v2_data);
351 event.private_data +=
352 sizeof(struct mpa_v2_data);
355 getname_local(cep->sock, &event.local_addr);
356 getname_peer(cep->sock, &event.remote_addr);
358 siw_dbg_cep(cep, "[QP %u]: id 0x%p, reason=%d, status=%d\n",
359 cep->qp ? qp_id(cep->qp) : -1, id, reason, status);
361 return id->event_handler(id, &event);
367 * Drops established LLP connection if present and not already
368 * scheduled for dropping. Called from user context, SQ workqueue
369 * or receive IRQ. Caller signals if socket can be immediately
370 * closed (basically, if not in IRQ).
372 void siw_qp_cm_drop(struct siw_qp *qp, int schedule)
374 struct siw_cep *cep = qp->cep;
376 qp->rx_stream.rx_suspend = 1;
377 qp->tx_ctx.tx_suspend = 1;
383 siw_cm_queue_work(cep, SIW_CM_WORK_CLOSE_LLP);
385 siw_cep_set_inuse(cep);
387 if (cep->state == SIW_EPSTATE_CLOSED) {
388 siw_dbg_cep(cep, "already closed\n");
391 siw_dbg_cep(cep, "immediate close, state %d\n", cep->state);
393 if (qp->term_info.valid)
394 siw_send_terminate(qp);
397 switch (cep->state) {
398 case SIW_EPSTATE_AWAIT_MPAREP:
399 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
403 case SIW_EPSTATE_RDMA_MODE:
404 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0);
407 case SIW_EPSTATE_IDLE:
408 case SIW_EPSTATE_LISTENING:
409 case SIW_EPSTATE_CONNECTING:
410 case SIW_EPSTATE_AWAIT_MPAREQ:
411 case SIW_EPSTATE_RECVD_MPAREQ:
412 case SIW_EPSTATE_CLOSED:
416 cep->cm_id->rem_ref(cep->cm_id);
420 cep->state = SIW_EPSTATE_CLOSED;
423 siw_socket_disassoc(cep->sock);
425 * Immediately close socket
427 sock_release(cep->sock);
435 siw_cep_set_free(cep);
439 void siw_cep_put(struct siw_cep *cep)
441 WARN_ON(kref_read(&cep->ref) < 1);
442 kref_put(&cep->ref, __siw_cep_dealloc);
445 void siw_cep_get(struct siw_cep *cep)
451 * Expects params->pd_len in host byte order
453 static int siw_send_mpareqrep(struct siw_cep *cep, const void *pdata, u8 pd_len)
455 struct socket *s = cep->sock;
456 struct mpa_rr *rr = &cep->mpa.hdr;
463 memset(&msg, 0, sizeof(msg));
465 iov[iovec_num].iov_base = rr;
466 iov[iovec_num].iov_len = sizeof(*rr);
467 mpa_len = sizeof(*rr);
469 if (cep->enhanced_rdma_conn_est) {
471 iov[iovec_num].iov_base = &cep->mpa.v2_ctrl;
472 iov[iovec_num].iov_len = sizeof(cep->mpa.v2_ctrl);
473 mpa_len += sizeof(cep->mpa.v2_ctrl);
477 iov[iovec_num].iov_base = (char *)pdata;
478 iov[iovec_num].iov_len = pd_len;
481 if (cep->enhanced_rdma_conn_est)
482 pd_len += sizeof(cep->mpa.v2_ctrl);
484 rr->params.pd_len = cpu_to_be16(pd_len);
486 rv = kernel_sendmsg(s, &msg, iov, iovec_num + 1, mpa_len);
488 return rv < 0 ? rv : 0;
492 * Receive MPA Request/Reply header.
494 * Returns 0 if complete MPA Request/Reply header including
495 * eventual private data was received. Returns -EAGAIN if
496 * header was partially received or negative error code otherwise.
498 * Context: May be called in process context only
500 static int siw_recv_mpa_rr(struct siw_cep *cep)
502 struct mpa_rr *hdr = &cep->mpa.hdr;
503 struct socket *s = cep->sock;
507 if (cep->mpa.bytes_rcvd < sizeof(struct mpa_rr)) {
508 rcvd = ksock_recv(s, (char *)hdr + cep->mpa.bytes_rcvd,
509 sizeof(struct mpa_rr) - cep->mpa.bytes_rcvd,
512 return -ECONNABORTED;
514 cep->mpa.bytes_rcvd += rcvd;
516 if (cep->mpa.bytes_rcvd < sizeof(struct mpa_rr))
519 if (be16_to_cpu(hdr->params.pd_len) > MPA_MAX_PRIVDATA)
522 pd_len = be16_to_cpu(hdr->params.pd_len);
525 * At least the MPA Request/Reply header (frame not including
526 * private data) has been received.
527 * Receive (or continue receiving) any private data.
529 to_rcv = pd_len - (cep->mpa.bytes_rcvd - sizeof(struct mpa_rr));
533 * We must have hdr->params.pd_len == 0 and thus received a
534 * complete MPA Request/Reply frame.
535 * Check against peer protocol violation.
539 rcvd = ksock_recv(s, (char *)&word, sizeof(word), MSG_DONTWAIT);
544 siw_dbg_cep(cep, "peer EOF\n");
548 siw_dbg_cep(cep, "error: %d\n", rcvd);
551 siw_dbg_cep(cep, "peer sent extra data: %d\n", rcvd);
557 * At this point, we must have hdr->params.pd_len != 0.
558 * A private data buffer gets allocated if hdr->params.pd_len != 0.
560 if (!cep->mpa.pdata) {
561 cep->mpa.pdata = kmalloc(pd_len + 4, GFP_KERNEL);
566 s, cep->mpa.pdata + cep->mpa.bytes_rcvd - sizeof(struct mpa_rr),
567 to_rcv + 4, MSG_DONTWAIT);
575 cep->mpa.bytes_rcvd += rcvd;
577 if (to_rcv == rcvd) {
578 siw_dbg_cep(cep, "%d bytes private data received\n", pd_len);
587 * Read MPA Request from socket and signal new connection to IWCM
588 * if success. Caller must hold lock on corresponding listening CEP.
590 static int siw_proc_mpareq(struct siw_cep *cep)
596 rv = siw_recv_mpa_rr(cep);
602 version = __mpa_rr_revision(req->params.bits);
603 pd_len = be16_to_cpu(req->params.pd_len);
605 if (version > MPA_REVISION_2)
606 /* allow for 0, 1, and 2 only */
609 if (memcmp(req->key, MPA_KEY_REQ, 16))
612 /* Prepare for sending MPA reply */
613 memcpy(req->key, MPA_KEY_REP, 16);
615 if (version == MPA_REVISION_2 &&
616 (req->params.bits & MPA_RR_FLAG_ENHANCED)) {
618 * MPA version 2 must signal IRD/ORD values and P2P mode
619 * in private data if header flag MPA_RR_FLAG_ENHANCED
622 if (pd_len < sizeof(struct mpa_v2_data))
625 cep->enhanced_rdma_conn_est = true;
628 /* MPA Markers: currently not supported. Marker TX to be added. */
629 if (req->params.bits & MPA_RR_FLAG_MARKERS)
632 if (req->params.bits & MPA_RR_FLAG_CRC) {
634 * RFC 5044, page 27: CRC MUST be used if peer requests it.
635 * siw specific: 'mpa_crc_strict' parameter to reject
636 * connection with CRC if local CRC off enforced by
637 * 'mpa_crc_strict' module parameter.
639 if (!mpa_crc_required && mpa_crc_strict)
642 /* Enable CRC if requested by module parameter */
643 if (mpa_crc_required)
644 req->params.bits |= MPA_RR_FLAG_CRC;
646 if (cep->enhanced_rdma_conn_est) {
647 struct mpa_v2_data *v2 = (struct mpa_v2_data *)cep->mpa.pdata;
650 * Peer requested ORD becomes requested local IRD,
651 * peer requested IRD becomes requested local ORD.
652 * IRD and ORD get limited by global maximum values.
654 cep->ord = ntohs(v2->ird) & MPA_IRD_ORD_MASK;
655 cep->ord = min(cep->ord, SIW_MAX_ORD_QP);
656 cep->ird = ntohs(v2->ord) & MPA_IRD_ORD_MASK;
657 cep->ird = min(cep->ird, SIW_MAX_IRD_QP);
659 /* May get overwritten by locally negotiated values */
660 cep->mpa.v2_ctrl.ird = htons(cep->ird);
661 cep->mpa.v2_ctrl.ord = htons(cep->ord);
664 * Support for peer sent zero length Write or Read to
665 * let local side enter RTS. Writes are preferred.
666 * Sends would require pre-posting a Receive and are
668 * Propose zero length Write if none of Read and Write
671 if (v2->ird & MPA_V2_PEER_TO_PEER) {
672 cep->mpa.v2_ctrl.ird |= MPA_V2_PEER_TO_PEER;
674 if (v2->ord & MPA_V2_RDMA_WRITE_RTR)
675 cep->mpa.v2_ctrl.ord |= MPA_V2_RDMA_WRITE_RTR;
676 else if (v2->ord & MPA_V2_RDMA_READ_RTR)
677 cep->mpa.v2_ctrl.ord |= MPA_V2_RDMA_READ_RTR;
679 cep->mpa.v2_ctrl.ord |= MPA_V2_RDMA_WRITE_RTR;
683 cep->state = SIW_EPSTATE_RECVD_MPAREQ;
685 /* Keep reference until IWCM accepts/rejects */
687 rv = siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REQUEST, 0);
694 siw_dbg_cep(cep, "reject: crc %d:%d:%d, m %d:%d\n",
695 req->params.bits & MPA_RR_FLAG_CRC ? 1 : 0,
696 mpa_crc_required, mpa_crc_strict,
697 req->params.bits & MPA_RR_FLAG_MARKERS ? 1 : 0, 0);
699 req->params.bits &= ~MPA_RR_FLAG_MARKERS;
700 req->params.bits |= MPA_RR_FLAG_REJECT;
702 if (!mpa_crc_required && mpa_crc_strict)
703 req->params.bits &= ~MPA_RR_FLAG_CRC;
706 kfree(cep->mpa.pdata);
708 cep->mpa.pdata = NULL;
710 siw_send_mpareqrep(cep, NULL, 0);
715 static int siw_proc_mpareply(struct siw_cep *cep)
717 struct siw_qp_attrs qp_attrs;
718 enum siw_qp_attr_mask qp_attr_mask;
719 struct siw_qp *qp = cep->qp;
724 bool ird_insufficient = false;
725 enum mpa_v2_ctrl mpa_p2p_mode = MPA_V2_RDMA_NO_RTR;
727 rv = siw_recv_mpa_rr(cep);
729 siw_cancel_mpatimer(cep);
735 if (__mpa_rr_revision(rep->params.bits) > MPA_REVISION_2) {
736 /* allow for 0, 1, and 2 only */
740 if (memcmp(rep->key, MPA_KEY_REP, 16)) {
741 siw_init_terminate(qp, TERM_ERROR_LAYER_LLP, LLP_ETYPE_MPA,
742 LLP_ECODE_INVALID_REQ_RESP, 0);
743 siw_send_terminate(qp);
747 if (rep->params.bits & MPA_RR_FLAG_REJECT) {
748 siw_dbg_cep(cep, "got mpa reject\n");
749 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -ECONNRESET);
753 if (try_gso && rep->params.bits & MPA_RR_FLAG_GSO_EXP) {
754 siw_dbg_cep(cep, "peer allows GSO on TX\n");
755 qp->tx_ctx.gso_seg_limit = 0;
757 if ((rep->params.bits & MPA_RR_FLAG_MARKERS) ||
758 (mpa_crc_required && !(rep->params.bits & MPA_RR_FLAG_CRC)) ||
759 (mpa_crc_strict && !mpa_crc_required &&
760 (rep->params.bits & MPA_RR_FLAG_CRC))) {
761 siw_dbg_cep(cep, "reply unsupp: crc %d:%d:%d, m %d:%d\n",
762 rep->params.bits & MPA_RR_FLAG_CRC ? 1 : 0,
763 mpa_crc_required, mpa_crc_strict,
764 rep->params.bits & MPA_RR_FLAG_MARKERS ? 1 : 0, 0);
766 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -ECONNREFUSED);
770 if (cep->enhanced_rdma_conn_est) {
771 struct mpa_v2_data *v2;
773 if (__mpa_rr_revision(rep->params.bits) < MPA_REVISION_2 ||
774 !(rep->params.bits & MPA_RR_FLAG_ENHANCED)) {
776 * Protocol failure: The responder MUST reply with
777 * MPA version 2 and MUST set MPA_RR_FLAG_ENHANCED.
779 siw_dbg_cep(cep, "mpa reply error: vers %d, enhcd %d\n",
780 __mpa_rr_revision(rep->params.bits),
781 rep->params.bits & MPA_RR_FLAG_ENHANCED ?
785 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
789 v2 = (struct mpa_v2_data *)cep->mpa.pdata;
790 rep_ird = ntohs(v2->ird) & MPA_IRD_ORD_MASK;
791 rep_ord = ntohs(v2->ord) & MPA_IRD_ORD_MASK;
793 if (cep->ird < rep_ord &&
794 (relaxed_ird_negotiation == false ||
795 rep_ord > cep->sdev->attrs.max_ird)) {
796 siw_dbg_cep(cep, "ird %d, rep_ord %d, max_ord %d\n",
798 cep->sdev->attrs.max_ord);
799 ird_insufficient = true;
801 if (cep->ord > rep_ird && relaxed_ird_negotiation == false) {
802 siw_dbg_cep(cep, "ord %d, rep_ird %d\n", cep->ord,
804 ird_insufficient = true;
807 * Always report negotiated peer values to user,
808 * even if IRD/ORD negotiation failed
813 if (ird_insufficient) {
815 * If the initiator IRD is insuffient for the
816 * responder ORD, send a TERM.
818 siw_init_terminate(qp, TERM_ERROR_LAYER_LLP,
820 LLP_ECODE_INSUFFICIENT_IRD, 0);
821 siw_send_terminate(qp);
825 if (cep->mpa.v2_ctrl_req.ird & MPA_V2_PEER_TO_PEER)
827 cep->mpa.v2_ctrl_req.ord &
828 (MPA_V2_RDMA_WRITE_RTR | MPA_V2_RDMA_READ_RTR);
831 * Check if we requested P2P mode, and if peer agrees
833 if (mpa_p2p_mode != MPA_V2_RDMA_NO_RTR) {
834 if ((mpa_p2p_mode & v2->ord) == 0) {
836 * We requested RTR mode(s), but the peer
837 * did not pick any mode we support.
840 "rtr mode: req %2x, got %2x\n",
842 v2->ord & (MPA_V2_RDMA_WRITE_RTR |
843 MPA_V2_RDMA_READ_RTR));
845 siw_init_terminate(qp, TERM_ERROR_LAYER_LLP,
847 LLP_ECODE_NO_MATCHING_RTR,
849 siw_send_terminate(qp);
853 mpa_p2p_mode = v2->ord & (MPA_V2_RDMA_WRITE_RTR |
854 MPA_V2_RDMA_READ_RTR);
857 memset(&qp_attrs, 0, sizeof(qp_attrs));
859 if (rep->params.bits & MPA_RR_FLAG_CRC)
860 qp_attrs.flags = SIW_MPA_CRC;
862 qp_attrs.irq_size = cep->ird;
863 qp_attrs.orq_size = cep->ord;
864 qp_attrs.sk = cep->sock;
865 qp_attrs.state = SIW_QP_STATE_RTS;
867 qp_attr_mask = SIW_QP_ATTR_STATE | SIW_QP_ATTR_LLP_HANDLE |
868 SIW_QP_ATTR_ORD | SIW_QP_ATTR_IRD | SIW_QP_ATTR_MPA;
870 /* Move socket RX/TX under QP control */
871 down_write(&qp->state_lock);
872 if (qp->attrs.state > SIW_QP_STATE_RTR) {
874 up_write(&qp->state_lock);
877 rv = siw_qp_modify(qp, &qp_attrs, qp_attr_mask);
879 siw_qp_socket_assoc(cep, qp);
881 up_write(&qp->state_lock);
883 /* Send extra RDMA frame to trigger peer RTS if negotiated */
884 if (mpa_p2p_mode != MPA_V2_RDMA_NO_RTR) {
885 rv = siw_qp_mpa_rts(qp, mpa_p2p_mode);
890 rv = siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 0);
892 cep->state = SIW_EPSTATE_RDMA_MODE;
898 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -EINVAL);
904 * siw_accept_newconn - accept an incoming pending connection
907 static void siw_accept_newconn(struct siw_cep *cep)
909 struct socket *s = cep->sock;
910 struct socket *new_s = NULL;
911 struct siw_cep *new_cep = NULL;
912 int rv = 0; /* debug only. should disappear */
914 if (cep->state != SIW_EPSTATE_LISTENING)
917 new_cep = siw_cep_alloc(cep->sdev);
922 * 4: Allocate a sufficient number of work elements
923 * to allow concurrent handling of local + peer close
924 * events, MPA header processing + MPA timeout.
926 if (siw_cm_alloc_work(new_cep, 4) != 0)
930 * Copy saved socket callbacks from listening CEP
931 * and assign new socket with new CEP
933 new_cep->sk_state_change = cep->sk_state_change;
934 new_cep->sk_data_ready = cep->sk_data_ready;
935 new_cep->sk_write_space = cep->sk_write_space;
936 new_cep->sk_error_report = cep->sk_error_report;
938 rv = kernel_accept(s, &new_s, O_NONBLOCK);
941 * Connection already aborted by peer..?
943 siw_dbg_cep(cep, "kernel_accept() error: %d\n", rv);
946 new_cep->sock = new_s;
947 siw_cep_get(new_cep);
948 new_s->sk->sk_user_data = new_cep;
950 siw_dbg_cep(cep, "listen socket 0x%p, new 0x%p\n", s, new_s);
952 if (siw_tcp_nagle == false) {
955 rv = kernel_setsockopt(new_s, SOL_TCP, TCP_NODELAY,
956 (char *)&val, sizeof(val));
958 siw_dbg_cep(cep, "setsockopt NODELAY error: %d\n", rv);
962 new_cep->state = SIW_EPSTATE_AWAIT_MPAREQ;
964 rv = siw_cm_queue_work(new_cep, SIW_CM_WORK_MPATIMEOUT);
968 * See siw_proc_mpareq() etc. for the use of new_cep->listen_cep.
970 new_cep->listen_cep = cep;
973 if (atomic_read(&new_s->sk->sk_rmem_alloc)) {
975 * MPA REQ already queued
977 siw_dbg_cep(cep, "immediate mpa request\n");
979 siw_cep_set_inuse(new_cep);
980 rv = siw_proc_mpareq(new_cep);
981 siw_cep_set_free(new_cep);
985 new_cep->listen_cep = NULL;
994 siw_cep_put(new_cep);
997 siw_socket_disassoc(new_s);
999 new_cep->sock = NULL;
1001 siw_dbg_cep(cep, "error %d\n", rv);
1004 static void siw_cm_work_handler(struct work_struct *w)
1006 struct siw_cm_work *work;
1007 struct siw_cep *cep;
1008 int release_cep = 0, rv = 0;
1010 work = container_of(w, struct siw_cm_work, work.work);
1013 siw_dbg_cep(cep, "[QP %u]: work type: %d, state %d\n",
1014 cep->qp ? qp_id(cep->qp) : -1, work->type, cep->state);
1016 siw_cep_set_inuse(cep);
1018 switch (work->type) {
1019 case SIW_CM_WORK_ACCEPT:
1020 siw_accept_newconn(cep);
1023 case SIW_CM_WORK_READ_MPAHDR:
1024 if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) {
1025 if (cep->listen_cep) {
1026 siw_cep_set_inuse(cep->listen_cep);
1028 if (cep->listen_cep->state ==
1029 SIW_EPSTATE_LISTENING)
1030 rv = siw_proc_mpareq(cep);
1034 siw_cep_set_free(cep->listen_cep);
1036 if (rv != -EAGAIN) {
1037 siw_cep_put(cep->listen_cep);
1038 cep->listen_cep = NULL;
1043 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) {
1044 rv = siw_proc_mpareply(cep);
1047 * CEP already moved out of MPA handshake.
1048 * any connection management already done.
1049 * silently ignore the mpa packet.
1051 if (cep->state == SIW_EPSTATE_RDMA_MODE) {
1052 cep->sock->sk->sk_data_ready(cep->sock->sk);
1053 siw_dbg_cep(cep, "already in RDMA mode");
1055 siw_dbg_cep(cep, "out of state: %d\n",
1059 if (rv && rv != EAGAIN)
1063 case SIW_CM_WORK_CLOSE_LLP:
1065 * QP scheduled LLP close
1067 if (cep->qp && cep->qp->term_info.valid)
1068 siw_send_terminate(cep->qp);
1071 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0);
1076 case SIW_CM_WORK_PEER_CLOSE:
1078 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) {
1080 * MPA reply not received, but connection drop
1082 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
1084 } else if (cep->state == SIW_EPSTATE_RDMA_MODE) {
1086 * NOTE: IW_CM_EVENT_DISCONNECT is given just
1087 * to transition IWCM into CLOSING.
1089 siw_cm_upcall(cep, IW_CM_EVENT_DISCONNECT, 0);
1090 siw_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0);
1093 * for other states there is no connection
1094 * known to the IWCM.
1097 if (cep->state == SIW_EPSTATE_RECVD_MPAREQ) {
1099 * Wait for the ulp/CM to call accept/reject
1102 "mpa req recvd, wait for ULP\n");
1103 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) {
1105 * Socket close before MPA request received.
1107 siw_dbg_cep(cep, "no mpareq: drop listener\n");
1108 siw_cep_put(cep->listen_cep);
1109 cep->listen_cep = NULL;
1115 case SIW_CM_WORK_MPATIMEOUT:
1116 cep->mpa_timer = NULL;
1118 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP) {
1120 * MPA request timed out:
1121 * Hide any partially received private data and signal
1124 cep->mpa.hdr.params.pd_len = 0;
1127 siw_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
1131 } else if (cep->state == SIW_EPSTATE_AWAIT_MPAREQ) {
1133 * No MPA request received after peer TCP stream setup.
1135 if (cep->listen_cep) {
1136 siw_cep_put(cep->listen_cep);
1137 cep->listen_cep = NULL;
1144 WARN(1, "Undefined CM work type: %d\n", work->type);
1148 "release: timer=%s, QP[%u], id 0x%p\n",
1149 cep->mpa_timer ? "y" : "n",
1150 cep->qp ? qp_id(cep->qp) : -1, cep->cm_id);
1152 siw_cancel_mpatimer(cep);
1154 cep->state = SIW_EPSTATE_CLOSED;
1157 struct siw_qp *qp = cep->qp;
1159 * Serialize a potential race with application
1160 * closing the QP and calling siw_qp_cm_drop()
1163 siw_cep_set_free(cep);
1165 siw_qp_llp_close(qp);
1168 siw_cep_set_inuse(cep);
1173 siw_socket_disassoc(cep->sock);
1174 sock_release(cep->sock);
1178 cep->cm_id->rem_ref(cep->cm_id);
1183 siw_cep_set_free(cep);
1188 static struct workqueue_struct *siw_cm_wq;
1190 int siw_cm_queue_work(struct siw_cep *cep, enum siw_work_type type)
1192 struct siw_cm_work *work = siw_get_work(cep);
1193 unsigned long delay = 0;
1196 siw_dbg_cep(cep, "failed with no work available\n");
1204 INIT_DELAYED_WORK(&work->work, siw_cm_work_handler);
1206 if (type == SIW_CM_WORK_MPATIMEOUT) {
1207 cep->mpa_timer = work;
1209 if (cep->state == SIW_EPSTATE_AWAIT_MPAREP)
1210 delay = MPAREQ_TIMEOUT;
1212 delay = MPAREP_TIMEOUT;
1214 siw_dbg_cep(cep, "[QP %u]: work type: %d, work 0x%p, timeout %lu\n",
1215 cep->qp ? qp_id(cep->qp) : -1, type, work, delay);
1217 queue_delayed_work(siw_cm_wq, &work->work, delay);
1222 static void siw_cm_llp_data_ready(struct sock *sk)
1224 struct siw_cep *cep;
1226 read_lock(&sk->sk_callback_lock);
1228 cep = sk_to_cep(sk);
1233 siw_dbg_cep(cep, "state: %d\n", cep->state);
1235 switch (cep->state) {
1236 case SIW_EPSTATE_RDMA_MODE:
1238 case SIW_EPSTATE_LISTENING:
1241 case SIW_EPSTATE_AWAIT_MPAREQ:
1243 case SIW_EPSTATE_AWAIT_MPAREP:
1244 siw_cm_queue_work(cep, SIW_CM_WORK_READ_MPAHDR);
1248 siw_dbg_cep(cep, "unexpected data, state %d\n", cep->state);
1252 read_unlock(&sk->sk_callback_lock);
1255 static void siw_cm_llp_write_space(struct sock *sk)
1257 struct siw_cep *cep = sk_to_cep(sk);
1260 siw_dbg_cep(cep, "state: %d\n", cep->state);
1263 static void siw_cm_llp_error_report(struct sock *sk)
1265 struct siw_cep *cep = sk_to_cep(sk);
1268 siw_dbg_cep(cep, "error %d, socket state: %d, cep state: %d\n",
1269 sk->sk_err, sk->sk_state, cep->state);
1270 cep->sk_error_report(sk);
1274 static void siw_cm_llp_state_change(struct sock *sk)
1276 struct siw_cep *cep;
1277 void (*orig_state_change)(struct sock *s);
1279 read_lock(&sk->sk_callback_lock);
1281 cep = sk_to_cep(sk);
1283 /* endpoint already disassociated */
1284 read_unlock(&sk->sk_callback_lock);
1287 orig_state_change = cep->sk_state_change;
1289 siw_dbg_cep(cep, "state: %d\n", cep->state);
1291 switch (sk->sk_state) {
1292 case TCP_ESTABLISHED:
1294 * handle accepting socket as special case where only
1295 * new connection is possible
1297 siw_cm_queue_work(cep, SIW_CM_WORK_ACCEPT);
1301 case TCP_CLOSE_WAIT:
1303 cep->qp->tx_ctx.tx_suspend = 1;
1304 siw_cm_queue_work(cep, SIW_CM_WORK_PEER_CLOSE);
1308 siw_dbg_cep(cep, "unexpected socket state %d\n", sk->sk_state);
1310 read_unlock(&sk->sk_callback_lock);
1311 orig_state_change(sk);
1314 static int kernel_bindconnect(struct socket *s, struct sockaddr *laddr,
1315 struct sockaddr *raddr)
1317 int rv, flags = 0, s_val = 1;
1318 size_t size = laddr->sa_family == AF_INET ?
1319 sizeof(struct sockaddr_in) : sizeof(struct sockaddr_in6);
1322 * Make address available again asap.
1324 rv = kernel_setsockopt(s, SOL_SOCKET, SO_REUSEADDR, (char *)&s_val,
1329 rv = s->ops->bind(s, laddr, size);
1333 rv = s->ops->connect(s, raddr, size, flags);
1335 return rv < 0 ? rv : 0;
1338 int siw_connect(struct iw_cm_id *id, struct iw_cm_conn_param *params)
1340 struct siw_device *sdev = to_siw_dev(id->device);
1342 struct siw_cep *cep = NULL;
1343 struct socket *s = NULL;
1344 struct sockaddr *laddr = (struct sockaddr *)&id->local_addr,
1345 *raddr = (struct sockaddr *)&id->remote_addr;
1346 bool p2p_mode = peer_to_peer, v4 = true;
1347 u16 pd_len = params->private_data_len;
1348 int version = mpa_version, rv;
1350 if (pd_len > MPA_MAX_PRIVDATA)
1353 if (params->ird > sdev->attrs.max_ird ||
1354 params->ord > sdev->attrs.max_ord)
1357 if (laddr->sa_family == AF_INET6)
1359 else if (laddr->sa_family != AF_INET)
1360 return -EAFNOSUPPORT;
1363 * Respect any iwarp port mapping: Use mapped remote address
1364 * if valid. Local address must not be mapped, since siw
1365 * uses kernel TCP stack.
1367 if ((v4 && to_sockaddr_in(id->remote_addr).sin_port != 0) ||
1368 to_sockaddr_in6(id->remote_addr).sin6_port != 0)
1369 raddr = (struct sockaddr *)&id->m_remote_addr;
1371 qp = siw_qp_id2obj(sdev, params->qpn);
1373 WARN(1, "[QP %u] does not exist\n", params->qpn);
1379 "id 0x%p, pd_len %d, laddr %pI4 %d, raddr %pI4 %d\n",
1381 &((struct sockaddr_in *)(laddr))->sin_addr,
1382 ntohs(((struct sockaddr_in *)(laddr))->sin_port),
1383 &((struct sockaddr_in *)(raddr))->sin_addr,
1384 ntohs(((struct sockaddr_in *)(raddr))->sin_port));
1387 "id 0x%p, pd_len %d, laddr %pI6 %d, raddr %pI6 %d\n",
1389 &((struct sockaddr_in6 *)(laddr))->sin6_addr,
1390 ntohs(((struct sockaddr_in6 *)(laddr))->sin6_port),
1391 &((struct sockaddr_in6 *)(raddr))->sin6_addr,
1392 ntohs(((struct sockaddr_in6 *)(raddr))->sin6_port));
1394 rv = sock_create(v4 ? AF_INET : AF_INET6, SOCK_STREAM, IPPROTO_TCP, &s);
1399 * NOTE: For simplification, connect() is called in blocking
1400 * mode. Might be reconsidered for async connection setup at
1403 rv = kernel_bindconnect(s, laddr, raddr);
1405 siw_dbg_qp(qp, "kernel_bindconnect: error %d\n", rv);
1408 if (siw_tcp_nagle == false) {
1411 rv = kernel_setsockopt(s, SOL_TCP, TCP_NODELAY, (char *)&val,
1414 siw_dbg_qp(qp, "setsockopt NODELAY error: %d\n", rv);
1418 cep = siw_cep_alloc(sdev);
1423 siw_cep_set_inuse(cep);
1425 /* Associate QP with CEP */
1429 /* siw_qp_get(qp) already done by QP lookup */
1436 * 4: Allocate a sufficient number of work elements
1437 * to allow concurrent handling of local + peer close
1438 * events, MPA header processing + MPA timeout.
1440 rv = siw_cm_alloc_work(cep, 4);
1445 cep->ird = params->ird;
1446 cep->ord = params->ord;
1448 if (p2p_mode && cep->ord == 0)
1451 cep->state = SIW_EPSTATE_CONNECTING;
1454 * Associate CEP with socket
1456 siw_cep_socket_assoc(cep, s);
1458 cep->state = SIW_EPSTATE_AWAIT_MPAREP;
1461 * Set MPA Request bits: CRC if required, no MPA Markers,
1462 * MPA Rev. according to module parameter 'mpa_version', Key 'Request'.
1464 cep->mpa.hdr.params.bits = 0;
1465 if (version > MPA_REVISION_2) {
1466 pr_warn("Setting MPA version to %u\n", MPA_REVISION_2);
1467 version = MPA_REVISION_2;
1468 /* Adjust also module parameter */
1469 mpa_version = MPA_REVISION_2;
1471 __mpa_rr_set_revision(&cep->mpa.hdr.params.bits, version);
1474 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_GSO_EXP;
1476 if (mpa_crc_required)
1477 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_CRC;
1480 * If MPA version == 2:
1481 * o Include ORD and IRD.
1482 * o Indicate peer-to-peer mode, if required by module
1483 * parameter 'peer_to_peer'.
1485 if (version == MPA_REVISION_2) {
1486 cep->enhanced_rdma_conn_est = true;
1487 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_ENHANCED;
1489 cep->mpa.v2_ctrl.ird = htons(cep->ird);
1490 cep->mpa.v2_ctrl.ord = htons(cep->ord);
1493 cep->mpa.v2_ctrl.ird |= MPA_V2_PEER_TO_PEER;
1494 cep->mpa.v2_ctrl.ord |= rtr_type;
1496 /* Remember own P2P mode requested */
1497 cep->mpa.v2_ctrl_req.ird = cep->mpa.v2_ctrl.ird;
1498 cep->mpa.v2_ctrl_req.ord = cep->mpa.v2_ctrl.ord;
1500 memcpy(cep->mpa.hdr.key, MPA_KEY_REQ, 16);
1502 rv = siw_send_mpareqrep(cep, params->private_data, pd_len);
1504 * Reset private data.
1506 cep->mpa.hdr.params.pd_len = 0;
1509 rv = siw_cm_queue_work(cep, SIW_CM_WORK_MPATIMEOUT);
1511 siw_dbg_cep(cep, "id 0x%p, [QP %u]: exit\n", id,
1513 siw_cep_set_free(cep);
1518 siw_dbg_qp(qp, "failed: %d\n", rv);
1521 siw_socket_disassoc(s);
1534 cep->state = SIW_EPSTATE_CLOSED;
1536 siw_cep_set_free(cep);
1549 * siw_accept - Let SoftiWARP accept an RDMA connection request
1551 * @id: New connection management id to be used for accepted
1552 * connection request
1553 * @params: Connection parameters provided by ULP for accepting connection
1555 * Transition QP to RTS state, associate new CM id @id with accepted CEP
1556 * and get prepared for TCP input by installing socket callbacks.
1557 * Then send MPA Reply and generate the "connection established" event.
1558 * Socket callbacks must be installed before sending MPA Reply, because
1559 * the latter may cause a first RDMA message to arrive from the RDMA Initiator
1560 * side very quickly, at which time the socket callbacks must be ready.
1562 int siw_accept(struct iw_cm_id *id, struct iw_cm_conn_param *params)
1564 struct siw_device *sdev = to_siw_dev(id->device);
1565 struct siw_cep *cep = (struct siw_cep *)id->provider_data;
1567 struct siw_qp_attrs qp_attrs;
1568 int rv, max_priv_data = MPA_MAX_PRIVDATA;
1569 bool wait_for_peer_rts = false;
1571 siw_cep_set_inuse(cep);
1574 /* Free lingering inbound private data */
1575 if (cep->mpa.hdr.params.pd_len) {
1576 cep->mpa.hdr.params.pd_len = 0;
1577 kfree(cep->mpa.pdata);
1578 cep->mpa.pdata = NULL;
1580 siw_cancel_mpatimer(cep);
1582 if (cep->state != SIW_EPSTATE_RECVD_MPAREQ) {
1583 siw_dbg_cep(cep, "id 0x%p: out of state\n", id);
1585 siw_cep_set_free(cep);
1590 qp = siw_qp_id2obj(sdev, params->qpn);
1592 WARN(1, "[QP %d] does not exist\n", params->qpn);
1593 siw_cep_set_free(cep);
1598 down_write(&qp->state_lock);
1599 if (qp->attrs.state > SIW_QP_STATE_RTR) {
1601 up_write(&qp->state_lock);
1604 siw_dbg_cep(cep, "id 0x%p\n", id);
1606 if (try_gso && cep->mpa.hdr.params.bits & MPA_RR_FLAG_GSO_EXP) {
1607 siw_dbg_cep(cep, "peer allows GSO on TX\n");
1608 qp->tx_ctx.gso_seg_limit = 0;
1610 if (params->ord > sdev->attrs.max_ord ||
1611 params->ird > sdev->attrs.max_ird) {
1614 "id 0x%p, [QP %u]: ord %d (max %d), ird %d (max %d)\n",
1615 id, qp_id(qp), params->ord, sdev->attrs.max_ord,
1616 params->ird, sdev->attrs.max_ird);
1618 up_write(&qp->state_lock);
1621 if (cep->enhanced_rdma_conn_est)
1622 max_priv_data -= sizeof(struct mpa_v2_data);
1624 if (params->private_data_len > max_priv_data) {
1627 "id 0x%p, [QP %u]: private data length: %d (max %d)\n",
1628 id, qp_id(qp), params->private_data_len, max_priv_data);
1630 up_write(&qp->state_lock);
1633 if (cep->enhanced_rdma_conn_est) {
1634 if (params->ord > cep->ord) {
1635 if (relaxed_ird_negotiation) {
1636 params->ord = cep->ord;
1638 cep->ird = params->ird;
1639 cep->ord = params->ord;
1641 up_write(&qp->state_lock);
1645 if (params->ird < cep->ird) {
1646 if (relaxed_ird_negotiation &&
1647 cep->ird <= sdev->attrs.max_ird)
1648 params->ird = cep->ird;
1651 up_write(&qp->state_lock);
1655 if (cep->mpa.v2_ctrl.ord &
1656 (MPA_V2_RDMA_WRITE_RTR | MPA_V2_RDMA_READ_RTR))
1657 wait_for_peer_rts = true;
1659 * Signal back negotiated IRD and ORD values
1661 cep->mpa.v2_ctrl.ord =
1662 htons(params->ord & MPA_IRD_ORD_MASK) |
1663 (cep->mpa.v2_ctrl.ord & ~MPA_V2_MASK_IRD_ORD);
1664 cep->mpa.v2_ctrl.ird =
1665 htons(params->ird & MPA_IRD_ORD_MASK) |
1666 (cep->mpa.v2_ctrl.ird & ~MPA_V2_MASK_IRD_ORD);
1668 cep->ird = params->ird;
1669 cep->ord = params->ord;
1674 memset(&qp_attrs, 0, sizeof(qp_attrs));
1675 qp_attrs.orq_size = cep->ord;
1676 qp_attrs.irq_size = cep->ird;
1677 qp_attrs.sk = cep->sock;
1678 if (cep->mpa.hdr.params.bits & MPA_RR_FLAG_CRC)
1679 qp_attrs.flags = SIW_MPA_CRC;
1680 qp_attrs.state = SIW_QP_STATE_RTS;
1682 siw_dbg_cep(cep, "id 0x%p, [QP%u]: moving to rts\n", id, qp_id(qp));
1684 /* Associate QP with CEP */
1688 /* siw_qp_get(qp) already done by QP lookup */
1691 cep->state = SIW_EPSTATE_RDMA_MODE;
1693 /* Move socket RX/TX under QP control */
1694 rv = siw_qp_modify(qp, &qp_attrs,
1695 SIW_QP_ATTR_STATE | SIW_QP_ATTR_LLP_HANDLE |
1696 SIW_QP_ATTR_ORD | SIW_QP_ATTR_IRD |
1698 up_write(&qp->state_lock);
1703 siw_dbg_cep(cep, "id 0x%p, [QP %u]: send mpa reply, %d byte pdata\n",
1704 id, qp_id(qp), params->private_data_len);
1706 rv = siw_send_mpareqrep(cep, params->private_data,
1707 params->private_data_len);
1711 if (wait_for_peer_rts) {
1712 siw_sk_assign_rtr_upcalls(cep);
1714 siw_qp_socket_assoc(cep, qp);
1715 rv = siw_cm_upcall(cep, IW_CM_EVENT_ESTABLISHED, 0);
1719 siw_cep_set_free(cep);
1723 siw_socket_disassoc(cep->sock);
1724 sock_release(cep->sock);
1727 cep->state = SIW_EPSTATE_CLOSED;
1730 cep->cm_id->rem_ref(id);
1740 siw_cep_set_free(cep);
1749 * Local connection reject case. Send private data back to peer,
1750 * close connection and dereference connection id.
1752 int siw_reject(struct iw_cm_id *id, const void *pdata, u8 pd_len)
1754 struct siw_cep *cep = (struct siw_cep *)id->provider_data;
1756 siw_cep_set_inuse(cep);
1759 siw_cancel_mpatimer(cep);
1761 if (cep->state != SIW_EPSTATE_RECVD_MPAREQ) {
1762 siw_dbg_cep(cep, "id 0x%p: out of state\n", id);
1764 siw_cep_set_free(cep);
1765 siw_cep_put(cep); /* put last reference */
1769 siw_dbg_cep(cep, "id 0x%p, cep->state %d, pd_len %d\n", id, cep->state,
1772 if (__mpa_rr_revision(cep->mpa.hdr.params.bits) >= MPA_REVISION_1) {
1773 cep->mpa.hdr.params.bits |= MPA_RR_FLAG_REJECT; /* reject */
1774 siw_send_mpareqrep(cep, pdata, pd_len);
1776 siw_socket_disassoc(cep->sock);
1777 sock_release(cep->sock);
1780 cep->state = SIW_EPSTATE_CLOSED;
1782 siw_cep_set_free(cep);
1788 static int siw_listen_address(struct iw_cm_id *id, int backlog,
1789 struct sockaddr *laddr, int addr_family)
1792 struct siw_cep *cep = NULL;
1793 struct siw_device *sdev = to_siw_dev(id->device);
1796 rv = sock_create(addr_family, SOCK_STREAM, IPPROTO_TCP, &s);
1801 * Allow binding local port when still in TIME_WAIT from last close.
1804 rv = kernel_setsockopt(s, SOL_SOCKET, SO_REUSEADDR, (char *)&s_val,
1807 siw_dbg(id->device, "id 0x%p: setsockopt error: %d\n", id, rv);
1810 rv = s->ops->bind(s, laddr, addr_family == AF_INET ?
1811 sizeof(struct sockaddr_in) :
1812 sizeof(struct sockaddr_in6));
1814 siw_dbg(id->device, "id 0x%p: socket bind error: %d\n", id, rv);
1817 cep = siw_cep_alloc(sdev);
1822 siw_cep_socket_assoc(cep, s);
1824 rv = siw_cm_alloc_work(cep, backlog);
1827 "id 0x%p: alloc_work error %d, backlog %d\n", id,
1831 rv = s->ops->listen(s, backlog);
1833 siw_dbg(id->device, "id 0x%p: listen error %d\n", id, rv);
1840 * In case of a wildcard rdma_listen on a multi-homed device,
1841 * a listener's IWCM id is associated with more than one listening CEP.
1843 * We currently use id->provider_data in three different ways:
1845 * o For a listener's IWCM id, id->provider_data points to
1846 * the list_head of the list of listening CEPs.
1847 * Uses: siw_create_listen(), siw_destroy_listen()
1849 * o For each accepted passive-side IWCM id, id->provider_data
1850 * points to the CEP itself. This is a consequence of
1851 * - siw_cm_upcall() setting event.provider_data = cep and
1852 * - the IWCM's cm_conn_req_handler() setting provider_data of the
1853 * new passive-side IWCM id equal to event.provider_data
1854 * Uses: siw_accept(), siw_reject()
1856 * o For an active-side IWCM id, id->provider_data is not used at all.
1859 if (!id->provider_data) {
1861 kmalloc(sizeof(struct list_head), GFP_KERNEL);
1862 if (!id->provider_data) {
1866 INIT_LIST_HEAD((struct list_head *)id->provider_data);
1868 list_add_tail(&cep->listenq, (struct list_head *)id->provider_data);
1869 cep->state = SIW_EPSTATE_LISTENING;
1871 if (addr_family == AF_INET)
1872 siw_dbg(id->device, "Listen at laddr %pI4 %u\n",
1873 &(((struct sockaddr_in *)laddr)->sin_addr),
1874 ((struct sockaddr_in *)laddr)->sin_port);
1876 siw_dbg(id->device, "Listen at laddr %pI6 %u\n",
1877 &(((struct sockaddr_in6 *)laddr)->sin6_addr),
1878 ((struct sockaddr_in6 *)laddr)->sin6_port);
1883 siw_dbg(id->device, "failed: %d\n", rv);
1886 siw_cep_set_inuse(cep);
1889 cep->cm_id->rem_ref(cep->cm_id);
1893 siw_socket_disassoc(s);
1894 cep->state = SIW_EPSTATE_CLOSED;
1896 siw_cep_set_free(cep);
1904 static void siw_drop_listeners(struct iw_cm_id *id)
1906 struct list_head *p, *tmp;
1909 * In case of a wildcard rdma_listen on a multi-homed device,
1910 * a listener's IWCM id is associated with more than one listening CEP.
1912 list_for_each_safe(p, tmp, (struct list_head *)id->provider_data) {
1913 struct siw_cep *cep = list_entry(p, struct siw_cep, listenq);
1917 siw_dbg_cep(cep, "id 0x%p: drop cep, state %d\n", id,
1920 siw_cep_set_inuse(cep);
1923 cep->cm_id->rem_ref(cep->cm_id);
1927 siw_socket_disassoc(cep->sock);
1928 sock_release(cep->sock);
1931 cep->state = SIW_EPSTATE_CLOSED;
1932 siw_cep_set_free(cep);
1938 * siw_create_listen - Create resources for a listener's IWCM ID @id
1940 * Listens on the socket addresses id->local_addr and id->remote_addr.
1942 * If the listener's @id provides a specific local IP address, at most one
1943 * listening socket is created and associated with @id.
1945 * If the listener's @id provides the wildcard (zero) local IP address,
1946 * a separate listen is performed for each local IP address of the device
1947 * by creating a listening socket and binding to that local IP address.
1950 int siw_create_listen(struct iw_cm_id *id, int backlog)
1952 struct net_device *dev = to_siw_dev(id->device)->netdev;
1953 int rv = 0, listeners = 0;
1955 siw_dbg(id->device, "id 0x%p: backlog %d\n", id, backlog);
1958 * For each attached address of the interface, create a
1959 * listening socket, if id->local_addr is the wildcard
1960 * IP address or matches the IP address.
1962 if (id->local_addr.ss_family == AF_INET) {
1963 struct in_device *in_dev = in_dev_get(dev);
1964 struct sockaddr_in s_laddr, *s_raddr;
1965 const struct in_ifaddr *ifa;
1967 memcpy(&s_laddr, &id->local_addr, sizeof(s_laddr));
1968 s_raddr = (struct sockaddr_in *)&id->remote_addr;
1971 "id 0x%p: laddr %pI4:%d, raddr %pI4:%d\n",
1972 id, &s_laddr.sin_addr, ntohs(s_laddr.sin_port),
1973 &s_raddr->sin_addr, ntohs(s_raddr->sin_port));
1976 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1977 if (ipv4_is_zeronet(s_laddr.sin_addr.s_addr) ||
1978 s_laddr.sin_addr.s_addr == ifa->ifa_address) {
1979 s_laddr.sin_addr.s_addr = ifa->ifa_address;
1981 rv = siw_listen_address(id, backlog,
1982 (struct sockaddr *)&s_laddr,
1990 } else if (id->local_addr.ss_family == AF_INET6) {
1991 struct inet6_dev *in6_dev = in6_dev_get(dev);
1992 struct inet6_ifaddr *ifp;
1993 struct sockaddr_in6 *s_laddr = &to_sockaddr_in6(id->local_addr),
1994 *s_raddr = &to_sockaddr_in6(id->remote_addr);
1997 "id 0x%p: laddr %pI6:%d, raddr %pI6:%d\n",
1998 id, &s_laddr->sin6_addr, ntohs(s_laddr->sin6_port),
1999 &s_raddr->sin6_addr, ntohs(s_raddr->sin6_port));
2001 read_lock_bh(&in6_dev->lock);
2002 list_for_each_entry(ifp, &in6_dev->addr_list, if_list) {
2003 struct sockaddr_in6 bind_addr;
2005 if (ipv6_addr_any(&s_laddr->sin6_addr) ||
2006 ipv6_addr_equal(&s_laddr->sin6_addr, &ifp->addr)) {
2007 bind_addr.sin6_family = AF_INET6;
2008 bind_addr.sin6_port = s_laddr->sin6_port;
2009 bind_addr.sin6_flowinfo = 0;
2010 bind_addr.sin6_addr = ifp->addr;
2011 bind_addr.sin6_scope_id = dev->ifindex;
2013 rv = siw_listen_address(id, backlog,
2014 (struct sockaddr *)&bind_addr,
2020 read_unlock_bh(&in6_dev->lock);
2022 in6_dev_put(in6_dev);
2024 return -EAFNOSUPPORT;
2031 siw_dbg(id->device, "id 0x%p: %s\n", id, rv ? "FAIL" : "OK");
2036 int siw_destroy_listen(struct iw_cm_id *id)
2038 siw_dbg(id->device, "id 0x%p\n", id);
2040 if (!id->provider_data) {
2041 siw_dbg(id->device, "id 0x%p: no cep(s)\n", id);
2044 siw_drop_listeners(id);
2045 kfree(id->provider_data);
2046 id->provider_data = NULL;
2051 int siw_cm_init(void)
2054 * create_single_workqueue for strict ordering
2056 siw_cm_wq = create_singlethread_workqueue("siw_cm_wq");
2063 void siw_cm_exit(void)
2066 flush_workqueue(siw_cm_wq);
2067 destroy_workqueue(siw_cm_wq);