1 // SPDX-License-Identifier: GPL-2.0-only
3 * Intel MIC Platform Software Stack (MPSS)
5 * Copyright(c) 2014 Intel Corporation.
9 #include <linux/scif.h>
10 #include "scif_main.h"
13 static const char * const scif_ep_states[] = {
25 enum conn_async_state {
26 ASYNC_CONN_IDLE = 1, /* ep setup for async connect */
27 ASYNC_CONN_INPROGRESS, /* async connect in progress */
28 ASYNC_CONN_FLUSH_WORK /* async work flush in progress */
32 * File operations for anonymous inode file associated with a SCIF endpoint,
33 * used in kernel mode SCIF poll. Kernel mode SCIF poll calls portions of the
34 * poll API in the kernel and these take in a struct file *. Since a struct
35 * file is not available to kernel mode SCIF, it uses an anonymous file for
38 const struct file_operations scif_anon_fops = {
42 scif_epd_t scif_open(void)
44 struct scif_endpt *ep;
48 ep = kzalloc(sizeof(*ep), GFP_KERNEL);
52 ep->qp_info.qp = kzalloc(sizeof(*ep->qp_info.qp), GFP_KERNEL);
56 err = scif_anon_inode_getfile(ep);
60 spin_lock_init(&ep->lock);
61 mutex_init(&ep->sendlock);
62 mutex_init(&ep->recvlock);
65 ep->state = SCIFEP_UNBOUND;
66 dev_dbg(scif_info.mdev.this_device,
67 "SCIFAPI open: ep %p success\n", ep);
71 kfree(ep->qp_info.qp);
77 EXPORT_SYMBOL_GPL(scif_open);
80 * scif_disconnect_ep - Disconnects the endpoint if found
81 * @epd: The end point returned from scif_open()
83 static struct scif_endpt *scif_disconnect_ep(struct scif_endpt *ep)
86 struct scif_endpt *fep = NULL;
87 struct scif_endpt *tmpep;
88 struct list_head *pos, *tmpq;
92 * Wake up any threads blocked in send()/recv() before closing
93 * out the connection. Grabbing and releasing the send/recv lock
94 * will ensure that any blocked senders/receivers have exited for
95 * Ring 0 endpoints. It is a Ring 0 bug to call send/recv after
96 * close. Ring 3 endpoints are not affected since close will not
97 * be called while there are IOCTLs executing.
99 wake_up_interruptible(&ep->sendwq);
100 wake_up_interruptible(&ep->recvwq);
101 mutex_lock(&ep->sendlock);
102 mutex_unlock(&ep->sendlock);
103 mutex_lock(&ep->recvlock);
104 mutex_unlock(&ep->recvlock);
106 /* Remove from the connected list */
107 mutex_lock(&scif_info.connlock);
108 list_for_each_safe(pos, tmpq, &scif_info.connected) {
109 tmpep = list_entry(pos, struct scif_endpt, list);
113 spin_lock(&ep->lock);
120 * The other side has completed the disconnect before
121 * the end point can be removed from the list. Therefore
122 * the ep lock is not locked, traverse the disconnected
123 * list to find the endpoint and release the conn lock.
125 list_for_each_safe(pos, tmpq, &scif_info.disconnected) {
126 tmpep = list_entry(pos, struct scif_endpt, list);
132 mutex_unlock(&scif_info.connlock);
136 init_completion(&ep->discon);
137 msg.uop = SCIF_DISCNCT;
140 msg.payload[0] = (u64)ep;
141 msg.payload[1] = ep->remote_ep;
143 err = scif_nodeqp_send(ep->remote_dev, &msg);
144 spin_unlock(&ep->lock);
145 mutex_unlock(&scif_info.connlock);
148 /* Wait for the remote node to respond with SCIF_DISCNT_ACK */
149 wait_for_completion_timeout(&ep->discon,
150 SCIF_NODE_ALIVE_TIMEOUT);
154 int scif_close(scif_epd_t epd)
156 struct scif_endpt *ep = (struct scif_endpt *)epd;
157 struct scif_endpt *tmpep;
158 struct list_head *pos, *tmpq;
159 enum scif_epd_state oldstate;
162 dev_dbg(scif_info.mdev.this_device, "SCIFAPI close: ep %p %s\n",
163 ep, scif_ep_states[ep->state]);
165 spin_lock(&ep->lock);
166 flush_conn = (ep->conn_async_state == ASYNC_CONN_INPROGRESS);
167 spin_unlock(&ep->lock);
170 flush_work(&scif_info.conn_work);
172 spin_lock(&ep->lock);
173 oldstate = ep->state;
175 ep->state = SCIFEP_CLOSING;
179 dev_err(scif_info.mdev.this_device,
180 "SCIFAPI close: zombie state unexpected\n");
182 case SCIFEP_DISCONNECTED:
183 spin_unlock(&ep->lock);
184 scif_unregister_all_windows(epd);
185 /* Remove from the disconnected list */
186 mutex_lock(&scif_info.connlock);
187 list_for_each_safe(pos, tmpq, &scif_info.disconnected) {
188 tmpep = list_entry(pos, struct scif_endpt, list);
194 mutex_unlock(&scif_info.connlock);
198 case SCIFEP_CONNECTING:
199 spin_unlock(&ep->lock);
202 case SCIFEP_CONNECTED:
205 spin_unlock(&ep->lock);
206 scif_unregister_all_windows(epd);
207 scif_disconnect_ep(ep);
210 case SCIFEP_LISTENING:
211 case SCIFEP_CLLISTEN:
213 struct scif_conreq *conreq;
215 struct scif_endpt *aep;
217 spin_unlock(&ep->lock);
218 mutex_lock(&scif_info.eplock);
220 /* remove from listen list */
221 list_for_each_safe(pos, tmpq, &scif_info.listen) {
222 tmpep = list_entry(pos, struct scif_endpt, list);
226 /* Remove any dangling accepts */
227 while (ep->acceptcnt) {
228 aep = list_first_entry(&ep->li_accept,
229 struct scif_endpt, liacceptlist);
230 list_del(&aep->liacceptlist);
231 scif_put_port(aep->port.port);
232 list_for_each_safe(pos, tmpq, &scif_info.uaccept) {
233 tmpep = list_entry(pos, struct scif_endpt,
240 mutex_unlock(&scif_info.eplock);
241 mutex_lock(&scif_info.connlock);
242 list_for_each_safe(pos, tmpq, &scif_info.connected) {
243 tmpep = list_entry(pos,
244 struct scif_endpt, list);
250 list_for_each_safe(pos, tmpq, &scif_info.disconnected) {
251 tmpep = list_entry(pos,
252 struct scif_endpt, list);
258 mutex_unlock(&scif_info.connlock);
259 scif_teardown_ep(aep);
260 mutex_lock(&scif_info.eplock);
261 scif_add_epd_to_zombie_list(aep, SCIF_EPLOCK_HELD);
265 spin_lock(&ep->lock);
266 mutex_unlock(&scif_info.eplock);
268 /* Remove and reject any pending connection requests. */
269 while (ep->conreqcnt) {
270 conreq = list_first_entry(&ep->conlist,
271 struct scif_conreq, list);
272 list_del(&conreq->list);
274 msg.uop = SCIF_CNCT_REJ;
275 msg.dst.node = conreq->msg.src.node;
276 msg.dst.port = conreq->msg.src.port;
277 msg.payload[0] = conreq->msg.payload[0];
278 msg.payload[1] = conreq->msg.payload[1];
280 * No Error Handling on purpose for scif_nodeqp_send().
281 * If the remote node is lost we still want free the
282 * connection requests on the self node.
284 scif_nodeqp_send(&scif_dev[conreq->msg.src.node],
290 spin_unlock(&ep->lock);
291 /* If a kSCIF accept is waiting wake it up */
292 wake_up_interruptible(&ep->conwq);
296 scif_put_port(ep->port.port);
297 scif_anon_inode_fput(ep);
298 scif_teardown_ep(ep);
299 scif_add_epd_to_zombie_list(ep, !SCIF_EPLOCK_HELD);
302 EXPORT_SYMBOL_GPL(scif_close);
305 * scif_flush() - Wakes up any blocking accepts. The endpoint will no longer
306 * accept new connections.
307 * @epd: The end point returned from scif_open()
309 int __scif_flush(scif_epd_t epd)
311 struct scif_endpt *ep = (struct scif_endpt *)epd;
314 case SCIFEP_LISTENING:
316 ep->state = SCIFEP_CLLISTEN;
318 /* If an accept is waiting wake it up */
319 wake_up_interruptible(&ep->conwq);
328 int scif_bind(scif_epd_t epd, u16 pn)
330 struct scif_endpt *ep = (struct scif_endpt *)epd;
334 dev_dbg(scif_info.mdev.this_device,
335 "SCIFAPI bind: ep %p %s requested port number %d\n",
336 ep, scif_ep_states[ep->state], pn);
339 * Similar to IETF RFC 1700, SCIF ports below
340 * SCIF_ADMIN_PORT_END can only be bound by system (or root)
341 * processes or by processes executed by privileged users.
343 if (pn < SCIF_ADMIN_PORT_END && !capable(CAP_SYS_ADMIN)) {
345 goto scif_bind_admin_exit;
349 spin_lock(&ep->lock);
350 if (ep->state == SCIFEP_BOUND) {
353 } else if (ep->state != SCIFEP_UNBOUND) {
359 tmp = scif_rsrv_port(pn);
365 ret = scif_get_new_port();
371 ep->state = SCIFEP_BOUND;
372 ep->port.node = scif_info.nodeid;
374 ep->conn_async_state = ASYNC_CONN_IDLE;
376 dev_dbg(scif_info.mdev.this_device,
377 "SCIFAPI bind: bound to port number %d\n", pn);
379 spin_unlock(&ep->lock);
380 scif_bind_admin_exit:
383 EXPORT_SYMBOL_GPL(scif_bind);
385 int scif_listen(scif_epd_t epd, int backlog)
387 struct scif_endpt *ep = (struct scif_endpt *)epd;
389 dev_dbg(scif_info.mdev.this_device,
390 "SCIFAPI listen: ep %p %s\n", ep, scif_ep_states[ep->state]);
391 spin_lock(&ep->lock);
395 case SCIFEP_CLLISTEN:
397 case SCIFEP_DISCONNECTED:
398 spin_unlock(&ep->lock);
400 case SCIFEP_LISTENING:
401 case SCIFEP_CONNECTED:
402 case SCIFEP_CONNECTING:
404 spin_unlock(&ep->lock);
410 ep->state = SCIFEP_LISTENING;
411 ep->backlog = backlog;
415 INIT_LIST_HEAD(&ep->conlist);
416 init_waitqueue_head(&ep->conwq);
417 INIT_LIST_HEAD(&ep->li_accept);
418 spin_unlock(&ep->lock);
421 * Listen status is complete so delete the qp information not needed
422 * on a listen before placing on the list of listening ep's
424 scif_teardown_ep(ep);
425 ep->qp_info.qp = NULL;
427 mutex_lock(&scif_info.eplock);
428 list_add_tail(&ep->list, &scif_info.listen);
429 mutex_unlock(&scif_info.eplock);
432 EXPORT_SYMBOL_GPL(scif_listen);
435 ************************************************************************
436 * SCIF connection flow:
438 * 1) A SCIF listening endpoint can call scif_accept(..) to wait for SCIF
439 * connections via a SCIF_CNCT_REQ message
440 * 2) A SCIF endpoint can initiate a SCIF connection by calling
441 * scif_connect(..) which calls scif_setup_qp_connect(..) which
442 * allocates the local qp for the endpoint ring buffer and then sends
443 * a SCIF_CNCT_REQ to the remote node and waits for a SCIF_CNCT_GNT or
444 * a SCIF_CNCT_REJ message
445 * 3) The peer node handles a SCIF_CNCT_REQ via scif_cnctreq_resp(..) which
446 * wakes up any threads blocked in step 1 or sends a SCIF_CNCT_REJ
448 * 4) A thread blocked waiting for incoming connections allocates its local
449 * endpoint QP and ring buffer following which it sends a SCIF_CNCT_GNT
450 * and waits for a SCIF_CNCT_GNT(N)ACK. If the allocation fails then
451 * the node sends a SCIF_CNCT_REJ message
452 * 5) Upon receipt of a SCIF_CNCT_GNT or a SCIF_CNCT_REJ message the
453 * connecting endpoint is woken up as part of handling
454 * scif_cnctgnt_resp(..) following which it maps the remote endpoints'
455 * QP, updates its outbound QP and sends a SCIF_CNCT_GNTACK message on
456 * success or a SCIF_CNCT_GNTNACK message on failure and completes
457 * the scif_connect(..) API
458 * 6) Upon receipt of a SCIF_CNCT_GNT(N)ACK the accepting endpoint blocked
459 * in step 4 is woken up and completes the scif_accept(..) API
460 * 7) The SCIF connection is now established between the two SCIF endpoints.
462 static int scif_conn_func(struct scif_endpt *ep)
466 struct device *spdev;
468 err = scif_reserve_dma_chan(ep);
470 dev_err(&ep->remote_dev->sdev->dev,
471 "%s %d err %d\n", __func__, __LINE__, err);
472 ep->state = SCIFEP_BOUND;
473 goto connect_error_simple;
475 /* Initiate the first part of the endpoint QP setup */
476 err = scif_setup_qp_connect(ep->qp_info.qp, &ep->qp_info.qp_offset,
477 SCIF_ENDPT_QP_SIZE, ep->remote_dev);
479 dev_err(&ep->remote_dev->sdev->dev,
480 "%s err %d qp_offset 0x%llx\n",
481 __func__, err, ep->qp_info.qp_offset);
482 ep->state = SCIFEP_BOUND;
483 goto connect_error_simple;
486 spdev = scif_get_peer_dev(ep->remote_dev);
488 err = PTR_ERR(spdev);
491 /* Format connect message and send it */
493 msg.dst = ep->conn_port;
494 msg.uop = SCIF_CNCT_REQ;
495 msg.payload[0] = (u64)ep;
496 msg.payload[1] = ep->qp_info.qp_offset;
497 err = _scif_nodeqp_send(ep->remote_dev, &msg);
499 goto connect_error_dec;
500 scif_put_peer_dev(spdev);
502 * Wait for the remote node to respond with SCIF_CNCT_GNT or
503 * SCIF_CNCT_REJ message.
505 err = wait_event_timeout(ep->conwq, ep->state != SCIFEP_CONNECTING,
506 SCIF_NODE_ALIVE_TIMEOUT);
508 dev_err(&ep->remote_dev->sdev->dev,
509 "%s %d timeout\n", __func__, __LINE__);
510 ep->state = SCIFEP_BOUND;
512 spdev = scif_get_peer_dev(ep->remote_dev);
514 err = PTR_ERR(spdev);
517 if (ep->state == SCIFEP_MAPPING) {
518 err = scif_setup_qp_connect_response(ep->remote_dev,
520 ep->qp_info.gnt_pld);
522 * If the resource to map the queue are not available then
523 * we need to tell the other side to terminate the accept
526 dev_err(&ep->remote_dev->sdev->dev,
527 "%s %d err %d\n", __func__, __LINE__, err);
528 msg.uop = SCIF_CNCT_GNTNACK;
529 msg.payload[0] = ep->remote_ep;
530 _scif_nodeqp_send(ep->remote_dev, &msg);
531 ep->state = SCIFEP_BOUND;
532 goto connect_error_dec;
535 msg.uop = SCIF_CNCT_GNTACK;
536 msg.payload[0] = ep->remote_ep;
537 err = _scif_nodeqp_send(ep->remote_dev, &msg);
539 ep->state = SCIFEP_BOUND;
540 goto connect_error_dec;
542 ep->state = SCIFEP_CONNECTED;
543 mutex_lock(&scif_info.connlock);
544 list_add_tail(&ep->list, &scif_info.connected);
545 mutex_unlock(&scif_info.connlock);
546 dev_dbg(&ep->remote_dev->sdev->dev,
547 "SCIFAPI connect: ep %p connected\n", ep);
548 } else if (ep->state == SCIFEP_BOUND) {
549 dev_dbg(&ep->remote_dev->sdev->dev,
550 "SCIFAPI connect: ep %p connection refused\n", ep);
552 goto connect_error_dec;
554 scif_put_peer_dev(spdev);
557 scif_put_peer_dev(spdev);
559 scif_cleanup_ep_qp(ep);
560 connect_error_simple:
567 * Workqueue handler for servicing non-blocking SCIF connect
570 void scif_conn_handler(struct work_struct *work)
572 struct scif_endpt *ep;
576 spin_lock(&scif_info.nb_connect_lock);
577 if (!list_empty(&scif_info.nb_connect_list)) {
578 ep = list_first_entry(&scif_info.nb_connect_list,
579 struct scif_endpt, conn_list);
580 list_del(&ep->conn_list);
582 spin_unlock(&scif_info.nb_connect_lock);
584 ep->conn_err = scif_conn_func(ep);
585 wake_up_interruptible(&ep->conn_pend_wq);
590 int __scif_connect(scif_epd_t epd, struct scif_port_id *dst, bool non_block)
592 struct scif_endpt *ep = (struct scif_endpt *)epd;
594 struct scif_dev *remote_dev;
595 struct device *spdev;
597 dev_dbg(scif_info.mdev.this_device, "SCIFAPI connect: ep %p %s\n", ep,
598 scif_ep_states[ep->state]);
600 if (!scif_dev || dst->node > scif_info.maxid)
605 remote_dev = &scif_dev[dst->node];
606 spdev = scif_get_peer_dev(remote_dev);
608 err = PTR_ERR(spdev);
612 spin_lock(&ep->lock);
618 case SCIFEP_DISCONNECTED:
619 if (ep->conn_async_state == ASYNC_CONN_INPROGRESS)
620 ep->conn_async_state = ASYNC_CONN_FLUSH_WORK;
624 case SCIFEP_LISTENING:
625 case SCIFEP_CLLISTEN:
628 case SCIFEP_CONNECTING:
630 if (ep->conn_async_state == ASYNC_CONN_INPROGRESS)
635 case SCIFEP_CONNECTED:
636 if (ep->conn_async_state == ASYNC_CONN_INPROGRESS)
637 ep->conn_async_state = ASYNC_CONN_FLUSH_WORK;
642 err = scif_get_new_port();
646 ep->port.node = scif_info.nodeid;
647 ep->conn_async_state = ASYNC_CONN_IDLE;
651 * If a non-blocking connect has been already initiated
652 * (conn_async_state is either ASYNC_CONN_INPROGRESS or
653 * ASYNC_CONN_FLUSH_WORK), the end point could end up in
654 * SCIF_BOUND due an error in the connection process
655 * (e.g., connection refused) If conn_async_state is
656 * ASYNC_CONN_INPROGRESS - transition to ASYNC_CONN_FLUSH_WORK
657 * so that the error status can be collected. If the state is
658 * already ASYNC_CONN_FLUSH_WORK - then set the error to
659 * EINPROGRESS since some other thread is waiting to collect
662 if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) {
663 ep->conn_async_state = ASYNC_CONN_FLUSH_WORK;
664 } else if (ep->conn_async_state == ASYNC_CONN_FLUSH_WORK) {
667 ep->conn_port = *dst;
668 init_waitqueue_head(&ep->sendwq);
669 init_waitqueue_head(&ep->recvwq);
670 init_waitqueue_head(&ep->conwq);
671 ep->conn_async_state = 0;
673 if (unlikely(non_block))
674 ep->conn_async_state = ASYNC_CONN_INPROGRESS;
679 if (err || ep->conn_async_state == ASYNC_CONN_FLUSH_WORK)
680 goto connect_simple_unlock1;
682 ep->state = SCIFEP_CONNECTING;
683 ep->remote_dev = &scif_dev[dst->node];
684 ep->qp_info.qp->magic = SCIFEP_MAGIC;
685 if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) {
686 init_waitqueue_head(&ep->conn_pend_wq);
687 spin_lock(&scif_info.nb_connect_lock);
688 list_add_tail(&ep->conn_list, &scif_info.nb_connect_list);
689 spin_unlock(&scif_info.nb_connect_lock);
691 schedule_work(&scif_info.conn_work);
693 connect_simple_unlock1:
694 spin_unlock(&ep->lock);
695 scif_put_peer_dev(spdev);
698 } else if (ep->conn_async_state == ASYNC_CONN_FLUSH_WORK) {
699 flush_work(&scif_info.conn_work);
701 spin_lock(&ep->lock);
702 ep->conn_async_state = ASYNC_CONN_IDLE;
703 spin_unlock(&ep->lock);
705 err = scif_conn_func(ep);
710 int scif_connect(scif_epd_t epd, struct scif_port_id *dst)
712 return __scif_connect(epd, dst, false);
714 EXPORT_SYMBOL_GPL(scif_connect);
717 * scif_accept() - Accept a connection request from the remote node
719 * The function accepts a connection request from the remote node. Successful
720 * complete is indicate by a new end point being created and passed back
721 * to the caller for future reference.
723 * Upon successful complete a zero will be returned and the peer information
726 * If the end point is not in the listening state -EINVAL will be returned.
728 * If during the connection sequence resource allocation fails the -ENOMEM
731 * If the function is called with the ASYNC flag set and no connection requests
732 * are pending it will return -EAGAIN.
734 * If the remote side is not sending any connection requests the caller may
735 * terminate this function with a signal. If so a -EINTR will be returned.
737 int scif_accept(scif_epd_t epd, struct scif_port_id *peer,
738 scif_epd_t *newepd, int flags)
740 struct scif_endpt *lep = (struct scif_endpt *)epd;
741 struct scif_endpt *cep;
742 struct scif_conreq *conreq;
745 struct device *spdev;
747 dev_dbg(scif_info.mdev.this_device,
748 "SCIFAPI accept: ep %p %s\n", lep, scif_ep_states[lep->state]);
750 if (flags & ~SCIF_ACCEPT_SYNC)
753 if (!peer || !newepd)
757 spin_lock(&lep->lock);
758 if (lep->state != SCIFEP_LISTENING) {
759 spin_unlock(&lep->lock);
763 if (!lep->conreqcnt && !(flags & SCIF_ACCEPT_SYNC)) {
764 /* No connection request present and we do not want to wait */
765 spin_unlock(&lep->lock);
769 lep->files = current->files;
771 spin_unlock(&lep->lock);
772 /* Wait for the remote node to send us a SCIF_CNCT_REQ */
773 err = wait_event_interruptible(lep->conwq,
775 (lep->state != SCIFEP_LISTENING)));
779 if (lep->state != SCIFEP_LISTENING)
782 spin_lock(&lep->lock);
785 goto retry_connection;
787 /* Get the first connect request off the list */
788 conreq = list_first_entry(&lep->conlist, struct scif_conreq, list);
789 list_del(&conreq->list);
791 spin_unlock(&lep->lock);
793 /* Fill in the peer information */
794 peer->node = conreq->msg.src.node;
795 peer->port = conreq->msg.src.port;
797 cep = kzalloc(sizeof(*cep), GFP_KERNEL);
800 goto scif_accept_error_epalloc;
802 spin_lock_init(&cep->lock);
803 mutex_init(&cep->sendlock);
804 mutex_init(&cep->recvlock);
805 cep->state = SCIFEP_CONNECTING;
806 cep->remote_dev = &scif_dev[peer->node];
807 cep->remote_ep = conreq->msg.payload[0];
809 scif_rma_ep_init(cep);
811 err = scif_reserve_dma_chan(cep);
813 dev_err(scif_info.mdev.this_device,
814 "%s %d err %d\n", __func__, __LINE__, err);
815 goto scif_accept_error_qpalloc;
818 cep->qp_info.qp = kzalloc(sizeof(*cep->qp_info.qp), GFP_KERNEL);
819 if (!cep->qp_info.qp) {
821 goto scif_accept_error_qpalloc;
824 err = scif_anon_inode_getfile(cep);
826 goto scif_accept_error_anon_inode;
828 cep->qp_info.qp->magic = SCIFEP_MAGIC;
829 spdev = scif_get_peer_dev(cep->remote_dev);
831 err = PTR_ERR(spdev);
832 goto scif_accept_error_map;
834 err = scif_setup_qp_accept(cep->qp_info.qp, &cep->qp_info.qp_offset,
835 conreq->msg.payload[1], SCIF_ENDPT_QP_SIZE,
838 dev_dbg(&cep->remote_dev->sdev->dev,
839 "SCIFAPI accept: ep %p new %p scif_setup_qp_accept %d qp_offset 0x%llx\n",
840 lep, cep, err, cep->qp_info.qp_offset);
841 scif_put_peer_dev(spdev);
842 goto scif_accept_error_map;
845 cep->port.node = lep->port.node;
846 cep->port.port = lep->port.port;
847 cep->peer.node = peer->node;
848 cep->peer.port = peer->port;
849 init_waitqueue_head(&cep->sendwq);
850 init_waitqueue_head(&cep->recvwq);
851 init_waitqueue_head(&cep->conwq);
853 msg.uop = SCIF_CNCT_GNT;
855 msg.payload[0] = cep->remote_ep;
856 msg.payload[1] = cep->qp_info.qp_offset;
857 msg.payload[2] = (u64)cep;
859 err = _scif_nodeqp_send(cep->remote_dev, &msg);
860 scif_put_peer_dev(spdev);
862 goto scif_accept_error_map;
864 /* Wait for the remote node to respond with SCIF_CNCT_GNT(N)ACK */
865 err = wait_event_timeout(cep->conwq, cep->state != SCIFEP_CONNECTING,
866 SCIF_NODE_ACCEPT_TIMEOUT);
867 if (!err && scifdev_alive(cep))
869 err = !err ? -ENODEV : 0;
871 goto scif_accept_error_map;
874 spin_lock(&cep->lock);
876 if (cep->state == SCIFEP_CLOSING) {
878 * Remote failed to allocate resources and NAKed the grant.
879 * There is at this point nothing referencing the new end point.
881 spin_unlock(&cep->lock);
882 scif_teardown_ep(cep);
885 /* If call with sync flag then go back and wait. */
886 if (flags & SCIF_ACCEPT_SYNC) {
887 spin_lock(&lep->lock);
888 goto retry_connection;
893 scif_get_port(cep->port.port);
894 *newepd = (scif_epd_t)cep;
895 spin_unlock(&cep->lock);
897 scif_accept_error_map:
898 scif_anon_inode_fput(cep);
899 scif_accept_error_anon_inode:
900 scif_teardown_ep(cep);
901 scif_accept_error_qpalloc:
903 scif_accept_error_epalloc:
904 msg.uop = SCIF_CNCT_REJ;
905 msg.dst.node = conreq->msg.src.node;
906 msg.dst.port = conreq->msg.src.port;
907 msg.payload[0] = conreq->msg.payload[0];
908 msg.payload[1] = conreq->msg.payload[1];
909 scif_nodeqp_send(&scif_dev[conreq->msg.src.node], &msg);
913 EXPORT_SYMBOL_GPL(scif_accept);
916 * scif_msg_param_check:
917 * @epd: The end point returned from scif_open()
918 * @len: Length to receive
919 * @flags: blocking or non blocking
921 * Validate parameters for messaging APIs scif_send(..)/scif_recv(..).
923 static inline int scif_msg_param_check(scif_epd_t epd, int len, int flags)
929 if (flags && (!(flags & SCIF_RECV_BLOCK)))
936 static int _scif_send(scif_epd_t epd, void *msg, int len, int flags)
938 struct scif_endpt *ep = (struct scif_endpt *)epd;
939 struct scifmsg notif_msg;
940 int curr_xfer_len = 0, sent_len = 0, write_count;
942 struct scif_qp *qp = ep->qp_info.qp;
944 if (flags & SCIF_SEND_BLOCK)
947 spin_lock(&ep->lock);
948 while (sent_len != len && SCIFEP_CONNECTED == ep->state) {
949 write_count = scif_rb_space(&qp->outbound_q);
951 /* Best effort to send as much data as possible */
952 curr_xfer_len = min(len - sent_len, write_count);
953 ret = scif_rb_write(&qp->outbound_q, msg,
957 /* Success. Update write pointer */
958 scif_rb_commit(&qp->outbound_q);
960 * Send a notification to the peer about the
961 * produced data message.
963 notif_msg.src = ep->port;
964 notif_msg.uop = SCIF_CLIENT_SENT;
965 notif_msg.payload[0] = ep->remote_ep;
966 ret = _scif_nodeqp_send(ep->remote_dev, ¬if_msg);
969 sent_len += curr_xfer_len;
970 msg = msg + curr_xfer_len;
973 curr_xfer_len = min(len - sent_len, SCIF_ENDPT_QP_SIZE - 1);
974 /* Not enough RB space. return for the Non Blocking case */
975 if (!(flags & SCIF_SEND_BLOCK))
978 spin_unlock(&ep->lock);
979 /* Wait for a SCIF_CLIENT_RCVD message in the Blocking case */
981 wait_event_interruptible(ep->sendwq,
982 (SCIFEP_CONNECTED != ep->state) ||
983 (scif_rb_space(&qp->outbound_q) >=
985 spin_lock(&ep->lock);
991 else if (!ret && SCIFEP_CONNECTED != ep->state)
992 ret = SCIFEP_DISCONNECTED == ep->state ?
993 -ECONNRESET : -ENOTCONN;
994 spin_unlock(&ep->lock);
998 static int _scif_recv(scif_epd_t epd, void *msg, int len, int flags)
1001 struct scif_endpt *ep = (struct scif_endpt *)epd;
1002 struct scifmsg notif_msg;
1003 int curr_recv_len = 0, remaining_len = len, read_count;
1005 struct scif_qp *qp = ep->qp_info.qp;
1007 if (flags & SCIF_RECV_BLOCK)
1009 spin_lock(&ep->lock);
1010 while (remaining_len && (SCIFEP_CONNECTED == ep->state ||
1011 SCIFEP_DISCONNECTED == ep->state)) {
1012 read_count = scif_rb_count(&qp->inbound_q, remaining_len);
1015 * Best effort to recv as much data as there
1016 * are bytes to read in the RB particularly
1017 * important for the Non Blocking case.
1019 curr_recv_len = min(remaining_len, read_count);
1020 read_size = scif_rb_get_next(&qp->inbound_q,
1021 msg, curr_recv_len);
1022 if (ep->state == SCIFEP_CONNECTED) {
1024 * Update the read pointer only if the endpoint
1025 * is still connected else the read pointer
1026 * might no longer exist since the peer has
1029 scif_rb_update_read_ptr(&qp->inbound_q);
1031 * Send a notification to the peer about the
1032 * consumed data message only if the EP is in
1033 * SCIFEP_CONNECTED state.
1035 notif_msg.src = ep->port;
1036 notif_msg.uop = SCIF_CLIENT_RCVD;
1037 notif_msg.payload[0] = ep->remote_ep;
1038 ret = _scif_nodeqp_send(ep->remote_dev,
1043 remaining_len -= curr_recv_len;
1044 msg = msg + curr_recv_len;
1048 * Bail out now if the EP is in SCIFEP_DISCONNECTED state else
1049 * we will keep looping forever.
1051 if (ep->state == SCIFEP_DISCONNECTED)
1054 * Return in the Non Blocking case if there is no data
1055 * to read in this iteration.
1057 if (!(flags & SCIF_RECV_BLOCK))
1059 curr_recv_len = min(remaining_len, SCIF_ENDPT_QP_SIZE - 1);
1060 spin_unlock(&ep->lock);
1062 * Wait for a SCIF_CLIENT_SEND message in the blocking case
1063 * or until other side disconnects.
1066 wait_event_interruptible(ep->recvwq,
1067 SCIFEP_CONNECTED != ep->state ||
1068 scif_rb_count(&qp->inbound_q,
1071 spin_lock(&ep->lock);
1075 if (len - remaining_len)
1076 ret = len - remaining_len;
1077 else if (!ret && ep->state != SCIFEP_CONNECTED)
1078 ret = ep->state == SCIFEP_DISCONNECTED ?
1079 -ECONNRESET : -ENOTCONN;
1080 spin_unlock(&ep->lock);
1085 * scif_user_send() - Send data to connection queue
1086 * @epd: The end point returned from scif_open()
1087 * @msg: Address to place data
1088 * @len: Length to receive
1089 * @flags: blocking or non blocking
1091 * This function is called from the driver IOCTL entry point
1092 * only and is a wrapper for _scif_send().
1094 int scif_user_send(scif_epd_t epd, void __user *msg, int len, int flags)
1096 struct scif_endpt *ep = (struct scif_endpt *)epd;
1101 int chunk_len = min(len, (1 << (MAX_ORDER + PAGE_SHIFT - 1)));
1103 dev_dbg(scif_info.mdev.this_device,
1104 "SCIFAPI send (U): ep %p %s\n", ep, scif_ep_states[ep->state]);
1108 err = scif_msg_param_check(epd, len, flags);
1112 tmp = kmalloc(chunk_len, GFP_KERNEL);
1118 * Grabbing the lock before breaking up the transfer in
1119 * multiple chunks is required to ensure that messages do
1120 * not get fragmented and reordered.
1122 mutex_lock(&ep->sendlock);
1123 while (sent_len != len) {
1124 loop_len = len - sent_len;
1125 loop_len = min(chunk_len, loop_len);
1126 if (copy_from_user(tmp, msg, loop_len)) {
1130 err = _scif_send(epd, tmp, loop_len, flags);
1135 if (err != loop_len)
1139 mutex_unlock(&ep->sendlock);
1142 return err < 0 ? err : sent_len;
1146 * scif_user_recv() - Receive data from connection queue
1147 * @epd: The end point returned from scif_open()
1148 * @msg: Address to place data
1149 * @len: Length to receive
1150 * @flags: blocking or non blocking
1152 * This function is called from the driver IOCTL entry point
1153 * only and is a wrapper for _scif_recv().
1155 int scif_user_recv(scif_epd_t epd, void __user *msg, int len, int flags)
1157 struct scif_endpt *ep = (struct scif_endpt *)epd;
1162 int chunk_len = min(len, (1 << (MAX_ORDER + PAGE_SHIFT - 1)));
1164 dev_dbg(scif_info.mdev.this_device,
1165 "SCIFAPI recv (U): ep %p %s\n", ep, scif_ep_states[ep->state]);
1169 err = scif_msg_param_check(epd, len, flags);
1173 tmp = kmalloc(chunk_len, GFP_KERNEL);
1179 * Grabbing the lock before breaking up the transfer in
1180 * multiple chunks is required to ensure that messages do
1181 * not get fragmented and reordered.
1183 mutex_lock(&ep->recvlock);
1184 while (recv_len != len) {
1185 loop_len = len - recv_len;
1186 loop_len = min(chunk_len, loop_len);
1187 err = _scif_recv(epd, tmp, loop_len, flags);
1190 if (copy_to_user(msg, tmp, err)) {
1196 if (err != loop_len)
1200 mutex_unlock(&ep->recvlock);
1203 return err < 0 ? err : recv_len;
1207 * scif_send() - Send data to connection queue
1208 * @epd: The end point returned from scif_open()
1209 * @msg: Address to place data
1210 * @len: Length to receive
1211 * @flags: blocking or non blocking
1213 * This function is called from the kernel mode only and is
1214 * a wrapper for _scif_send().
1216 int scif_send(scif_epd_t epd, void *msg, int len, int flags)
1218 struct scif_endpt *ep = (struct scif_endpt *)epd;
1221 dev_dbg(scif_info.mdev.this_device,
1222 "SCIFAPI send (K): ep %p %s\n", ep, scif_ep_states[ep->state]);
1226 ret = scif_msg_param_check(epd, len, flags);
1229 if (!ep->remote_dev)
1232 * Grab the mutex lock in the blocking case only
1233 * to ensure messages do not get fragmented/reordered.
1234 * The non blocking mode is protected using spin locks
1237 if (flags & SCIF_SEND_BLOCK)
1238 mutex_lock(&ep->sendlock);
1240 ret = _scif_send(epd, msg, len, flags);
1242 if (flags & SCIF_SEND_BLOCK)
1243 mutex_unlock(&ep->sendlock);
1246 EXPORT_SYMBOL_GPL(scif_send);
1249 * scif_recv() - Receive data from connection queue
1250 * @epd: The end point returned from scif_open()
1251 * @msg: Address to place data
1252 * @len: Length to receive
1253 * @flags: blocking or non blocking
1255 * This function is called from the kernel mode only and is
1256 * a wrapper for _scif_recv().
1258 int scif_recv(scif_epd_t epd, void *msg, int len, int flags)
1260 struct scif_endpt *ep = (struct scif_endpt *)epd;
1263 dev_dbg(scif_info.mdev.this_device,
1264 "SCIFAPI recv (K): ep %p %s\n", ep, scif_ep_states[ep->state]);
1268 ret = scif_msg_param_check(epd, len, flags);
1272 * Grab the mutex lock in the blocking case only
1273 * to ensure messages do not get fragmented/reordered.
1274 * The non blocking mode is protected using spin locks
1277 if (flags & SCIF_RECV_BLOCK)
1278 mutex_lock(&ep->recvlock);
1280 ret = _scif_recv(epd, msg, len, flags);
1282 if (flags & SCIF_RECV_BLOCK)
1283 mutex_unlock(&ep->recvlock);
1287 EXPORT_SYMBOL_GPL(scif_recv);
1289 static inline void _scif_poll_wait(struct file *f, wait_queue_head_t *wq,
1290 poll_table *p, struct scif_endpt *ep)
1293 * Because poll_wait makes a GFP_KERNEL allocation, give up the lock
1294 * and regrab it afterwards. Because the endpoint state might have
1295 * changed while the lock was given up, the state must be checked
1296 * again after re-acquiring the lock. The code in __scif_pollfd(..)
1299 spin_unlock(&ep->lock);
1300 poll_wait(f, wq, p);
1301 spin_lock(&ep->lock);
1305 __scif_pollfd(struct file *f, poll_table *wait, struct scif_endpt *ep)
1309 dev_dbg(scif_info.mdev.this_device,
1310 "SCIFAPI pollfd: ep %p %s\n", ep, scif_ep_states[ep->state]);
1312 spin_lock(&ep->lock);
1314 /* Endpoint is waiting for a non-blocking connect to complete */
1315 if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) {
1316 _scif_poll_wait(f, &ep->conn_pend_wq, wait, ep);
1317 if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) {
1318 if (ep->state == SCIFEP_CONNECTED ||
1319 ep->state == SCIFEP_DISCONNECTED ||
1326 /* Endpoint is listening for incoming connection requests */
1327 if (ep->state == SCIFEP_LISTENING) {
1328 _scif_poll_wait(f, &ep->conwq, wait, ep);
1329 if (ep->state == SCIFEP_LISTENING) {
1336 /* Endpoint is connected or disconnected */
1337 if (ep->state == SCIFEP_CONNECTED || ep->state == SCIFEP_DISCONNECTED) {
1338 if (poll_requested_events(wait) & EPOLLIN)
1339 _scif_poll_wait(f, &ep->recvwq, wait, ep);
1340 if (poll_requested_events(wait) & EPOLLOUT)
1341 _scif_poll_wait(f, &ep->sendwq, wait, ep);
1342 if (ep->state == SCIFEP_CONNECTED ||
1343 ep->state == SCIFEP_DISCONNECTED) {
1344 /* Data can be read without blocking */
1345 if (scif_rb_count(&ep->qp_info.qp->inbound_q, 1))
1347 /* Data can be written without blocking */
1348 if (scif_rb_space(&ep->qp_info.qp->outbound_q))
1350 /* Return EPOLLHUP if endpoint is disconnected */
1351 if (ep->state == SCIFEP_DISCONNECTED)
1357 /* Return EPOLLERR if the endpoint is in none of the above states */
1360 spin_unlock(&ep->lock);
1365 * scif_poll() - Kernel mode SCIF poll
1366 * @ufds: Array of scif_pollepd structures containing the end points
1367 * and events to poll on
1368 * @nfds: Size of the ufds array
1369 * @timeout_msecs: Timeout in msecs, -ve implies infinite timeout
1371 * The code flow in this function is based on do_poll(..) in select.c
1373 * Returns the number of endpoints which have pending events or 0 in
1374 * the event of a timeout. If a signal is used for wake up, -EINTR is
1378 scif_poll(struct scif_pollepd *ufds, unsigned int nfds, long timeout_msecs)
1380 struct poll_wqueues table;
1382 int i, count = 0, timed_out = timeout_msecs == 0;
1384 u64 timeout = timeout_msecs < 0 ? MAX_SCHEDULE_TIMEOUT
1385 : msecs_to_jiffies(timeout_msecs);
1387 poll_initwait(&table);
1390 for (i = 0; i < nfds; i++) {
1391 pt->_key = ufds[i].events | EPOLLERR | EPOLLHUP;
1392 mask = __scif_pollfd(ufds[i].epd->anon,
1394 mask &= ufds[i].events | EPOLLERR | EPOLLHUP;
1399 ufds[i].revents = mask;
1403 count = table.error;
1404 if (signal_pending(current))
1407 if (count || timed_out)
1410 if (!schedule_timeout_interruptible(timeout))
1413 poll_freewait(&table);
1416 EXPORT_SYMBOL_GPL(scif_poll);
1418 int scif_get_node_ids(u16 *nodes, int len, u16 *self)
1424 if (!scif_is_mgmt_node())
1425 scif_get_node_info();
1427 *self = scif_info.nodeid;
1428 mutex_lock(&scif_info.conflock);
1429 len = min_t(int, len, scif_info.total);
1430 for (node = 0; node <= scif_info.maxid; node++) {
1431 if (_scifdev_alive(&scif_dev[node])) {
1434 nodes[offset++] = node;
1437 dev_dbg(scif_info.mdev.this_device,
1438 "SCIFAPI get_node_ids total %d online %d filled in %d nodes\n",
1439 scif_info.total, online, offset);
1440 mutex_unlock(&scif_info.conflock);
1444 EXPORT_SYMBOL_GPL(scif_get_node_ids);
1446 static int scif_add_client_dev(struct device *dev, struct subsys_interface *si)
1448 struct scif_client *client =
1449 container_of(si, struct scif_client, si);
1450 struct scif_peer_dev *spdev =
1451 container_of(dev, struct scif_peer_dev, dev);
1454 client->probe(spdev);
1458 static void scif_remove_client_dev(struct device *dev,
1459 struct subsys_interface *si)
1461 struct scif_client *client =
1462 container_of(si, struct scif_client, si);
1463 struct scif_peer_dev *spdev =
1464 container_of(dev, struct scif_peer_dev, dev);
1467 client->remove(spdev);
1470 void scif_client_unregister(struct scif_client *client)
1472 subsys_interface_unregister(&client->si);
1474 EXPORT_SYMBOL_GPL(scif_client_unregister);
1476 int scif_client_register(struct scif_client *client)
1478 struct subsys_interface *si = &client->si;
1480 si->name = client->name;
1481 si->subsys = &scif_peer_bus;
1482 si->add_dev = scif_add_client_dev;
1483 si->remove_dev = scif_remove_client_dev;
1485 return subsys_interface_register(&client->si);
1487 EXPORT_SYMBOL_GPL(scif_client_register);