2 * Intel MIC Platform Software Stack (MPSS)
4 * Copyright(c) 2014 Intel Corporation.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2, as
8 * published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License for more details.
18 #include "../bus/scif_bus.h"
19 #include "scif_peer_bus.h"
20 #include "scif_main.h"
21 #include "scif_nodeqp.h"
25 ************************************************************************
26 * SCIF node Queue Pair (QP) setup flow:
28 * 1) SCIF driver gets probed with a scif_hw_dev via the scif_hw_bus
29 * 2) scif_setup_qp(..) allocates the local qp and calls
30 * scif_setup_qp_connect(..) which allocates and maps the local
31 * buffer for the inbound QP
32 * 3) The local node updates the device page with the DMA address of the QP
33 * 4) A delayed work is scheduled (qp_dwork) which periodically reads if
34 * the peer node has updated its QP DMA address
35 * 5) Once a valid non zero address is found in the QP DMA address field
36 * in the device page, the local node maps the remote node's QP,
37 * updates its outbound QP and sends a SCIF_INIT message to the peer
38 * 6) The SCIF_INIT message is received by the peer node QP interrupt bottom
39 * half handler by calling scif_init(..)
40 * 7) scif_init(..) registers a new SCIF peer node by calling
41 * scif_peer_register_device(..) which signifies the addition of a new
43 * 8) On the mgmt node, P2P network setup/teardown is initiated if all the
44 * remote nodes are online via scif_p2p_setup(..)
45 * 9) For P2P setup, the host maps the remote nodes' aperture and memory
46 * bars and sends a SCIF_NODE_ADD message to both nodes
47 * 10) As part of scif_nodeadd, both nodes set up their local inbound
48 * QPs and send a SCIF_NODE_ADD_ACK to the mgmt node
49 * 11) As part of scif_node_add_ack(..) the mgmt node forwards the
50 * SCIF_NODE_ADD_ACK to the remote nodes
51 * 12) As part of scif_node_add_ack(..) the remote nodes update their
52 * outbound QPs, make sure they can access memory on the remote node
53 * and then add a new SCIF peer node by calling
54 * scif_peer_register_device(..) which signifies the addition of a new
56 * 13) The SCIF network is now established across all nodes.
58 ************************************************************************
59 * SCIF node QP teardown flow (initiated by non mgmt node):
61 * 1) SCIF driver gets a remove callback with a scif_hw_dev via the scif_hw_bus
62 * 2) The device page QP DMA address field is updated with 0x0
63 * 3) A non mgmt node now cleans up all local data structures and sends a
64 * SCIF_EXIT message to the peer and waits for a SCIF_EXIT_ACK
65 * 4) As part of scif_exit(..) handling scif_disconnect_node(..) is called
66 * 5) scif_disconnect_node(..) sends a SCIF_NODE_REMOVE message to all the
67 * peers and waits for a SCIF_NODE_REMOVE_ACK
68 * 6) As part of scif_node_remove(..) a remote node unregisters the peer
69 * node from the SCIF network and sends a SCIF_NODE_REMOVE_ACK
70 * 7) When the mgmt node has received all the SCIF_NODE_REMOVE_ACKs
71 * it sends itself a node remove message whose handling cleans up local
72 * data structures and unregisters the peer node from the SCIF network
73 * 8) The mgmt node sends a SCIF_EXIT_ACK
74 * 9) Upon receipt of the SCIF_EXIT_ACK the node initiating the teardown
75 * completes the SCIF remove routine
76 * 10) The SCIF network is now torn down for the node initiating the
79 ************************************************************************
80 * SCIF node QP teardown flow (initiated by mgmt node):
82 * 1) SCIF driver gets a remove callback with a scif_hw_dev via the scif_hw_bus
83 * 2) The device page QP DMA address field is updated with 0x0
84 * 3) The mgmt node calls scif_disconnect_node(..)
85 * 4) scif_disconnect_node(..) sends a SCIF_NODE_REMOVE message to all the peers
86 * and waits for a SCIF_NODE_REMOVE_ACK
87 * 5) As part of scif_node_remove(..) a remote node unregisters the peer
88 * node from the SCIF network and sends a SCIF_NODE_REMOVE_ACK
89 * 6) When the mgmt node has received all the SCIF_NODE_REMOVE_ACKs
90 * it unregisters the peer node from the SCIF network
91 * 7) The mgmt node sends a SCIF_EXIT message and waits for a SCIF_EXIT_ACK.
92 * 8) A non mgmt node upon receipt of a SCIF_EXIT message calls scif_stop(..)
93 * which would clean up local data structures for all SCIF nodes and
94 * then send a SCIF_EXIT_ACK back to the mgmt node
95 * 9) Upon receipt of the SCIF_EXIT_ACK the the mgmt node sends itself a node
96 * remove message whose handling cleans up local data structures and
97 * destroys any P2P mappings.
98 * 10) The SCIF hardware device for which a remove callback was received is now
99 * disconnected from the SCIF network.
102 * Initializes "local" data structures for the QP. Allocates the QP
103 * ring buffer (rb) and initializes the "in bound" queue.
105 int scif_setup_qp_connect(struct scif_qp *qp, dma_addr_t *qp_offset,
106 int local_size, struct scif_dev *scifdev)
108 void *local_q = NULL;
112 spin_lock_init(&qp->send_lock);
113 spin_lock_init(&qp->recv_lock);
115 local_q = kzalloc(local_size, GFP_KERNEL);
120 err = scif_map_single(&qp->local_buf, local_q, scifdev, local_size);
124 * To setup the inbound_q, the buffer lives locally, the read pointer
125 * is remote and the write pointer is local.
127 scif_rb_init(&qp->inbound_q,
130 local_q, get_count_order(local_size));
132 * The read pointer is NULL initially and it is unsafe to use the ring
133 * buffer til this changes!
135 qp->inbound_q.read_ptr = NULL;
136 err = scif_map_single(qp_offset, qp,
137 scifdev, sizeof(struct scif_qp));
140 qp->local_qp = *qp_offset;
143 scif_unmap_single(qp->local_buf, scifdev, local_size);
150 /* When the other side has already done it's allocation, this is called */
151 int scif_setup_qp_accept(struct scif_qp *qp, dma_addr_t *qp_offset,
152 dma_addr_t phys, int local_size,
153 struct scif_dev *scifdev)
157 struct scif_qp *remote_qp;
161 spin_lock_init(&qp->send_lock);
162 spin_lock_init(&qp->recv_lock);
163 /* Start by figuring out where we need to point */
164 remote_qp = scif_ioremap(phys, sizeof(struct scif_qp), scifdev);
167 qp->remote_qp = remote_qp;
168 if (qp->remote_qp->magic != SCIFEP_MAGIC) {
172 qp->remote_buf = remote_qp->local_buf;
173 remote_size = qp->remote_qp->inbound_q.size;
174 remote_q = scif_ioremap(qp->remote_buf, remote_size, scifdev);
179 qp->remote_qp->local_write = 0;
181 * To setup the outbound_q, the buffer lives in remote memory,
182 * the read pointer is local, the write pointer is remote
184 scif_rb_init(&qp->outbound_q,
186 &qp->remote_qp->local_write,
188 get_count_order(remote_size));
189 local_q = kzalloc(local_size, GFP_KERNEL);
194 err = scif_map_single(&qp->local_buf, local_q, scifdev, local_size);
197 qp->remote_qp->local_read = 0;
199 * To setup the inbound_q, the buffer lives locally, the read pointer
200 * is remote and the write pointer is local
202 scif_rb_init(&qp->inbound_q,
203 &qp->remote_qp->local_read,
205 local_q, get_count_order(local_size));
206 err = scif_map_single(qp_offset, qp, scifdev,
207 sizeof(struct scif_qp));
210 qp->local_qp = *qp_offset;
213 scif_unmap_single(qp->local_buf, scifdev, local_size);
218 scif_iounmap(remote_q, remote_size, scifdev);
219 qp->outbound_q.rb_base = NULL;
221 scif_iounmap(qp->remote_qp, sizeof(struct scif_qp), scifdev);
222 qp->remote_qp = NULL;
226 int scif_setup_qp_connect_response(struct scif_dev *scifdev,
227 struct scif_qp *qp, u64 payload)
232 phys_addr_t tmp_phys;
234 qp->remote_qp = scif_ioremap(payload, sizeof(struct scif_qp), scifdev);
236 if (!qp->remote_qp) {
241 if (qp->remote_qp->magic != SCIFEP_MAGIC) {
242 dev_err(&scifdev->sdev->dev,
243 "SCIFEP_MAGIC mismatch between self %d remote %d\n",
244 scif_dev[scif_info.nodeid].node, scifdev->node);
249 tmp_phys = qp->remote_qp->local_buf;
250 remote_size = qp->remote_qp->inbound_q.size;
251 r_buf = scif_ioremap(tmp_phys, remote_size, scifdev);
257 scif_rb_init(&qp->outbound_q,
259 &qp->remote_qp->local_write,
261 get_count_order(remote_size));
263 * resetup the inbound_q now that we know where the
264 * inbound_read really is.
266 scif_rb_init(&qp->inbound_q,
267 &qp->remote_qp->local_read,
269 qp->inbound_q.rb_base,
270 get_count_order(qp->inbound_q.size));
275 static __always_inline void
276 scif_send_msg_intr(struct scif_dev *scifdev)
278 struct scif_hw_dev *sdev = scifdev->sdev;
280 if (scifdev_is_p2p(scifdev))
281 sdev->hw_ops->send_p2p_intr(sdev, scifdev->rdb, &scifdev->mmio);
283 sdev->hw_ops->send_intr(sdev, scifdev->rdb);
286 int scif_qp_response(phys_addr_t phys, struct scif_dev *scifdev)
291 err = scif_setup_qp_connect_response(scifdev, scifdev->qpairs, phys);
294 * Now that everything is setup and mapped, we're ready
295 * to tell the peer about our queue's location
298 msg.dst.node = scifdev->node;
299 err = scif_nodeqp_send(scifdev, &msg);
304 void scif_send_exit(struct scif_dev *scifdev)
309 scifdev->exit = OP_IN_PROGRESS;
311 msg.src.node = scif_info.nodeid;
312 msg.dst.node = scifdev->node;
313 ret = scif_nodeqp_send(scifdev, &msg);
316 /* Wait for a SCIF_EXIT_ACK message */
317 wait_event_timeout(scif_info.exitwq, scifdev->exit == OP_COMPLETED,
318 SCIF_NODE_ALIVE_TIMEOUT);
320 scifdev->exit = OP_IDLE;
323 int scif_setup_qp(struct scif_dev *scifdev)
329 local_size = SCIF_NODE_QP_SIZE;
331 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
336 qp->magic = SCIFEP_MAGIC;
337 scifdev->qpairs = qp;
338 err = scif_setup_qp_connect(qp, &scifdev->qp_dma_addr,
339 local_size, scifdev);
343 * We're as setup as we can be. The inbound_q is setup, w/o a usable
344 * outbound q. When we get a message, the read_ptr will be updated,
345 * and we will pull the message.
349 kfree(scifdev->qpairs);
350 scifdev->qpairs = NULL;
354 static void scif_p2p_freesg(struct scatterlist *sg)
359 static struct scatterlist *
360 scif_p2p_setsg(phys_addr_t pa, int page_size, int page_cnt)
362 struct scatterlist *sg;
366 sg = kcalloc(page_cnt, sizeof(struct scatterlist), GFP_KERNEL);
369 sg_init_table(sg, page_cnt);
370 for (i = 0; i < page_cnt; i++) {
371 page = pfn_to_page(pa >> PAGE_SHIFT);
372 sg_set_page(&sg[i], page, page_size, 0);
378 /* Init p2p mappings required to access peerdev from scifdev */
379 static struct scif_p2p_info *
380 scif_init_p2p_info(struct scif_dev *scifdev, struct scif_dev *peerdev)
382 struct scif_p2p_info *p2p;
383 int num_mmio_pages, num_aper_pages, sg_page_shift, err, num_aper_chunks;
384 struct scif_hw_dev *psdev = peerdev->sdev;
385 struct scif_hw_dev *sdev = scifdev->sdev;
387 num_mmio_pages = psdev->mmio->len >> PAGE_SHIFT;
388 num_aper_pages = psdev->aper->len >> PAGE_SHIFT;
390 p2p = kzalloc(sizeof(*p2p), GFP_KERNEL);
393 p2p->ppi_sg[SCIF_PPI_MMIO] = scif_p2p_setsg(psdev->mmio->pa,
394 PAGE_SIZE, num_mmio_pages);
395 if (!p2p->ppi_sg[SCIF_PPI_MMIO])
397 p2p->sg_nentries[SCIF_PPI_MMIO] = num_mmio_pages;
398 sg_page_shift = get_order(min(psdev->aper->len, (u64)(1 << 30)));
399 num_aper_chunks = num_aper_pages >> (sg_page_shift - PAGE_SHIFT);
400 p2p->ppi_sg[SCIF_PPI_APER] = scif_p2p_setsg(psdev->aper->pa,
403 p2p->sg_nentries[SCIF_PPI_APER] = num_aper_chunks;
404 err = dma_map_sg(&sdev->dev, p2p->ppi_sg[SCIF_PPI_MMIO],
405 num_mmio_pages, PCI_DMA_BIDIRECTIONAL);
406 if (err != num_mmio_pages)
408 err = dma_map_sg(&sdev->dev, p2p->ppi_sg[SCIF_PPI_APER],
409 num_aper_chunks, PCI_DMA_BIDIRECTIONAL);
410 if (err != num_aper_chunks)
412 p2p->ppi_da[SCIF_PPI_MMIO] = sg_dma_address(p2p->ppi_sg[SCIF_PPI_MMIO]);
413 p2p->ppi_da[SCIF_PPI_APER] = sg_dma_address(p2p->ppi_sg[SCIF_PPI_APER]);
414 p2p->ppi_len[SCIF_PPI_MMIO] = num_mmio_pages;
415 p2p->ppi_len[SCIF_PPI_APER] = num_aper_pages;
416 p2p->ppi_peer_id = peerdev->node;
419 dma_unmap_sg(&sdev->dev, p2p->ppi_sg[SCIF_PPI_MMIO],
420 p2p->sg_nentries[SCIF_PPI_MMIO], DMA_BIDIRECTIONAL);
422 scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_MMIO]);
423 scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_APER]);
430 * scif_node_connect: Respond to SCIF_NODE_CONNECT interrupt message
431 * @dst: Destination node
433 * Connect the src and dst node by setting up the p2p connection
434 * between them. Management node here acts like a proxy.
436 static void scif_node_connect(struct scif_dev *scifdev, int dst)
438 struct scif_dev *dev_j = scifdev;
439 struct scif_dev *dev_i = NULL;
440 struct scif_p2p_info *p2p_ij = NULL; /* bus addr for j from i */
441 struct scif_p2p_info *p2p_ji = NULL; /* bus addr for i from j */
442 struct scif_p2p_info *p2p;
443 struct list_head *pos, *tmp;
448 if (dst < 1 || dst > scif_info.maxid)
451 dev_i = &scif_dev[dst];
453 if (!_scifdev_alive(dev_i))
456 * If the p2p connection is already setup or in the process of setting
457 * up then just ignore this request. The requested node will get
458 * informed by SCIF_NODE_ADD_ACK or SCIF_NODE_ADD_NACK
460 if (!list_empty(&dev_i->p2p)) {
461 list_for_each_safe(pos, tmp, &dev_i->p2p) {
462 p2p = list_entry(pos, struct scif_p2p_info, ppi_list);
463 if (p2p->ppi_peer_id == dev_j->node)
467 p2p_ij = scif_init_p2p_info(dev_i, dev_j);
470 p2p_ji = scif_init_p2p_info(dev_j, dev_i);
473 list_add_tail(&p2p_ij->ppi_list, &dev_i->p2p);
474 list_add_tail(&p2p_ji->ppi_list, &dev_j->p2p);
477 * Send a SCIF_NODE_ADD to dev_i, pass it its bus address
480 msg.uop = SCIF_NODE_ADD;
481 msg.src.node = dev_j->node;
482 msg.dst.node = dev_i->node;
484 msg.payload[0] = p2p_ji->ppi_da[SCIF_PPI_APER];
485 msg.payload[1] = p2p_ij->ppi_da[SCIF_PPI_MMIO];
486 msg.payload[2] = p2p_ij->ppi_da[SCIF_PPI_APER];
487 msg.payload[3] = p2p_ij->ppi_len[SCIF_PPI_APER] << PAGE_SHIFT;
489 err = scif_nodeqp_send(dev_i, &msg);
491 dev_err(&scifdev->sdev->dev,
492 "%s %d error %d\n", __func__, __LINE__, err);
496 /* Same as above but to dev_j */
497 msg.uop = SCIF_NODE_ADD;
498 msg.src.node = dev_i->node;
499 msg.dst.node = dev_j->node;
501 tmppayload = msg.payload[0];
502 msg.payload[0] = msg.payload[2];
503 msg.payload[2] = tmppayload;
504 msg.payload[1] = p2p_ji->ppi_da[SCIF_PPI_MMIO];
505 msg.payload[3] = p2p_ji->ppi_len[SCIF_PPI_APER] << PAGE_SHIFT;
507 scif_nodeqp_send(dev_j, &msg);
510 static void scif_p2p_setup(void)
514 if (!scif_info.p2p_enable)
517 for (i = 1; i <= scif_info.maxid; i++)
518 if (!_scifdev_alive(&scif_dev[i]))
521 for (i = 1; i <= scif_info.maxid; i++) {
522 for (j = 1; j <= scif_info.maxid; j++) {
523 struct scif_dev *scifdev = &scif_dev[i];
527 scif_node_connect(scifdev, j);
532 void scif_qp_response_ack(struct work_struct *work)
534 struct scif_dev *scifdev = container_of(work, struct scif_dev,
536 struct scif_peer_dev *spdev;
538 /* Drop the INIT message if it has already been received */
539 if (_scifdev_alive(scifdev))
542 spdev = scif_peer_register_device(scifdev);
546 if (scif_is_mgmt_node()) {
547 mutex_lock(&scif_info.conflock);
549 mutex_unlock(&scif_info.conflock);
553 static char *message_types[] = {"BAD",
559 "SCIF_NODE_ADD_NACK",
571 "SCIF_GET_NODE_INFO"};
574 scif_display_message(struct scif_dev *scifdev, struct scifmsg *msg,
577 if (!scif_info.en_msg_log)
579 if (msg->uop > SCIF_MAX_MSG) {
580 dev_err(&scifdev->sdev->dev,
581 "%s: unknown msg type %d\n", label, msg->uop);
584 dev_info(&scifdev->sdev->dev,
585 "%s: msg type %s, src %d:%d, dest %d:%d payload 0x%llx:0x%llx:0x%llx:0x%llx\n",
586 label, message_types[msg->uop], msg->src.node, msg->src.port,
587 msg->dst.node, msg->dst.port, msg->payload[0], msg->payload[1],
588 msg->payload[2], msg->payload[3]);
591 int _scif_nodeqp_send(struct scif_dev *scifdev, struct scifmsg *msg)
593 struct scif_qp *qp = scifdev->qpairs;
594 int err = -ENOMEM, loop_cnt = 0;
596 scif_display_message(scifdev, msg, "Sent");
601 spin_lock(&qp->send_lock);
603 while ((err = scif_rb_write(&qp->outbound_q,
604 msg, sizeof(struct scifmsg)))) {
606 #define SCIF_NODEQP_SEND_TO_MSEC (3 * 1000)
607 if (loop_cnt++ > (SCIF_NODEQP_SEND_TO_MSEC)) {
613 scif_rb_commit(&qp->outbound_q);
614 spin_unlock(&qp->send_lock);
616 if (scifdev_self(scifdev))
618 * For loopback we need to emulate an interrupt by
619 * queuing work for the queue handling real node
622 queue_work(scifdev->intr_wq, &scifdev->intr_bh);
624 scif_send_msg_intr(scifdev);
628 dev_dbg(&scifdev->sdev->dev,
629 "%s %d error %d uop %d\n",
630 __func__, __LINE__, err, msg->uop);
635 * scif_nodeqp_send - Send a message on the node queue pair
636 * @scifdev: Scif Device.
637 * @msg: The message to be sent.
639 int scif_nodeqp_send(struct scif_dev *scifdev, struct scifmsg *msg)
642 struct device *spdev = NULL;
644 if (msg->uop > SCIF_EXIT_ACK) {
645 /* Dont send messages once the exit flow has begun */
646 if (OP_IDLE != scifdev->exit)
648 spdev = scif_get_peer_dev(scifdev);
650 err = PTR_ERR(spdev);
654 err = _scif_nodeqp_send(scifdev, msg);
655 if (msg->uop > SCIF_EXIT_ACK)
656 scif_put_peer_dev(spdev);
663 * Work queue handler for servicing miscellaneous SCIF tasks.
665 * 1) Cleanup of zombie endpoints.
667 void scif_misc_handler(struct work_struct *work)
669 scif_cleanup_zombie_epd();
673 * scif_init() - Respond to SCIF_INIT interrupt message
674 * @scifdev: Remote SCIF device node
675 * @msg: Interrupt message
677 static __always_inline void
678 scif_init(struct scif_dev *scifdev, struct scifmsg *msg)
681 * Allow the thread waiting for device page updates for the peer QP DMA
682 * address to complete initializing the inbound_q.
684 flush_delayed_work(&scifdev->qp_dwork);
686 * Delegate the peer device registration to a workqueue, otherwise if
687 * SCIF client probe (called during peer device registration) calls
688 * scif_connect(..), it will block the message processing thread causing
691 schedule_work(&scifdev->init_msg_work);
695 * scif_exit() - Respond to SCIF_EXIT interrupt message
696 * @scifdev: Remote SCIF device node
697 * @msg: Interrupt message
699 * This function stops the SCIF interface for the node which sent
700 * the SCIF_EXIT message and starts waiting for that node to
701 * resetup the queue pair again.
703 static __always_inline void
704 scif_exit(struct scif_dev *scifdev, struct scifmsg *unused)
706 scifdev->exit_ack_pending = true;
707 if (scif_is_mgmt_node())
708 scif_disconnect_node(scifdev->node, false);
711 schedule_delayed_work(&scifdev->qp_dwork,
712 msecs_to_jiffies(1000));
716 * scif_exitack() - Respond to SCIF_EXIT_ACK interrupt message
717 * @scifdev: Remote SCIF device node
718 * @msg: Interrupt message
721 static __always_inline void
722 scif_exit_ack(struct scif_dev *scifdev, struct scifmsg *unused)
724 scifdev->exit = OP_COMPLETED;
725 wake_up(&scif_info.exitwq);
729 * scif_node_add() - Respond to SCIF_NODE_ADD interrupt message
730 * @scifdev: Remote SCIF device node
731 * @msg: Interrupt message
733 * When the mgmt node driver has finished initializing a MIC node queue pair it
734 * marks the node as online. It then looks for all currently online MIC cards
735 * and send a SCIF_NODE_ADD message to identify the ID of the new card for
736 * peer to peer initialization
738 * The local node allocates its incoming queue and sends its address in the
739 * SCIF_NODE_ADD_ACK message back to the mgmt node, the mgmt node "reflects"
740 * this message to the new node
742 static __always_inline void
743 scif_node_add(struct scif_dev *scifdev, struct scifmsg *msg)
745 struct scif_dev *newdev;
746 dma_addr_t qp_offset;
748 struct scif_hw_dev *sdev;
750 dev_dbg(&scifdev->sdev->dev,
751 "Scifdev %d:%d received NODE_ADD msg for node %d\n",
752 scifdev->node, msg->dst.node, msg->src.node);
753 dev_dbg(&scifdev->sdev->dev,
754 "Remote address for this node's aperture %llx\n",
756 newdev = &scif_dev[msg->src.node];
757 newdev->node = msg->src.node;
758 newdev->sdev = scif_dev[SCIF_MGMT_NODE].sdev;
761 if (scif_setup_intr_wq(newdev)) {
762 dev_err(&scifdev->sdev->dev,
763 "failed to setup interrupts for %d\n", msg->src.node);
764 goto interrupt_setup_error;
766 newdev->mmio.va = ioremap_nocache(msg->payload[1], sdev->mmio->len);
767 if (!newdev->mmio.va) {
768 dev_err(&scifdev->sdev->dev,
769 "failed to map mmio for %d\n", msg->src.node);
772 newdev->qpairs = kzalloc(sizeof(*newdev->qpairs), GFP_KERNEL);
776 * Set the base address of the remote node's memory since it gets
779 newdev->base_addr = msg->payload[0];
781 qp_connect = scif_setup_qp_connect(newdev->qpairs, &qp_offset,
782 SCIF_NODE_QP_SIZE, newdev);
784 dev_err(&scifdev->sdev->dev,
785 "failed to setup qp_connect %d\n", qp_connect);
786 goto qp_connect_error;
789 newdev->db = sdev->hw_ops->next_db(sdev);
790 newdev->cookie = sdev->hw_ops->request_irq(sdev, scif_intr_handler,
793 if (IS_ERR(newdev->cookie))
794 goto qp_connect_error;
795 newdev->qpairs->magic = SCIFEP_MAGIC;
796 newdev->qpairs->qp_state = SCIF_QP_OFFLINE;
798 msg->uop = SCIF_NODE_ADD_ACK;
799 msg->dst.node = msg->src.node;
800 msg->src.node = scif_info.nodeid;
801 msg->payload[0] = qp_offset;
802 msg->payload[2] = newdev->db;
803 scif_nodeqp_send(&scif_dev[SCIF_MGMT_NODE], msg);
806 kfree(newdev->qpairs);
807 newdev->qpairs = NULL;
809 iounmap(newdev->mmio.va);
810 newdev->mmio.va = NULL;
812 interrupt_setup_error:
813 dev_err(&scifdev->sdev->dev,
814 "node add failed for node %d\n", msg->src.node);
815 msg->uop = SCIF_NODE_ADD_NACK;
816 msg->dst.node = msg->src.node;
817 msg->src.node = scif_info.nodeid;
818 scif_nodeqp_send(&scif_dev[SCIF_MGMT_NODE], msg);
821 void scif_poll_qp_state(struct work_struct *work)
823 #define SCIF_NODE_QP_RETRY 100
824 #define SCIF_NODE_QP_TIMEOUT 100
825 struct scif_dev *peerdev = container_of(work, struct scif_dev,
827 struct scif_qp *qp = &peerdev->qpairs[0];
829 if (qp->qp_state != SCIF_QP_ONLINE ||
830 qp->remote_qp->qp_state != SCIF_QP_ONLINE) {
831 if (peerdev->p2p_retry++ == SCIF_NODE_QP_RETRY) {
832 dev_err(&peerdev->sdev->dev,
833 "Warning: QP check timeout with state %d\n",
837 schedule_delayed_work(&peerdev->p2p_dwork,
838 msecs_to_jiffies(SCIF_NODE_QP_TIMEOUT));
841 scif_peer_register_device(peerdev);
844 dev_err(&peerdev->sdev->dev,
845 "%s %d remote node %d offline, state = 0x%x\n",
846 __func__, __LINE__, peerdev->node, qp->qp_state);
847 qp->remote_qp->qp_state = SCIF_QP_OFFLINE;
848 scif_cleanup_scifdev(peerdev);
852 * scif_node_add_ack() - Respond to SCIF_NODE_ADD_ACK interrupt message
853 * @scifdev: Remote SCIF device node
854 * @msg: Interrupt message
856 * After a MIC node receives the SCIF_NODE_ADD_ACK message it send this
857 * message to the mgmt node to confirm the sequence is finished.
860 static __always_inline void
861 scif_node_add_ack(struct scif_dev *scifdev, struct scifmsg *msg)
863 struct scif_dev *peerdev;
865 struct scif_dev *dst_dev = &scif_dev[msg->dst.node];
867 dev_dbg(&scifdev->sdev->dev,
868 "Scifdev %d received SCIF_NODE_ADD_ACK msg src %d dst %d\n",
869 scifdev->node, msg->src.node, msg->dst.node);
870 dev_dbg(&scifdev->sdev->dev,
871 "payload %llx %llx %llx %llx\n", msg->payload[0],
872 msg->payload[1], msg->payload[2], msg->payload[3]);
873 if (scif_is_mgmt_node()) {
875 * the lock serializes with scif_qp_response_ack. The mgmt node
876 * is forwarding the NODE_ADD_ACK message from src to dst we
877 * need to make sure that the dst has already received a
878 * NODE_ADD for src and setup its end of the qp to dst
880 mutex_lock(&scif_info.conflock);
881 msg->payload[1] = scif_info.maxid;
882 scif_nodeqp_send(dst_dev, msg);
883 mutex_unlock(&scif_info.conflock);
886 peerdev = &scif_dev[msg->src.node];
887 peerdev->sdev = scif_dev[SCIF_MGMT_NODE].sdev;
888 peerdev->node = msg->src.node;
890 qp = &peerdev->qpairs[0];
892 if ((scif_setup_qp_connect_response(peerdev, &peerdev->qpairs[0],
895 peerdev->rdb = msg->payload[2];
896 qp->remote_qp->qp_state = SCIF_QP_ONLINE;
897 schedule_delayed_work(&peerdev->p2p_dwork, 0);
900 scif_cleanup_scifdev(peerdev);
904 * scif_node_add_nack: Respond to SCIF_NODE_ADD_NACK interrupt message
905 * @msg: Interrupt message
907 * SCIF_NODE_ADD failed, so inform the waiting wq.
909 static __always_inline void
910 scif_node_add_nack(struct scif_dev *scifdev, struct scifmsg *msg)
912 if (scif_is_mgmt_node()) {
913 struct scif_dev *dst_dev = &scif_dev[msg->dst.node];
915 dev_dbg(&scifdev->sdev->dev,
916 "SCIF_NODE_ADD_NACK received from %d\n", scifdev->node);
917 scif_nodeqp_send(dst_dev, msg);
922 * scif_node_remove: Handle SCIF_NODE_REMOVE message
923 * @msg: Interrupt message
925 * Handle node removal.
927 static __always_inline void
928 scif_node_remove(struct scif_dev *scifdev, struct scifmsg *msg)
930 int node = msg->payload[0];
931 struct scif_dev *scdev = &scif_dev[node];
933 scdev->node_remove_ack_pending = true;
934 scif_handle_remove_node(node);
938 * scif_node_remove_ack: Handle SCIF_NODE_REMOVE_ACK message
939 * @msg: Interrupt message
941 * The peer has acked a SCIF_NODE_REMOVE message.
943 static __always_inline void
944 scif_node_remove_ack(struct scif_dev *scifdev, struct scifmsg *msg)
946 struct scif_dev *sdev = &scif_dev[msg->payload[0]];
948 atomic_inc(&sdev->disconn_rescnt);
949 wake_up(&sdev->disconn_wq);
953 * scif_get_node_info: Respond to SCIF_GET_NODE_INFO interrupt message
954 * @msg: Interrupt message
956 * Retrieve node info i.e maxid and total from the mgmt node.
958 static __always_inline void
959 scif_get_node_info_resp(struct scif_dev *scifdev, struct scifmsg *msg)
961 if (scif_is_mgmt_node()) {
962 swap(msg->dst.node, msg->src.node);
963 mutex_lock(&scif_info.conflock);
964 msg->payload[1] = scif_info.maxid;
965 msg->payload[2] = scif_info.total;
966 mutex_unlock(&scif_info.conflock);
967 scif_nodeqp_send(scifdev, msg);
969 struct completion *node_info =
970 (struct completion *)msg->payload[3];
972 mutex_lock(&scif_info.conflock);
973 scif_info.maxid = msg->payload[1];
974 scif_info.total = msg->payload[2];
975 complete_all(node_info);
976 mutex_unlock(&scif_info.conflock);
981 scif_msg_unknown(struct scif_dev *scifdev, struct scifmsg *msg)
983 /* Bogus Node Qp Message? */
984 dev_err(&scifdev->sdev->dev,
985 "Unknown message 0x%xn scifdev->node 0x%x\n",
986 msg->uop, scifdev->node);
989 static void (*scif_intr_func[SCIF_MAX_MSG + 1])
990 (struct scif_dev *, struct scifmsg *msg) = {
991 scif_msg_unknown, /* Error */
992 scif_init, /* SCIF_INIT */
993 scif_exit, /* SCIF_EXIT */
994 scif_exit_ack, /* SCIF_EXIT_ACK */
995 scif_node_add, /* SCIF_NODE_ADD */
996 scif_node_add_ack, /* SCIF_NODE_ADD_ACK */
997 scif_node_add_nack, /* SCIF_NODE_ADD_NACK */
998 scif_node_remove, /* SCIF_NODE_REMOVE */
999 scif_node_remove_ack, /* SCIF_NODE_REMOVE_ACK */
1000 scif_cnctreq, /* SCIF_CNCT_REQ */
1001 scif_cnctgnt, /* SCIF_CNCT_GNT */
1002 scif_cnctgnt_ack, /* SCIF_CNCT_GNTACK */
1003 scif_cnctgnt_nack, /* SCIF_CNCT_GNTNACK */
1004 scif_cnctrej, /* SCIF_CNCT_REJ */
1005 scif_discnct, /* SCIF_DISCNCT */
1006 scif_discnt_ack, /* SCIF_DISCNT_ACK */
1007 scif_clientsend, /* SCIF_CLIENT_SENT */
1008 scif_clientrcvd, /* SCIF_CLIENT_RCVD */
1009 scif_get_node_info_resp,/* SCIF_GET_NODE_INFO */
1013 * scif_nodeqp_msg_handler() - Common handler for node messages
1014 * @scifdev: Remote device to respond to
1015 * @qp: Remote memory pointer
1016 * @msg: The message to be handled.
1018 * This routine calls the appropriate routine to handle a Node Qp
1021 static int scif_max_msg_id = SCIF_MAX_MSG;
1024 scif_nodeqp_msg_handler(struct scif_dev *scifdev,
1025 struct scif_qp *qp, struct scifmsg *msg)
1027 scif_display_message(scifdev, msg, "Rcvd");
1029 if (msg->uop > (u32)scif_max_msg_id) {
1030 /* Bogus Node Qp Message? */
1031 dev_err(&scifdev->sdev->dev,
1032 "Unknown message 0x%xn scifdev->node 0x%x\n",
1033 msg->uop, scifdev->node);
1037 scif_intr_func[msg->uop](scifdev, msg);
1041 * scif_nodeqp_intrhandler() - Interrupt handler for node messages
1042 * @scifdev: Remote device to respond to
1043 * @qp: Remote memory pointer
1045 * This routine is triggered by the interrupt mechanism. It reads
1046 * messages from the node queue RB and calls the Node QP Message handling
1049 void scif_nodeqp_intrhandler(struct scif_dev *scifdev, struct scif_qp *qp)
1055 read_size = scif_rb_get_next(&qp->inbound_q, &msg, sizeof(msg));
1058 scif_nodeqp_msg_handler(scifdev, qp, &msg);
1060 * The node queue pair is unmapped so skip the read pointer
1061 * update after receipt of a SCIF_EXIT_ACK
1063 if (SCIF_EXIT_ACK == msg.uop)
1065 scif_rb_update_read_ptr(&qp->inbound_q);
1070 * scif_loopb_wq_handler - Loopback Workqueue Handler.
1071 * @work: loop back work
1073 * This work queue routine is invoked by the loopback work queue handler.
1074 * It grabs the recv lock, dequeues any available messages from the head
1075 * of the loopback message list, calls the node QP message handler,
1076 * waits for it to return, then frees up this message and dequeues more
1077 * elements of the list if available.
1079 static void scif_loopb_wq_handler(struct work_struct *unused)
1081 struct scif_dev *scifdev = scif_info.loopb_dev;
1082 struct scif_qp *qp = scifdev->qpairs;
1083 struct scif_loopb_msg *msg;
1087 spin_lock(&qp->recv_lock);
1088 if (!list_empty(&scif_info.loopb_recv_q)) {
1089 msg = list_first_entry(&scif_info.loopb_recv_q,
1090 struct scif_loopb_msg,
1092 list_del(&msg->list);
1094 spin_unlock(&qp->recv_lock);
1097 scif_nodeqp_msg_handler(scifdev, qp, &msg->msg);
1104 * scif_loopb_msg_handler() - Workqueue handler for loopback messages.
1105 * @scifdev: SCIF device
1108 * This work queue routine is triggered when a loopback message is received.
1110 * We need special handling for receiving Node Qp messages on a loopback SCIF
1111 * device via two workqueues for receiving messages.
1113 * The reason we need the extra workqueue which is not required with *normal*
1114 * non-loopback SCIF devices is the potential classic deadlock described below:
1116 * Thread A tries to send a message on a loopback SCIF device and blocks since
1117 * there is no space in the RB while it has the send_lock held or another
1118 * lock called lock X for example.
1120 * Thread B: The Loopback Node QP message receive workqueue receives the message
1121 * and tries to send a message (eg an ACK) to the loopback SCIF device. It tries
1122 * to grab the send lock again or lock X and deadlocks with Thread A. The RB
1123 * cannot be drained any further due to this classic deadlock.
1125 * In order to avoid deadlocks as mentioned above we have an extra level of
1126 * indirection achieved by having two workqueues.
1127 * 1) The first workqueue whose handler is scif_loopb_msg_handler reads
1128 * messages from the Node QP RB, adds them to a list and queues work for the
1131 * 2) The second workqueue whose handler is scif_loopb_wq_handler dequeues
1132 * messages from the list, handles them, frees up the memory and dequeues
1133 * more elements from the list if possible.
1136 scif_loopb_msg_handler(struct scif_dev *scifdev, struct scif_qp *qp)
1139 struct scif_loopb_msg *msg;
1142 msg = kmalloc(sizeof(*msg), GFP_KERNEL);
1145 read_size = scif_rb_get_next(&qp->inbound_q, &msg->msg,
1146 sizeof(struct scifmsg));
1147 if (read_size != sizeof(struct scifmsg)) {
1149 scif_rb_update_read_ptr(&qp->inbound_q);
1152 spin_lock(&qp->recv_lock);
1153 list_add_tail(&msg->list, &scif_info.loopb_recv_q);
1154 spin_unlock(&qp->recv_lock);
1155 queue_work(scif_info.loopb_wq, &scif_info.loopb_work);
1156 scif_rb_update_read_ptr(&qp->inbound_q);
1157 } while (read_size == sizeof(struct scifmsg));
1162 * scif_setup_loopback_qp - One time setup work for Loopback Node Qp.
1163 * @scifdev: SCIF device
1165 * Sets up the required loopback workqueues, queue pairs and ring buffers
1167 int scif_setup_loopback_qp(struct scif_dev *scifdev)
1172 struct scif_peer_dev *spdev;
1174 err = scif_setup_intr_wq(scifdev);
1177 INIT_LIST_HEAD(&scif_info.loopb_recv_q);
1178 snprintf(scif_info.loopb_wqname, sizeof(scif_info.loopb_wqname),
1179 "SCIF LOOPB %d", scifdev->node);
1180 scif_info.loopb_wq =
1181 alloc_ordered_workqueue(scif_info.loopb_wqname, 0);
1182 if (!scif_info.loopb_wq) {
1186 INIT_WORK(&scif_info.loopb_work, scif_loopb_wq_handler);
1187 /* Allocate Self Qpair */
1188 scifdev->qpairs = kzalloc(sizeof(*scifdev->qpairs), GFP_KERNEL);
1189 if (!scifdev->qpairs) {
1191 goto destroy_loopb_wq;
1194 qp = scifdev->qpairs;
1195 qp->magic = SCIFEP_MAGIC;
1196 spin_lock_init(&qp->send_lock);
1197 spin_lock_init(&qp->recv_lock);
1199 local_q = kzalloc(SCIF_NODE_QP_SIZE, GFP_KERNEL);
1205 * For loopback the inbound_q and outbound_q are essentially the same
1206 * since the Node sends a message on the loopback interface to the
1207 * outbound_q which is then received on the inbound_q.
1209 scif_rb_init(&qp->outbound_q,
1212 local_q, get_count_order(SCIF_NODE_QP_SIZE));
1214 scif_rb_init(&qp->inbound_q,
1217 local_q, get_count_order(SCIF_NODE_QP_SIZE));
1218 scif_info.nodeid = scifdev->node;
1219 spdev = scif_peer_register_device(scifdev);
1220 if (IS_ERR(spdev)) {
1221 err = PTR_ERR(spdev);
1224 scif_info.loopb_dev = scifdev;
1229 kfree(scifdev->qpairs);
1231 destroy_workqueue(scif_info.loopb_wq);
1233 scif_destroy_intr_wq(scifdev);
1239 * scif_destroy_loopback_qp - One time uninit work for Loopback Node Qp
1240 * @scifdev: SCIF device
1242 * Destroys the workqueues and frees up the Ring Buffer and Queue Pair memory.
1244 int scif_destroy_loopback_qp(struct scif_dev *scifdev)
1246 struct scif_peer_dev *spdev;
1249 spdev = rcu_dereference(scifdev->spdev);
1252 scif_peer_unregister_device(spdev);
1253 destroy_workqueue(scif_info.loopb_wq);
1254 scif_destroy_intr_wq(scifdev);
1255 kfree(scifdev->qpairs->outbound_q.rb_base);
1256 kfree(scifdev->qpairs);
1257 scifdev->sdev = NULL;
1258 scif_info.loopb_dev = NULL;
1262 void scif_destroy_p2p(struct scif_dev *scifdev)
1264 struct scif_dev *peer_dev;
1265 struct scif_p2p_info *p2p;
1266 struct list_head *pos, *tmp;
1269 mutex_lock(&scif_info.conflock);
1270 /* Free P2P mappings in the given node for all its peer nodes */
1271 list_for_each_safe(pos, tmp, &scifdev->p2p) {
1272 p2p = list_entry(pos, struct scif_p2p_info, ppi_list);
1273 dma_unmap_sg(&scifdev->sdev->dev, p2p->ppi_sg[SCIF_PPI_MMIO],
1274 p2p->sg_nentries[SCIF_PPI_MMIO],
1276 dma_unmap_sg(&scifdev->sdev->dev, p2p->ppi_sg[SCIF_PPI_APER],
1277 p2p->sg_nentries[SCIF_PPI_APER],
1279 scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_MMIO]);
1280 scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_APER]);
1285 /* Free P2P mapping created in the peer nodes for the given node */
1286 for (bd = SCIF_MGMT_NODE + 1; bd <= scif_info.maxid; bd++) {
1287 peer_dev = &scif_dev[bd];
1288 list_for_each_safe(pos, tmp, &peer_dev->p2p) {
1289 p2p = list_entry(pos, struct scif_p2p_info, ppi_list);
1290 if (p2p->ppi_peer_id == scifdev->node) {
1291 dma_unmap_sg(&peer_dev->sdev->dev,
1292 p2p->ppi_sg[SCIF_PPI_MMIO],
1293 p2p->sg_nentries[SCIF_PPI_MMIO],
1295 dma_unmap_sg(&peer_dev->sdev->dev,
1296 p2p->ppi_sg[SCIF_PPI_APER],
1297 p2p->sg_nentries[SCIF_PPI_APER],
1299 scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_MMIO]);
1300 scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_APER]);
1306 mutex_unlock(&scif_info.conflock);