1/*
2 * Copyright (c) 2006 Mellanox Technologies. All rights reserved
3 *
4 * This software is available to you under a choice of one of two
5 * licenses.  You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 *     Redistribution and use in source and binary forms, with or
11 *     without modification, are permitted provided that the following
12 *     conditions are met:
13 *
14 *      - Redistributions of source code must retain the above
15 *        copyright notice, this list of conditions and the following
16 *        disclaimer.
17 *
18 *      - Redistributions in binary form must reproduce the above
19 *        copyright notice, this list of conditions and the following
20 *        disclaimer in the documentation and/or other materials
21 *        provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33#include <rdma/ib_cm.h>
34#include <net/dst.h>
35#include <net/icmp.h>
36#include <linux/icmpv6.h>
37#include <linux/delay.h>
38#include <linux/slab.h>
39#include <linux/vmalloc.h>
40#include <linux/moduleparam.h>
41
42#include "ipoib.h"
43
44int ipoib_max_conn_qp = 128;
45
46module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
47MODULE_PARM_DESC(max_nonsrq_conn_qp,
48		 "Max number of connected-mode QPs per interface "
49		 "(applied only if shared receive queue is not available)");
50
51#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
52static int data_debug_level;
53
54module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
55MODULE_PARM_DESC(cm_data_debug_level,
56		 "Enable data path debug tracing for connected mode if > 0");
57#endif
58
59#define IPOIB_CM_IETF_ID 0x1000000000000000ULL
60
61#define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
62#define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
63#define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
64#define IPOIB_CM_RX_UPDATE_MASK (0x3)
65
66static struct ib_qp_attr ipoib_cm_err_attr = {
67	.qp_state = IB_QPS_ERR
68};
69
70#define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
71
72static struct ib_send_wr ipoib_cm_rx_drain_wr = {
73	.wr_id = IPOIB_CM_RX_DRAIN_WRID,
74	.opcode = IB_WR_SEND,
75};
76
77static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
78			       struct ib_cm_event *event);
79
80static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
81				  u64 mapping[IPOIB_CM_RX_SG])
82{
83	int i;
84
85	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
86
87	for (i = 0; i < frags; ++i)
88		ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
89}
90
91static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
92{
93	struct ipoib_dev_priv *priv = netdev_priv(dev);
94	struct ib_recv_wr *bad_wr;
95	int i, ret;
96
97	priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
98
99	for (i = 0; i < priv->cm.num_frags; ++i)
100		priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
101
102	ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
103	if (unlikely(ret)) {
104		ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
105		ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
106				      priv->cm.srq_ring[id].mapping);
107		dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
108		priv->cm.srq_ring[id].skb = NULL;
109	}
110
111	return ret;
112}
113
114static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
115					struct ipoib_cm_rx *rx,
116					struct ib_recv_wr *wr,
117					struct ib_sge *sge, int id)
118{
119	struct ipoib_dev_priv *priv = netdev_priv(dev);
120	struct ib_recv_wr *bad_wr;
121	int i, ret;
122
123	wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
124
125	for (i = 0; i < IPOIB_CM_RX_SG; ++i)
126		sge[i].addr = rx->rx_ring[id].mapping[i];
127
128	ret = ib_post_recv(rx->qp, wr, &bad_wr);
129	if (unlikely(ret)) {
130		ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
131		ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
132				      rx->rx_ring[id].mapping);
133		dev_kfree_skb_any(rx->rx_ring[id].skb);
134		rx->rx_ring[id].skb = NULL;
135	}
136
137	return ret;
138}
139
140static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
141					     struct ipoib_cm_rx_buf *rx_ring,
142					     int id, int frags,
143					     u64 mapping[IPOIB_CM_RX_SG])
144{
145	struct ipoib_dev_priv *priv = netdev_priv(dev);
146	struct sk_buff *skb;
147	int i;
148
149	skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
150	if (unlikely(!skb))
151		return NULL;
152
153	/*
154	 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
155	 * IP header to a multiple of 16.
156	 */
157	skb_reserve(skb, 12);
158
159	mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
160				       DMA_FROM_DEVICE);
161	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
162		dev_kfree_skb_any(skb);
163		return NULL;
164	}
165
166	for (i = 0; i < frags; i++) {
167		struct page *page = alloc_page(GFP_ATOMIC);
168
169		if (!page)
170			goto partial_error;
171		skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
172
173		mapping[i + 1] = ib_dma_map_page(priv->ca, page,
174						 0, PAGE_SIZE, DMA_FROM_DEVICE);
175		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
176			goto partial_error;
177	}
178
179	rx_ring[id].skb = skb;
180	return skb;
181
182partial_error:
183
184	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
185
186	for (; i > 0; --i)
187		ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
188
189	dev_kfree_skb_any(skb);
190	return NULL;
191}
192
193static void ipoib_cm_free_rx_ring(struct net_device *dev,
194				  struct ipoib_cm_rx_buf *rx_ring)
195{
196	struct ipoib_dev_priv *priv = netdev_priv(dev);
197	int i;
198
199	for (i = 0; i < ipoib_recvq_size; ++i)
200		if (rx_ring[i].skb) {
201			ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
202					      rx_ring[i].mapping);
203			dev_kfree_skb_any(rx_ring[i].skb);
204		}
205
206	vfree(rx_ring);
207}
208
209static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
210{
211	struct ib_send_wr *bad_wr;
212	struct ipoib_cm_rx *p;
213
214	/* We only reserved 1 extra slot in CQ for drain WRs, so
215	 * make sure we have at most 1 outstanding WR. */
216	if (list_empty(&priv->cm.rx_flush_list) ||
217	    !list_empty(&priv->cm.rx_drain_list))
218		return;
219
220	/*
221	 * QPs on flush list are error state.  This way, a "flush
222	 * error" WC will be immediately generated for each WR we post.
223	 */
224	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
225	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
226		ipoib_warn(priv, "failed to post drain wr\n");
227
228	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
229}
230
231static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
232{
233	struct ipoib_cm_rx *p = ctx;
234	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
235	unsigned long flags;
236
237	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
238		return;
239
240	spin_lock_irqsave(&priv->lock, flags);
241	list_move(&p->list, &priv->cm.rx_flush_list);
242	p->state = IPOIB_CM_RX_FLUSH;
243	ipoib_cm_start_rx_drain(priv);
244	spin_unlock_irqrestore(&priv->lock, flags);
245}
246
247static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
248					   struct ipoib_cm_rx *p)
249{
250	struct ipoib_dev_priv *priv = netdev_priv(dev);
251	struct ib_qp_init_attr attr = {
252		.event_handler = ipoib_cm_rx_event_handler,
253		.send_cq = priv->recv_cq, /* For drain WR */
254		.recv_cq = priv->recv_cq,
255		.srq = priv->cm.srq,
256		.cap.max_send_wr = 1, /* For drain WR */
257		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
258		.sq_sig_type = IB_SIGNAL_ALL_WR,
259		.qp_type = IB_QPT_RC,
260		.qp_context = p,
261	};
262
263	if (!ipoib_cm_has_srq(dev)) {
264		attr.cap.max_recv_wr  = ipoib_recvq_size;
265		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
266	}
267
268	return ib_create_qp(priv->pd, &attr);
269}
270
271static int ipoib_cm_modify_rx_qp(struct net_device *dev,
272				 struct ib_cm_id *cm_id, struct ib_qp *qp,
273				 unsigned psn)
274{
275	struct ipoib_dev_priv *priv = netdev_priv(dev);
276	struct ib_qp_attr qp_attr;
277	int qp_attr_mask, ret;
278
279	qp_attr.qp_state = IB_QPS_INIT;
280	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
281	if (ret) {
282		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
283		return ret;
284	}
285	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
286	if (ret) {
287		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
288		return ret;
289	}
290	qp_attr.qp_state = IB_QPS_RTR;
291	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
292	if (ret) {
293		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
294		return ret;
295	}
296	qp_attr.rq_psn = psn;
297	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
298	if (ret) {
299		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
300		return ret;
301	}
302
303	/*
304	 * Current Mellanox HCA firmware won't generate completions
305	 * with error for drain WRs unless the QP has been moved to
306	 * RTS first. This work-around leaves a window where a QP has
307	 * moved to error asynchronously, but this will eventually get
308	 * fixed in firmware, so let's not error out if modify QP
309	 * fails.
310	 */
311	qp_attr.qp_state = IB_QPS_RTS;
312	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
313	if (ret) {
314		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
315		return 0;
316	}
317	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
318	if (ret) {
319		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
320		return 0;
321	}
322
323	return 0;
324}
325
326static void ipoib_cm_init_rx_wr(struct net_device *dev,
327				struct ib_recv_wr *wr,
328				struct ib_sge *sge)
329{
330	struct ipoib_dev_priv *priv = netdev_priv(dev);
331	int i;
332
333	for (i = 0; i < priv->cm.num_frags; ++i)
334		sge[i].lkey = priv->mr->lkey;
335
336	sge[0].length = IPOIB_CM_HEAD_SIZE;
337	for (i = 1; i < priv->cm.num_frags; ++i)
338		sge[i].length = PAGE_SIZE;
339
340	wr->next    = NULL;
341	wr->sg_list = sge;
342	wr->num_sge = priv->cm.num_frags;
343}
344
345static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
346				   struct ipoib_cm_rx *rx)
347{
348	struct ipoib_dev_priv *priv = netdev_priv(dev);
349	struct {
350		struct ib_recv_wr wr;
351		struct ib_sge sge[IPOIB_CM_RX_SG];
352	} *t;
353	int ret;
354	int i;
355
356	rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
357	if (!rx->rx_ring) {
358		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
359		       priv->ca->name, ipoib_recvq_size);
360		return -ENOMEM;
361	}
362
363	t = kmalloc(sizeof *t, GFP_KERNEL);
364	if (!t) {
365		ret = -ENOMEM;
366		goto err_free;
367	}
368
369	ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
370
371	spin_lock_irq(&priv->lock);
372
373	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
374		spin_unlock_irq(&priv->lock);
375		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
376		ret = -EINVAL;
377		goto err_free;
378	} else
379		++priv->cm.nonsrq_conn_qp;
380
381	spin_unlock_irq(&priv->lock);
382
383	for (i = 0; i < ipoib_recvq_size; ++i) {
384		if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
385					   rx->rx_ring[i].mapping)) {
386			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
387				ret = -ENOMEM;
388				goto err_count;
389		}
390		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
391		if (ret) {
392			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
393				   "failed for buf %d\n", i);
394			ret = -EIO;
395			goto err_count;
396		}
397	}
398
399	rx->recv_count = ipoib_recvq_size;
400
401	kfree(t);
402
403	return 0;
404
405err_count:
406	spin_lock_irq(&priv->lock);
407	--priv->cm.nonsrq_conn_qp;
408	spin_unlock_irq(&priv->lock);
409
410err_free:
411	kfree(t);
412	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
413
414	return ret;
415}
416
417static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
418			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
419			     unsigned psn)
420{
421	struct ipoib_dev_priv *priv = netdev_priv(dev);
422	struct ipoib_cm_data data = {};
423	struct ib_cm_rep_param rep = {};
424
425	data.qpn = cpu_to_be32(priv->qp->qp_num);
426	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
427
428	rep.private_data = &data;
429	rep.private_data_len = sizeof data;
430	rep.flow_control = 0;
431	rep.rnr_retry_count = req->rnr_retry_count;
432	rep.srq = ipoib_cm_has_srq(dev);
433	rep.qp_num = qp->qp_num;
434	rep.starting_psn = psn;
435	return ib_send_cm_rep(cm_id, &rep);
436}
437
438static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
439{
440	struct net_device *dev = cm_id->context;
441	struct ipoib_dev_priv *priv = netdev_priv(dev);
442	struct ipoib_cm_rx *p;
443	unsigned psn;
444	int ret;
445
446	ipoib_dbg(priv, "REQ arrived\n");
447	p = kzalloc(sizeof *p, GFP_KERNEL);
448	if (!p)
449		return -ENOMEM;
450	p->dev = dev;
451	p->id = cm_id;
452	cm_id->context = p;
453	p->state = IPOIB_CM_RX_LIVE;
454	p->jiffies = jiffies;
455	INIT_LIST_HEAD(&p->list);
456
457	p->qp = ipoib_cm_create_rx_qp(dev, p);
458	if (IS_ERR(p->qp)) {
459		ret = PTR_ERR(p->qp);
460		goto err_qp;
461	}
462
463	psn = random32() & 0xffffff;
464	ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
465	if (ret)
466		goto err_modify;
467
468	if (!ipoib_cm_has_srq(dev)) {
469		ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
470		if (ret)
471			goto err_modify;
472	}
473
474	spin_lock_irq(&priv->lock);
475	queue_delayed_work(ipoib_workqueue,
476			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
477	/* Add this entry to passive ids list head, but do not re-add it
478	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
479	p->jiffies = jiffies;
480	if (p->state == IPOIB_CM_RX_LIVE)
481		list_move(&p->list, &priv->cm.passive_ids);
482	spin_unlock_irq(&priv->lock);
483
484	ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
485	if (ret) {
486		ipoib_warn(priv, "failed to send REP: %d\n", ret);
487		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
488			ipoib_warn(priv, "unable to move qp to error state\n");
489	}
490	return 0;
491
492err_modify:
493	ib_destroy_qp(p->qp);
494err_qp:
495	kfree(p);
496	return ret;
497}
498
499static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
500			       struct ib_cm_event *event)
501{
502	struct ipoib_cm_rx *p;
503	struct ipoib_dev_priv *priv;
504
505	switch (event->event) {
506	case IB_CM_REQ_RECEIVED:
507		return ipoib_cm_req_handler(cm_id, event);
508	case IB_CM_DREQ_RECEIVED:
509		p = cm_id->context;
510		ib_send_cm_drep(cm_id, NULL, 0);
511		/* Fall through */
512	case IB_CM_REJ_RECEIVED:
513		p = cm_id->context;
514		priv = netdev_priv(p->dev);
515		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
516			ipoib_warn(priv, "unable to move qp to error state\n");
517		/* Fall through */
518	default:
519		return 0;
520	}
521}
522/* Adjust length of skb with fragments to match received data */
523static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
524			  unsigned int length, struct sk_buff *toskb)
525{
526	int i, num_frags;
527	unsigned int size;
528
529	/* put header into skb */
530	size = min(length, hdr_space);
531	skb->tail += size;
532	skb->len += size;
533	length -= size;
534
535	num_frags = skb_shinfo(skb)->nr_frags;
536	for (i = 0; i < num_frags; i++) {
537		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
538
539		if (length == 0) {
540			/* don't need this page */
541			skb_fill_page_desc(toskb, i, skb_frag_page(frag),
542					   0, PAGE_SIZE);
543			--skb_shinfo(skb)->nr_frags;
544		} else {
545			size = min(length, (unsigned) PAGE_SIZE);
546
547			skb_frag_size_set(frag, size);
548			skb->data_len += size;
549			skb->truesize += size;
550			skb->len += size;
551			length -= size;
552		}
553	}
554}
555
556void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
557{
558	struct ipoib_dev_priv *priv = netdev_priv(dev);
559	struct ipoib_cm_rx_buf *rx_ring;
560	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
561	struct sk_buff *skb, *newskb;
562	struct ipoib_cm_rx *p;
563	unsigned long flags;
564	u64 mapping[IPOIB_CM_RX_SG];
565	int frags;
566	int has_srq;
567	struct sk_buff *small_skb;
568
569	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
570		       wr_id, wc->status);
571
572	if (unlikely(wr_id >= ipoib_recvq_size)) {
573		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
574			spin_lock_irqsave(&priv->lock, flags);
575			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
576			ipoib_cm_start_rx_drain(priv);
577			queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
578			spin_unlock_irqrestore(&priv->lock, flags);
579		} else
580			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
581				   wr_id, ipoib_recvq_size);
582		return;
583	}
584
585	p = wc->qp->qp_context;
586
587	has_srq = ipoib_cm_has_srq(dev);
588	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
589
590	skb = rx_ring[wr_id].skb;
591
592	if (unlikely(wc->status != IB_WC_SUCCESS)) {
593		ipoib_dbg(priv, "cm recv error "
594			   "(status=%d, wrid=%d vend_err %x)\n",
595			   wc->status, wr_id, wc->vendor_err);
596		++dev->stats.rx_dropped;
597		if (has_srq)
598			goto repost;
599		else {
600			if (!--p->recv_count) {
601				spin_lock_irqsave(&priv->lock, flags);
602				list_move(&p->list, &priv->cm.rx_reap_list);
603				spin_unlock_irqrestore(&priv->lock, flags);
604				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
605			}
606			return;
607		}
608	}
609
610	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
611		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
612			spin_lock_irqsave(&priv->lock, flags);
613			p->jiffies = jiffies;
614			/* Move this entry to list head, but do not re-add it
615			 * if it has been moved out of list. */
616			if (p->state == IPOIB_CM_RX_LIVE)
617				list_move(&p->list, &priv->cm.passive_ids);
618			spin_unlock_irqrestore(&priv->lock, flags);
619		}
620	}
621
622	if (wc->byte_len < IPOIB_CM_COPYBREAK) {
623		int dlen = wc->byte_len;
624
625		small_skb = dev_alloc_skb(dlen + 12);
626		if (small_skb) {
627			skb_reserve(small_skb, 12);
628			ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
629						   dlen, DMA_FROM_DEVICE);
630			skb_copy_from_linear_data(skb, small_skb->data, dlen);
631			ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
632						      dlen, DMA_FROM_DEVICE);
633			skb_put(small_skb, dlen);
634			skb = small_skb;
635			goto copied;
636		}
637	}
638
639	frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
640					      (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
641
642	newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
643	if (unlikely(!newskb)) {
644		/*
645		 * If we can't allocate a new RX buffer, dump
646		 * this packet and reuse the old buffer.
647		 */
648		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
649		++dev->stats.rx_dropped;
650		goto repost;
651	}
652
653	ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
654	memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
655
656	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
657		       wc->byte_len, wc->slid);
658
659	skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
660
661copied:
662	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
663	skb_reset_mac_header(skb);
664	skb_pull(skb, IPOIB_ENCAP_LEN);
665
666	++dev->stats.rx_packets;
667	dev->stats.rx_bytes += skb->len;
668
669	skb->dev = dev;
670	/* XXX get correct PACKET_ type here */
671	skb->pkt_type = PACKET_HOST;
672	netif_receive_skb(skb);
673
674repost:
675	if (has_srq) {
676		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
677			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
678				   "for buf %d\n", wr_id);
679	} else {
680		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
681							  &priv->cm.rx_wr,
682							  priv->cm.rx_sge,
683							  wr_id))) {
684			--p->recv_count;
685			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
686				   "for buf %d\n", wr_id);
687		}
688	}
689}
690
691static inline int post_send(struct ipoib_dev_priv *priv,
692			    struct ipoib_cm_tx *tx,
693			    unsigned int wr_id,
694			    u64 addr, int len)
695{
696	struct ib_send_wr *bad_wr;
697
698	priv->tx_sge[0].addr          = addr;
699	priv->tx_sge[0].length        = len;
700
701	priv->tx_wr.num_sge	= 1;
702	priv->tx_wr.wr_id	= wr_id | IPOIB_OP_CM;
703
704	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
705}
706
707void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
708{
709	struct ipoib_dev_priv *priv = netdev_priv(dev);
710	struct ipoib_cm_tx_buf *tx_req;
711	u64 addr;
712	int rc;
713
714	if (unlikely(skb->len > tx->mtu)) {
715		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
716			   skb->len, tx->mtu);
717		++dev->stats.tx_dropped;
718		++dev->stats.tx_errors;
719		ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
720		return;
721	}
722
723	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
724		       tx->tx_head, skb->len, tx->qp->qp_num);
725
726	/*
727	 * We put the skb into the tx_ring _before_ we call post_send()
728	 * because it's entirely possible that the completion handler will
729	 * run before we execute anything after the post_send().  That
730	 * means we have to make sure everything is properly recorded and
731	 * our state is consistent before we call post_send().
732	 */
733	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
734	tx_req->skb = skb;
735	addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
736	if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
737		++dev->stats.tx_errors;
738		dev_kfree_skb_any(skb);
739		return;
740	}
741
742	tx_req->mapping = addr;
743
744	rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
745		       addr, skb->len);
746	if (unlikely(rc)) {
747		ipoib_warn(priv, "post_send failed, error %d\n", rc);
748		++dev->stats.tx_errors;
749		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
750		dev_kfree_skb_any(skb);
751	} else {
752		dev->trans_start = jiffies;
753		++tx->tx_head;
754
755		if (++priv->tx_outstanding == ipoib_sendq_size) {
756			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
757				  tx->qp->qp_num);
758			if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
759				ipoib_warn(priv, "request notify on send CQ failed\n");
760			netif_stop_queue(dev);
761		}
762	}
763}
764
765void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
766{
767	struct ipoib_dev_priv *priv = netdev_priv(dev);
768	struct ipoib_cm_tx *tx = wc->qp->qp_context;
769	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
770	struct ipoib_cm_tx_buf *tx_req;
771	unsigned long flags;
772
773	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
774		       wr_id, wc->status);
775
776	if (unlikely(wr_id >= ipoib_sendq_size)) {
777		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
778			   wr_id, ipoib_sendq_size);
779		return;
780	}
781
782	tx_req = &tx->tx_ring[wr_id];
783
784	ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
785
786	/* FIXME: is this right? Shouldn't we only increment on success? */
787	++dev->stats.tx_packets;
788	dev->stats.tx_bytes += tx_req->skb->len;
789
790	dev_kfree_skb_any(tx_req->skb);
791
792	netif_tx_lock(dev);
793
794	++tx->tx_tail;
795	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
796	    netif_queue_stopped(dev) &&
797	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
798		netif_wake_queue(dev);
799
800	if (wc->status != IB_WC_SUCCESS &&
801	    wc->status != IB_WC_WR_FLUSH_ERR) {
802		struct ipoib_neigh *neigh;
803
804		ipoib_dbg(priv, "failed cm send event "
805			   "(status=%d, wrid=%d vend_err %x)\n",
806			   wc->status, wr_id, wc->vendor_err);
807
808		spin_lock_irqsave(&priv->lock, flags);
809		neigh = tx->neigh;
810
811		if (neigh) {
812			neigh->cm = NULL;
813			list_del(&neigh->list);
814			if (neigh->ah)
815				ipoib_put_ah(neigh->ah);
816			ipoib_neigh_free(dev, neigh);
817
818			tx->neigh = NULL;
819		}
820
821		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
822			list_move(&tx->list, &priv->cm.reap_list);
823			queue_work(ipoib_workqueue, &priv->cm.reap_task);
824		}
825
826		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
827
828		spin_unlock_irqrestore(&priv->lock, flags);
829	}
830
831	netif_tx_unlock(dev);
832}
833
834int ipoib_cm_dev_open(struct net_device *dev)
835{
836	struct ipoib_dev_priv *priv = netdev_priv(dev);
837	int ret;
838
839	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
840		return 0;
841
842	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
843	if (IS_ERR(priv->cm.id)) {
844		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
845		ret = PTR_ERR(priv->cm.id);
846		goto err_cm;
847	}
848
849	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
850			   0, NULL);
851	if (ret) {
852		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
853		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
854		goto err_listen;
855	}
856
857	return 0;
858
859err_listen:
860	ib_destroy_cm_id(priv->cm.id);
861err_cm:
862	priv->cm.id = NULL;
863	return ret;
864}
865
866static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
867{
868	struct ipoib_dev_priv *priv = netdev_priv(dev);
869	struct ipoib_cm_rx *rx, *n;
870	LIST_HEAD(list);
871
872	spin_lock_irq(&priv->lock);
873	list_splice_init(&priv->cm.rx_reap_list, &list);
874	spin_unlock_irq(&priv->lock);
875
876	list_for_each_entry_safe(rx, n, &list, list) {
877		ib_destroy_cm_id(rx->id);
878		ib_destroy_qp(rx->qp);
879		if (!ipoib_cm_has_srq(dev)) {
880			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
881			spin_lock_irq(&priv->lock);
882			--priv->cm.nonsrq_conn_qp;
883			spin_unlock_irq(&priv->lock);
884		}
885		kfree(rx);
886	}
887}
888
889void ipoib_cm_dev_stop(struct net_device *dev)
890{
891	struct ipoib_dev_priv *priv = netdev_priv(dev);
892	struct ipoib_cm_rx *p;
893	unsigned long begin;
894	int ret;
895
896	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
897		return;
898
899	ib_destroy_cm_id(priv->cm.id);
900	priv->cm.id = NULL;
901
902	spin_lock_irq(&priv->lock);
903	while (!list_empty(&priv->cm.passive_ids)) {
904		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
905		list_move(&p->list, &priv->cm.rx_error_list);
906		p->state = IPOIB_CM_RX_ERROR;
907		spin_unlock_irq(&priv->lock);
908		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
909		if (ret)
910			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
911		spin_lock_irq(&priv->lock);
912	}
913
914	/* Wait for all RX to be drained */
915	begin = jiffies;
916
917	while (!list_empty(&priv->cm.rx_error_list) ||
918	       !list_empty(&priv->cm.rx_flush_list) ||
919	       !list_empty(&priv->cm.rx_drain_list)) {
920		if (time_after(jiffies, begin + 5 * HZ)) {
921			ipoib_warn(priv, "RX drain timing out\n");
922
923			/*
924			 * assume the HW is wedged and just free up everything.
925			 */
926			list_splice_init(&priv->cm.rx_flush_list,
927					 &priv->cm.rx_reap_list);
928			list_splice_init(&priv->cm.rx_error_list,
929					 &priv->cm.rx_reap_list);
930			list_splice_init(&priv->cm.rx_drain_list,
931					 &priv->cm.rx_reap_list);
932			break;
933		}
934		spin_unlock_irq(&priv->lock);
935		msleep(1);
936		ipoib_drain_cq(dev);
937		spin_lock_irq(&priv->lock);
938	}
939
940	spin_unlock_irq(&priv->lock);
941
942	ipoib_cm_free_rx_reap_list(dev);
943
944	cancel_delayed_work(&priv->cm.stale_task);
945}
946
947static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
948{
949	struct ipoib_cm_tx *p = cm_id->context;
950	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
951	struct ipoib_cm_data *data = event->private_data;
952	struct sk_buff_head skqueue;
953	struct ib_qp_attr qp_attr;
954	int qp_attr_mask, ret;
955	struct sk_buff *skb;
956
957	p->mtu = be32_to_cpu(data->mtu);
958
959	if (p->mtu <= IPOIB_ENCAP_LEN) {
960		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
961			   p->mtu, IPOIB_ENCAP_LEN);
962		return -EINVAL;
963	}
964
965	qp_attr.qp_state = IB_QPS_RTR;
966	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
967	if (ret) {
968		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
969		return ret;
970	}
971
972	qp_attr.rq_psn = 0 /* FIXME */;
973	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
974	if (ret) {
975		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
976		return ret;
977	}
978
979	qp_attr.qp_state = IB_QPS_RTS;
980	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
981	if (ret) {
982		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
983		return ret;
984	}
985	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
986	if (ret) {
987		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
988		return ret;
989	}
990
991	skb_queue_head_init(&skqueue);
992
993	spin_lock_irq(&priv->lock);
994	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
995	if (p->neigh)
996		while ((skb = __skb_dequeue(&p->neigh->queue)))
997			__skb_queue_tail(&skqueue, skb);
998	spin_unlock_irq(&priv->lock);
999
1000	while ((skb = __skb_dequeue(&skqueue))) {
1001		skb->dev = p->dev;
1002		if (dev_queue_xmit(skb))
1003			ipoib_warn(priv, "dev_queue_xmit failed "
1004				   "to requeue packet\n");
1005	}
1006
1007	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1008	if (ret) {
1009		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1010		return ret;
1011	}
1012	return 0;
1013}
1014
1015static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1016{
1017	struct ipoib_dev_priv *priv = netdev_priv(dev);
1018	struct ib_qp_init_attr attr = {
1019		.send_cq		= priv->recv_cq,
1020		.recv_cq		= priv->recv_cq,
1021		.srq			= priv->cm.srq,
1022		.cap.max_send_wr	= ipoib_sendq_size,
1023		.cap.max_send_sge	= 1,
1024		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1025		.qp_type		= IB_QPT_RC,
1026		.qp_context		= tx
1027	};
1028
1029	return ib_create_qp(priv->pd, &attr);
1030}
1031
1032static int ipoib_cm_send_req(struct net_device *dev,
1033			     struct ib_cm_id *id, struct ib_qp *qp,
1034			     u32 qpn,
1035			     struct ib_sa_path_rec *pathrec)
1036{
1037	struct ipoib_dev_priv *priv = netdev_priv(dev);
1038	struct ipoib_cm_data data = {};
1039	struct ib_cm_req_param req = {};
1040
1041	data.qpn = cpu_to_be32(priv->qp->qp_num);
1042	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1043
1044	req.primary_path		= pathrec;
1045	req.alternate_path		= NULL;
1046	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1047	req.qp_num			= qp->qp_num;
1048	req.qp_type			= qp->qp_type;
1049	req.private_data		= &data;
1050	req.private_data_len		= sizeof data;
1051	req.flow_control		= 0;
1052
1053	req.starting_psn		= 0; /* FIXME */
1054
1055	/*
1056	 * Pick some arbitrary defaults here; we could make these
1057	 * module parameters if anyone cared about setting them.
1058	 */
1059	req.responder_resources		= 4;
1060	req.remote_cm_response_timeout	= 20;
1061	req.local_cm_response_timeout	= 20;
1062	req.retry_count			= 0; /* RFC draft warns against retries */
1063	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1064	req.max_cm_retries		= 15;
1065	req.srq				= ipoib_cm_has_srq(dev);
1066	return ib_send_cm_req(id, &req);
1067}
1068
1069static int ipoib_cm_modify_tx_init(struct net_device *dev,
1070				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1071{
1072	struct ipoib_dev_priv *priv = netdev_priv(dev);
1073	struct ib_qp_attr qp_attr;
1074	int qp_attr_mask, ret;
1075	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1076	if (ret) {
1077		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1078		return ret;
1079	}
1080
1081	qp_attr.qp_state = IB_QPS_INIT;
1082	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1083	qp_attr.port_num = priv->port;
1084	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1085
1086	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1087	if (ret) {
1088		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1089		return ret;
1090	}
1091	return 0;
1092}
1093
1094static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1095			    struct ib_sa_path_rec *pathrec)
1096{
1097	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1098	int ret;
1099
1100	p->tx_ring = vzalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1101	if (!p->tx_ring) {
1102		ipoib_warn(priv, "failed to allocate tx ring\n");
1103		ret = -ENOMEM;
1104		goto err_tx;
1105	}
1106
1107	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1108	if (IS_ERR(p->qp)) {
1109		ret = PTR_ERR(p->qp);
1110		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1111		goto err_qp;
1112	}
1113
1114	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1115	if (IS_ERR(p->id)) {
1116		ret = PTR_ERR(p->id);
1117		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1118		goto err_id;
1119	}
1120
1121	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1122	if (ret) {
1123		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1124		goto err_modify;
1125	}
1126
1127	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1128	if (ret) {
1129		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1130		goto err_send_cm;
1131	}
1132
1133	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1134		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1135
1136	return 0;
1137
1138err_send_cm:
1139err_modify:
1140	ib_destroy_cm_id(p->id);
1141err_id:
1142	p->id = NULL;
1143	ib_destroy_qp(p->qp);
1144err_qp:
1145	p->qp = NULL;
1146	vfree(p->tx_ring);
1147err_tx:
1148	return ret;
1149}
1150
1151static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1152{
1153	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1154	struct ipoib_cm_tx_buf *tx_req;
1155	unsigned long begin;
1156
1157	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1158		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1159
1160	if (p->id)
1161		ib_destroy_cm_id(p->id);
1162
1163	if (p->tx_ring) {
1164		/* Wait for all sends to complete */
1165		begin = jiffies;
1166		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1167			if (time_after(jiffies, begin + 5 * HZ)) {
1168				ipoib_warn(priv, "timing out; %d sends not completed\n",
1169					   p->tx_head - p->tx_tail);
1170				goto timeout;
1171			}
1172
1173			msleep(1);
1174		}
1175	}
1176
1177timeout:
1178
1179	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1180		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1181		ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1182				    DMA_TO_DEVICE);
1183		dev_kfree_skb_any(tx_req->skb);
1184		++p->tx_tail;
1185		netif_tx_lock_bh(p->dev);
1186		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1187		    netif_queue_stopped(p->dev) &&
1188		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1189			netif_wake_queue(p->dev);
1190		netif_tx_unlock_bh(p->dev);
1191	}
1192
1193	if (p->qp)
1194		ib_destroy_qp(p->qp);
1195
1196	vfree(p->tx_ring);
1197	kfree(p);
1198}
1199
1200static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1201			       struct ib_cm_event *event)
1202{
1203	struct ipoib_cm_tx *tx = cm_id->context;
1204	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1205	struct net_device *dev = priv->dev;
1206	struct ipoib_neigh *neigh;
1207	unsigned long flags;
1208	int ret;
1209
1210	switch (event->event) {
1211	case IB_CM_DREQ_RECEIVED:
1212		ipoib_dbg(priv, "DREQ received.\n");
1213		ib_send_cm_drep(cm_id, NULL, 0);
1214		break;
1215	case IB_CM_REP_RECEIVED:
1216		ipoib_dbg(priv, "REP received.\n");
1217		ret = ipoib_cm_rep_handler(cm_id, event);
1218		if (ret)
1219			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1220				       NULL, 0, NULL, 0);
1221		break;
1222	case IB_CM_REQ_ERROR:
1223	case IB_CM_REJ_RECEIVED:
1224	case IB_CM_TIMEWAIT_EXIT:
1225		ipoib_dbg(priv, "CM error %d.\n", event->event);
1226		netif_tx_lock_bh(dev);
1227		spin_lock_irqsave(&priv->lock, flags);
1228		neigh = tx->neigh;
1229
1230		if (neigh) {
1231			neigh->cm = NULL;
1232			list_del(&neigh->list);
1233			if (neigh->ah)
1234				ipoib_put_ah(neigh->ah);
1235			ipoib_neigh_free(dev, neigh);
1236
1237			tx->neigh = NULL;
1238		}
1239
1240		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1241			list_move(&tx->list, &priv->cm.reap_list);
1242			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1243		}
1244
1245		spin_unlock_irqrestore(&priv->lock, flags);
1246		netif_tx_unlock_bh(dev);
1247		break;
1248	default:
1249		break;
1250	}
1251
1252	return 0;
1253}
1254
1255struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1256				       struct ipoib_neigh *neigh)
1257{
1258	struct ipoib_dev_priv *priv = netdev_priv(dev);
1259	struct ipoib_cm_tx *tx;
1260
1261	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1262	if (!tx)
1263		return NULL;
1264
1265	neigh->cm = tx;
1266	tx->neigh = neigh;
1267	tx->path = path;
1268	tx->dev = dev;
1269	list_add(&tx->list, &priv->cm.start_list);
1270	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1271	queue_work(ipoib_workqueue, &priv->cm.start_task);
1272	return tx;
1273}
1274
1275void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1276{
1277	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1278	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1279		list_move(&tx->list, &priv->cm.reap_list);
1280		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1281		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1282			  tx->neigh->dgid.raw);
1283		tx->neigh = NULL;
1284	}
1285}
1286
1287static void ipoib_cm_tx_start(struct work_struct *work)
1288{
1289	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1290						   cm.start_task);
1291	struct net_device *dev = priv->dev;
1292	struct ipoib_neigh *neigh;
1293	struct ipoib_cm_tx *p;
1294	unsigned long flags;
1295	int ret;
1296
1297	struct ib_sa_path_rec pathrec;
1298	u32 qpn;
1299
1300	netif_tx_lock_bh(dev);
1301	spin_lock_irqsave(&priv->lock, flags);
1302
1303	while (!list_empty(&priv->cm.start_list)) {
1304		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1305		list_del_init(&p->list);
1306		neigh = p->neigh;
1307		qpn = IPOIB_QPN(neigh->neighbour->ha);
1308		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1309
1310		spin_unlock_irqrestore(&priv->lock, flags);
1311		netif_tx_unlock_bh(dev);
1312
1313		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1314
1315		netif_tx_lock_bh(dev);
1316		spin_lock_irqsave(&priv->lock, flags);
1317
1318		if (ret) {
1319			neigh = p->neigh;
1320			if (neigh) {
1321				neigh->cm = NULL;
1322				list_del(&neigh->list);
1323				if (neigh->ah)
1324					ipoib_put_ah(neigh->ah);
1325				ipoib_neigh_free(dev, neigh);
1326			}
1327			list_del(&p->list);
1328			kfree(p);
1329		}
1330	}
1331
1332	spin_unlock_irqrestore(&priv->lock, flags);
1333	netif_tx_unlock_bh(dev);
1334}
1335
1336static void ipoib_cm_tx_reap(struct work_struct *work)
1337{
1338	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1339						   cm.reap_task);
1340	struct net_device *dev = priv->dev;
1341	struct ipoib_cm_tx *p;
1342	unsigned long flags;
1343
1344	netif_tx_lock_bh(dev);
1345	spin_lock_irqsave(&priv->lock, flags);
1346
1347	while (!list_empty(&priv->cm.reap_list)) {
1348		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1349		list_del(&p->list);
1350		spin_unlock_irqrestore(&priv->lock, flags);
1351		netif_tx_unlock_bh(dev);
1352		ipoib_cm_tx_destroy(p);
1353		netif_tx_lock_bh(dev);
1354		spin_lock_irqsave(&priv->lock, flags);
1355	}
1356
1357	spin_unlock_irqrestore(&priv->lock, flags);
1358	netif_tx_unlock_bh(dev);
1359}
1360
1361static void ipoib_cm_skb_reap(struct work_struct *work)
1362{
1363	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1364						   cm.skb_task);
1365	struct net_device *dev = priv->dev;
1366	struct sk_buff *skb;
1367	unsigned long flags;
1368	unsigned mtu = priv->mcast_mtu;
1369
1370	netif_tx_lock_bh(dev);
1371	spin_lock_irqsave(&priv->lock, flags);
1372
1373	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1374		spin_unlock_irqrestore(&priv->lock, flags);
1375		netif_tx_unlock_bh(dev);
1376
1377		if (skb->protocol == htons(ETH_P_IP))
1378			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1379#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1380		else if (skb->protocol == htons(ETH_P_IPV6))
1381			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1382#endif
1383		dev_kfree_skb_any(skb);
1384
1385		netif_tx_lock_bh(dev);
1386		spin_lock_irqsave(&priv->lock, flags);
1387	}
1388
1389	spin_unlock_irqrestore(&priv->lock, flags);
1390	netif_tx_unlock_bh(dev);
1391}
1392
1393void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1394			   unsigned int mtu)
1395{
1396	struct ipoib_dev_priv *priv = netdev_priv(dev);
1397	int e = skb_queue_empty(&priv->cm.skb_queue);
1398
1399	if (skb_dst(skb))
1400		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu);
1401
1402	skb_queue_tail(&priv->cm.skb_queue, skb);
1403	if (e)
1404		queue_work(ipoib_workqueue, &priv->cm.skb_task);
1405}
1406
1407static void ipoib_cm_rx_reap(struct work_struct *work)
1408{
1409	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1410						cm.rx_reap_task)->dev);
1411}
1412
1413static void ipoib_cm_stale_task(struct work_struct *work)
1414{
1415	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1416						   cm.stale_task.work);
1417	struct ipoib_cm_rx *p;
1418	int ret;
1419
1420	spin_lock_irq(&priv->lock);
1421	while (!list_empty(&priv->cm.passive_ids)) {
1422		/* List is sorted by LRU, start from tail,
1423		 * stop when we see a recently used entry */
1424		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1425		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1426			break;
1427		list_move(&p->list, &priv->cm.rx_error_list);
1428		p->state = IPOIB_CM_RX_ERROR;
1429		spin_unlock_irq(&priv->lock);
1430		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1431		if (ret)
1432			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1433		spin_lock_irq(&priv->lock);
1434	}
1435
1436	if (!list_empty(&priv->cm.passive_ids))
1437		queue_delayed_work(ipoib_workqueue,
1438				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1439	spin_unlock_irq(&priv->lock);
1440}
1441
1442
1443static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1444			 char *buf)
1445{
1446	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1447
1448	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1449		return sprintf(buf, "connected\n");
1450	else
1451		return sprintf(buf, "datagram\n");
1452}
1453
1454static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1455			const char *buf, size_t count)
1456{
1457	struct net_device *dev = to_net_dev(d);
1458	struct ipoib_dev_priv *priv = netdev_priv(dev);
1459
1460	if (!rtnl_trylock())
1461		return restart_syscall();
1462
1463	/* flush paths if we switch modes so that connections are restarted */
1464	if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1465		set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1466		ipoib_warn(priv, "enabling connected mode "
1467			   "will cause multicast packet drops\n");
1468		netdev_update_features(dev);
1469		rtnl_unlock();
1470		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1471
1472		ipoib_flush_paths(dev);
1473		return count;
1474	}
1475
1476	if (!strcmp(buf, "datagram\n")) {
1477		clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1478		netdev_update_features(dev);
1479		dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1480		rtnl_unlock();
1481		ipoib_flush_paths(dev);
1482
1483		return count;
1484	}
1485	rtnl_unlock();
1486
1487	return -EINVAL;
1488}
1489
1490static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1491
1492int ipoib_cm_add_mode_attr(struct net_device *dev)
1493{
1494	return device_create_file(&dev->dev, &dev_attr_mode);
1495}
1496
1497static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1498{
1499	struct ipoib_dev_priv *priv = netdev_priv(dev);
1500	struct ib_srq_init_attr srq_init_attr = {
1501		.srq_type = IB_SRQT_BASIC,
1502		.attr = {
1503			.max_wr  = ipoib_recvq_size,
1504			.max_sge = max_sge
1505		}
1506	};
1507
1508	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1509	if (IS_ERR(priv->cm.srq)) {
1510		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1511			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1512			       priv->ca->name, PTR_ERR(priv->cm.srq));
1513		priv->cm.srq = NULL;
1514		return;
1515	}
1516
1517	priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1518	if (!priv->cm.srq_ring) {
1519		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1520		       priv->ca->name, ipoib_recvq_size);
1521		ib_destroy_srq(priv->cm.srq);
1522		priv->cm.srq = NULL;
1523		return;
1524	}
1525
1526}
1527
1528int ipoib_cm_dev_init(struct net_device *dev)
1529{
1530	struct ipoib_dev_priv *priv = netdev_priv(dev);
1531	int i, ret;
1532	struct ib_device_attr attr;
1533
1534	INIT_LIST_HEAD(&priv->cm.passive_ids);
1535	INIT_LIST_HEAD(&priv->cm.reap_list);
1536	INIT_LIST_HEAD(&priv->cm.start_list);
1537	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1538	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1539	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1540	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1541	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1542	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1543	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1544	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1545	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1546
1547	skb_queue_head_init(&priv->cm.skb_queue);
1548
1549	ret = ib_query_device(priv->ca, &attr);
1550	if (ret) {
1551		printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1552		return ret;
1553	}
1554
1555	ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1556
1557	attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1558	ipoib_cm_create_srq(dev, attr.max_srq_sge);
1559	if (ipoib_cm_has_srq(dev)) {
1560		priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1561		priv->cm.num_frags  = attr.max_srq_sge;
1562		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1563			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1564	} else {
1565		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1566		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1567	}
1568
1569	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1570
1571	if (ipoib_cm_has_srq(dev)) {
1572		for (i = 0; i < ipoib_recvq_size; ++i) {
1573			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1574						   priv->cm.num_frags - 1,
1575						   priv->cm.srq_ring[i].mapping)) {
1576				ipoib_warn(priv, "failed to allocate "
1577					   "receive buffer %d\n", i);
1578				ipoib_cm_dev_cleanup(dev);
1579				return -ENOMEM;
1580			}
1581
1582			if (ipoib_cm_post_receive_srq(dev, i)) {
1583				ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1584					   "failed for buf %d\n", i);
1585				ipoib_cm_dev_cleanup(dev);
1586				return -EIO;
1587			}
1588		}
1589	}
1590
1591	priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1592	return 0;
1593}
1594
1595void ipoib_cm_dev_cleanup(struct net_device *dev)
1596{
1597	struct ipoib_dev_priv *priv = netdev_priv(dev);
1598	int ret;
1599
1600	if (!priv->cm.srq)
1601		return;
1602
1603	ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1604
1605	ret = ib_destroy_srq(priv->cm.srq);
1606	if (ret)
1607		ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1608
1609	priv->cm.srq = NULL;
1610	if (!priv->cm.srq_ring)
1611		return;
1612
1613	ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1614	priv->cm.srq_ring = NULL;
1615}
1616