[go: nahoru, domu]

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					     gfp_t gfp)
145{
146	struct ipoib_dev_priv *priv = netdev_priv(dev);
147	struct sk_buff *skb;
148	int i;
149
150	skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
151	if (unlikely(!skb))
152		return NULL;
153
154	/*
155	 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
156	 * IP header to a multiple of 16.
157	 */
158	skb_reserve(skb, 12);
159
160	mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
161				       DMA_FROM_DEVICE);
162	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
163		dev_kfree_skb_any(skb);
164		return NULL;
165	}
166
167	for (i = 0; i < frags; i++) {
168		struct page *page = alloc_page(gfp);
169
170		if (!page)
171			goto partial_error;
172		skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
173
174		mapping[i + 1] = ib_dma_map_page(priv->ca, page,
175						 0, PAGE_SIZE, DMA_FROM_DEVICE);
176		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
177			goto partial_error;
178	}
179
180	rx_ring[id].skb = skb;
181	return skb;
182
183partial_error:
184
185	ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
186
187	for (; i > 0; --i)
188		ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
189
190	dev_kfree_skb_any(skb);
191	return NULL;
192}
193
194static void ipoib_cm_free_rx_ring(struct net_device *dev,
195				  struct ipoib_cm_rx_buf *rx_ring)
196{
197	struct ipoib_dev_priv *priv = netdev_priv(dev);
198	int i;
199
200	for (i = 0; i < ipoib_recvq_size; ++i)
201		if (rx_ring[i].skb) {
202			ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
203					      rx_ring[i].mapping);
204			dev_kfree_skb_any(rx_ring[i].skb);
205		}
206
207	vfree(rx_ring);
208}
209
210static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
211{
212	struct ib_send_wr *bad_wr;
213	struct ipoib_cm_rx *p;
214
215	/* We only reserved 1 extra slot in CQ for drain WRs, so
216	 * make sure we have at most 1 outstanding WR. */
217	if (list_empty(&priv->cm.rx_flush_list) ||
218	    !list_empty(&priv->cm.rx_drain_list))
219		return;
220
221	/*
222	 * QPs on flush list are error state.  This way, a "flush
223	 * error" WC will be immediately generated for each WR we post.
224	 */
225	p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
226	if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
227		ipoib_warn(priv, "failed to post drain wr\n");
228
229	list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
230}
231
232static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
233{
234	struct ipoib_cm_rx *p = ctx;
235	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
236	unsigned long flags;
237
238	if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
239		return;
240
241	spin_lock_irqsave(&priv->lock, flags);
242	list_move(&p->list, &priv->cm.rx_flush_list);
243	p->state = IPOIB_CM_RX_FLUSH;
244	ipoib_cm_start_rx_drain(priv);
245	spin_unlock_irqrestore(&priv->lock, flags);
246}
247
248static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
249					   struct ipoib_cm_rx *p)
250{
251	struct ipoib_dev_priv *priv = netdev_priv(dev);
252	struct ib_qp_init_attr attr = {
253		.event_handler = ipoib_cm_rx_event_handler,
254		.send_cq = priv->recv_cq, /* For drain WR */
255		.recv_cq = priv->recv_cq,
256		.srq = priv->cm.srq,
257		.cap.max_send_wr = 1, /* For drain WR */
258		.cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
259		.sq_sig_type = IB_SIGNAL_ALL_WR,
260		.qp_type = IB_QPT_RC,
261		.qp_context = p,
262	};
263
264	if (!ipoib_cm_has_srq(dev)) {
265		attr.cap.max_recv_wr  = ipoib_recvq_size;
266		attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
267	}
268
269	return ib_create_qp(priv->pd, &attr);
270}
271
272static int ipoib_cm_modify_rx_qp(struct net_device *dev,
273				 struct ib_cm_id *cm_id, struct ib_qp *qp,
274				 unsigned psn)
275{
276	struct ipoib_dev_priv *priv = netdev_priv(dev);
277	struct ib_qp_attr qp_attr;
278	int qp_attr_mask, ret;
279
280	qp_attr.qp_state = IB_QPS_INIT;
281	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
282	if (ret) {
283		ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
284		return ret;
285	}
286	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
287	if (ret) {
288		ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
289		return ret;
290	}
291	qp_attr.qp_state = IB_QPS_RTR;
292	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
293	if (ret) {
294		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
295		return ret;
296	}
297	qp_attr.rq_psn = psn;
298	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
299	if (ret) {
300		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
301		return ret;
302	}
303
304	/*
305	 * Current Mellanox HCA firmware won't generate completions
306	 * with error for drain WRs unless the QP has been moved to
307	 * RTS first. This work-around leaves a window where a QP has
308	 * moved to error asynchronously, but this will eventually get
309	 * fixed in firmware, so let's not error out if modify QP
310	 * fails.
311	 */
312	qp_attr.qp_state = IB_QPS_RTS;
313	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
314	if (ret) {
315		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
316		return 0;
317	}
318	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
319	if (ret) {
320		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
321		return 0;
322	}
323
324	return 0;
325}
326
327static void ipoib_cm_init_rx_wr(struct net_device *dev,
328				struct ib_recv_wr *wr,
329				struct ib_sge *sge)
330{
331	struct ipoib_dev_priv *priv = netdev_priv(dev);
332	int i;
333
334	for (i = 0; i < priv->cm.num_frags; ++i)
335		sge[i].lkey = priv->mr->lkey;
336
337	sge[0].length = IPOIB_CM_HEAD_SIZE;
338	for (i = 1; i < priv->cm.num_frags; ++i)
339		sge[i].length = PAGE_SIZE;
340
341	wr->next    = NULL;
342	wr->sg_list = sge;
343	wr->num_sge = priv->cm.num_frags;
344}
345
346static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
347				   struct ipoib_cm_rx *rx)
348{
349	struct ipoib_dev_priv *priv = netdev_priv(dev);
350	struct {
351		struct ib_recv_wr wr;
352		struct ib_sge sge[IPOIB_CM_RX_SG];
353	} *t;
354	int ret;
355	int i;
356
357	rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
358	if (!rx->rx_ring) {
359		printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
360		       priv->ca->name, ipoib_recvq_size);
361		return -ENOMEM;
362	}
363
364	t = kmalloc(sizeof *t, GFP_KERNEL);
365	if (!t) {
366		ret = -ENOMEM;
367		goto err_free;
368	}
369
370	ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
371
372	spin_lock_irq(&priv->lock);
373
374	if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
375		spin_unlock_irq(&priv->lock);
376		ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
377		ret = -EINVAL;
378		goto err_free;
379	} else
380		++priv->cm.nonsrq_conn_qp;
381
382	spin_unlock_irq(&priv->lock);
383
384	for (i = 0; i < ipoib_recvq_size; ++i) {
385		if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
386					   rx->rx_ring[i].mapping,
387					   GFP_KERNEL)) {
388			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
389				ret = -ENOMEM;
390				goto err_count;
391		}
392		ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
393		if (ret) {
394			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
395				   "failed for buf %d\n", i);
396			ret = -EIO;
397			goto err_count;
398		}
399	}
400
401	rx->recv_count = ipoib_recvq_size;
402
403	kfree(t);
404
405	return 0;
406
407err_count:
408	spin_lock_irq(&priv->lock);
409	--priv->cm.nonsrq_conn_qp;
410	spin_unlock_irq(&priv->lock);
411
412err_free:
413	kfree(t);
414	ipoib_cm_free_rx_ring(dev, rx->rx_ring);
415
416	return ret;
417}
418
419static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
420			     struct ib_qp *qp, struct ib_cm_req_event_param *req,
421			     unsigned psn)
422{
423	struct ipoib_dev_priv *priv = netdev_priv(dev);
424	struct ipoib_cm_data data = {};
425	struct ib_cm_rep_param rep = {};
426
427	data.qpn = cpu_to_be32(priv->qp->qp_num);
428	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
429
430	rep.private_data = &data;
431	rep.private_data_len = sizeof data;
432	rep.flow_control = 0;
433	rep.rnr_retry_count = req->rnr_retry_count;
434	rep.srq = ipoib_cm_has_srq(dev);
435	rep.qp_num = qp->qp_num;
436	rep.starting_psn = psn;
437	return ib_send_cm_rep(cm_id, &rep);
438}
439
440static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
441{
442	struct net_device *dev = cm_id->context;
443	struct ipoib_dev_priv *priv = netdev_priv(dev);
444	struct ipoib_cm_rx *p;
445	unsigned psn;
446	int ret;
447
448	ipoib_dbg(priv, "REQ arrived\n");
449	p = kzalloc(sizeof *p, GFP_KERNEL);
450	if (!p)
451		return -ENOMEM;
452	p->dev = dev;
453	p->id = cm_id;
454	cm_id->context = p;
455	p->state = IPOIB_CM_RX_LIVE;
456	p->jiffies = jiffies;
457	INIT_LIST_HEAD(&p->list);
458
459	p->qp = ipoib_cm_create_rx_qp(dev, p);
460	if (IS_ERR(p->qp)) {
461		ret = PTR_ERR(p->qp);
462		goto err_qp;
463	}
464
465	psn = prandom_u32() & 0xffffff;
466	ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
467	if (ret)
468		goto err_modify;
469
470	if (!ipoib_cm_has_srq(dev)) {
471		ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
472		if (ret)
473			goto err_modify;
474	}
475
476	spin_lock_irq(&priv->lock);
477	queue_delayed_work(ipoib_workqueue,
478			   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
479	/* Add this entry to passive ids list head, but do not re-add it
480	 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
481	p->jiffies = jiffies;
482	if (p->state == IPOIB_CM_RX_LIVE)
483		list_move(&p->list, &priv->cm.passive_ids);
484	spin_unlock_irq(&priv->lock);
485
486	ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
487	if (ret) {
488		ipoib_warn(priv, "failed to send REP: %d\n", ret);
489		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
490			ipoib_warn(priv, "unable to move qp to error state\n");
491	}
492	return 0;
493
494err_modify:
495	ib_destroy_qp(p->qp);
496err_qp:
497	kfree(p);
498	return ret;
499}
500
501static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
502			       struct ib_cm_event *event)
503{
504	struct ipoib_cm_rx *p;
505	struct ipoib_dev_priv *priv;
506
507	switch (event->event) {
508	case IB_CM_REQ_RECEIVED:
509		return ipoib_cm_req_handler(cm_id, event);
510	case IB_CM_DREQ_RECEIVED:
511		p = cm_id->context;
512		ib_send_cm_drep(cm_id, NULL, 0);
513		/* Fall through */
514	case IB_CM_REJ_RECEIVED:
515		p = cm_id->context;
516		priv = netdev_priv(p->dev);
517		if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
518			ipoib_warn(priv, "unable to move qp to error state\n");
519		/* Fall through */
520	default:
521		return 0;
522	}
523}
524/* Adjust length of skb with fragments to match received data */
525static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
526			  unsigned int length, struct sk_buff *toskb)
527{
528	int i, num_frags;
529	unsigned int size;
530
531	/* put header into skb */
532	size = min(length, hdr_space);
533	skb->tail += size;
534	skb->len += size;
535	length -= size;
536
537	num_frags = skb_shinfo(skb)->nr_frags;
538	for (i = 0; i < num_frags; i++) {
539		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
540
541		if (length == 0) {
542			/* don't need this page */
543			skb_fill_page_desc(toskb, i, skb_frag_page(frag),
544					   0, PAGE_SIZE);
545			--skb_shinfo(skb)->nr_frags;
546		} else {
547			size = min(length, (unsigned) PAGE_SIZE);
548
549			skb_frag_size_set(frag, size);
550			skb->data_len += size;
551			skb->truesize += size;
552			skb->len += size;
553			length -= size;
554		}
555	}
556}
557
558void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
559{
560	struct ipoib_dev_priv *priv = netdev_priv(dev);
561	struct ipoib_cm_rx_buf *rx_ring;
562	unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
563	struct sk_buff *skb, *newskb;
564	struct ipoib_cm_rx *p;
565	unsigned long flags;
566	u64 mapping[IPOIB_CM_RX_SG];
567	int frags;
568	int has_srq;
569	struct sk_buff *small_skb;
570
571	ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
572		       wr_id, wc->status);
573
574	if (unlikely(wr_id >= ipoib_recvq_size)) {
575		if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
576			spin_lock_irqsave(&priv->lock, flags);
577			list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
578			ipoib_cm_start_rx_drain(priv);
579			queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
580			spin_unlock_irqrestore(&priv->lock, flags);
581		} else
582			ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
583				   wr_id, ipoib_recvq_size);
584		return;
585	}
586
587	p = wc->qp->qp_context;
588
589	has_srq = ipoib_cm_has_srq(dev);
590	rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
591
592	skb = rx_ring[wr_id].skb;
593
594	if (unlikely(wc->status != IB_WC_SUCCESS)) {
595		ipoib_dbg(priv, "cm recv error "
596			   "(status=%d, wrid=%d vend_err %x)\n",
597			   wc->status, wr_id, wc->vendor_err);
598		++dev->stats.rx_dropped;
599		if (has_srq)
600			goto repost;
601		else {
602			if (!--p->recv_count) {
603				spin_lock_irqsave(&priv->lock, flags);
604				list_move(&p->list, &priv->cm.rx_reap_list);
605				spin_unlock_irqrestore(&priv->lock, flags);
606				queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
607			}
608			return;
609		}
610	}
611
612	if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
613		if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
614			spin_lock_irqsave(&priv->lock, flags);
615			p->jiffies = jiffies;
616			/* Move this entry to list head, but do not re-add it
617			 * if it has been moved out of list. */
618			if (p->state == IPOIB_CM_RX_LIVE)
619				list_move(&p->list, &priv->cm.passive_ids);
620			spin_unlock_irqrestore(&priv->lock, flags);
621		}
622	}
623
624	if (wc->byte_len < IPOIB_CM_COPYBREAK) {
625		int dlen = wc->byte_len;
626
627		small_skb = dev_alloc_skb(dlen + 12);
628		if (small_skb) {
629			skb_reserve(small_skb, 12);
630			ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
631						   dlen, DMA_FROM_DEVICE);
632			skb_copy_from_linear_data(skb, small_skb->data, dlen);
633			ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
634						      dlen, DMA_FROM_DEVICE);
635			skb_put(small_skb, dlen);
636			skb = small_skb;
637			goto copied;
638		}
639	}
640
641	frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
642					      (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
643
644	newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
645				       mapping, GFP_ATOMIC);
646	if (unlikely(!newskb)) {
647		/*
648		 * If we can't allocate a new RX buffer, dump
649		 * this packet and reuse the old buffer.
650		 */
651		ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
652		++dev->stats.rx_dropped;
653		goto repost;
654	}
655
656	ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
657	memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
658
659	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
660		       wc->byte_len, wc->slid);
661
662	skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
663
664copied:
665	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
666	skb_reset_mac_header(skb);
667	skb_pull(skb, IPOIB_ENCAP_LEN);
668
669	++dev->stats.rx_packets;
670	dev->stats.rx_bytes += skb->len;
671
672	skb->dev = dev;
673	/* XXX get correct PACKET_ type here */
674	skb->pkt_type = PACKET_HOST;
675	netif_receive_skb(skb);
676
677repost:
678	if (has_srq) {
679		if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
680			ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
681				   "for buf %d\n", wr_id);
682	} else {
683		if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
684							  &priv->cm.rx_wr,
685							  priv->cm.rx_sge,
686							  wr_id))) {
687			--p->recv_count;
688			ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
689				   "for buf %d\n", wr_id);
690		}
691	}
692}
693
694static inline int post_send(struct ipoib_dev_priv *priv,
695			    struct ipoib_cm_tx *tx,
696			    unsigned int wr_id,
697			    u64 addr, int len)
698{
699	struct ib_send_wr *bad_wr;
700
701	priv->tx_sge[0].addr          = addr;
702	priv->tx_sge[0].length        = len;
703
704	priv->tx_wr.num_sge	= 1;
705	priv->tx_wr.wr_id	= wr_id | IPOIB_OP_CM;
706
707	return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
708}
709
710void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
711{
712	struct ipoib_dev_priv *priv = netdev_priv(dev);
713	struct ipoib_cm_tx_buf *tx_req;
714	u64 addr;
715	int rc;
716
717	if (unlikely(skb->len > tx->mtu)) {
718		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
719			   skb->len, tx->mtu);
720		++dev->stats.tx_dropped;
721		++dev->stats.tx_errors;
722		ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
723		return;
724	}
725
726	ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
727		       tx->tx_head, skb->len, tx->qp->qp_num);
728
729	/*
730	 * We put the skb into the tx_ring _before_ we call post_send()
731	 * because it's entirely possible that the completion handler will
732	 * run before we execute anything after the post_send().  That
733	 * means we have to make sure everything is properly recorded and
734	 * our state is consistent before we call post_send().
735	 */
736	tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
737	tx_req->skb = skb;
738	addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
739	if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
740		++dev->stats.tx_errors;
741		dev_kfree_skb_any(skb);
742		return;
743	}
744
745	tx_req->mapping = addr;
746
747	skb_orphan(skb);
748	skb_dst_drop(skb);
749
750	rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
751		       addr, skb->len);
752	if (unlikely(rc)) {
753		ipoib_warn(priv, "post_send failed, error %d\n", rc);
754		++dev->stats.tx_errors;
755		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
756		dev_kfree_skb_any(skb);
757	} else {
758		dev->trans_start = jiffies;
759		++tx->tx_head;
760
761		if (++priv->tx_outstanding == ipoib_sendq_size) {
762			ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
763				  tx->qp->qp_num);
764			netif_stop_queue(dev);
765			rc = ib_req_notify_cq(priv->send_cq,
766				IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
767			if (rc < 0)
768				ipoib_warn(priv, "request notify on send CQ failed\n");
769			else if (rc)
770				ipoib_send_comp_handler(priv->send_cq, dev);
771		}
772	}
773}
774
775void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
776{
777	struct ipoib_dev_priv *priv = netdev_priv(dev);
778	struct ipoib_cm_tx *tx = wc->qp->qp_context;
779	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
780	struct ipoib_cm_tx_buf *tx_req;
781	unsigned long flags;
782
783	ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
784		       wr_id, wc->status);
785
786	if (unlikely(wr_id >= ipoib_sendq_size)) {
787		ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
788			   wr_id, ipoib_sendq_size);
789		return;
790	}
791
792	tx_req = &tx->tx_ring[wr_id];
793
794	ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
795
796	/* FIXME: is this right? Shouldn't we only increment on success? */
797	++dev->stats.tx_packets;
798	dev->stats.tx_bytes += tx_req->skb->len;
799
800	dev_kfree_skb_any(tx_req->skb);
801
802	netif_tx_lock(dev);
803
804	++tx->tx_tail;
805	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
806	    netif_queue_stopped(dev) &&
807	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
808		netif_wake_queue(dev);
809
810	if (wc->status != IB_WC_SUCCESS &&
811	    wc->status != IB_WC_WR_FLUSH_ERR) {
812		struct ipoib_neigh *neigh;
813
814		ipoib_dbg(priv, "failed cm send event "
815			   "(status=%d, wrid=%d vend_err %x)\n",
816			   wc->status, wr_id, wc->vendor_err);
817
818		spin_lock_irqsave(&priv->lock, flags);
819		neigh = tx->neigh;
820
821		if (neigh) {
822			neigh->cm = NULL;
823			ipoib_neigh_free(neigh);
824
825			tx->neigh = NULL;
826		}
827
828		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
829			list_move(&tx->list, &priv->cm.reap_list);
830			queue_work(ipoib_workqueue, &priv->cm.reap_task);
831		}
832
833		clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
834
835		spin_unlock_irqrestore(&priv->lock, flags);
836	}
837
838	netif_tx_unlock(dev);
839}
840
841int ipoib_cm_dev_open(struct net_device *dev)
842{
843	struct ipoib_dev_priv *priv = netdev_priv(dev);
844	int ret;
845
846	if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
847		return 0;
848
849	priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
850	if (IS_ERR(priv->cm.id)) {
851		printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
852		ret = PTR_ERR(priv->cm.id);
853		goto err_cm;
854	}
855
856	ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
857			   0, NULL);
858	if (ret) {
859		printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
860		       IPOIB_CM_IETF_ID | priv->qp->qp_num);
861		goto err_listen;
862	}
863
864	return 0;
865
866err_listen:
867	ib_destroy_cm_id(priv->cm.id);
868err_cm:
869	priv->cm.id = NULL;
870	return ret;
871}
872
873static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
874{
875	struct ipoib_dev_priv *priv = netdev_priv(dev);
876	struct ipoib_cm_rx *rx, *n;
877	LIST_HEAD(list);
878
879	spin_lock_irq(&priv->lock);
880	list_splice_init(&priv->cm.rx_reap_list, &list);
881	spin_unlock_irq(&priv->lock);
882
883	list_for_each_entry_safe(rx, n, &list, list) {
884		ib_destroy_cm_id(rx->id);
885		ib_destroy_qp(rx->qp);
886		if (!ipoib_cm_has_srq(dev)) {
887			ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
888			spin_lock_irq(&priv->lock);
889			--priv->cm.nonsrq_conn_qp;
890			spin_unlock_irq(&priv->lock);
891		}
892		kfree(rx);
893	}
894}
895
896void ipoib_cm_dev_stop(struct net_device *dev)
897{
898	struct ipoib_dev_priv *priv = netdev_priv(dev);
899	struct ipoib_cm_rx *p;
900	unsigned long begin;
901	int ret;
902
903	if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
904		return;
905
906	ib_destroy_cm_id(priv->cm.id);
907	priv->cm.id = NULL;
908
909	spin_lock_irq(&priv->lock);
910	while (!list_empty(&priv->cm.passive_ids)) {
911		p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
912		list_move(&p->list, &priv->cm.rx_error_list);
913		p->state = IPOIB_CM_RX_ERROR;
914		spin_unlock_irq(&priv->lock);
915		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
916		if (ret)
917			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
918		spin_lock_irq(&priv->lock);
919	}
920
921	/* Wait for all RX to be drained */
922	begin = jiffies;
923
924	while (!list_empty(&priv->cm.rx_error_list) ||
925	       !list_empty(&priv->cm.rx_flush_list) ||
926	       !list_empty(&priv->cm.rx_drain_list)) {
927		if (time_after(jiffies, begin + 5 * HZ)) {
928			ipoib_warn(priv, "RX drain timing out\n");
929
930			/*
931			 * assume the HW is wedged and just free up everything.
932			 */
933			list_splice_init(&priv->cm.rx_flush_list,
934					 &priv->cm.rx_reap_list);
935			list_splice_init(&priv->cm.rx_error_list,
936					 &priv->cm.rx_reap_list);
937			list_splice_init(&priv->cm.rx_drain_list,
938					 &priv->cm.rx_reap_list);
939			break;
940		}
941		spin_unlock_irq(&priv->lock);
942		msleep(1);
943		ipoib_drain_cq(dev);
944		spin_lock_irq(&priv->lock);
945	}
946
947	spin_unlock_irq(&priv->lock);
948
949	ipoib_cm_free_rx_reap_list(dev);
950
951	cancel_delayed_work(&priv->cm.stale_task);
952}
953
954static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
955{
956	struct ipoib_cm_tx *p = cm_id->context;
957	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
958	struct ipoib_cm_data *data = event->private_data;
959	struct sk_buff_head skqueue;
960	struct ib_qp_attr qp_attr;
961	int qp_attr_mask, ret;
962	struct sk_buff *skb;
963
964	p->mtu = be32_to_cpu(data->mtu);
965
966	if (p->mtu <= IPOIB_ENCAP_LEN) {
967		ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
968			   p->mtu, IPOIB_ENCAP_LEN);
969		return -EINVAL;
970	}
971
972	qp_attr.qp_state = IB_QPS_RTR;
973	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
974	if (ret) {
975		ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
976		return ret;
977	}
978
979	qp_attr.rq_psn = 0 /* FIXME */;
980	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
981	if (ret) {
982		ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
983		return ret;
984	}
985
986	qp_attr.qp_state = IB_QPS_RTS;
987	ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
988	if (ret) {
989		ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
990		return ret;
991	}
992	ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
993	if (ret) {
994		ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
995		return ret;
996	}
997
998	skb_queue_head_init(&skqueue);
999
1000	spin_lock_irq(&priv->lock);
1001	set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1002	if (p->neigh)
1003		while ((skb = __skb_dequeue(&p->neigh->queue)))
1004			__skb_queue_tail(&skqueue, skb);
1005	spin_unlock_irq(&priv->lock);
1006
1007	while ((skb = __skb_dequeue(&skqueue))) {
1008		skb->dev = p->dev;
1009		if (dev_queue_xmit(skb))
1010			ipoib_warn(priv, "dev_queue_xmit failed "
1011				   "to requeue packet\n");
1012	}
1013
1014	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1015	if (ret) {
1016		ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1017		return ret;
1018	}
1019	return 0;
1020}
1021
1022static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1023{
1024	struct ipoib_dev_priv *priv = netdev_priv(dev);
1025	struct ib_qp_init_attr attr = {
1026		.send_cq		= priv->recv_cq,
1027		.recv_cq		= priv->recv_cq,
1028		.srq			= priv->cm.srq,
1029		.cap.max_send_wr	= ipoib_sendq_size,
1030		.cap.max_send_sge	= 1,
1031		.sq_sig_type		= IB_SIGNAL_ALL_WR,
1032		.qp_type		= IB_QPT_RC,
1033		.qp_context		= tx,
1034		.create_flags		= IB_QP_CREATE_USE_GFP_NOIO
1035	};
1036
1037	struct ib_qp *tx_qp;
1038
1039	tx_qp = ib_create_qp(priv->pd, &attr);
1040	if (PTR_ERR(tx_qp) == -EINVAL) {
1041		ipoib_warn(priv, "can't use GFP_NOIO for QPs on device %s, using GFP_KERNEL\n",
1042			   priv->ca->name);
1043		attr.create_flags &= ~IB_QP_CREATE_USE_GFP_NOIO;
1044		tx_qp = ib_create_qp(priv->pd, &attr);
1045	}
1046	return tx_qp;
1047}
1048
1049static int ipoib_cm_send_req(struct net_device *dev,
1050			     struct ib_cm_id *id, struct ib_qp *qp,
1051			     u32 qpn,
1052			     struct ib_sa_path_rec *pathrec)
1053{
1054	struct ipoib_dev_priv *priv = netdev_priv(dev);
1055	struct ipoib_cm_data data = {};
1056	struct ib_cm_req_param req = {};
1057
1058	data.qpn = cpu_to_be32(priv->qp->qp_num);
1059	data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1060
1061	req.primary_path		= pathrec;
1062	req.alternate_path		= NULL;
1063	req.service_id			= cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1064	req.qp_num			= qp->qp_num;
1065	req.qp_type			= qp->qp_type;
1066	req.private_data		= &data;
1067	req.private_data_len		= sizeof data;
1068	req.flow_control		= 0;
1069
1070	req.starting_psn		= 0; /* FIXME */
1071
1072	/*
1073	 * Pick some arbitrary defaults here; we could make these
1074	 * module parameters if anyone cared about setting them.
1075	 */
1076	req.responder_resources		= 4;
1077	req.remote_cm_response_timeout	= 20;
1078	req.local_cm_response_timeout	= 20;
1079	req.retry_count			= 0; /* RFC draft warns against retries */
1080	req.rnr_retry_count		= 0; /* RFC draft warns against retries */
1081	req.max_cm_retries		= 15;
1082	req.srq				= ipoib_cm_has_srq(dev);
1083	return ib_send_cm_req(id, &req);
1084}
1085
1086static int ipoib_cm_modify_tx_init(struct net_device *dev,
1087				  struct ib_cm_id *cm_id, struct ib_qp *qp)
1088{
1089	struct ipoib_dev_priv *priv = netdev_priv(dev);
1090	struct ib_qp_attr qp_attr;
1091	int qp_attr_mask, ret;
1092	ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1093	if (ret) {
1094		ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1095		return ret;
1096	}
1097
1098	qp_attr.qp_state = IB_QPS_INIT;
1099	qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1100	qp_attr.port_num = priv->port;
1101	qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1102
1103	ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1104	if (ret) {
1105		ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1106		return ret;
1107	}
1108	return 0;
1109}
1110
1111static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1112			    struct ib_sa_path_rec *pathrec)
1113{
1114	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1115	int ret;
1116
1117	p->tx_ring = __vmalloc(ipoib_sendq_size * sizeof *p->tx_ring,
1118			       GFP_NOIO, PAGE_KERNEL);
1119	if (!p->tx_ring) {
1120		ipoib_warn(priv, "failed to allocate tx ring\n");
1121		ret = -ENOMEM;
1122		goto err_tx;
1123	}
1124	memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1125
1126	p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1127	if (IS_ERR(p->qp)) {
1128		ret = PTR_ERR(p->qp);
1129		ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1130		goto err_qp;
1131	}
1132
1133	p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1134	if (IS_ERR(p->id)) {
1135		ret = PTR_ERR(p->id);
1136		ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1137		goto err_id;
1138	}
1139
1140	ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1141	if (ret) {
1142		ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1143		goto err_modify;
1144	}
1145
1146	ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1147	if (ret) {
1148		ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1149		goto err_send_cm;
1150	}
1151
1152	ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1153		  p->qp->qp_num, pathrec->dgid.raw, qpn);
1154
1155	return 0;
1156
1157err_send_cm:
1158err_modify:
1159	ib_destroy_cm_id(p->id);
1160err_id:
1161	p->id = NULL;
1162	ib_destroy_qp(p->qp);
1163err_qp:
1164	p->qp = NULL;
1165	vfree(p->tx_ring);
1166err_tx:
1167	return ret;
1168}
1169
1170static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1171{
1172	struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1173	struct ipoib_cm_tx_buf *tx_req;
1174	unsigned long begin;
1175
1176	ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1177		  p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1178
1179	if (p->id)
1180		ib_destroy_cm_id(p->id);
1181
1182	if (p->tx_ring) {
1183		/* Wait for all sends to complete */
1184		begin = jiffies;
1185		while ((int) p->tx_tail - (int) p->tx_head < 0) {
1186			if (time_after(jiffies, begin + 5 * HZ)) {
1187				ipoib_warn(priv, "timing out; %d sends not completed\n",
1188					   p->tx_head - p->tx_tail);
1189				goto timeout;
1190			}
1191
1192			msleep(1);
1193		}
1194	}
1195
1196timeout:
1197
1198	while ((int) p->tx_tail - (int) p->tx_head < 0) {
1199		tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1200		ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1201				    DMA_TO_DEVICE);
1202		dev_kfree_skb_any(tx_req->skb);
1203		++p->tx_tail;
1204		netif_tx_lock_bh(p->dev);
1205		if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1206		    netif_queue_stopped(p->dev) &&
1207		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1208			netif_wake_queue(p->dev);
1209		netif_tx_unlock_bh(p->dev);
1210	}
1211
1212	if (p->qp)
1213		ib_destroy_qp(p->qp);
1214
1215	vfree(p->tx_ring);
1216	kfree(p);
1217}
1218
1219static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1220			       struct ib_cm_event *event)
1221{
1222	struct ipoib_cm_tx *tx = cm_id->context;
1223	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1224	struct net_device *dev = priv->dev;
1225	struct ipoib_neigh *neigh;
1226	unsigned long flags;
1227	int ret;
1228
1229	switch (event->event) {
1230	case IB_CM_DREQ_RECEIVED:
1231		ipoib_dbg(priv, "DREQ received.\n");
1232		ib_send_cm_drep(cm_id, NULL, 0);
1233		break;
1234	case IB_CM_REP_RECEIVED:
1235		ipoib_dbg(priv, "REP received.\n");
1236		ret = ipoib_cm_rep_handler(cm_id, event);
1237		if (ret)
1238			ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1239				       NULL, 0, NULL, 0);
1240		break;
1241	case IB_CM_REQ_ERROR:
1242	case IB_CM_REJ_RECEIVED:
1243	case IB_CM_TIMEWAIT_EXIT:
1244		ipoib_dbg(priv, "CM error %d.\n", event->event);
1245		netif_tx_lock_bh(dev);
1246		spin_lock_irqsave(&priv->lock, flags);
1247		neigh = tx->neigh;
1248
1249		if (neigh) {
1250			neigh->cm = NULL;
1251			ipoib_neigh_free(neigh);
1252
1253			tx->neigh = NULL;
1254		}
1255
1256		if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1257			list_move(&tx->list, &priv->cm.reap_list);
1258			queue_work(ipoib_workqueue, &priv->cm.reap_task);
1259		}
1260
1261		spin_unlock_irqrestore(&priv->lock, flags);
1262		netif_tx_unlock_bh(dev);
1263		break;
1264	default:
1265		break;
1266	}
1267
1268	return 0;
1269}
1270
1271struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1272				       struct ipoib_neigh *neigh)
1273{
1274	struct ipoib_dev_priv *priv = netdev_priv(dev);
1275	struct ipoib_cm_tx *tx;
1276
1277	tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1278	if (!tx)
1279		return NULL;
1280
1281	neigh->cm = tx;
1282	tx->neigh = neigh;
1283	tx->path = path;
1284	tx->dev = dev;
1285	list_add(&tx->list, &priv->cm.start_list);
1286	set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1287	queue_work(ipoib_workqueue, &priv->cm.start_task);
1288	return tx;
1289}
1290
1291void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1292{
1293	struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1294	unsigned long flags;
1295	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1296		spin_lock_irqsave(&priv->lock, flags);
1297		list_move(&tx->list, &priv->cm.reap_list);
1298		queue_work(ipoib_workqueue, &priv->cm.reap_task);
1299		ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1300			  tx->neigh->daddr + 4);
1301		tx->neigh = NULL;
1302		spin_unlock_irqrestore(&priv->lock, flags);
1303	}
1304}
1305
1306static void ipoib_cm_tx_start(struct work_struct *work)
1307{
1308	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1309						   cm.start_task);
1310	struct net_device *dev = priv->dev;
1311	struct ipoib_neigh *neigh;
1312	struct ipoib_cm_tx *p;
1313	unsigned long flags;
1314	int ret;
1315
1316	struct ib_sa_path_rec pathrec;
1317	u32 qpn;
1318
1319	netif_tx_lock_bh(dev);
1320	spin_lock_irqsave(&priv->lock, flags);
1321
1322	while (!list_empty(&priv->cm.start_list)) {
1323		p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1324		list_del_init(&p->list);
1325		neigh = p->neigh;
1326		qpn = IPOIB_QPN(neigh->daddr);
1327		memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1328
1329		spin_unlock_irqrestore(&priv->lock, flags);
1330		netif_tx_unlock_bh(dev);
1331
1332		ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1333
1334		netif_tx_lock_bh(dev);
1335		spin_lock_irqsave(&priv->lock, flags);
1336
1337		if (ret) {
1338			neigh = p->neigh;
1339			if (neigh) {
1340				neigh->cm = NULL;
1341				ipoib_neigh_free(neigh);
1342			}
1343			list_del(&p->list);
1344			kfree(p);
1345		}
1346	}
1347
1348	spin_unlock_irqrestore(&priv->lock, flags);
1349	netif_tx_unlock_bh(dev);
1350}
1351
1352static void ipoib_cm_tx_reap(struct work_struct *work)
1353{
1354	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1355						   cm.reap_task);
1356	struct net_device *dev = priv->dev;
1357	struct ipoib_cm_tx *p;
1358	unsigned long flags;
1359
1360	netif_tx_lock_bh(dev);
1361	spin_lock_irqsave(&priv->lock, flags);
1362
1363	while (!list_empty(&priv->cm.reap_list)) {
1364		p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1365		list_del(&p->list);
1366		spin_unlock_irqrestore(&priv->lock, flags);
1367		netif_tx_unlock_bh(dev);
1368		ipoib_cm_tx_destroy(p);
1369		netif_tx_lock_bh(dev);
1370		spin_lock_irqsave(&priv->lock, flags);
1371	}
1372
1373	spin_unlock_irqrestore(&priv->lock, flags);
1374	netif_tx_unlock_bh(dev);
1375}
1376
1377static void ipoib_cm_skb_reap(struct work_struct *work)
1378{
1379	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1380						   cm.skb_task);
1381	struct net_device *dev = priv->dev;
1382	struct sk_buff *skb;
1383	unsigned long flags;
1384	unsigned mtu = priv->mcast_mtu;
1385
1386	netif_tx_lock_bh(dev);
1387	spin_lock_irqsave(&priv->lock, flags);
1388
1389	while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1390		spin_unlock_irqrestore(&priv->lock, flags);
1391		netif_tx_unlock_bh(dev);
1392
1393		if (skb->protocol == htons(ETH_P_IP))
1394			icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1395#if IS_ENABLED(CONFIG_IPV6)
1396		else if (skb->protocol == htons(ETH_P_IPV6))
1397			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1398#endif
1399		dev_kfree_skb_any(skb);
1400
1401		netif_tx_lock_bh(dev);
1402		spin_lock_irqsave(&priv->lock, flags);
1403	}
1404
1405	spin_unlock_irqrestore(&priv->lock, flags);
1406	netif_tx_unlock_bh(dev);
1407}
1408
1409void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1410			   unsigned int mtu)
1411{
1412	struct ipoib_dev_priv *priv = netdev_priv(dev);
1413	int e = skb_queue_empty(&priv->cm.skb_queue);
1414
1415	if (skb_dst(skb))
1416		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
1417
1418	skb_queue_tail(&priv->cm.skb_queue, skb);
1419	if (e)
1420		queue_work(ipoib_workqueue, &priv->cm.skb_task);
1421}
1422
1423static void ipoib_cm_rx_reap(struct work_struct *work)
1424{
1425	ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1426						cm.rx_reap_task)->dev);
1427}
1428
1429static void ipoib_cm_stale_task(struct work_struct *work)
1430{
1431	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1432						   cm.stale_task.work);
1433	struct ipoib_cm_rx *p;
1434	int ret;
1435
1436	spin_lock_irq(&priv->lock);
1437	while (!list_empty(&priv->cm.passive_ids)) {
1438		/* List is sorted by LRU, start from tail,
1439		 * stop when we see a recently used entry */
1440		p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1441		if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1442			break;
1443		list_move(&p->list, &priv->cm.rx_error_list);
1444		p->state = IPOIB_CM_RX_ERROR;
1445		spin_unlock_irq(&priv->lock);
1446		ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1447		if (ret)
1448			ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1449		spin_lock_irq(&priv->lock);
1450	}
1451
1452	if (!list_empty(&priv->cm.passive_ids))
1453		queue_delayed_work(ipoib_workqueue,
1454				   &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1455	spin_unlock_irq(&priv->lock);
1456}
1457
1458
1459static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1460			 char *buf)
1461{
1462	struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1463
1464	if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1465		return sprintf(buf, "connected\n");
1466	else
1467		return sprintf(buf, "datagram\n");
1468}
1469
1470static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1471			const char *buf, size_t count)
1472{
1473	struct net_device *dev = to_net_dev(d);
1474	int ret;
1475
1476	if (!rtnl_trylock())
1477		return restart_syscall();
1478
1479	ret = ipoib_set_mode(dev, buf);
1480
1481	rtnl_unlock();
1482
1483	if (!ret)
1484		return count;
1485
1486	return ret;
1487}
1488
1489static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1490
1491int ipoib_cm_add_mode_attr(struct net_device *dev)
1492{
1493	return device_create_file(&dev->dev, &dev_attr_mode);
1494}
1495
1496static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1497{
1498	struct ipoib_dev_priv *priv = netdev_priv(dev);
1499	struct ib_srq_init_attr srq_init_attr = {
1500		.srq_type = IB_SRQT_BASIC,
1501		.attr = {
1502			.max_wr  = ipoib_recvq_size,
1503			.max_sge = max_sge
1504		}
1505	};
1506
1507	priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1508	if (IS_ERR(priv->cm.srq)) {
1509		if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1510			printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1511			       priv->ca->name, PTR_ERR(priv->cm.srq));
1512		priv->cm.srq = NULL;
1513		return;
1514	}
1515
1516	priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1517	if (!priv->cm.srq_ring) {
1518		printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1519		       priv->ca->name, ipoib_recvq_size);
1520		ib_destroy_srq(priv->cm.srq);
1521		priv->cm.srq = NULL;
1522		return;
1523	}
1524
1525}
1526
1527int ipoib_cm_dev_init(struct net_device *dev)
1528{
1529	struct ipoib_dev_priv *priv = netdev_priv(dev);
1530	int i, ret;
1531	struct ib_device_attr attr;
1532
1533	INIT_LIST_HEAD(&priv->cm.passive_ids);
1534	INIT_LIST_HEAD(&priv->cm.reap_list);
1535	INIT_LIST_HEAD(&priv->cm.start_list);
1536	INIT_LIST_HEAD(&priv->cm.rx_error_list);
1537	INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1538	INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1539	INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1540	INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1541	INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1542	INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1543	INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1544	INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1545
1546	skb_queue_head_init(&priv->cm.skb_queue);
1547
1548	ret = ib_query_device(priv->ca, &attr);
1549	if (ret) {
1550		printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1551		return ret;
1552	}
1553
1554	ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1555
1556	attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1557	ipoib_cm_create_srq(dev, attr.max_srq_sge);
1558	if (ipoib_cm_has_srq(dev)) {
1559		priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1560		priv->cm.num_frags  = attr.max_srq_sge;
1561		ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1562			  priv->cm.max_cm_mtu, priv->cm.num_frags);
1563	} else {
1564		priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1565		priv->cm.num_frags  = IPOIB_CM_RX_SG;
1566	}
1567
1568	ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1569
1570	if (ipoib_cm_has_srq(dev)) {
1571		for (i = 0; i < ipoib_recvq_size; ++i) {
1572			if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1573						   priv->cm.num_frags - 1,
1574						   priv->cm.srq_ring[i].mapping,
1575						   GFP_KERNEL)) {
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