Merge branch 'for-linus' of master.kernel.org:/pub/scm/linux/kernel/git/roland/infiniband
[sfrench/cifs-2.6.git] / drivers / infiniband / ulp / ipoib / ipoib_cm.c
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  * $Id$
33  */
34
35 #include <rdma/ib_cm.h>
36 #include <rdma/ib_cache.h>
37 #include <net/dst.h>
38 #include <net/icmp.h>
39 #include <linux/icmpv6.h>
40 #include <linux/delay.h>
41
42 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
43 static int data_debug_level;
44
45 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
46 MODULE_PARM_DESC(cm_data_debug_level,
47                  "Enable data path debug tracing for connected mode if > 0");
48 #endif
49
50 #include "ipoib.h"
51
52 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
53
54 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
55 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
56 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
57 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
58
59 struct ipoib_cm_id {
60         struct ib_cm_id *id;
61         int flags;
62         u32 remote_qpn;
63         u32 remote_mtu;
64 };
65
66 static struct ib_qp_attr ipoib_cm_err_attr = {
67         .qp_state = IB_QPS_ERR
68 };
69
70 #define IPOIB_CM_RX_DRAIN_WRID 0x7fffffff
71
72 static struct ib_recv_wr ipoib_cm_rx_drain_wr = {
73         .wr_id = IPOIB_CM_RX_DRAIN_WRID
74 };
75
76 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
77                                struct ib_cm_event *event);
78
79 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
80                                   u64 mapping[IPOIB_CM_RX_SG])
81 {
82         int i;
83
84         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
85
86         for (i = 0; i < frags; ++i)
87                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
88 }
89
90 static int ipoib_cm_post_receive(struct net_device *dev, int id)
91 {
92         struct ipoib_dev_priv *priv = netdev_priv(dev);
93         struct ib_recv_wr *bad_wr;
94         int i, ret;
95
96         priv->cm.rx_wr.wr_id = id | IPOIB_CM_OP_SRQ;
97
98         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
99                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
100
101         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
102         if (unlikely(ret)) {
103                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
104                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
105                                       priv->cm.srq_ring[id].mapping);
106                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
107                 priv->cm.srq_ring[id].skb = NULL;
108         }
109
110         return ret;
111 }
112
113 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, int id, int frags,
114                                              u64 mapping[IPOIB_CM_RX_SG])
115 {
116         struct ipoib_dev_priv *priv = netdev_priv(dev);
117         struct sk_buff *skb;
118         int i;
119
120         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
121         if (unlikely(!skb))
122                 return NULL;
123
124         /*
125          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
126          * IP header to a multiple of 16.
127          */
128         skb_reserve(skb, 12);
129
130         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
131                                        DMA_FROM_DEVICE);
132         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
133                 dev_kfree_skb_any(skb);
134                 return NULL;
135         }
136
137         for (i = 0; i < frags; i++) {
138                 struct page *page = alloc_page(GFP_ATOMIC);
139
140                 if (!page)
141                         goto partial_error;
142                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
143
144                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
145                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
146                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
147                         goto partial_error;
148         }
149
150         priv->cm.srq_ring[id].skb = skb;
151         return skb;
152
153 partial_error:
154
155         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
156
157         for (; i >= 0; --i)
158                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
159
160         dev_kfree_skb_any(skb);
161         return NULL;
162 }
163
164 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv* priv)
165 {
166         struct ib_recv_wr *bad_wr;
167
168         /* rx_drain_qp send queue depth is 1, so
169          * make sure we have at most 1 outstanding WR. */
170         if (list_empty(&priv->cm.rx_flush_list) ||
171             !list_empty(&priv->cm.rx_drain_list))
172                 return;
173
174         if (ib_post_recv(priv->cm.rx_drain_qp, &ipoib_cm_rx_drain_wr, &bad_wr))
175                 ipoib_warn(priv, "failed to post rx_drain wr\n");
176
177         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
178 }
179
180 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
181 {
182         struct ipoib_cm_rx *p = ctx;
183         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
184         unsigned long flags;
185
186         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
187                 return;
188
189         spin_lock_irqsave(&priv->lock, flags);
190         list_move(&p->list, &priv->cm.rx_flush_list);
191         p->state = IPOIB_CM_RX_FLUSH;
192         ipoib_cm_start_rx_drain(priv);
193         spin_unlock_irqrestore(&priv->lock, flags);
194 }
195
196 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
197                                            struct ipoib_cm_rx *p)
198 {
199         struct ipoib_dev_priv *priv = netdev_priv(dev);
200         struct ib_qp_init_attr attr = {
201                 .event_handler = ipoib_cm_rx_event_handler,
202                 .send_cq = priv->cq, /* does not matter, we never send anything */
203                 .recv_cq = priv->cq,
204                 .srq = priv->cm.srq,
205                 .cap.max_send_wr = 1, /* FIXME: 0 Seems not to work */
206                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
207                 .sq_sig_type = IB_SIGNAL_ALL_WR,
208                 .qp_type = IB_QPT_RC,
209                 .qp_context = p,
210         };
211         return ib_create_qp(priv->pd, &attr);
212 }
213
214 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
215                                   struct ib_cm_id *cm_id, struct ib_qp *qp,
216                                   unsigned psn)
217 {
218         struct ipoib_dev_priv *priv = netdev_priv(dev);
219         struct ib_qp_attr qp_attr;
220         int qp_attr_mask, ret;
221
222         qp_attr.qp_state = IB_QPS_INIT;
223         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
224         if (ret) {
225                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
226                 return ret;
227         }
228         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
229         if (ret) {
230                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
231                 return ret;
232         }
233         qp_attr.qp_state = IB_QPS_RTR;
234         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
235         if (ret) {
236                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
237                 return ret;
238         }
239         qp_attr.rq_psn = psn;
240         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
241         if (ret) {
242                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
243                 return ret;
244         }
245         return 0;
246 }
247
248 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
249                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
250                              unsigned psn)
251 {
252         struct ipoib_dev_priv *priv = netdev_priv(dev);
253         struct ipoib_cm_data data = {};
254         struct ib_cm_rep_param rep = {};
255
256         data.qpn = cpu_to_be32(priv->qp->qp_num);
257         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
258
259         rep.private_data = &data;
260         rep.private_data_len = sizeof data;
261         rep.flow_control = 0;
262         rep.rnr_retry_count = req->rnr_retry_count;
263         rep.target_ack_delay = 20; /* FIXME */
264         rep.srq = 1;
265         rep.qp_num = qp->qp_num;
266         rep.starting_psn = psn;
267         return ib_send_cm_rep(cm_id, &rep);
268 }
269
270 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
271 {
272         struct net_device *dev = cm_id->context;
273         struct ipoib_dev_priv *priv = netdev_priv(dev);
274         struct ipoib_cm_rx *p;
275         unsigned psn;
276         int ret;
277
278         ipoib_dbg(priv, "REQ arrived\n");
279         p = kzalloc(sizeof *p, GFP_KERNEL);
280         if (!p)
281                 return -ENOMEM;
282         p->dev = dev;
283         p->id = cm_id;
284         p->qp = ipoib_cm_create_rx_qp(dev, p);
285         if (IS_ERR(p->qp)) {
286                 ret = PTR_ERR(p->qp);
287                 goto err_qp;
288         }
289
290         psn = random32() & 0xffffff;
291         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
292         if (ret)
293                 goto err_modify;
294
295         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
296         if (ret) {
297                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
298                 goto err_rep;
299         }
300
301         cm_id->context = p;
302         p->jiffies = jiffies;
303         p->state = IPOIB_CM_RX_LIVE;
304         spin_lock_irq(&priv->lock);
305         if (list_empty(&priv->cm.passive_ids))
306                 queue_delayed_work(ipoib_workqueue,
307                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
308         list_add(&p->list, &priv->cm.passive_ids);
309         spin_unlock_irq(&priv->lock);
310         return 0;
311
312 err_rep:
313 err_modify:
314         ib_destroy_qp(p->qp);
315 err_qp:
316         kfree(p);
317         return ret;
318 }
319
320 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
321                                struct ib_cm_event *event)
322 {
323         struct ipoib_cm_rx *p;
324         struct ipoib_dev_priv *priv;
325
326         switch (event->event) {
327         case IB_CM_REQ_RECEIVED:
328                 return ipoib_cm_req_handler(cm_id, event);
329         case IB_CM_DREQ_RECEIVED:
330                 p = cm_id->context;
331                 ib_send_cm_drep(cm_id, NULL, 0);
332                 /* Fall through */
333         case IB_CM_REJ_RECEIVED:
334                 p = cm_id->context;
335                 priv = netdev_priv(p->dev);
336                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
337                         ipoib_warn(priv, "unable to move qp to error state\n");
338                 /* Fall through */
339         default:
340                 return 0;
341         }
342 }
343 /* Adjust length of skb with fragments to match received data */
344 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
345                           unsigned int length, struct sk_buff *toskb)
346 {
347         int i, num_frags;
348         unsigned int size;
349
350         /* put header into skb */
351         size = min(length, hdr_space);
352         skb->tail += size;
353         skb->len += size;
354         length -= size;
355
356         num_frags = skb_shinfo(skb)->nr_frags;
357         for (i = 0; i < num_frags; i++) {
358                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
359
360                 if (length == 0) {
361                         /* don't need this page */
362                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
363                         --skb_shinfo(skb)->nr_frags;
364                 } else {
365                         size = min(length, (unsigned) PAGE_SIZE);
366
367                         frag->size = size;
368                         skb->data_len += size;
369                         skb->truesize += size;
370                         skb->len += size;
371                         length -= size;
372                 }
373         }
374 }
375
376 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
377 {
378         struct ipoib_dev_priv *priv = netdev_priv(dev);
379         unsigned int wr_id = wc->wr_id & ~IPOIB_CM_OP_SRQ;
380         struct sk_buff *skb, *newskb;
381         struct ipoib_cm_rx *p;
382         unsigned long flags;
383         u64 mapping[IPOIB_CM_RX_SG];
384         int frags;
385
386         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
387                        wr_id, wc->status);
388
389         if (unlikely(wr_id >= ipoib_recvq_size)) {
390                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~IPOIB_CM_OP_SRQ)) {
391                         spin_lock_irqsave(&priv->lock, flags);
392                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
393                         ipoib_cm_start_rx_drain(priv);
394                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
395                         spin_unlock_irqrestore(&priv->lock, flags);
396                 } else
397                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
398                                    wr_id, ipoib_recvq_size);
399                 return;
400         }
401
402         skb  = priv->cm.srq_ring[wr_id].skb;
403
404         if (unlikely(wc->status != IB_WC_SUCCESS)) {
405                 ipoib_dbg(priv, "cm recv error "
406                            "(status=%d, wrid=%d vend_err %x)\n",
407                            wc->status, wr_id, wc->vendor_err);
408                 ++priv->stats.rx_dropped;
409                 goto repost;
410         }
411
412         if (!likely(wr_id & IPOIB_CM_RX_UPDATE_MASK)) {
413                 p = wc->qp->qp_context;
414                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
415                         spin_lock_irqsave(&priv->lock, flags);
416                         p->jiffies = jiffies;
417                         /* Move this entry to list head, but do not re-add it
418                          * if it has been moved out of list. */
419                         if (p->state == IPOIB_CM_RX_LIVE)
420                                 list_move(&p->list, &priv->cm.passive_ids);
421                         spin_unlock_irqrestore(&priv->lock, flags);
422                 }
423         }
424
425         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
426                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
427
428         newskb = ipoib_cm_alloc_rx_skb(dev, wr_id, frags, mapping);
429         if (unlikely(!newskb)) {
430                 /*
431                  * If we can't allocate a new RX buffer, dump
432                  * this packet and reuse the old buffer.
433                  */
434                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
435                 ++priv->stats.rx_dropped;
436                 goto repost;
437         }
438
439         ipoib_cm_dma_unmap_rx(priv, frags, priv->cm.srq_ring[wr_id].mapping);
440         memcpy(priv->cm.srq_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
441
442         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
443                        wc->byte_len, wc->slid);
444
445         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
446
447         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
448         skb_reset_mac_header(skb);
449         skb_pull(skb, IPOIB_ENCAP_LEN);
450
451         dev->last_rx = jiffies;
452         ++priv->stats.rx_packets;
453         priv->stats.rx_bytes += skb->len;
454
455         skb->dev = dev;
456         /* XXX get correct PACKET_ type here */
457         skb->pkt_type = PACKET_HOST;
458         netif_receive_skb(skb);
459
460 repost:
461         if (unlikely(ipoib_cm_post_receive(dev, wr_id)))
462                 ipoib_warn(priv, "ipoib_cm_post_receive failed "
463                            "for buf %d\n", wr_id);
464 }
465
466 static inline int post_send(struct ipoib_dev_priv *priv,
467                             struct ipoib_cm_tx *tx,
468                             unsigned int wr_id,
469                             u64 addr, int len)
470 {
471         struct ib_send_wr *bad_wr;
472
473         priv->tx_sge.addr             = addr;
474         priv->tx_sge.length           = len;
475
476         priv->tx_wr.wr_id             = wr_id;
477
478         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
479 }
480
481 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
482 {
483         struct ipoib_dev_priv *priv = netdev_priv(dev);
484         struct ipoib_tx_buf *tx_req;
485         u64 addr;
486
487         if (unlikely(skb->len > tx->mtu)) {
488                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
489                            skb->len, tx->mtu);
490                 ++priv->stats.tx_dropped;
491                 ++priv->stats.tx_errors;
492                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
493                 return;
494         }
495
496         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
497                        tx->tx_head, skb->len, tx->qp->qp_num);
498
499         /*
500          * We put the skb into the tx_ring _before_ we call post_send()
501          * because it's entirely possible that the completion handler will
502          * run before we execute anything after the post_send().  That
503          * means we have to make sure everything is properly recorded and
504          * our state is consistent before we call post_send().
505          */
506         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
507         tx_req->skb = skb;
508         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
509         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
510                 ++priv->stats.tx_errors;
511                 dev_kfree_skb_any(skb);
512                 return;
513         }
514
515         tx_req->mapping = addr;
516
517         if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
518                                 addr, skb->len))) {
519                 ipoib_warn(priv, "post_send failed\n");
520                 ++priv->stats.tx_errors;
521                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
522                 dev_kfree_skb_any(skb);
523         } else {
524                 dev->trans_start = jiffies;
525                 ++tx->tx_head;
526
527                 if (tx->tx_head - tx->tx_tail == ipoib_sendq_size) {
528                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
529                                   tx->qp->qp_num);
530                         netif_stop_queue(dev);
531                         set_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
532                 }
533         }
534 }
535
536 static void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ipoib_cm_tx *tx,
537                                   struct ib_wc *wc)
538 {
539         struct ipoib_dev_priv *priv = netdev_priv(dev);
540         unsigned int wr_id = wc->wr_id;
541         struct ipoib_tx_buf *tx_req;
542         unsigned long flags;
543
544         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
545                        wr_id, wc->status);
546
547         if (unlikely(wr_id >= ipoib_sendq_size)) {
548                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
549                            wr_id, ipoib_sendq_size);
550                 return;
551         }
552
553         tx_req = &tx->tx_ring[wr_id];
554
555         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
556
557         /* FIXME: is this right? Shouldn't we only increment on success? */
558         ++priv->stats.tx_packets;
559         priv->stats.tx_bytes += tx_req->skb->len;
560
561         dev_kfree_skb_any(tx_req->skb);
562
563         spin_lock_irqsave(&priv->tx_lock, flags);
564         ++tx->tx_tail;
565         if (unlikely(test_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags)) &&
566             tx->tx_head - tx->tx_tail <= ipoib_sendq_size >> 1) {
567                 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
568                 netif_wake_queue(dev);
569         }
570
571         if (wc->status != IB_WC_SUCCESS &&
572             wc->status != IB_WC_WR_FLUSH_ERR) {
573                 struct ipoib_neigh *neigh;
574
575                 ipoib_dbg(priv, "failed cm send event "
576                            "(status=%d, wrid=%d vend_err %x)\n",
577                            wc->status, wr_id, wc->vendor_err);
578
579                 spin_lock(&priv->lock);
580                 neigh = tx->neigh;
581
582                 if (neigh) {
583                         neigh->cm = NULL;
584                         list_del(&neigh->list);
585                         if (neigh->ah)
586                                 ipoib_put_ah(neigh->ah);
587                         ipoib_neigh_free(dev, neigh);
588
589                         tx->neigh = NULL;
590                 }
591
592                 /* queue would be re-started anyway when TX is destroyed,
593                  * but it makes sense to do it ASAP here. */
594                 if (test_and_clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags))
595                         netif_wake_queue(dev);
596
597                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
598                         list_move(&tx->list, &priv->cm.reap_list);
599                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
600                 }
601
602                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
603
604                 spin_unlock(&priv->lock);
605         }
606
607         spin_unlock_irqrestore(&priv->tx_lock, flags);
608 }
609
610 static void ipoib_cm_tx_completion(struct ib_cq *cq, void *tx_ptr)
611 {
612         struct ipoib_cm_tx *tx = tx_ptr;
613         int n, i;
614
615         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
616         do {
617                 n = ib_poll_cq(cq, IPOIB_NUM_WC, tx->ibwc);
618                 for (i = 0; i < n; ++i)
619                         ipoib_cm_handle_tx_wc(tx->dev, tx, tx->ibwc + i);
620         } while (n == IPOIB_NUM_WC);
621 }
622
623 int ipoib_cm_dev_open(struct net_device *dev)
624 {
625         struct ipoib_dev_priv *priv = netdev_priv(dev);
626         struct ib_qp_init_attr qp_init_attr = {
627                 .send_cq = priv->cq,   /* does not matter, we never send anything */
628                 .recv_cq = priv->cq,
629                 .cap.max_send_wr = 1,  /* FIXME: 0 Seems not to work */
630                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
631                 .cap.max_recv_wr = 1,
632                 .cap.max_recv_sge = 1, /* FIXME: 0 Seems not to work */
633                 .sq_sig_type = IB_SIGNAL_ALL_WR,
634                 .qp_type = IB_QPT_UC,
635         };
636         int ret;
637
638         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
639                 return 0;
640
641         priv->cm.rx_drain_qp = ib_create_qp(priv->pd, &qp_init_attr);
642         if (IS_ERR(priv->cm.rx_drain_qp)) {
643                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
644                 ret = PTR_ERR(priv->cm.rx_drain_qp);
645                 return ret;
646         }
647
648         /*
649          * We put the QP in error state directly.  This way, a "flush
650          * error" WC will be immediately generated for each WR we post.
651          */
652         ret = ib_modify_qp(priv->cm.rx_drain_qp, &ipoib_cm_err_attr, IB_QP_STATE);
653         if (ret) {
654                 ipoib_warn(priv, "failed to modify drain QP to error: %d\n", ret);
655                 goto err_qp;
656         }
657
658         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
659         if (IS_ERR(priv->cm.id)) {
660                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
661                 ret = PTR_ERR(priv->cm.id);
662                 goto err_cm;
663         }
664
665         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
666                            0, NULL);
667         if (ret) {
668                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
669                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
670                 goto err_listen;
671         }
672
673         return 0;
674
675 err_listen:
676         ib_destroy_cm_id(priv->cm.id);
677 err_cm:
678         priv->cm.id = NULL;
679 err_qp:
680         ib_destroy_qp(priv->cm.rx_drain_qp);
681         return ret;
682 }
683
684 void ipoib_cm_dev_stop(struct net_device *dev)
685 {
686         struct ipoib_dev_priv *priv = netdev_priv(dev);
687         struct ipoib_cm_rx *p, *n;
688         unsigned long begin;
689         LIST_HEAD(list);
690         int ret;
691
692         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
693                 return;
694
695         ib_destroy_cm_id(priv->cm.id);
696         priv->cm.id = NULL;
697
698         spin_lock_irq(&priv->lock);
699         while (!list_empty(&priv->cm.passive_ids)) {
700                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
701                 list_move(&p->list, &priv->cm.rx_error_list);
702                 p->state = IPOIB_CM_RX_ERROR;
703                 spin_unlock_irq(&priv->lock);
704                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
705                 if (ret)
706                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
707                 spin_lock_irq(&priv->lock);
708         }
709
710         /* Wait for all RX to be drained */
711         begin = jiffies;
712
713         while (!list_empty(&priv->cm.rx_error_list) ||
714                !list_empty(&priv->cm.rx_flush_list) ||
715                !list_empty(&priv->cm.rx_drain_list)) {
716                 if (time_after(jiffies, begin + 5 * HZ)) {
717                         ipoib_warn(priv, "RX drain timing out\n");
718
719                         /*
720                          * assume the HW is wedged and just free up everything.
721                          */
722                         list_splice_init(&priv->cm.rx_flush_list, &list);
723                         list_splice_init(&priv->cm.rx_error_list, &list);
724                         list_splice_init(&priv->cm.rx_drain_list, &list);
725                         break;
726                 }
727                 spin_unlock_irq(&priv->lock);
728                 msleep(1);
729                 ipoib_drain_cq(dev);
730                 spin_lock_irq(&priv->lock);
731         }
732
733         list_splice_init(&priv->cm.rx_reap_list, &list);
734
735         spin_unlock_irq(&priv->lock);
736
737         list_for_each_entry_safe(p, n, &list, list) {
738                 ib_destroy_cm_id(p->id);
739                 ib_destroy_qp(p->qp);
740                 kfree(p);
741         }
742
743         ib_destroy_qp(priv->cm.rx_drain_qp);
744         cancel_delayed_work(&priv->cm.stale_task);
745 }
746
747 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
748 {
749         struct ipoib_cm_tx *p = cm_id->context;
750         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
751         struct ipoib_cm_data *data = event->private_data;
752         struct sk_buff_head skqueue;
753         struct ib_qp_attr qp_attr;
754         int qp_attr_mask, ret;
755         struct sk_buff *skb;
756
757         p->mtu = be32_to_cpu(data->mtu);
758
759         if (p->mtu < priv->dev->mtu + IPOIB_ENCAP_LEN) {
760                 ipoib_warn(priv, "Rejecting connection: mtu %d < device mtu %d + 4\n",
761                            p->mtu, priv->dev->mtu);
762                 return -EINVAL;
763         }
764
765         qp_attr.qp_state = IB_QPS_RTR;
766         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
767         if (ret) {
768                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
769                 return ret;
770         }
771
772         qp_attr.rq_psn = 0 /* FIXME */;
773         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
774         if (ret) {
775                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
776                 return ret;
777         }
778
779         qp_attr.qp_state = IB_QPS_RTS;
780         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
781         if (ret) {
782                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
783                 return ret;
784         }
785         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
786         if (ret) {
787                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
788                 return ret;
789         }
790
791         skb_queue_head_init(&skqueue);
792
793         spin_lock_irq(&priv->lock);
794         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
795         if (p->neigh)
796                 while ((skb = __skb_dequeue(&p->neigh->queue)))
797                         __skb_queue_tail(&skqueue, skb);
798         spin_unlock_irq(&priv->lock);
799
800         while ((skb = __skb_dequeue(&skqueue))) {
801                 skb->dev = p->dev;
802                 if (dev_queue_xmit(skb))
803                         ipoib_warn(priv, "dev_queue_xmit failed "
804                                    "to requeue packet\n");
805         }
806
807         ret = ib_send_cm_rtu(cm_id, NULL, 0);
808         if (ret) {
809                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
810                 return ret;
811         }
812         return 0;
813 }
814
815 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ib_cq *cq)
816 {
817         struct ipoib_dev_priv *priv = netdev_priv(dev);
818         struct ib_qp_init_attr attr = {};
819         attr.recv_cq = priv->cq;
820         attr.srq = priv->cm.srq;
821         attr.cap.max_send_wr = ipoib_sendq_size;
822         attr.cap.max_send_sge = 1;
823         attr.sq_sig_type = IB_SIGNAL_ALL_WR;
824         attr.qp_type = IB_QPT_RC;
825         attr.send_cq = cq;
826         return ib_create_qp(priv->pd, &attr);
827 }
828
829 static int ipoib_cm_send_req(struct net_device *dev,
830                              struct ib_cm_id *id, struct ib_qp *qp,
831                              u32 qpn,
832                              struct ib_sa_path_rec *pathrec)
833 {
834         struct ipoib_dev_priv *priv = netdev_priv(dev);
835         struct ipoib_cm_data data = {};
836         struct ib_cm_req_param req = {};
837
838         data.qpn = cpu_to_be32(priv->qp->qp_num);
839         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
840
841         req.primary_path              = pathrec;
842         req.alternate_path            = NULL;
843         req.service_id                = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
844         req.qp_num                    = qp->qp_num;
845         req.qp_type                   = qp->qp_type;
846         req.private_data              = &data;
847         req.private_data_len          = sizeof data;
848         req.flow_control              = 0;
849
850         req.starting_psn              = 0; /* FIXME */
851
852         /*
853          * Pick some arbitrary defaults here; we could make these
854          * module parameters if anyone cared about setting them.
855          */
856         req.responder_resources       = 4;
857         req.remote_cm_response_timeout = 20;
858         req.local_cm_response_timeout  = 20;
859         req.retry_count               = 0; /* RFC draft warns against retries */
860         req.rnr_retry_count           = 0; /* RFC draft warns against retries */
861         req.max_cm_retries            = 15;
862         req.srq                       = 1;
863         return ib_send_cm_req(id, &req);
864 }
865
866 static int ipoib_cm_modify_tx_init(struct net_device *dev,
867                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
868 {
869         struct ipoib_dev_priv *priv = netdev_priv(dev);
870         struct ib_qp_attr qp_attr;
871         int qp_attr_mask, ret;
872         ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
873         if (ret) {
874                 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
875                 return ret;
876         }
877
878         qp_attr.qp_state = IB_QPS_INIT;
879         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
880         qp_attr.port_num = priv->port;
881         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
882
883         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
884         if (ret) {
885                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
886                 return ret;
887         }
888         return 0;
889 }
890
891 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
892                             struct ib_sa_path_rec *pathrec)
893 {
894         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
895         int ret;
896
897         p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
898                                 GFP_KERNEL);
899         if (!p->tx_ring) {
900                 ipoib_warn(priv, "failed to allocate tx ring\n");
901                 ret = -ENOMEM;
902                 goto err_tx;
903         }
904
905         p->cq = ib_create_cq(priv->ca, ipoib_cm_tx_completion, NULL, p,
906                              ipoib_sendq_size + 1, 0);
907         if (IS_ERR(p->cq)) {
908                 ret = PTR_ERR(p->cq);
909                 ipoib_warn(priv, "failed to allocate tx cq: %d\n", ret);
910                 goto err_cq;
911         }
912
913         ret = ib_req_notify_cq(p->cq, IB_CQ_NEXT_COMP);
914         if (ret) {
915                 ipoib_warn(priv, "failed to request completion notification: %d\n", ret);
916                 goto err_req_notify;
917         }
918
919         p->qp = ipoib_cm_create_tx_qp(p->dev, p->cq);
920         if (IS_ERR(p->qp)) {
921                 ret = PTR_ERR(p->qp);
922                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
923                 goto err_qp;
924         }
925
926         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
927         if (IS_ERR(p->id)) {
928                 ret = PTR_ERR(p->id);
929                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
930                 goto err_id;
931         }
932
933         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
934         if (ret) {
935                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
936                 goto err_modify;
937         }
938
939         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
940         if (ret) {
941                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
942                 goto err_send_cm;
943         }
944
945         ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
946                   p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
947
948         return 0;
949
950 err_send_cm:
951 err_modify:
952         ib_destroy_cm_id(p->id);
953 err_id:
954         p->id = NULL;
955         ib_destroy_qp(p->qp);
956 err_req_notify:
957 err_qp:
958         p->qp = NULL;
959         ib_destroy_cq(p->cq);
960 err_cq:
961         p->cq = NULL;
962 err_tx:
963         return ret;
964 }
965
966 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
967 {
968         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
969         struct ipoib_tx_buf *tx_req;
970
971         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
972                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
973
974         if (p->id)
975                 ib_destroy_cm_id(p->id);
976
977         if (p->qp)
978                 ib_destroy_qp(p->qp);
979
980         if (p->cq)
981                 ib_destroy_cq(p->cq);
982
983         if (test_bit(IPOIB_FLAG_NETIF_STOPPED, &p->flags))
984                 netif_wake_queue(p->dev);
985
986         if (p->tx_ring) {
987                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
988                         tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
989                         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
990                                          DMA_TO_DEVICE);
991                         dev_kfree_skb_any(tx_req->skb);
992                         ++p->tx_tail;
993                 }
994
995                 kfree(p->tx_ring);
996         }
997
998         kfree(p);
999 }
1000
1001 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1002                                struct ib_cm_event *event)
1003 {
1004         struct ipoib_cm_tx *tx = cm_id->context;
1005         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1006         struct net_device *dev = priv->dev;
1007         struct ipoib_neigh *neigh;
1008         int ret;
1009
1010         switch (event->event) {
1011         case IB_CM_DREQ_RECEIVED:
1012                 ipoib_dbg(priv, "DREQ received.\n");
1013                 ib_send_cm_drep(cm_id, NULL, 0);
1014                 break;
1015         case IB_CM_REP_RECEIVED:
1016                 ipoib_dbg(priv, "REP received.\n");
1017                 ret = ipoib_cm_rep_handler(cm_id, event);
1018                 if (ret)
1019                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1020                                        NULL, 0, NULL, 0);
1021                 break;
1022         case IB_CM_REQ_ERROR:
1023         case IB_CM_REJ_RECEIVED:
1024         case IB_CM_TIMEWAIT_EXIT:
1025                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1026                 spin_lock_irq(&priv->tx_lock);
1027                 spin_lock(&priv->lock);
1028                 neigh = tx->neigh;
1029
1030                 if (neigh) {
1031                         neigh->cm = NULL;
1032                         list_del(&neigh->list);
1033                         if (neigh->ah)
1034                                 ipoib_put_ah(neigh->ah);
1035                         ipoib_neigh_free(dev, neigh);
1036
1037                         tx->neigh = NULL;
1038                 }
1039
1040                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1041                         list_move(&tx->list, &priv->cm.reap_list);
1042                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1043                 }
1044
1045                 spin_unlock(&priv->lock);
1046                 spin_unlock_irq(&priv->tx_lock);
1047                 break;
1048         default:
1049                 break;
1050         }
1051
1052         return 0;
1053 }
1054
1055 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1056                                        struct ipoib_neigh *neigh)
1057 {
1058         struct ipoib_dev_priv *priv = netdev_priv(dev);
1059         struct ipoib_cm_tx *tx;
1060
1061         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1062         if (!tx)
1063                 return NULL;
1064
1065         neigh->cm = tx;
1066         tx->neigh = neigh;
1067         tx->path = path;
1068         tx->dev = dev;
1069         list_add(&tx->list, &priv->cm.start_list);
1070         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1071         queue_work(ipoib_workqueue, &priv->cm.start_task);
1072         return tx;
1073 }
1074
1075 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1076 {
1077         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1078         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1079                 list_move(&tx->list, &priv->cm.reap_list);
1080                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1081                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1082                           IPOIB_GID_ARG(tx->neigh->dgid));
1083                 tx->neigh = NULL;
1084         }
1085 }
1086
1087 static void ipoib_cm_tx_start(struct work_struct *work)
1088 {
1089         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1090                                                    cm.start_task);
1091         struct net_device *dev = priv->dev;
1092         struct ipoib_neigh *neigh;
1093         struct ipoib_cm_tx *p;
1094         unsigned long flags;
1095         int ret;
1096
1097         struct ib_sa_path_rec pathrec;
1098         u32 qpn;
1099
1100         spin_lock_irqsave(&priv->tx_lock, flags);
1101         spin_lock(&priv->lock);
1102         while (!list_empty(&priv->cm.start_list)) {
1103                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1104                 list_del_init(&p->list);
1105                 neigh = p->neigh;
1106                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1107                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1108                 spin_unlock(&priv->lock);
1109                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1110                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1111                 spin_lock_irqsave(&priv->tx_lock, flags);
1112                 spin_lock(&priv->lock);
1113                 if (ret) {
1114                         neigh = p->neigh;
1115                         if (neigh) {
1116                                 neigh->cm = NULL;
1117                                 list_del(&neigh->list);
1118                                 if (neigh->ah)
1119                                         ipoib_put_ah(neigh->ah);
1120                                 ipoib_neigh_free(dev, neigh);
1121                         }
1122                         list_del(&p->list);
1123                         kfree(p);
1124                 }
1125         }
1126         spin_unlock(&priv->lock);
1127         spin_unlock_irqrestore(&priv->tx_lock, flags);
1128 }
1129
1130 static void ipoib_cm_tx_reap(struct work_struct *work)
1131 {
1132         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1133                                                    cm.reap_task);
1134         struct ipoib_cm_tx *p;
1135
1136         spin_lock_irq(&priv->tx_lock);
1137         spin_lock(&priv->lock);
1138         while (!list_empty(&priv->cm.reap_list)) {
1139                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1140                 list_del(&p->list);
1141                 spin_unlock(&priv->lock);
1142                 spin_unlock_irq(&priv->tx_lock);
1143                 ipoib_cm_tx_destroy(p);
1144                 spin_lock_irq(&priv->tx_lock);
1145                 spin_lock(&priv->lock);
1146         }
1147         spin_unlock(&priv->lock);
1148         spin_unlock_irq(&priv->tx_lock);
1149 }
1150
1151 static void ipoib_cm_skb_reap(struct work_struct *work)
1152 {
1153         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1154                                                    cm.skb_task);
1155         struct net_device *dev = priv->dev;
1156         struct sk_buff *skb;
1157
1158         unsigned mtu = priv->mcast_mtu;
1159
1160         spin_lock_irq(&priv->tx_lock);
1161         spin_lock(&priv->lock);
1162         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1163                 spin_unlock(&priv->lock);
1164                 spin_unlock_irq(&priv->tx_lock);
1165                 if (skb->protocol == htons(ETH_P_IP))
1166                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1167 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1168                 else if (skb->protocol == htons(ETH_P_IPV6))
1169                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, dev);
1170 #endif
1171                 dev_kfree_skb_any(skb);
1172                 spin_lock_irq(&priv->tx_lock);
1173                 spin_lock(&priv->lock);
1174         }
1175         spin_unlock(&priv->lock);
1176         spin_unlock_irq(&priv->tx_lock);
1177 }
1178
1179 void ipoib_cm_skb_too_long(struct net_device* dev, struct sk_buff *skb,
1180                            unsigned int mtu)
1181 {
1182         struct ipoib_dev_priv *priv = netdev_priv(dev);
1183         int e = skb_queue_empty(&priv->cm.skb_queue);
1184
1185         if (skb->dst)
1186                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1187
1188         skb_queue_tail(&priv->cm.skb_queue, skb);
1189         if (e)
1190                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1191 }
1192
1193 static void ipoib_cm_rx_reap(struct work_struct *work)
1194 {
1195         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1196                                                    cm.rx_reap_task);
1197         struct ipoib_cm_rx *p, *n;
1198         LIST_HEAD(list);
1199
1200         spin_lock_irq(&priv->lock);
1201         list_splice_init(&priv->cm.rx_reap_list, &list);
1202         spin_unlock_irq(&priv->lock);
1203
1204         list_for_each_entry_safe(p, n, &list, list) {
1205                 ib_destroy_cm_id(p->id);
1206                 ib_destroy_qp(p->qp);
1207                 kfree(p);
1208         }
1209 }
1210
1211 static void ipoib_cm_stale_task(struct work_struct *work)
1212 {
1213         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1214                                                    cm.stale_task.work);
1215         struct ipoib_cm_rx *p;
1216         int ret;
1217
1218         spin_lock_irq(&priv->lock);
1219         while (!list_empty(&priv->cm.passive_ids)) {
1220                 /* List is sorted by LRU, start from tail,
1221                  * stop when we see a recently used entry */
1222                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1223                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1224                         break;
1225                 list_move(&p->list, &priv->cm.rx_error_list);
1226                 p->state = IPOIB_CM_RX_ERROR;
1227                 spin_unlock_irq(&priv->lock);
1228                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1229                 if (ret)
1230                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1231                 spin_lock_irq(&priv->lock);
1232         }
1233
1234         if (!list_empty(&priv->cm.passive_ids))
1235                 queue_delayed_work(ipoib_workqueue,
1236                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1237         spin_unlock_irq(&priv->lock);
1238 }
1239
1240
1241 static ssize_t show_mode(struct device *d, struct device_attribute *attr, 
1242                          char *buf)
1243 {
1244         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1245
1246         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1247                 return sprintf(buf, "connected\n");
1248         else
1249                 return sprintf(buf, "datagram\n");
1250 }
1251
1252 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1253                         const char *buf, size_t count)
1254 {
1255         struct net_device *dev = to_net_dev(d);
1256         struct ipoib_dev_priv *priv = netdev_priv(dev);
1257
1258         /* flush paths if we switch modes so that connections are restarted */
1259         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1260                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1261                 ipoib_warn(priv, "enabling connected mode "
1262                            "will cause multicast packet drops\n");
1263                 ipoib_flush_paths(dev);
1264                 return count;
1265         }
1266
1267         if (!strcmp(buf, "datagram\n")) {
1268                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1269                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1270                 ipoib_flush_paths(dev);
1271                 return count;
1272         }
1273
1274         return -EINVAL;
1275 }
1276
1277 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1278
1279 int ipoib_cm_add_mode_attr(struct net_device *dev)
1280 {
1281         return device_create_file(&dev->dev, &dev_attr_mode);
1282 }
1283
1284 int ipoib_cm_dev_init(struct net_device *dev)
1285 {
1286         struct ipoib_dev_priv *priv = netdev_priv(dev);
1287         struct ib_srq_init_attr srq_init_attr = {
1288                 .attr = {
1289                         .max_wr  = ipoib_recvq_size,
1290                         .max_sge = IPOIB_CM_RX_SG
1291                 }
1292         };
1293         int ret, i;
1294
1295         INIT_LIST_HEAD(&priv->cm.passive_ids);
1296         INIT_LIST_HEAD(&priv->cm.reap_list);
1297         INIT_LIST_HEAD(&priv->cm.start_list);
1298         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1299         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1300         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1301         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1302         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1303         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1304         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1305         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1306         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1307
1308         skb_queue_head_init(&priv->cm.skb_queue);
1309
1310         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1311         if (IS_ERR(priv->cm.srq)) {
1312                 ret = PTR_ERR(priv->cm.srq);
1313                 priv->cm.srq = NULL;
1314                 return ret;
1315         }
1316
1317         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1318                                     GFP_KERNEL);
1319         if (!priv->cm.srq_ring) {
1320                 printk(KERN_WARNING "%s: failed to allocate CM ring (%d entries)\n",
1321                        priv->ca->name, ipoib_recvq_size);
1322                 ipoib_cm_dev_cleanup(dev);
1323                 return -ENOMEM;
1324         }
1325
1326         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
1327                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1328
1329         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1330         for (i = 1; i < IPOIB_CM_RX_SG; ++i)
1331                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1332         priv->cm.rx_wr.next = NULL;
1333         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1334         priv->cm.rx_wr.num_sge = IPOIB_CM_RX_SG;
1335
1336         for (i = 0; i < ipoib_recvq_size; ++i) {
1337                 if (!ipoib_cm_alloc_rx_skb(dev, i, IPOIB_CM_RX_SG - 1,
1338                                            priv->cm.srq_ring[i].mapping)) {
1339                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
1340                         ipoib_cm_dev_cleanup(dev);
1341                         return -ENOMEM;
1342                 }
1343                 if (ipoib_cm_post_receive(dev, i)) {
1344                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
1345                         ipoib_cm_dev_cleanup(dev);
1346                         return -EIO;
1347                 }
1348         }
1349
1350         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1351         return 0;
1352 }
1353
1354 void ipoib_cm_dev_cleanup(struct net_device *dev)
1355 {
1356         struct ipoib_dev_priv *priv = netdev_priv(dev);
1357         int i, ret;
1358
1359         if (!priv->cm.srq)
1360                 return;
1361
1362         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1363
1364         ret = ib_destroy_srq(priv->cm.srq);
1365         if (ret)
1366                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1367
1368         priv->cm.srq = NULL;
1369         if (!priv->cm.srq_ring)
1370                 return;
1371         for (i = 0; i < ipoib_recvq_size; ++i)
1372                 if (priv->cm.srq_ring[i].skb) {
1373                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
1374                                               priv->cm.srq_ring[i].mapping);
1375                         dev_kfree_skb_any(priv->cm.srq_ring[i].skb);
1376                         priv->cm.srq_ring[i].skb = NULL;
1377                 }
1378         kfree(priv->cm.srq_ring);
1379         priv->cm.srq_ring = NULL;
1380 }