Merge branch 'upstream-linus' of master.kernel.org:/pub/scm/linux/kernel/git/jgarzik...
[linux-2.6] / drivers / infiniband / ulp / ipoib / ipoib_ib.c
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  *
35  * $Id: ipoib_ib.c 1386 2004-12-27 16:23:17Z roland $
36  */
37
38 #include <linux/delay.h>
39 #include <linux/dma-mapping.h>
40
41 #include <rdma/ib_cache.h>
42
43 #include "ipoib.h"
44
45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
46 static int data_debug_level;
47
48 module_param(data_debug_level, int, 0644);
49 MODULE_PARM_DESC(data_debug_level,
50                  "Enable data path debug tracing if > 0");
51 #endif
52
53 static DEFINE_MUTEX(pkey_mutex);
54
55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
56                                  struct ib_pd *pd, struct ib_ah_attr *attr)
57 {
58         struct ipoib_ah *ah;
59
60         ah = kmalloc(sizeof *ah, GFP_KERNEL);
61         if (!ah)
62                 return NULL;
63
64         ah->dev       = dev;
65         ah->last_send = 0;
66         kref_init(&ah->ref);
67
68         ah->ah = ib_create_ah(pd, attr);
69         if (IS_ERR(ah->ah)) {
70                 kfree(ah);
71                 ah = NULL;
72         } else
73                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
74
75         return ah;
76 }
77
78 void ipoib_free_ah(struct kref *kref)
79 {
80         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
81         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
82
83         unsigned long flags;
84
85         spin_lock_irqsave(&priv->lock, flags);
86         list_add_tail(&ah->list, &priv->dead_ahs);
87         spin_unlock_irqrestore(&priv->lock, flags);
88 }
89
90 static int ipoib_ib_post_receive(struct net_device *dev, int id)
91 {
92         struct ipoib_dev_priv *priv = netdev_priv(dev);
93         struct ib_sge list;
94         struct ib_recv_wr param;
95         struct ib_recv_wr *bad_wr;
96         int ret;
97
98         list.addr     = priv->rx_ring[id].mapping;
99         list.length   = IPOIB_BUF_SIZE;
100         list.lkey     = priv->mr->lkey;
101
102         param.next    = NULL;
103         param.wr_id   = id | IPOIB_OP_RECV;
104         param.sg_list = &list;
105         param.num_sge = 1;
106
107         ret = ib_post_recv(priv->qp, &param, &bad_wr);
108         if (unlikely(ret)) {
109                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
110                 ib_dma_unmap_single(priv->ca, priv->rx_ring[id].mapping,
111                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
112                 dev_kfree_skb_any(priv->rx_ring[id].skb);
113                 priv->rx_ring[id].skb = NULL;
114         }
115
116         return ret;
117 }
118
119 static int ipoib_alloc_rx_skb(struct net_device *dev, int id)
120 {
121         struct ipoib_dev_priv *priv = netdev_priv(dev);
122         struct sk_buff *skb;
123         u64 addr;
124
125         skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
126         if (!skb)
127                 return -ENOMEM;
128
129         /*
130          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
131          * header.  So we need 4 more bytes to get to 48 and align the
132          * IP header to a multiple of 16.
133          */
134         skb_reserve(skb, 4);
135
136         addr = ib_dma_map_single(priv->ca, skb->data, IPOIB_BUF_SIZE,
137                                  DMA_FROM_DEVICE);
138         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
139                 dev_kfree_skb_any(skb);
140                 return -EIO;
141         }
142
143         priv->rx_ring[id].skb     = skb;
144         priv->rx_ring[id].mapping = addr;
145
146         return 0;
147 }
148
149 static int ipoib_ib_post_receives(struct net_device *dev)
150 {
151         struct ipoib_dev_priv *priv = netdev_priv(dev);
152         int i;
153
154         for (i = 0; i < ipoib_recvq_size; ++i) {
155                 if (ipoib_alloc_rx_skb(dev, i)) {
156                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
157                         return -ENOMEM;
158                 }
159                 if (ipoib_ib_post_receive(dev, i)) {
160                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
161                         return -EIO;
162                 }
163         }
164
165         return 0;
166 }
167
168 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
169 {
170         struct ipoib_dev_priv *priv = netdev_priv(dev);
171         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
172         struct sk_buff *skb;
173         u64 addr;
174
175         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
176                        wr_id, wc->status);
177
178         if (unlikely(wr_id >= ipoib_recvq_size)) {
179                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
180                            wr_id, ipoib_recvq_size);
181                 return;
182         }
183
184         skb  = priv->rx_ring[wr_id].skb;
185         addr = priv->rx_ring[wr_id].mapping;
186
187         if (unlikely(wc->status != IB_WC_SUCCESS)) {
188                 if (wc->status != IB_WC_WR_FLUSH_ERR)
189                         ipoib_warn(priv, "failed recv event "
190                                    "(status=%d, wrid=%d vend_err %x)\n",
191                                    wc->status, wr_id, wc->vendor_err);
192                 ib_dma_unmap_single(priv->ca, addr,
193                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
194                 dev_kfree_skb_any(skb);
195                 priv->rx_ring[wr_id].skb = NULL;
196                 return;
197         }
198
199         /*
200          * If we can't allocate a new RX buffer, dump
201          * this packet and reuse the old buffer.
202          */
203         if (unlikely(ipoib_alloc_rx_skb(dev, wr_id))) {
204                 ++priv->stats.rx_dropped;
205                 goto repost;
206         }
207
208         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
209                        wc->byte_len, wc->slid);
210
211         ib_dma_unmap_single(priv->ca, addr, IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
212
213         skb_put(skb, wc->byte_len);
214         skb_pull(skb, IB_GRH_BYTES);
215
216         if (wc->slid != priv->local_lid ||
217             wc->src_qp != priv->qp->qp_num) {
218                 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
219                 skb_reset_mac_header(skb);
220                 skb_pull(skb, IPOIB_ENCAP_LEN);
221
222                 dev->last_rx = jiffies;
223                 ++priv->stats.rx_packets;
224                 priv->stats.rx_bytes += skb->len;
225
226                 skb->dev = dev;
227                 /* XXX get correct PACKET_ type here */
228                 skb->pkt_type = PACKET_HOST;
229                 netif_receive_skb(skb);
230         } else {
231                 ipoib_dbg_data(priv, "dropping loopback packet\n");
232                 dev_kfree_skb_any(skb);
233         }
234
235 repost:
236         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
237                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
238                            "for buf %d\n", wr_id);
239 }
240
241 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
242 {
243         struct ipoib_dev_priv *priv = netdev_priv(dev);
244         unsigned int wr_id = wc->wr_id;
245         struct ipoib_tx_buf *tx_req;
246         unsigned long flags;
247
248         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
249                        wr_id, wc->status);
250
251         if (unlikely(wr_id >= ipoib_sendq_size)) {
252                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
253                            wr_id, ipoib_sendq_size);
254                 return;
255         }
256
257         tx_req = &priv->tx_ring[wr_id];
258
259         ib_dma_unmap_single(priv->ca, tx_req->mapping,
260                             tx_req->skb->len, DMA_TO_DEVICE);
261
262         ++priv->stats.tx_packets;
263         priv->stats.tx_bytes += tx_req->skb->len;
264
265         dev_kfree_skb_any(tx_req->skb);
266
267         spin_lock_irqsave(&priv->tx_lock, flags);
268         ++priv->tx_tail;
269         if (unlikely(test_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags)) &&
270             priv->tx_head - priv->tx_tail <= ipoib_sendq_size >> 1) {
271                 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
272                 netif_wake_queue(dev);
273         }
274         spin_unlock_irqrestore(&priv->tx_lock, flags);
275
276         if (wc->status != IB_WC_SUCCESS &&
277             wc->status != IB_WC_WR_FLUSH_ERR)
278                 ipoib_warn(priv, "failed send event "
279                            "(status=%d, wrid=%d vend_err %x)\n",
280                            wc->status, wr_id, wc->vendor_err);
281 }
282
283 int ipoib_poll(struct net_device *dev, int *budget)
284 {
285         struct ipoib_dev_priv *priv = netdev_priv(dev);
286         int max = min(*budget, dev->quota);
287         int done;
288         int t;
289         int empty;
290         int n, i;
291
292         done  = 0;
293         empty = 0;
294
295         while (max) {
296                 t = min(IPOIB_NUM_WC, max);
297                 n = ib_poll_cq(priv->cq, t, priv->ibwc);
298
299                 for (i = 0; i < n; ++i) {
300                         struct ib_wc *wc = priv->ibwc + i;
301
302                         if (wc->wr_id & IPOIB_CM_OP_SRQ) {
303                                 ++done;
304                                 --max;
305                                 ipoib_cm_handle_rx_wc(dev, wc);
306                         } else if (wc->wr_id & IPOIB_OP_RECV) {
307                                 ++done;
308                                 --max;
309                                 ipoib_ib_handle_rx_wc(dev, wc);
310                         } else
311                                 ipoib_ib_handle_tx_wc(dev, wc);
312                 }
313
314                 if (n != t) {
315                         empty = 1;
316                         break;
317                 }
318         }
319
320         dev->quota -= done;
321         *budget    -= done;
322
323         if (empty) {
324                 netif_rx_complete(dev);
325                 if (unlikely(ib_req_notify_cq(priv->cq,
326                                               IB_CQ_NEXT_COMP |
327                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
328                     netif_rx_reschedule(dev, 0))
329                         return 1;
330
331                 return 0;
332         }
333
334         return 1;
335 }
336
337 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
338 {
339         netif_rx_schedule(dev_ptr);
340 }
341
342 static inline int post_send(struct ipoib_dev_priv *priv,
343                             unsigned int wr_id,
344                             struct ib_ah *address, u32 qpn,
345                             u64 addr, int len)
346 {
347         struct ib_send_wr *bad_wr;
348
349         priv->tx_sge.addr             = addr;
350         priv->tx_sge.length           = len;
351
352         priv->tx_wr.wr_id             = wr_id;
353         priv->tx_wr.wr.ud.remote_qpn  = qpn;
354         priv->tx_wr.wr.ud.ah          = address;
355
356         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
357 }
358
359 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
360                 struct ipoib_ah *address, u32 qpn)
361 {
362         struct ipoib_dev_priv *priv = netdev_priv(dev);
363         struct ipoib_tx_buf *tx_req;
364         u64 addr;
365
366         if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
367                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
368                            skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
369                 ++priv->stats.tx_dropped;
370                 ++priv->stats.tx_errors;
371                 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
372                 return;
373         }
374
375         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
376                        skb->len, address, qpn);
377
378         /*
379          * We put the skb into the tx_ring _before_ we call post_send()
380          * because it's entirely possible that the completion handler will
381          * run before we execute anything after the post_send().  That
382          * means we have to make sure everything is properly recorded and
383          * our state is consistent before we call post_send().
384          */
385         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
386         tx_req->skb = skb;
387         addr = ib_dma_map_single(priv->ca, skb->data, skb->len,
388                                  DMA_TO_DEVICE);
389         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
390                 ++priv->stats.tx_errors;
391                 dev_kfree_skb_any(skb);
392                 return;
393         }
394         tx_req->mapping = addr;
395
396         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
397                                address->ah, qpn, addr, skb->len))) {
398                 ipoib_warn(priv, "post_send failed\n");
399                 ++priv->stats.tx_errors;
400                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
401                 dev_kfree_skb_any(skb);
402         } else {
403                 dev->trans_start = jiffies;
404
405                 address->last_send = priv->tx_head;
406                 ++priv->tx_head;
407
408                 if (priv->tx_head - priv->tx_tail == ipoib_sendq_size) {
409                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
410                         netif_stop_queue(dev);
411                         set_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
412                 }
413         }
414 }
415
416 static void __ipoib_reap_ah(struct net_device *dev)
417 {
418         struct ipoib_dev_priv *priv = netdev_priv(dev);
419         struct ipoib_ah *ah, *tah;
420         LIST_HEAD(remove_list);
421
422         spin_lock_irq(&priv->tx_lock);
423         spin_lock(&priv->lock);
424         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
425                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
426                         list_del(&ah->list);
427                         ib_destroy_ah(ah->ah);
428                         kfree(ah);
429                 }
430         spin_unlock(&priv->lock);
431         spin_unlock_irq(&priv->tx_lock);
432 }
433
434 void ipoib_reap_ah(struct work_struct *work)
435 {
436         struct ipoib_dev_priv *priv =
437                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
438         struct net_device *dev = priv->dev;
439
440         __ipoib_reap_ah(dev);
441
442         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
443                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
444 }
445
446 int ipoib_ib_dev_open(struct net_device *dev)
447 {
448         struct ipoib_dev_priv *priv = netdev_priv(dev);
449         int ret;
450
451         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
452                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
453                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
454                 return -1;
455         }
456         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
457
458         ret = ipoib_init_qp(dev);
459         if (ret) {
460                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
461                 return -1;
462         }
463
464         ret = ipoib_ib_post_receives(dev);
465         if (ret) {
466                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
467                 ipoib_ib_dev_stop(dev, 1);
468                 return -1;
469         }
470
471         ret = ipoib_cm_dev_open(dev);
472         if (ret) {
473                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
474                 ipoib_ib_dev_stop(dev, 1);
475                 return -1;
476         }
477
478         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
479         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
480
481         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
482
483         return 0;
484 }
485
486 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
487 {
488         struct ipoib_dev_priv *priv = netdev_priv(dev);
489         u16 pkey_index = 0;
490
491         if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
492                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
493         else
494                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
495 }
496
497 int ipoib_ib_dev_up(struct net_device *dev)
498 {
499         struct ipoib_dev_priv *priv = netdev_priv(dev);
500
501         ipoib_pkey_dev_check_presence(dev);
502
503         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
504                 ipoib_dbg(priv, "PKEY is not assigned.\n");
505                 return 0;
506         }
507
508         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
509
510         return ipoib_mcast_start_thread(dev);
511 }
512
513 int ipoib_ib_dev_down(struct net_device *dev, int flush)
514 {
515         struct ipoib_dev_priv *priv = netdev_priv(dev);
516
517         ipoib_dbg(priv, "downing ib_dev\n");
518
519         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
520         netif_carrier_off(dev);
521
522         /* Shutdown the P_Key thread if still active */
523         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
524                 mutex_lock(&pkey_mutex);
525                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
526                 cancel_delayed_work(&priv->pkey_poll_task);
527                 mutex_unlock(&pkey_mutex);
528                 if (flush)
529                         flush_workqueue(ipoib_workqueue);
530         }
531
532         ipoib_mcast_stop_thread(dev, flush);
533         ipoib_mcast_dev_flush(dev);
534
535         ipoib_flush_paths(dev);
536
537         return 0;
538 }
539
540 static int recvs_pending(struct net_device *dev)
541 {
542         struct ipoib_dev_priv *priv = netdev_priv(dev);
543         int pending = 0;
544         int i;
545
546         for (i = 0; i < ipoib_recvq_size; ++i)
547                 if (priv->rx_ring[i].skb)
548                         ++pending;
549
550         return pending;
551 }
552
553 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
554 {
555         struct ipoib_dev_priv *priv = netdev_priv(dev);
556         struct ib_qp_attr qp_attr;
557         unsigned long begin;
558         struct ipoib_tx_buf *tx_req;
559         int i, n;
560
561         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
562         netif_poll_disable(dev);
563
564         ipoib_cm_dev_stop(dev);
565
566         /*
567          * Move our QP to the error state and then reinitialize in
568          * when all work requests have completed or have been flushed.
569          */
570         qp_attr.qp_state = IB_QPS_ERR;
571         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
572                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
573
574         /* Wait for all sends and receives to complete */
575         begin = jiffies;
576
577         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
578                 if (time_after(jiffies, begin + 5 * HZ)) {
579                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
580                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
581
582                         /*
583                          * assume the HW is wedged and just free up
584                          * all our pending work requests.
585                          */
586                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
587                                 tx_req = &priv->tx_ring[priv->tx_tail &
588                                                         (ipoib_sendq_size - 1)];
589                                 ib_dma_unmap_single(priv->ca,
590                                                     tx_req->mapping,
591                                                     tx_req->skb->len,
592                                                     DMA_TO_DEVICE);
593                                 dev_kfree_skb_any(tx_req->skb);
594                                 ++priv->tx_tail;
595                         }
596
597                         for (i = 0; i < ipoib_recvq_size; ++i) {
598                                 struct ipoib_rx_buf *rx_req;
599
600                                 rx_req = &priv->rx_ring[i];
601                                 if (!rx_req->skb)
602                                         continue;
603                                 ib_dma_unmap_single(priv->ca,
604                                                     rx_req->mapping,
605                                                     IPOIB_BUF_SIZE,
606                                                     DMA_FROM_DEVICE);
607                                 dev_kfree_skb_any(rx_req->skb);
608                                 rx_req->skb = NULL;
609                         }
610
611                         goto timeout;
612                 }
613
614                 do {
615                         n = ib_poll_cq(priv->cq, IPOIB_NUM_WC, priv->ibwc);
616                         for (i = 0; i < n; ++i) {
617                                 if (priv->ibwc[i].wr_id & IPOIB_CM_OP_SRQ)
618                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
619                                 else if (priv->ibwc[i].wr_id & IPOIB_OP_RECV)
620                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
621                                 else
622                                         ipoib_ib_handle_tx_wc(dev, priv->ibwc + i);
623                         }
624                 } while (n == IPOIB_NUM_WC);
625
626                 msleep(1);
627         }
628
629         ipoib_dbg(priv, "All sends and receives done.\n");
630
631 timeout:
632         qp_attr.qp_state = IB_QPS_RESET;
633         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
634                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
635
636         /* Wait for all AHs to be reaped */
637         set_bit(IPOIB_STOP_REAPER, &priv->flags);
638         cancel_delayed_work(&priv->ah_reap_task);
639         if (flush)
640                 flush_workqueue(ipoib_workqueue);
641
642         begin = jiffies;
643
644         while (!list_empty(&priv->dead_ahs)) {
645                 __ipoib_reap_ah(dev);
646
647                 if (time_after(jiffies, begin + HZ)) {
648                         ipoib_warn(priv, "timing out; will leak address handles\n");
649                         break;
650                 }
651
652                 msleep(1);
653         }
654
655         netif_poll_enable(dev);
656         ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP);
657
658         return 0;
659 }
660
661 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
662 {
663         struct ipoib_dev_priv *priv = netdev_priv(dev);
664
665         priv->ca = ca;
666         priv->port = port;
667         priv->qp = NULL;
668
669         if (ipoib_transport_dev_init(dev, ca)) {
670                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
671                 return -ENODEV;
672         }
673
674         if (dev->flags & IFF_UP) {
675                 if (ipoib_ib_dev_open(dev)) {
676                         ipoib_transport_dev_cleanup(dev);
677                         return -ENODEV;
678                 }
679         }
680
681         return 0;
682 }
683
684 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, int pkey_event)
685 {
686         struct ipoib_dev_priv *cpriv;
687         struct net_device *dev = priv->dev;
688         u16 new_index;
689
690         mutex_lock(&priv->vlan_mutex);
691
692         /*
693          * Flush any child interfaces too -- they might be up even if
694          * the parent is down.
695          */
696         list_for_each_entry(cpriv, &priv->child_intfs, list)
697                 __ipoib_ib_dev_flush(cpriv, pkey_event);
698
699         mutex_unlock(&priv->vlan_mutex);
700
701         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
702                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
703                 return;
704         }
705
706         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
707                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
708                 return;
709         }
710
711         if (pkey_event) {
712                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
713                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
714                         ipoib_ib_dev_down(dev, 0);
715                         ipoib_pkey_dev_delay_open(dev);
716                         return;
717                 }
718                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
719
720                 /* restart QP only if P_Key index is changed */
721                 if (new_index == priv->pkey_index) {
722                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
723                         return;
724                 }
725                 priv->pkey_index = new_index;
726         }
727
728         ipoib_dbg(priv, "flushing\n");
729
730         ipoib_ib_dev_down(dev, 0);
731
732         if (pkey_event) {
733                 ipoib_ib_dev_stop(dev, 0);
734                 ipoib_ib_dev_open(dev);
735         }
736
737         /*
738          * The device could have been brought down between the start and when
739          * we get here, don't bring it back up if it's not configured up
740          */
741         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
742                 ipoib_ib_dev_up(dev);
743                 ipoib_mcast_restart_task(&priv->restart_task);
744         }
745 }
746
747 void ipoib_ib_dev_flush(struct work_struct *work)
748 {
749         struct ipoib_dev_priv *priv =
750                 container_of(work, struct ipoib_dev_priv, flush_task);
751
752         ipoib_dbg(priv, "Flushing %s\n", priv->dev->name);
753         __ipoib_ib_dev_flush(priv, 0);
754 }
755
756 void ipoib_pkey_event(struct work_struct *work)
757 {
758         struct ipoib_dev_priv *priv =
759                 container_of(work, struct ipoib_dev_priv, pkey_event_task);
760
761         ipoib_dbg(priv, "Flushing %s and restarting its QP\n", priv->dev->name);
762         __ipoib_ib_dev_flush(priv, 1);
763 }
764
765 void ipoib_ib_dev_cleanup(struct net_device *dev)
766 {
767         struct ipoib_dev_priv *priv = netdev_priv(dev);
768
769         ipoib_dbg(priv, "cleaning up ib_dev\n");
770
771         ipoib_mcast_stop_thread(dev, 1);
772         ipoib_mcast_dev_flush(dev);
773
774         ipoib_transport_dev_cleanup(dev);
775 }
776
777 /*
778  * Delayed P_Key Assigment Interim Support
779  *
780  * The following is initial implementation of delayed P_Key assigment
781  * mechanism. It is using the same approach implemented for the multicast
782  * group join. The single goal of this implementation is to quickly address
783  * Bug #2507. This implementation will probably be removed when the P_Key
784  * change async notification is available.
785  */
786
787 void ipoib_pkey_poll(struct work_struct *work)
788 {
789         struct ipoib_dev_priv *priv =
790                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
791         struct net_device *dev = priv->dev;
792
793         ipoib_pkey_dev_check_presence(dev);
794
795         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
796                 ipoib_open(dev);
797         else {
798                 mutex_lock(&pkey_mutex);
799                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
800                         queue_delayed_work(ipoib_workqueue,
801                                            &priv->pkey_poll_task,
802                                            HZ);
803                 mutex_unlock(&pkey_mutex);
804         }
805 }
806
807 int ipoib_pkey_dev_delay_open(struct net_device *dev)
808 {
809         struct ipoib_dev_priv *priv = netdev_priv(dev);
810
811         /* Look for the interface pkey value in the IB Port P_Key table and */
812         /* set the interface pkey assigment flag                            */
813         ipoib_pkey_dev_check_presence(dev);
814
815         /* P_Key value not assigned yet - start polling */
816         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
817                 mutex_lock(&pkey_mutex);
818                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
819                 queue_delayed_work(ipoib_workqueue,
820                                    &priv->pkey_poll_task,
821                                    HZ);
822                 mutex_unlock(&pkey_mutex);
823                 return 1;
824         }
825
826         return 0;
827 }