Pull style into test branch
[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 #define IPOIB_OP_RECV   (1ul << 31)
54
55 static DEFINE_MUTEX(pkey_mutex);
56
57 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
58                                  struct ib_pd *pd, struct ib_ah_attr *attr)
59 {
60         struct ipoib_ah *ah;
61
62         ah = kmalloc(sizeof *ah, GFP_KERNEL);
63         if (!ah)
64                 return NULL;
65
66         ah->dev       = dev;
67         ah->last_send = 0;
68         kref_init(&ah->ref);
69
70         ah->ah = ib_create_ah(pd, attr);
71         if (IS_ERR(ah->ah)) {
72                 kfree(ah);
73                 ah = NULL;
74         } else
75                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
76
77         return ah;
78 }
79
80 void ipoib_free_ah(struct kref *kref)
81 {
82         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
83         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
84
85         unsigned long flags;
86
87         spin_lock_irqsave(&priv->lock, flags);
88         list_add_tail(&ah->list, &priv->dead_ahs);
89         spin_unlock_irqrestore(&priv->lock, flags);
90 }
91
92 static int ipoib_ib_post_receive(struct net_device *dev, int id)
93 {
94         struct ipoib_dev_priv *priv = netdev_priv(dev);
95         struct ib_sge list;
96         struct ib_recv_wr param;
97         struct ib_recv_wr *bad_wr;
98         int ret;
99
100         list.addr     = priv->rx_ring[id].mapping;
101         list.length   = IPOIB_BUF_SIZE;
102         list.lkey     = priv->mr->lkey;
103
104         param.next    = NULL;
105         param.wr_id   = id | IPOIB_OP_RECV;
106         param.sg_list = &list;
107         param.num_sge = 1;
108
109         ret = ib_post_recv(priv->qp, &param, &bad_wr);
110         if (unlikely(ret)) {
111                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
112                 ib_dma_unmap_single(priv->ca, priv->rx_ring[id].mapping,
113                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
114                 dev_kfree_skb_any(priv->rx_ring[id].skb);
115                 priv->rx_ring[id].skb = NULL;
116         }
117
118         return ret;
119 }
120
121 static int ipoib_alloc_rx_skb(struct net_device *dev, int id)
122 {
123         struct ipoib_dev_priv *priv = netdev_priv(dev);
124         struct sk_buff *skb;
125         u64 addr;
126
127         skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
128         if (!skb)
129                 return -ENOMEM;
130
131         /*
132          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
133          * header.  So we need 4 more bytes to get to 48 and align the
134          * IP header to a multiple of 16.
135          */
136         skb_reserve(skb, 4);
137
138         addr = ib_dma_map_single(priv->ca, skb->data, IPOIB_BUF_SIZE,
139                                  DMA_FROM_DEVICE);
140         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
141                 dev_kfree_skb_any(skb);
142                 return -EIO;
143         }
144
145         priv->rx_ring[id].skb     = skb;
146         priv->rx_ring[id].mapping = addr;
147
148         return 0;
149 }
150
151 static int ipoib_ib_post_receives(struct net_device *dev)
152 {
153         struct ipoib_dev_priv *priv = netdev_priv(dev);
154         int i;
155
156         for (i = 0; i < ipoib_recvq_size; ++i) {
157                 if (ipoib_alloc_rx_skb(dev, i)) {
158                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
159                         return -ENOMEM;
160                 }
161                 if (ipoib_ib_post_receive(dev, i)) {
162                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
163                         return -EIO;
164                 }
165         }
166
167         return 0;
168 }
169
170 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
171 {
172         struct ipoib_dev_priv *priv = netdev_priv(dev);
173         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
174         struct sk_buff *skb;
175         u64 addr;
176
177         ipoib_dbg_data(priv, "recv completion: id %d, op %d, status: %d\n",
178                        wr_id, wc->opcode, wc->status);
179
180         if (unlikely(wr_id >= ipoib_recvq_size)) {
181                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
182                            wr_id, ipoib_recvq_size);
183                 return;
184         }
185
186         skb  = priv->rx_ring[wr_id].skb;
187         addr = priv->rx_ring[wr_id].mapping;
188
189         if (unlikely(wc->status != IB_WC_SUCCESS)) {
190                 if (wc->status != IB_WC_WR_FLUSH_ERR)
191                         ipoib_warn(priv, "failed recv event "
192                                    "(status=%d, wrid=%d vend_err %x)\n",
193                                    wc->status, wr_id, wc->vendor_err);
194                 ib_dma_unmap_single(priv->ca, addr,
195                                     IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
196                 dev_kfree_skb_any(skb);
197                 priv->rx_ring[wr_id].skb = NULL;
198                 return;
199         }
200
201         /*
202          * If we can't allocate a new RX buffer, dump
203          * this packet and reuse the old buffer.
204          */
205         if (unlikely(ipoib_alloc_rx_skb(dev, wr_id))) {
206                 ++priv->stats.rx_dropped;
207                 goto repost;
208         }
209
210         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
211                        wc->byte_len, wc->slid);
212
213         ib_dma_unmap_single(priv->ca, addr, IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
214
215         skb_put(skb, wc->byte_len);
216         skb_pull(skb, IB_GRH_BYTES);
217
218         if (wc->slid != priv->local_lid ||
219             wc->src_qp != priv->qp->qp_num) {
220                 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
221                 skb->mac.raw = skb->data;
222                 skb_pull(skb, IPOIB_ENCAP_LEN);
223
224                 dev->last_rx = jiffies;
225                 ++priv->stats.rx_packets;
226                 priv->stats.rx_bytes += skb->len;
227
228                 skb->dev = dev;
229                 /* XXX get correct PACKET_ type here */
230                 skb->pkt_type = PACKET_HOST;
231                 netif_rx_ni(skb);
232         } else {
233                 ipoib_dbg_data(priv, "dropping loopback packet\n");
234                 dev_kfree_skb_any(skb);
235         }
236
237 repost:
238         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
239                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
240                            "for buf %d\n", wr_id);
241 }
242
243 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
244 {
245         struct ipoib_dev_priv *priv = netdev_priv(dev);
246         unsigned int wr_id = wc->wr_id;
247         struct ipoib_tx_buf *tx_req;
248         unsigned long flags;
249
250         ipoib_dbg_data(priv, "send completion: id %d, op %d, status: %d\n",
251                        wr_id, wc->opcode, wc->status);
252
253         if (unlikely(wr_id >= ipoib_sendq_size)) {
254                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
255                            wr_id, ipoib_sendq_size);
256                 return;
257         }
258
259         tx_req = &priv->tx_ring[wr_id];
260
261         ib_dma_unmap_single(priv->ca, tx_req->mapping,
262                             tx_req->skb->len, DMA_TO_DEVICE);
263
264         ++priv->stats.tx_packets;
265         priv->stats.tx_bytes += tx_req->skb->len;
266
267         dev_kfree_skb_any(tx_req->skb);
268
269         spin_lock_irqsave(&priv->tx_lock, flags);
270         ++priv->tx_tail;
271         if (netif_queue_stopped(dev) &&
272             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags) &&
273             priv->tx_head - priv->tx_tail <= ipoib_sendq_size >> 1)
274                 netif_wake_queue(dev);
275         spin_unlock_irqrestore(&priv->tx_lock, flags);
276
277         if (wc->status != IB_WC_SUCCESS &&
278             wc->status != IB_WC_WR_FLUSH_ERR)
279                 ipoib_warn(priv, "failed send event "
280                            "(status=%d, wrid=%d vend_err %x)\n",
281                            wc->status, wr_id, wc->vendor_err);
282 }
283
284 static void ipoib_ib_handle_wc(struct net_device *dev, struct ib_wc *wc)
285 {
286         if (wc->wr_id & IPOIB_OP_RECV)
287                 ipoib_ib_handle_rx_wc(dev, wc);
288         else
289                 ipoib_ib_handle_tx_wc(dev, wc);
290 }
291
292 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
293 {
294         struct net_device *dev = (struct net_device *) dev_ptr;
295         struct ipoib_dev_priv *priv = netdev_priv(dev);
296         int n, i;
297
298         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
299         do {
300                 n = ib_poll_cq(cq, IPOIB_NUM_WC, priv->ibwc);
301                 for (i = 0; i < n; ++i)
302                         ipoib_ib_handle_wc(dev, priv->ibwc + i);
303         } while (n == IPOIB_NUM_WC);
304 }
305
306 static inline int post_send(struct ipoib_dev_priv *priv,
307                             unsigned int wr_id,
308                             struct ib_ah *address, u32 qpn,
309                             u64 addr, int len)
310 {
311         struct ib_send_wr *bad_wr;
312
313         priv->tx_sge.addr             = addr;
314         priv->tx_sge.length           = len;
315
316         priv->tx_wr.wr_id             = wr_id;
317         priv->tx_wr.wr.ud.remote_qpn  = qpn;
318         priv->tx_wr.wr.ud.ah          = address;
319
320         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
321 }
322
323 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
324                 struct ipoib_ah *address, u32 qpn)
325 {
326         struct ipoib_dev_priv *priv = netdev_priv(dev);
327         struct ipoib_tx_buf *tx_req;
328         u64 addr;
329
330         if (unlikely(skb->len > dev->mtu + INFINIBAND_ALEN)) {
331                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
332                            skb->len, dev->mtu + INFINIBAND_ALEN);
333                 ++priv->stats.tx_dropped;
334                 ++priv->stats.tx_errors;
335                 dev_kfree_skb_any(skb);
336                 return;
337         }
338
339         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
340                        skb->len, address, qpn);
341
342         /*
343          * We put the skb into the tx_ring _before_ we call post_send()
344          * because it's entirely possible that the completion handler will
345          * run before we execute anything after the post_send().  That
346          * means we have to make sure everything is properly recorded and
347          * our state is consistent before we call post_send().
348          */
349         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
350         tx_req->skb = skb;
351         addr = ib_dma_map_single(priv->ca, skb->data, skb->len,
352                                  DMA_TO_DEVICE);
353         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
354                 ++priv->stats.tx_errors;
355                 dev_kfree_skb_any(skb);
356                 return;
357         }
358         tx_req->mapping = addr;
359
360         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
361                                address->ah, qpn, addr, skb->len))) {
362                 ipoib_warn(priv, "post_send failed\n");
363                 ++priv->stats.tx_errors;
364                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
365                 dev_kfree_skb_any(skb);
366         } else {
367                 dev->trans_start = jiffies;
368
369                 address->last_send = priv->tx_head;
370                 ++priv->tx_head;
371
372                 if (priv->tx_head - priv->tx_tail == ipoib_sendq_size) {
373                         ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
374                         netif_stop_queue(dev);
375                 }
376         }
377 }
378
379 static void __ipoib_reap_ah(struct net_device *dev)
380 {
381         struct ipoib_dev_priv *priv = netdev_priv(dev);
382         struct ipoib_ah *ah, *tah;
383         LIST_HEAD(remove_list);
384
385         spin_lock_irq(&priv->tx_lock);
386         spin_lock(&priv->lock);
387         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
388                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
389                         list_del(&ah->list);
390                         ib_destroy_ah(ah->ah);
391                         kfree(ah);
392                 }
393         spin_unlock(&priv->lock);
394         spin_unlock_irq(&priv->tx_lock);
395 }
396
397 void ipoib_reap_ah(struct work_struct *work)
398 {
399         struct ipoib_dev_priv *priv =
400                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
401         struct net_device *dev = priv->dev;
402
403         __ipoib_reap_ah(dev);
404
405         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
406                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
407 }
408
409 int ipoib_ib_dev_open(struct net_device *dev)
410 {
411         struct ipoib_dev_priv *priv = netdev_priv(dev);
412         int ret;
413
414         ret = ipoib_init_qp(dev);
415         if (ret) {
416                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
417                 return -1;
418         }
419
420         ret = ipoib_ib_post_receives(dev);
421         if (ret) {
422                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
423                 ipoib_ib_dev_stop(dev);
424                 return -1;
425         }
426
427         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
428         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
429
430         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
431
432         return 0;
433 }
434
435 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
436 {
437         struct ipoib_dev_priv *priv = netdev_priv(dev);
438         u16 pkey_index = 0;
439
440         if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
441                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
442         else
443                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
444 }
445
446 int ipoib_ib_dev_up(struct net_device *dev)
447 {
448         struct ipoib_dev_priv *priv = netdev_priv(dev);
449
450         ipoib_pkey_dev_check_presence(dev);
451
452         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
453                 ipoib_dbg(priv, "PKEY is not assigned.\n");
454                 return 0;
455         }
456
457         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
458
459         return ipoib_mcast_start_thread(dev);
460 }
461
462 int ipoib_ib_dev_down(struct net_device *dev, int flush)
463 {
464         struct ipoib_dev_priv *priv = netdev_priv(dev);
465
466         ipoib_dbg(priv, "downing ib_dev\n");
467
468         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
469         netif_carrier_off(dev);
470
471         /* Shutdown the P_Key thread if still active */
472         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
473                 mutex_lock(&pkey_mutex);
474                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
475                 cancel_delayed_work(&priv->pkey_task);
476                 mutex_unlock(&pkey_mutex);
477                 if (flush)
478                         flush_workqueue(ipoib_workqueue);
479         }
480
481         ipoib_mcast_stop_thread(dev, flush);
482         ipoib_mcast_dev_flush(dev);
483
484         ipoib_flush_paths(dev);
485
486         return 0;
487 }
488
489 static int recvs_pending(struct net_device *dev)
490 {
491         struct ipoib_dev_priv *priv = netdev_priv(dev);
492         int pending = 0;
493         int i;
494
495         for (i = 0; i < ipoib_recvq_size; ++i)
496                 if (priv->rx_ring[i].skb)
497                         ++pending;
498
499         return pending;
500 }
501
502 int ipoib_ib_dev_stop(struct net_device *dev)
503 {
504         struct ipoib_dev_priv *priv = netdev_priv(dev);
505         struct ib_qp_attr qp_attr;
506         unsigned long begin;
507         struct ipoib_tx_buf *tx_req;
508         int i;
509
510         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
511
512         /*
513          * Move our QP to the error state and then reinitialize in
514          * when all work requests have completed or have been flushed.
515          */
516         qp_attr.qp_state = IB_QPS_ERR;
517         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
518                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
519
520         /* Wait for all sends and receives to complete */
521         begin = jiffies;
522
523         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
524                 if (time_after(jiffies, begin + 5 * HZ)) {
525                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
526                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
527
528                         /*
529                          * assume the HW is wedged and just free up
530                          * all our pending work requests.
531                          */
532                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
533                                 tx_req = &priv->tx_ring[priv->tx_tail &
534                                                         (ipoib_sendq_size - 1)];
535                                 ib_dma_unmap_single(priv->ca,
536                                                     tx_req->mapping,
537                                                     tx_req->skb->len,
538                                                     DMA_TO_DEVICE);
539                                 dev_kfree_skb_any(tx_req->skb);
540                                 ++priv->tx_tail;
541                         }
542
543                         for (i = 0; i < ipoib_recvq_size; ++i) {
544                                 struct ipoib_rx_buf *rx_req;
545
546                                 rx_req = &priv->rx_ring[i];
547                                 if (!rx_req->skb)
548                                         continue;
549                                 ib_dma_unmap_single(priv->ca,
550                                                     rx_req->mapping,
551                                                     IPOIB_BUF_SIZE,
552                                                     DMA_FROM_DEVICE);
553                                 dev_kfree_skb_any(rx_req->skb);
554                                 rx_req->skb = NULL;
555                         }
556
557                         goto timeout;
558                 }
559
560                 msleep(1);
561         }
562
563         ipoib_dbg(priv, "All sends and receives done.\n");
564
565 timeout:
566         qp_attr.qp_state = IB_QPS_RESET;
567         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
568                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
569
570         /* Wait for all AHs to be reaped */
571         set_bit(IPOIB_STOP_REAPER, &priv->flags);
572         cancel_delayed_work(&priv->ah_reap_task);
573         flush_workqueue(ipoib_workqueue);
574
575         begin = jiffies;
576
577         while (!list_empty(&priv->dead_ahs)) {
578                 __ipoib_reap_ah(dev);
579
580                 if (time_after(jiffies, begin + HZ)) {
581                         ipoib_warn(priv, "timing out; will leak address handles\n");
582                         break;
583                 }
584
585                 msleep(1);
586         }
587
588         return 0;
589 }
590
591 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
592 {
593         struct ipoib_dev_priv *priv = netdev_priv(dev);
594
595         priv->ca = ca;
596         priv->port = port;
597         priv->qp = NULL;
598
599         if (ipoib_transport_dev_init(dev, ca)) {
600                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
601                 return -ENODEV;
602         }
603
604         if (dev->flags & IFF_UP) {
605                 if (ipoib_ib_dev_open(dev)) {
606                         ipoib_transport_dev_cleanup(dev);
607                         return -ENODEV;
608                 }
609         }
610
611         return 0;
612 }
613
614 void ipoib_ib_dev_flush(struct work_struct *work)
615 {
616         struct ipoib_dev_priv *cpriv, *priv =
617                 container_of(work, struct ipoib_dev_priv, flush_task);
618         struct net_device *dev = priv->dev;
619
620         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) ) {
621                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
622                 return;
623         }
624
625         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
626                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
627                 return;
628         }
629
630         ipoib_dbg(priv, "flushing\n");
631
632         ipoib_ib_dev_down(dev, 0);
633
634         /*
635          * The device could have been brought down between the start and when
636          * we get here, don't bring it back up if it's not configured up
637          */
638         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
639                 ipoib_ib_dev_up(dev);
640                 ipoib_mcast_restart_task(&priv->restart_task);
641         }
642
643         mutex_lock(&priv->vlan_mutex);
644
645         /* Flush any child interfaces too */
646         list_for_each_entry(cpriv, &priv->child_intfs, list)
647                 ipoib_ib_dev_flush(&cpriv->flush_task);
648
649         mutex_unlock(&priv->vlan_mutex);
650 }
651
652 void ipoib_ib_dev_cleanup(struct net_device *dev)
653 {
654         struct ipoib_dev_priv *priv = netdev_priv(dev);
655
656         ipoib_dbg(priv, "cleaning up ib_dev\n");
657
658         ipoib_mcast_stop_thread(dev, 1);
659         ipoib_mcast_dev_flush(dev);
660
661         ipoib_transport_dev_cleanup(dev);
662 }
663
664 /*
665  * Delayed P_Key Assigment Interim Support
666  *
667  * The following is initial implementation of delayed P_Key assigment
668  * mechanism. It is using the same approach implemented for the multicast
669  * group join. The single goal of this implementation is to quickly address
670  * Bug #2507. This implementation will probably be removed when the P_Key
671  * change async notification is available.
672  */
673
674 void ipoib_pkey_poll(struct work_struct *work)
675 {
676         struct ipoib_dev_priv *priv =
677                 container_of(work, struct ipoib_dev_priv, pkey_task.work);
678         struct net_device *dev = priv->dev;
679
680         ipoib_pkey_dev_check_presence(dev);
681
682         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
683                 ipoib_open(dev);
684         else {
685                 mutex_lock(&pkey_mutex);
686                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
687                         queue_delayed_work(ipoib_workqueue,
688                                            &priv->pkey_task,
689                                            HZ);
690                 mutex_unlock(&pkey_mutex);
691         }
692 }
693
694 int ipoib_pkey_dev_delay_open(struct net_device *dev)
695 {
696         struct ipoib_dev_priv *priv = netdev_priv(dev);
697
698         /* Look for the interface pkey value in the IB Port P_Key table and */
699         /* set the interface pkey assigment flag                            */
700         ipoib_pkey_dev_check_presence(dev);
701
702         /* P_Key value not assigned yet - start polling */
703         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
704                 mutex_lock(&pkey_mutex);
705                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
706                 queue_delayed_work(ipoib_workqueue,
707                                    &priv->pkey_task,
708                                    HZ);
709                 mutex_unlock(&pkey_mutex);
710                 return 1;
711         }
712
713         return 0;
714 }