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 	 * Drop packets that this interface sent, ie multicast packets
201 	 * that the HCA has replicated.
202 	 */
203 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
204 		goto repost;
205 
206 	/*
207 	 * If we can't allocate a new RX buffer, dump
208 	 * this packet and reuse the old buffer.
209 	 */
210 	if (unlikely(ipoib_alloc_rx_skb(dev, wr_id))) {
211 		++dev->stats.rx_dropped;
212 		goto repost;
213 	}
214 
215 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
216 		       wc->byte_len, wc->slid);
217 
218 	ib_dma_unmap_single(priv->ca, addr, IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
219 
220 	skb_put(skb, wc->byte_len);
221 	skb_pull(skb, IB_GRH_BYTES);
222 
223 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
224 	skb_reset_mac_header(skb);
225 	skb_pull(skb, IPOIB_ENCAP_LEN);
226 
227 	dev->last_rx = jiffies;
228 	++dev->stats.rx_packets;
229 	dev->stats.rx_bytes += skb->len;
230 
231 	skb->dev = dev;
232 	/* XXX get correct PACKET_ type here */
233 	skb->pkt_type = PACKET_HOST;
234 	netif_receive_skb(skb);
235 
236 repost:
237 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
238 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
239 			   "for buf %d\n", wr_id);
240 }
241 
242 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
243 {
244 	struct ipoib_dev_priv *priv = netdev_priv(dev);
245 	unsigned int wr_id = wc->wr_id;
246 	struct ipoib_tx_buf *tx_req;
247 	unsigned long flags;
248 
249 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
250 		       wr_id, wc->status);
251 
252 	if (unlikely(wr_id >= ipoib_sendq_size)) {
253 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
254 			   wr_id, ipoib_sendq_size);
255 		return;
256 	}
257 
258 	tx_req = &priv->tx_ring[wr_id];
259 
260 	ib_dma_unmap_single(priv->ca, tx_req->mapping,
261 			    tx_req->skb->len, DMA_TO_DEVICE);
262 
263 	++dev->stats.tx_packets;
264 	dev->stats.tx_bytes += tx_req->skb->len;
265 
266 	dev_kfree_skb_any(tx_req->skb);
267 
268 	spin_lock_irqsave(&priv->tx_lock, flags);
269 	++priv->tx_tail;
270 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
271 	    netif_queue_stopped(dev) &&
272 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
273 		netif_wake_queue(dev);
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 napi_struct *napi, int budget)
284 {
285 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
286 	struct net_device *dev = priv->dev;
287 	int done;
288 	int t;
289 	int n, i;
290 
291 	done  = 0;
292 
293 poll_more:
294 	while (done < budget) {
295 		int max = (budget - done);
296 
297 		t = min(IPOIB_NUM_WC, max);
298 		n = ib_poll_cq(priv->cq, t, priv->ibwc);
299 
300 		for (i = 0; i < n; i++) {
301 			struct ib_wc *wc = priv->ibwc + i;
302 
303 			if (wc->wr_id & IPOIB_OP_RECV) {
304 				++done;
305 				if (wc->wr_id & IPOIB_OP_CM)
306 					ipoib_cm_handle_rx_wc(dev, wc);
307 				else
308 					ipoib_ib_handle_rx_wc(dev, wc);
309 			} else {
310 				if (wc->wr_id & IPOIB_OP_CM)
311 					ipoib_cm_handle_tx_wc(dev, wc);
312 				else
313 					ipoib_ib_handle_tx_wc(dev, wc);
314 			}
315 		}
316 
317 		if (n != t)
318 			break;
319 	}
320 
321 	if (done < budget) {
322 		netif_rx_complete(dev, napi);
323 		if (unlikely(ib_req_notify_cq(priv->cq,
324 					      IB_CQ_NEXT_COMP |
325 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
326 		    netif_rx_reschedule(dev, napi))
327 			goto poll_more;
328 	}
329 
330 	return done;
331 }
332 
333 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
334 {
335 	struct net_device *dev = dev_ptr;
336 	struct ipoib_dev_priv *priv = netdev_priv(dev);
337 
338 	netif_rx_schedule(dev, &priv->napi);
339 }
340 
341 static inline int post_send(struct ipoib_dev_priv *priv,
342 			    unsigned int wr_id,
343 			    struct ib_ah *address, u32 qpn,
344 			    u64 addr, int len)
345 {
346 	struct ib_send_wr *bad_wr;
347 
348 	priv->tx_sge.addr	      = addr;
349 	priv->tx_sge.length	      = len;
350 
351 	priv->tx_wr.wr_id	      = wr_id;
352 	priv->tx_wr.wr.ud.remote_qpn  = qpn;
353 	priv->tx_wr.wr.ud.ah	      = address;
354 
355 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
356 }
357 
358 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
359 		struct ipoib_ah *address, u32 qpn)
360 {
361 	struct ipoib_dev_priv *priv = netdev_priv(dev);
362 	struct ipoib_tx_buf *tx_req;
363 	u64 addr;
364 
365 	if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
366 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
367 			   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
368 		++dev->stats.tx_dropped;
369 		++dev->stats.tx_errors;
370 		ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
371 		return;
372 	}
373 
374 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
375 		       skb->len, address, qpn);
376 
377 	/*
378 	 * We put the skb into the tx_ring _before_ we call post_send()
379 	 * because it's entirely possible that the completion handler will
380 	 * run before we execute anything after the post_send().  That
381 	 * means we have to make sure everything is properly recorded and
382 	 * our state is consistent before we call post_send().
383 	 */
384 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
385 	tx_req->skb = skb;
386 	addr = ib_dma_map_single(priv->ca, skb->data, skb->len,
387 				 DMA_TO_DEVICE);
388 	if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
389 		++dev->stats.tx_errors;
390 		dev_kfree_skb_any(skb);
391 		return;
392 	}
393 	tx_req->mapping = addr;
394 
395 	if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
396 			       address->ah, qpn, addr, skb->len))) {
397 		ipoib_warn(priv, "post_send failed\n");
398 		++dev->stats.tx_errors;
399 		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
400 		dev_kfree_skb_any(skb);
401 	} else {
402 		dev->trans_start = jiffies;
403 
404 		address->last_send = priv->tx_head;
405 		++priv->tx_head;
406 
407 		if (++priv->tx_outstanding == ipoib_sendq_size) {
408 			ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
409 			netif_stop_queue(dev);
410 		}
411 	}
412 }
413 
414 static void __ipoib_reap_ah(struct net_device *dev)
415 {
416 	struct ipoib_dev_priv *priv = netdev_priv(dev);
417 	struct ipoib_ah *ah, *tah;
418 	LIST_HEAD(remove_list);
419 
420 	spin_lock_irq(&priv->tx_lock);
421 	spin_lock(&priv->lock);
422 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
423 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
424 			list_del(&ah->list);
425 			ib_destroy_ah(ah->ah);
426 			kfree(ah);
427 		}
428 	spin_unlock(&priv->lock);
429 	spin_unlock_irq(&priv->tx_lock);
430 }
431 
432 void ipoib_reap_ah(struct work_struct *work)
433 {
434 	struct ipoib_dev_priv *priv =
435 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
436 	struct net_device *dev = priv->dev;
437 
438 	__ipoib_reap_ah(dev);
439 
440 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
441 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
442 				   round_jiffies_relative(HZ));
443 }
444 
445 int ipoib_ib_dev_open(struct net_device *dev)
446 {
447 	struct ipoib_dev_priv *priv = netdev_priv(dev);
448 	int ret;
449 
450 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
451 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
452 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
453 		return -1;
454 	}
455 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
456 
457 	ret = ipoib_init_qp(dev);
458 	if (ret) {
459 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
460 		return -1;
461 	}
462 
463 	ret = ipoib_ib_post_receives(dev);
464 	if (ret) {
465 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
466 		ipoib_ib_dev_stop(dev, 1);
467 		return -1;
468 	}
469 
470 	ret = ipoib_cm_dev_open(dev);
471 	if (ret) {
472 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
473 		ipoib_ib_dev_stop(dev, 1);
474 		return -1;
475 	}
476 
477 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
478 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
479 			   round_jiffies_relative(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 void ipoib_drain_cq(struct net_device *dev)
554 {
555 	struct ipoib_dev_priv *priv = netdev_priv(dev);
556 	int i, n;
557 	do {
558 		n = ib_poll_cq(priv->cq, IPOIB_NUM_WC, priv->ibwc);
559 		for (i = 0; i < n; ++i) {
560 			/*
561 			 * Convert any successful completions to flush
562 			 * errors to avoid passing packets up the
563 			 * stack after bringing the device down.
564 			 */
565 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
566 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
567 
568 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
569 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
570 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
571 				else
572 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
573 			} else {
574 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
575 					ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
576 				else
577 					ipoib_ib_handle_tx_wc(dev, priv->ibwc + i);
578 			}
579 		}
580 	} while (n == IPOIB_NUM_WC);
581 }
582 
583 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
584 {
585 	struct ipoib_dev_priv *priv = netdev_priv(dev);
586 	struct ib_qp_attr qp_attr;
587 	unsigned long begin;
588 	struct ipoib_tx_buf *tx_req;
589 	int i;
590 
591 	clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
592 
593 	ipoib_cm_dev_stop(dev);
594 
595 	/*
596 	 * Move our QP to the error state and then reinitialize in
597 	 * when all work requests have completed or have been flushed.
598 	 */
599 	qp_attr.qp_state = IB_QPS_ERR;
600 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
601 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
602 
603 	/* Wait for all sends and receives to complete */
604 	begin = jiffies;
605 
606 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
607 		if (time_after(jiffies, begin + 5 * HZ)) {
608 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
609 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
610 
611 			/*
612 			 * assume the HW is wedged and just free up
613 			 * all our pending work requests.
614 			 */
615 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
616 				tx_req = &priv->tx_ring[priv->tx_tail &
617 							(ipoib_sendq_size - 1)];
618 				ib_dma_unmap_single(priv->ca,
619 						    tx_req->mapping,
620 						    tx_req->skb->len,
621 						    DMA_TO_DEVICE);
622 				dev_kfree_skb_any(tx_req->skb);
623 				++priv->tx_tail;
624 				--priv->tx_outstanding;
625 			}
626 
627 			for (i = 0; i < ipoib_recvq_size; ++i) {
628 				struct ipoib_rx_buf *rx_req;
629 
630 				rx_req = &priv->rx_ring[i];
631 				if (!rx_req->skb)
632 					continue;
633 				ib_dma_unmap_single(priv->ca,
634 						    rx_req->mapping,
635 						    IPOIB_BUF_SIZE,
636 						    DMA_FROM_DEVICE);
637 				dev_kfree_skb_any(rx_req->skb);
638 				rx_req->skb = NULL;
639 			}
640 
641 			goto timeout;
642 		}
643 
644 		ipoib_drain_cq(dev);
645 
646 		msleep(1);
647 	}
648 
649 	ipoib_dbg(priv, "All sends and receives done.\n");
650 
651 timeout:
652 	qp_attr.qp_state = IB_QPS_RESET;
653 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
654 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
655 
656 	/* Wait for all AHs to be reaped */
657 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
658 	cancel_delayed_work(&priv->ah_reap_task);
659 	if (flush)
660 		flush_workqueue(ipoib_workqueue);
661 
662 	begin = jiffies;
663 
664 	while (!list_empty(&priv->dead_ahs)) {
665 		__ipoib_reap_ah(dev);
666 
667 		if (time_after(jiffies, begin + HZ)) {
668 			ipoib_warn(priv, "timing out; will leak address handles\n");
669 			break;
670 		}
671 
672 		msleep(1);
673 	}
674 
675 	ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP);
676 
677 	return 0;
678 }
679 
680 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
681 {
682 	struct ipoib_dev_priv *priv = netdev_priv(dev);
683 
684 	priv->ca = ca;
685 	priv->port = port;
686 	priv->qp = NULL;
687 
688 	if (ipoib_transport_dev_init(dev, ca)) {
689 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
690 		return -ENODEV;
691 	}
692 
693 	if (dev->flags & IFF_UP) {
694 		if (ipoib_ib_dev_open(dev)) {
695 			ipoib_transport_dev_cleanup(dev);
696 			return -ENODEV;
697 		}
698 	}
699 
700 	return 0;
701 }
702 
703 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, int pkey_event)
704 {
705 	struct ipoib_dev_priv *cpriv;
706 	struct net_device *dev = priv->dev;
707 	u16 new_index;
708 
709 	mutex_lock(&priv->vlan_mutex);
710 
711 	/*
712 	 * Flush any child interfaces too -- they might be up even if
713 	 * the parent is down.
714 	 */
715 	list_for_each_entry(cpriv, &priv->child_intfs, list)
716 		__ipoib_ib_dev_flush(cpriv, pkey_event);
717 
718 	mutex_unlock(&priv->vlan_mutex);
719 
720 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
721 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
722 		return;
723 	}
724 
725 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
726 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
727 		return;
728 	}
729 
730 	if (pkey_event) {
731 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
732 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
733 			ipoib_ib_dev_down(dev, 0);
734 			ipoib_pkey_dev_delay_open(dev);
735 			return;
736 		}
737 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
738 
739 		/* restart QP only if P_Key index is changed */
740 		if (new_index == priv->pkey_index) {
741 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
742 			return;
743 		}
744 		priv->pkey_index = new_index;
745 	}
746 
747 	ipoib_dbg(priv, "flushing\n");
748 
749 	ipoib_ib_dev_down(dev, 0);
750 
751 	if (pkey_event) {
752 		ipoib_ib_dev_stop(dev, 0);
753 		ipoib_ib_dev_open(dev);
754 	}
755 
756 	/*
757 	 * The device could have been brought down between the start and when
758 	 * we get here, don't bring it back up if it's not configured up
759 	 */
760 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
761 		ipoib_ib_dev_up(dev);
762 		ipoib_mcast_restart_task(&priv->restart_task);
763 	}
764 }
765 
766 void ipoib_ib_dev_flush(struct work_struct *work)
767 {
768 	struct ipoib_dev_priv *priv =
769 		container_of(work, struct ipoib_dev_priv, flush_task);
770 
771 	ipoib_dbg(priv, "Flushing %s\n", priv->dev->name);
772 	__ipoib_ib_dev_flush(priv, 0);
773 }
774 
775 void ipoib_pkey_event(struct work_struct *work)
776 {
777 	struct ipoib_dev_priv *priv =
778 		container_of(work, struct ipoib_dev_priv, pkey_event_task);
779 
780 	ipoib_dbg(priv, "Flushing %s and restarting its QP\n", priv->dev->name);
781 	__ipoib_ib_dev_flush(priv, 1);
782 }
783 
784 void ipoib_ib_dev_cleanup(struct net_device *dev)
785 {
786 	struct ipoib_dev_priv *priv = netdev_priv(dev);
787 
788 	ipoib_dbg(priv, "cleaning up ib_dev\n");
789 
790 	ipoib_mcast_stop_thread(dev, 1);
791 	ipoib_mcast_dev_flush(dev);
792 
793 	ipoib_transport_dev_cleanup(dev);
794 }
795 
796 /*
797  * Delayed P_Key Assigment Interim Support
798  *
799  * The following is initial implementation of delayed P_Key assigment
800  * mechanism. It is using the same approach implemented for the multicast
801  * group join. The single goal of this implementation is to quickly address
802  * Bug #2507. This implementation will probably be removed when the P_Key
803  * change async notification is available.
804  */
805 
806 void ipoib_pkey_poll(struct work_struct *work)
807 {
808 	struct ipoib_dev_priv *priv =
809 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
810 	struct net_device *dev = priv->dev;
811 
812 	ipoib_pkey_dev_check_presence(dev);
813 
814 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
815 		ipoib_open(dev);
816 	else {
817 		mutex_lock(&pkey_mutex);
818 		if (!test_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 	}
824 }
825 
826 int ipoib_pkey_dev_delay_open(struct net_device *dev)
827 {
828 	struct ipoib_dev_priv *priv = netdev_priv(dev);
829 
830 	/* Look for the interface pkey value in the IB Port P_Key table and */
831 	/* set the interface pkey assigment flag                            */
832 	ipoib_pkey_dev_check_presence(dev);
833 
834 	/* P_Key value not assigned yet - start polling */
835 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
836 		mutex_lock(&pkey_mutex);
837 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
838 		queue_delayed_work(ipoib_workqueue,
839 				   &priv->pkey_poll_task,
840 				   HZ);
841 		mutex_unlock(&pkey_mutex);
842 		return 1;
843 	}
844 
845 	return 0;
846 }
847