xref: /openbmc/linux/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision 96de0e252cedffad61b3cb5e05662c591898e69a)
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(test_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags)) &&
271 	    priv->tx_head - priv->tx_tail <= ipoib_sendq_size >> 1) {
272 		clear_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
273 		netif_wake_queue(dev);
274 	}
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 int ipoib_poll(struct napi_struct *napi, int budget)
285 {
286 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
287 	struct net_device *dev = priv->dev;
288 	int done;
289 	int t;
290 	int n, i;
291 
292 	done  = 0;
293 
294 poll_more:
295 	while (done < budget) {
296 		int max = (budget - done);
297 
298 		t = min(IPOIB_NUM_WC, max);
299 		n = ib_poll_cq(priv->cq, t, priv->ibwc);
300 
301 		for (i = 0; i < n; i++) {
302 			struct ib_wc *wc = priv->ibwc + i;
303 
304 			if (wc->wr_id & IPOIB_CM_OP_SRQ) {
305 				++done;
306 				ipoib_cm_handle_rx_wc(dev, wc);
307 			} else if (wc->wr_id & IPOIB_OP_RECV) {
308 				++done;
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 			break;
316 	}
317 
318 	if (done < budget) {
319 		netif_rx_complete(dev, napi);
320 		if (unlikely(ib_req_notify_cq(priv->cq,
321 					      IB_CQ_NEXT_COMP |
322 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
323 		    netif_rx_reschedule(dev, napi))
324 			goto poll_more;
325 	}
326 
327 	return done;
328 }
329 
330 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
331 {
332 	struct net_device *dev = dev_ptr;
333 	struct ipoib_dev_priv *priv = netdev_priv(dev);
334 
335 	netif_rx_schedule(dev, &priv->napi);
336 }
337 
338 static inline int post_send(struct ipoib_dev_priv *priv,
339 			    unsigned int wr_id,
340 			    struct ib_ah *address, u32 qpn,
341 			    u64 addr, int len)
342 {
343 	struct ib_send_wr *bad_wr;
344 
345 	priv->tx_sge.addr             = addr;
346 	priv->tx_sge.length           = len;
347 
348 	priv->tx_wr.wr_id 	      = wr_id;
349 	priv->tx_wr.wr.ud.remote_qpn  = qpn;
350 	priv->tx_wr.wr.ud.ah 	      = address;
351 
352 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
353 }
354 
355 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
356 		struct ipoib_ah *address, u32 qpn)
357 {
358 	struct ipoib_dev_priv *priv = netdev_priv(dev);
359 	struct ipoib_tx_buf *tx_req;
360 	u64 addr;
361 
362 	if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
363 		ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
364 			   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
365 		++dev->stats.tx_dropped;
366 		++dev->stats.tx_errors;
367 		ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
368 		return;
369 	}
370 
371 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
372 		       skb->len, address, qpn);
373 
374 	/*
375 	 * We put the skb into the tx_ring _before_ we call post_send()
376 	 * because it's entirely possible that the completion handler will
377 	 * run before we execute anything after the post_send().  That
378 	 * means we have to make sure everything is properly recorded and
379 	 * our state is consistent before we call post_send().
380 	 */
381 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
382 	tx_req->skb = skb;
383 	addr = ib_dma_map_single(priv->ca, skb->data, skb->len,
384 				 DMA_TO_DEVICE);
385 	if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
386 		++dev->stats.tx_errors;
387 		dev_kfree_skb_any(skb);
388 		return;
389 	}
390 	tx_req->mapping = addr;
391 
392 	if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
393 			       address->ah, qpn, addr, skb->len))) {
394 		ipoib_warn(priv, "post_send failed\n");
395 		++dev->stats.tx_errors;
396 		ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
397 		dev_kfree_skb_any(skb);
398 	} else {
399 		dev->trans_start = jiffies;
400 
401 		address->last_send = priv->tx_head;
402 		++priv->tx_head;
403 
404 		if (priv->tx_head - priv->tx_tail == ipoib_sendq_size) {
405 			ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
406 			netif_stop_queue(dev);
407 			set_bit(IPOIB_FLAG_NETIF_STOPPED, &priv->flags);
408 		}
409 	}
410 }
411 
412 static void __ipoib_reap_ah(struct net_device *dev)
413 {
414 	struct ipoib_dev_priv *priv = netdev_priv(dev);
415 	struct ipoib_ah *ah, *tah;
416 	LIST_HEAD(remove_list);
417 
418 	spin_lock_irq(&priv->tx_lock);
419 	spin_lock(&priv->lock);
420 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
421 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
422 			list_del(&ah->list);
423 			ib_destroy_ah(ah->ah);
424 			kfree(ah);
425 		}
426 	spin_unlock(&priv->lock);
427 	spin_unlock_irq(&priv->tx_lock);
428 }
429 
430 void ipoib_reap_ah(struct work_struct *work)
431 {
432 	struct ipoib_dev_priv *priv =
433 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
434 	struct net_device *dev = priv->dev;
435 
436 	__ipoib_reap_ah(dev);
437 
438 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
439 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
440 }
441 
442 int ipoib_ib_dev_open(struct net_device *dev)
443 {
444 	struct ipoib_dev_priv *priv = netdev_priv(dev);
445 	int ret;
446 
447 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
448 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
449 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
450 		return -1;
451 	}
452 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
453 
454 	ret = ipoib_init_qp(dev);
455 	if (ret) {
456 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
457 		return -1;
458 	}
459 
460 	ret = ipoib_ib_post_receives(dev);
461 	if (ret) {
462 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
463 		ipoib_ib_dev_stop(dev, 1);
464 		return -1;
465 	}
466 
467 	ret = ipoib_cm_dev_open(dev);
468 	if (ret) {
469 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
470 		ipoib_ib_dev_stop(dev, 1);
471 		return -1;
472 	}
473 
474 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
475 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
476 
477 	set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
478 
479 	return 0;
480 }
481 
482 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
483 {
484 	struct ipoib_dev_priv *priv = netdev_priv(dev);
485 	u16 pkey_index = 0;
486 
487 	if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
488 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
489 	else
490 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
491 }
492 
493 int ipoib_ib_dev_up(struct net_device *dev)
494 {
495 	struct ipoib_dev_priv *priv = netdev_priv(dev);
496 
497 	ipoib_pkey_dev_check_presence(dev);
498 
499 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
500 		ipoib_dbg(priv, "PKEY is not assigned.\n");
501 		return 0;
502 	}
503 
504 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
505 
506 	return ipoib_mcast_start_thread(dev);
507 }
508 
509 int ipoib_ib_dev_down(struct net_device *dev, int flush)
510 {
511 	struct ipoib_dev_priv *priv = netdev_priv(dev);
512 
513 	ipoib_dbg(priv, "downing ib_dev\n");
514 
515 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
516 	netif_carrier_off(dev);
517 
518 	/* Shutdown the P_Key thread if still active */
519 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
520 		mutex_lock(&pkey_mutex);
521 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
522 		cancel_delayed_work(&priv->pkey_poll_task);
523 		mutex_unlock(&pkey_mutex);
524 		if (flush)
525 			flush_workqueue(ipoib_workqueue);
526 	}
527 
528 	ipoib_mcast_stop_thread(dev, flush);
529 	ipoib_mcast_dev_flush(dev);
530 
531 	ipoib_flush_paths(dev);
532 
533 	return 0;
534 }
535 
536 static int recvs_pending(struct net_device *dev)
537 {
538 	struct ipoib_dev_priv *priv = netdev_priv(dev);
539 	int pending = 0;
540 	int i;
541 
542 	for (i = 0; i < ipoib_recvq_size; ++i)
543 		if (priv->rx_ring[i].skb)
544 			++pending;
545 
546 	return pending;
547 }
548 
549 void ipoib_drain_cq(struct net_device *dev)
550 {
551 	struct ipoib_dev_priv *priv = netdev_priv(dev);
552 	int i, n;
553 	do {
554 		n = ib_poll_cq(priv->cq, IPOIB_NUM_WC, priv->ibwc);
555 		for (i = 0; i < n; ++i) {
556 			/*
557 			 * Convert any successful completions to flush
558 			 * errors to avoid passing packets up the
559 			 * stack after bringing the device down.
560 			 */
561 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
562 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
563 
564 			if (priv->ibwc[i].wr_id & IPOIB_CM_OP_SRQ)
565 				ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
566 			else if (priv->ibwc[i].wr_id & IPOIB_OP_RECV)
567 				ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
568 			else
569 				ipoib_ib_handle_tx_wc(dev, priv->ibwc + i);
570 		}
571 	} while (n == IPOIB_NUM_WC);
572 }
573 
574 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
575 {
576 	struct ipoib_dev_priv *priv = netdev_priv(dev);
577 	struct ib_qp_attr qp_attr;
578 	unsigned long begin;
579 	struct ipoib_tx_buf *tx_req;
580 	int i;
581 
582 	clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
583 
584 	ipoib_cm_dev_stop(dev);
585 
586 	/*
587 	 * Move our QP to the error state and then reinitialize in
588 	 * when all work requests have completed or have been flushed.
589 	 */
590 	qp_attr.qp_state = IB_QPS_ERR;
591 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
592 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
593 
594 	/* Wait for all sends and receives to complete */
595 	begin = jiffies;
596 
597 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
598 		if (time_after(jiffies, begin + 5 * HZ)) {
599 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
600 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
601 
602 			/*
603 			 * assume the HW is wedged and just free up
604 			 * all our pending work requests.
605 			 */
606 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
607 				tx_req = &priv->tx_ring[priv->tx_tail &
608 							(ipoib_sendq_size - 1)];
609 				ib_dma_unmap_single(priv->ca,
610 						    tx_req->mapping,
611 						    tx_req->skb->len,
612 						    DMA_TO_DEVICE);
613 				dev_kfree_skb_any(tx_req->skb);
614 				++priv->tx_tail;
615 			}
616 
617 			for (i = 0; i < ipoib_recvq_size; ++i) {
618 				struct ipoib_rx_buf *rx_req;
619 
620 				rx_req = &priv->rx_ring[i];
621 				if (!rx_req->skb)
622 					continue;
623 				ib_dma_unmap_single(priv->ca,
624 						    rx_req->mapping,
625 						    IPOIB_BUF_SIZE,
626 						    DMA_FROM_DEVICE);
627 				dev_kfree_skb_any(rx_req->skb);
628 				rx_req->skb = NULL;
629 			}
630 
631 			goto timeout;
632 		}
633 
634 		ipoib_drain_cq(dev);
635 
636 		msleep(1);
637 	}
638 
639 	ipoib_dbg(priv, "All sends and receives done.\n");
640 
641 timeout:
642 	qp_attr.qp_state = IB_QPS_RESET;
643 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
644 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
645 
646 	/* Wait for all AHs to be reaped */
647 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
648 	cancel_delayed_work(&priv->ah_reap_task);
649 	if (flush)
650 		flush_workqueue(ipoib_workqueue);
651 
652 	begin = jiffies;
653 
654 	while (!list_empty(&priv->dead_ahs)) {
655 		__ipoib_reap_ah(dev);
656 
657 		if (time_after(jiffies, begin + HZ)) {
658 			ipoib_warn(priv, "timing out; will leak address handles\n");
659 			break;
660 		}
661 
662 		msleep(1);
663 	}
664 
665 	ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP);
666 
667 	return 0;
668 }
669 
670 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
671 {
672 	struct ipoib_dev_priv *priv = netdev_priv(dev);
673 
674 	priv->ca = ca;
675 	priv->port = port;
676 	priv->qp = NULL;
677 
678 	if (ipoib_transport_dev_init(dev, ca)) {
679 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
680 		return -ENODEV;
681 	}
682 
683 	if (dev->flags & IFF_UP) {
684 		if (ipoib_ib_dev_open(dev)) {
685 			ipoib_transport_dev_cleanup(dev);
686 			return -ENODEV;
687 		}
688 	}
689 
690 	return 0;
691 }
692 
693 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, int pkey_event)
694 {
695 	struct ipoib_dev_priv *cpriv;
696 	struct net_device *dev = priv->dev;
697 	u16 new_index;
698 
699 	mutex_lock(&priv->vlan_mutex);
700 
701 	/*
702 	 * Flush any child interfaces too -- they might be up even if
703 	 * the parent is down.
704 	 */
705 	list_for_each_entry(cpriv, &priv->child_intfs, list)
706 		__ipoib_ib_dev_flush(cpriv, pkey_event);
707 
708 	mutex_unlock(&priv->vlan_mutex);
709 
710 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
711 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
712 		return;
713 	}
714 
715 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
716 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
717 		return;
718 	}
719 
720 	if (pkey_event) {
721 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
722 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
723 			ipoib_ib_dev_down(dev, 0);
724 			ipoib_pkey_dev_delay_open(dev);
725 			return;
726 		}
727 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
728 
729 		/* restart QP only if P_Key index is changed */
730 		if (new_index == priv->pkey_index) {
731 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
732 			return;
733 		}
734 		priv->pkey_index = new_index;
735 	}
736 
737 	ipoib_dbg(priv, "flushing\n");
738 
739 	ipoib_ib_dev_down(dev, 0);
740 
741 	if (pkey_event) {
742 		ipoib_ib_dev_stop(dev, 0);
743 		ipoib_ib_dev_open(dev);
744 	}
745 
746 	/*
747 	 * The device could have been brought down between the start and when
748 	 * we get here, don't bring it back up if it's not configured up
749 	 */
750 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
751 		ipoib_ib_dev_up(dev);
752 		ipoib_mcast_restart_task(&priv->restart_task);
753 	}
754 }
755 
756 void ipoib_ib_dev_flush(struct work_struct *work)
757 {
758 	struct ipoib_dev_priv *priv =
759 		container_of(work, struct ipoib_dev_priv, flush_task);
760 
761 	ipoib_dbg(priv, "Flushing %s\n", priv->dev->name);
762 	__ipoib_ib_dev_flush(priv, 0);
763 }
764 
765 void ipoib_pkey_event(struct work_struct *work)
766 {
767 	struct ipoib_dev_priv *priv =
768 		container_of(work, struct ipoib_dev_priv, pkey_event_task);
769 
770 	ipoib_dbg(priv, "Flushing %s and restarting its QP\n", priv->dev->name);
771 	__ipoib_ib_dev_flush(priv, 1);
772 }
773 
774 void ipoib_ib_dev_cleanup(struct net_device *dev)
775 {
776 	struct ipoib_dev_priv *priv = netdev_priv(dev);
777 
778 	ipoib_dbg(priv, "cleaning up ib_dev\n");
779 
780 	ipoib_mcast_stop_thread(dev, 1);
781 	ipoib_mcast_dev_flush(dev);
782 
783 	ipoib_transport_dev_cleanup(dev);
784 }
785 
786 /*
787  * Delayed P_Key Assigment Interim Support
788  *
789  * The following is initial implementation of delayed P_Key assigment
790  * mechanism. It is using the same approach implemented for the multicast
791  * group join. The single goal of this implementation is to quickly address
792  * Bug #2507. This implementation will probably be removed when the P_Key
793  * change async notification is available.
794  */
795 
796 void ipoib_pkey_poll(struct work_struct *work)
797 {
798 	struct ipoib_dev_priv *priv =
799 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
800 	struct net_device *dev = priv->dev;
801 
802 	ipoib_pkey_dev_check_presence(dev);
803 
804 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
805 		ipoib_open(dev);
806 	else {
807 		mutex_lock(&pkey_mutex);
808 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
809 			queue_delayed_work(ipoib_workqueue,
810 					   &priv->pkey_poll_task,
811 					   HZ);
812 		mutex_unlock(&pkey_mutex);
813 	}
814 }
815 
816 int ipoib_pkey_dev_delay_open(struct net_device *dev)
817 {
818 	struct ipoib_dev_priv *priv = netdev_priv(dev);
819 
820 	/* Look for the interface pkey value in the IB Port P_Key table and */
821 	/* set the interface pkey assigment flag                            */
822 	ipoib_pkey_dev_check_presence(dev);
823 
824 	/* P_Key value not assigned yet - start polling */
825 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
826 		mutex_lock(&pkey_mutex);
827 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
828 		queue_delayed_work(ipoib_workqueue,
829 				   &priv->pkey_poll_task,
830 				   HZ);
831 		mutex_unlock(&pkey_mutex);
832 		return 1;
833 	}
834 
835 	return 0;
836 }
837