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 
36 #include <linux/delay.h>
37 #include <linux/moduleparam.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
40 
41 #include <linux/ip.h>
42 #include <linux/tcp.h>
43 #include <rdma/ib_cache.h>
44 
45 #include "ipoib.h"
46 
47 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
48 static int data_debug_level;
49 
50 module_param(data_debug_level, int, 0644);
51 MODULE_PARM_DESC(data_debug_level,
52 		 "Enable data path debug tracing if > 0");
53 #endif
54 
55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
56 				 struct ib_pd *pd, struct rdma_ah_attr *attr)
57 {
58 	struct ipoib_ah *ah;
59 	struct ib_ah *vah;
60 
61 	ah = kmalloc(sizeof(*ah), GFP_KERNEL);
62 	if (!ah)
63 		return ERR_PTR(-ENOMEM);
64 
65 	ah->dev       = dev;
66 	ah->last_send = 0;
67 	kref_init(&ah->ref);
68 
69 	vah = rdma_create_ah(pd, attr, RDMA_CREATE_AH_SLEEPABLE);
70 	if (IS_ERR(vah)) {
71 		kfree(ah);
72 		ah = (struct ipoib_ah *)vah;
73 	} else {
74 		ah->ah = vah;
75 		ipoib_dbg(ipoib_priv(dev), "Created ah %p\n", ah->ah);
76 	}
77 
78 	return ah;
79 }
80 
81 void ipoib_free_ah(struct kref *kref)
82 {
83 	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
84 	struct ipoib_dev_priv *priv = ipoib_priv(ah->dev);
85 
86 	unsigned long flags;
87 
88 	spin_lock_irqsave(&priv->lock, flags);
89 	list_add_tail(&ah->list, &priv->dead_ahs);
90 	spin_unlock_irqrestore(&priv->lock, flags);
91 }
92 
93 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
94 				  u64 mapping[IPOIB_UD_RX_SG])
95 {
96 	ib_dma_unmap_single(priv->ca, mapping[0],
97 			    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
98 			    DMA_FROM_DEVICE);
99 }
100 
101 static int ipoib_ib_post_receive(struct net_device *dev, int id)
102 {
103 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
104 	int ret;
105 
106 	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
107 	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
108 	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
109 
110 
111 	ret = ib_post_recv(priv->qp, &priv->rx_wr, NULL);
112 	if (unlikely(ret)) {
113 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
114 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
115 		dev_kfree_skb_any(priv->rx_ring[id].skb);
116 		priv->rx_ring[id].skb = NULL;
117 	}
118 
119 	return ret;
120 }
121 
122 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
123 {
124 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
125 	struct sk_buff *skb;
126 	int buf_size;
127 	u64 *mapping;
128 
129 	buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
130 
131 	skb = dev_alloc_skb(buf_size + IPOIB_HARD_LEN);
132 	if (unlikely(!skb))
133 		return NULL;
134 
135 	/*
136 	 * the IP header will be at IPOIP_HARD_LEN + IB_GRH_BYTES, that is
137 	 * 64 bytes aligned
138 	 */
139 	skb_reserve(skb, sizeof(struct ipoib_pseudo_header));
140 
141 	mapping = priv->rx_ring[id].mapping;
142 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
143 				       DMA_FROM_DEVICE);
144 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
145 		goto error;
146 
147 	priv->rx_ring[id].skb = skb;
148 	return skb;
149 error:
150 	dev_kfree_skb_any(skb);
151 	return NULL;
152 }
153 
154 static int ipoib_ib_post_receives(struct net_device *dev)
155 {
156 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
157 	int i;
158 
159 	for (i = 0; i < ipoib_recvq_size; ++i) {
160 		if (!ipoib_alloc_rx_skb(dev, i)) {
161 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
162 			return -ENOMEM;
163 		}
164 		if (ipoib_ib_post_receive(dev, i)) {
165 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
166 			return -EIO;
167 		}
168 	}
169 
170 	return 0;
171 }
172 
173 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
174 {
175 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
176 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
177 	struct sk_buff *skb;
178 	u64 mapping[IPOIB_UD_RX_SG];
179 	union ib_gid *dgid;
180 	union ib_gid *sgid;
181 
182 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
183 		       wr_id, wc->status);
184 
185 	if (unlikely(wr_id >= ipoib_recvq_size)) {
186 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
187 			   wr_id, ipoib_recvq_size);
188 		return;
189 	}
190 
191 	skb  = priv->rx_ring[wr_id].skb;
192 
193 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
194 		if (wc->status != IB_WC_WR_FLUSH_ERR)
195 			ipoib_warn(priv,
196 				   "failed recv event (status=%d, wrid=%d vend_err %#x)\n",
197 				   wc->status, wr_id, wc->vendor_err);
198 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
199 		dev_kfree_skb_any(skb);
200 		priv->rx_ring[wr_id].skb = NULL;
201 		return;
202 	}
203 
204 	memcpy(mapping, priv->rx_ring[wr_id].mapping,
205 	       IPOIB_UD_RX_SG * sizeof(*mapping));
206 
207 	/*
208 	 * If we can't allocate a new RX buffer, dump
209 	 * this packet and reuse the old buffer.
210 	 */
211 	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
212 		++dev->stats.rx_dropped;
213 		goto repost;
214 	}
215 
216 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
217 		       wc->byte_len, wc->slid);
218 
219 	ipoib_ud_dma_unmap_rx(priv, mapping);
220 
221 	skb_put(skb, wc->byte_len);
222 
223 	/* First byte of dgid signals multicast when 0xff */
224 	dgid = &((struct ib_grh *)skb->data)->dgid;
225 
226 	if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
227 		skb->pkt_type = PACKET_HOST;
228 	else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
229 		skb->pkt_type = PACKET_BROADCAST;
230 	else
231 		skb->pkt_type = PACKET_MULTICAST;
232 
233 	sgid = &((struct ib_grh *)skb->data)->sgid;
234 
235 	/*
236 	 * Drop packets that this interface sent, ie multicast packets
237 	 * that the HCA has replicated.
238 	 */
239 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) {
240 		int need_repost = 1;
241 
242 		if ((wc->wc_flags & IB_WC_GRH) &&
243 		    sgid->global.interface_id != priv->local_gid.global.interface_id)
244 			need_repost = 0;
245 
246 		if (need_repost) {
247 			dev_kfree_skb_any(skb);
248 			goto repost;
249 		}
250 	}
251 
252 	skb_pull(skb, IB_GRH_BYTES);
253 
254 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
255 	skb_add_pseudo_hdr(skb);
256 
257 	++dev->stats.rx_packets;
258 	dev->stats.rx_bytes += skb->len;
259 	if (skb->pkt_type == PACKET_MULTICAST)
260 		dev->stats.multicast++;
261 
262 	skb->dev = dev;
263 	if ((dev->features & NETIF_F_RXCSUM) &&
264 			likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
265 		skb->ip_summed = CHECKSUM_UNNECESSARY;
266 
267 	napi_gro_receive(&priv->recv_napi, skb);
268 
269 repost:
270 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
271 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
272 			   "for buf %d\n", wr_id);
273 }
274 
275 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
276 {
277 	struct sk_buff *skb = tx_req->skb;
278 	u64 *mapping = tx_req->mapping;
279 	int i;
280 	int off;
281 
282 	if (skb_headlen(skb)) {
283 		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
284 					       DMA_TO_DEVICE);
285 		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
286 			return -EIO;
287 
288 		off = 1;
289 	} else
290 		off = 0;
291 
292 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
293 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
294 		mapping[i + off] = ib_dma_map_page(ca,
295 						 skb_frag_page(frag),
296 						 skb_frag_off(frag),
297 						 skb_frag_size(frag),
298 						 DMA_TO_DEVICE);
299 		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
300 			goto partial_error;
301 	}
302 	return 0;
303 
304 partial_error:
305 	for (; i > 0; --i) {
306 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
307 
308 		ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
309 	}
310 
311 	if (off)
312 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
313 
314 	return -EIO;
315 }
316 
317 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
318 			struct ipoib_tx_buf *tx_req)
319 {
320 	struct sk_buff *skb = tx_req->skb;
321 	u64 *mapping = tx_req->mapping;
322 	int i;
323 	int off;
324 
325 	if (skb_headlen(skb)) {
326 		ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
327 				    DMA_TO_DEVICE);
328 		off = 1;
329 	} else
330 		off = 0;
331 
332 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
333 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
334 
335 		ib_dma_unmap_page(priv->ca, mapping[i + off],
336 				  skb_frag_size(frag), DMA_TO_DEVICE);
337 	}
338 }
339 
340 /*
341  * As the result of a completion error the QP Can be transferred to SQE states.
342  * The function checks if the (send)QP is in SQE state and
343  * moves it back to RTS state, that in order to have it functional again.
344  */
345 static void ipoib_qp_state_validate_work(struct work_struct *work)
346 {
347 	struct ipoib_qp_state_validate *qp_work =
348 		container_of(work, struct ipoib_qp_state_validate, work);
349 
350 	struct ipoib_dev_priv *priv = qp_work->priv;
351 	struct ib_qp_attr qp_attr;
352 	struct ib_qp_init_attr query_init_attr;
353 	int ret;
354 
355 	ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
356 	if (ret) {
357 		ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
358 			   __func__, ret);
359 		goto free_res;
360 	}
361 	pr_info("%s: QP: 0x%x is in state: %d\n",
362 		__func__, priv->qp->qp_num, qp_attr.qp_state);
363 
364 	/* currently support only in SQE->RTS transition*/
365 	if (qp_attr.qp_state == IB_QPS_SQE) {
366 		qp_attr.qp_state = IB_QPS_RTS;
367 
368 		ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
369 		if (ret) {
370 			pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
371 				ret, priv->qp->qp_num);
372 			goto free_res;
373 		}
374 		pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
375 			__func__, priv->qp->qp_num);
376 	} else {
377 		pr_warn("QP (%d) will stay in state: %d\n",
378 			priv->qp->qp_num, qp_attr.qp_state);
379 	}
380 
381 free_res:
382 	kfree(qp_work);
383 }
384 
385 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
386 {
387 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
388 	unsigned int wr_id = wc->wr_id;
389 	struct ipoib_tx_buf *tx_req;
390 
391 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
392 		       wr_id, wc->status);
393 
394 	if (unlikely(wr_id >= ipoib_sendq_size)) {
395 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
396 			   wr_id, ipoib_sendq_size);
397 		return;
398 	}
399 
400 	tx_req = &priv->tx_ring[wr_id];
401 
402 	ipoib_dma_unmap_tx(priv, tx_req);
403 
404 	++dev->stats.tx_packets;
405 	dev->stats.tx_bytes += tx_req->skb->len;
406 
407 	dev_kfree_skb_any(tx_req->skb);
408 
409 	++priv->tx_tail;
410 	++priv->global_tx_tail;
411 
412 	if (unlikely(netif_queue_stopped(dev) &&
413 		     ((priv->global_tx_head - priv->global_tx_tail) <=
414 		      ipoib_sendq_size >> 1) &&
415 		     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
416 		netif_wake_queue(dev);
417 
418 	if (wc->status != IB_WC_SUCCESS &&
419 	    wc->status != IB_WC_WR_FLUSH_ERR) {
420 		struct ipoib_qp_state_validate *qp_work;
421 		ipoib_warn(priv,
422 			   "failed send event (status=%d, wrid=%d vend_err %#x)\n",
423 			   wc->status, wr_id, wc->vendor_err);
424 		qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
425 		if (!qp_work)
426 			return;
427 
428 		INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
429 		qp_work->priv = priv;
430 		queue_work(priv->wq, &qp_work->work);
431 	}
432 }
433 
434 static int poll_tx(struct ipoib_dev_priv *priv)
435 {
436 	int n, i;
437 	struct ib_wc *wc;
438 
439 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
440 	for (i = 0; i < n; ++i) {
441 		wc = priv->send_wc + i;
442 		if (wc->wr_id & IPOIB_OP_CM)
443 			ipoib_cm_handle_tx_wc(priv->dev, priv->send_wc + i);
444 		else
445 			ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
446 	}
447 	return n == MAX_SEND_CQE;
448 }
449 
450 int ipoib_rx_poll(struct napi_struct *napi, int budget)
451 {
452 	struct ipoib_dev_priv *priv =
453 		container_of(napi, struct ipoib_dev_priv, recv_napi);
454 	struct net_device *dev = priv->dev;
455 	int done;
456 	int t;
457 	int n, i;
458 
459 	done  = 0;
460 
461 poll_more:
462 	while (done < budget) {
463 		int max = (budget - done);
464 
465 		t = min(IPOIB_NUM_WC, max);
466 		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
467 
468 		for (i = 0; i < n; i++) {
469 			struct ib_wc *wc = priv->ibwc + i;
470 
471 			if (wc->wr_id & IPOIB_OP_RECV) {
472 				++done;
473 				if (wc->wr_id & IPOIB_OP_CM)
474 					ipoib_cm_handle_rx_wc(dev, wc);
475 				else
476 					ipoib_ib_handle_rx_wc(dev, wc);
477 			} else {
478 				pr_warn("%s: Got unexpected wqe id\n", __func__);
479 			}
480 		}
481 
482 		if (n != t)
483 			break;
484 	}
485 
486 	if (done < budget) {
487 		napi_complete(napi);
488 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
489 					      IB_CQ_NEXT_COMP |
490 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
491 		    napi_reschedule(napi))
492 			goto poll_more;
493 	}
494 
495 	return done;
496 }
497 
498 int ipoib_tx_poll(struct napi_struct *napi, int budget)
499 {
500 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv,
501 						   send_napi);
502 	struct net_device *dev = priv->dev;
503 	int n, i;
504 	struct ib_wc *wc;
505 
506 poll_more:
507 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
508 
509 	for (i = 0; i < n; i++) {
510 		wc = priv->send_wc + i;
511 		if (wc->wr_id & IPOIB_OP_CM)
512 			ipoib_cm_handle_tx_wc(dev, wc);
513 		else
514 			ipoib_ib_handle_tx_wc(dev, wc);
515 	}
516 
517 	if (n < budget) {
518 		napi_complete(napi);
519 		if (unlikely(ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
520 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
521 		    napi_reschedule(napi))
522 			goto poll_more;
523 	}
524 	return n < 0 ? 0 : n;
525 }
526 
527 void ipoib_ib_rx_completion(struct ib_cq *cq, void *ctx_ptr)
528 {
529 	struct ipoib_dev_priv *priv = ctx_ptr;
530 
531 	napi_schedule(&priv->recv_napi);
532 }
533 
534 void ipoib_ib_tx_completion(struct ib_cq *cq, void *ctx_ptr)
535 {
536 	struct ipoib_dev_priv *priv = ctx_ptr;
537 
538 	napi_schedule(&priv->send_napi);
539 }
540 
541 static inline int post_send(struct ipoib_dev_priv *priv,
542 			    unsigned int wr_id,
543 			    struct ib_ah *address, u32 dqpn,
544 			    struct ipoib_tx_buf *tx_req,
545 			    void *head, int hlen)
546 {
547 	struct sk_buff *skb = tx_req->skb;
548 
549 	ipoib_build_sge(priv, tx_req);
550 
551 	priv->tx_wr.wr.wr_id	= wr_id;
552 	priv->tx_wr.remote_qpn	= dqpn;
553 	priv->tx_wr.ah		= address;
554 
555 	if (head) {
556 		priv->tx_wr.mss		= skb_shinfo(skb)->gso_size;
557 		priv->tx_wr.header	= head;
558 		priv->tx_wr.hlen	= hlen;
559 		priv->tx_wr.wr.opcode	= IB_WR_LSO;
560 	} else
561 		priv->tx_wr.wr.opcode	= IB_WR_SEND;
562 
563 	return ib_post_send(priv->qp, &priv->tx_wr.wr, NULL);
564 }
565 
566 int ipoib_send(struct net_device *dev, struct sk_buff *skb,
567 	       struct ib_ah *address, u32 dqpn)
568 {
569 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
570 	struct ipoib_tx_buf *tx_req;
571 	int hlen, rc;
572 	void *phead;
573 	unsigned int usable_sge = priv->max_send_sge - !!skb_headlen(skb);
574 
575 	if (skb_is_gso(skb)) {
576 		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
577 		phead = skb->data;
578 		if (unlikely(!skb_pull(skb, hlen))) {
579 			ipoib_warn(priv, "linear data too small\n");
580 			++dev->stats.tx_dropped;
581 			++dev->stats.tx_errors;
582 			dev_kfree_skb_any(skb);
583 			return -1;
584 		}
585 	} else {
586 		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
587 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
588 				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
589 			++dev->stats.tx_dropped;
590 			++dev->stats.tx_errors;
591 			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
592 			return -1;
593 		}
594 		phead = NULL;
595 		hlen  = 0;
596 	}
597 	if (skb_shinfo(skb)->nr_frags > usable_sge) {
598 		if (skb_linearize(skb) < 0) {
599 			ipoib_warn(priv, "skb could not be linearized\n");
600 			++dev->stats.tx_dropped;
601 			++dev->stats.tx_errors;
602 			dev_kfree_skb_any(skb);
603 			return -1;
604 		}
605 		/* Does skb_linearize return ok without reducing nr_frags? */
606 		if (skb_shinfo(skb)->nr_frags > usable_sge) {
607 			ipoib_warn(priv, "too many frags after skb linearize\n");
608 			++dev->stats.tx_dropped;
609 			++dev->stats.tx_errors;
610 			dev_kfree_skb_any(skb);
611 			return -1;
612 		}
613 	}
614 
615 	ipoib_dbg_data(priv,
616 		       "sending packet, length=%d address=%p dqpn=0x%06x\n",
617 		       skb->len, address, dqpn);
618 
619 	/*
620 	 * We put the skb into the tx_ring _before_ we call post_send()
621 	 * because it's entirely possible that the completion handler will
622 	 * run before we execute anything after the post_send().  That
623 	 * means we have to make sure everything is properly recorded and
624 	 * our state is consistent before we call post_send().
625 	 */
626 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
627 	tx_req->skb = skb;
628 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
629 		++dev->stats.tx_errors;
630 		dev_kfree_skb_any(skb);
631 		return -1;
632 	}
633 
634 	if (skb->ip_summed == CHECKSUM_PARTIAL)
635 		priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
636 	else
637 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
638 	/* increase the tx_head after send success, but use it for queue state */
639 	if ((priv->global_tx_head - priv->global_tx_tail) ==
640 	    ipoib_sendq_size - 1) {
641 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
642 		netif_stop_queue(dev);
643 	}
644 
645 	skb_orphan(skb);
646 	skb_dst_drop(skb);
647 
648 	if (netif_queue_stopped(dev))
649 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
650 				     IB_CQ_REPORT_MISSED_EVENTS) < 0)
651 			ipoib_warn(priv, "request notify on send CQ failed\n");
652 
653 	rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
654 		       address, dqpn, tx_req, phead, hlen);
655 	if (unlikely(rc)) {
656 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
657 		++dev->stats.tx_errors;
658 		ipoib_dma_unmap_tx(priv, tx_req);
659 		dev_kfree_skb_any(skb);
660 		if (netif_queue_stopped(dev))
661 			netif_wake_queue(dev);
662 		rc = 0;
663 	} else {
664 		netif_trans_update(dev);
665 
666 		rc = priv->tx_head;
667 		++priv->tx_head;
668 		++priv->global_tx_head;
669 	}
670 	return rc;
671 }
672 
673 static void __ipoib_reap_ah(struct net_device *dev)
674 {
675 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
676 	struct ipoib_ah *ah, *tah;
677 	unsigned long flags;
678 
679 	netif_tx_lock_bh(dev);
680 	spin_lock_irqsave(&priv->lock, flags);
681 
682 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
683 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
684 			list_del(&ah->list);
685 			rdma_destroy_ah(ah->ah, 0);
686 			kfree(ah);
687 		}
688 
689 	spin_unlock_irqrestore(&priv->lock, flags);
690 	netif_tx_unlock_bh(dev);
691 }
692 
693 void ipoib_reap_ah(struct work_struct *work)
694 {
695 	struct ipoib_dev_priv *priv =
696 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
697 	struct net_device *dev = priv->dev;
698 
699 	__ipoib_reap_ah(dev);
700 
701 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
702 		queue_delayed_work(priv->wq, &priv->ah_reap_task,
703 				   round_jiffies_relative(HZ));
704 }
705 
706 static void ipoib_flush_ah(struct net_device *dev)
707 {
708 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
709 
710 	cancel_delayed_work(&priv->ah_reap_task);
711 	flush_workqueue(priv->wq);
712 	ipoib_reap_ah(&priv->ah_reap_task.work);
713 }
714 
715 static void ipoib_stop_ah(struct net_device *dev)
716 {
717 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
718 
719 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
720 	ipoib_flush_ah(dev);
721 }
722 
723 static int recvs_pending(struct net_device *dev)
724 {
725 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
726 	int pending = 0;
727 	int i;
728 
729 	for (i = 0; i < ipoib_recvq_size; ++i)
730 		if (priv->rx_ring[i].skb)
731 			++pending;
732 
733 	return pending;
734 }
735 
736 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,
737 					struct ib_qp *qp,
738 					enum ib_qp_state new_state)
739 {
740 	struct ib_qp_attr qp_attr;
741 	struct ib_qp_init_attr query_init_attr;
742 	int ret;
743 
744 	ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);
745 	if (ret) {
746 		ipoib_warn(priv, "%s: Failed to query QP\n", __func__);
747 		return;
748 	}
749 	/* print according to the new-state and the previous state.*/
750 	if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET)
751 		ipoib_dbg(priv, "Failed modify QP, IB_QPS_RESET to IB_QPS_ERR, acceptable\n");
752 	else
753 		ipoib_warn(priv, "Failed to modify QP to state: %d from state: %d\n",
754 			   new_state, qp_attr.qp_state);
755 }
756 
757 static void ipoib_napi_enable(struct net_device *dev)
758 {
759 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
760 
761 	napi_enable(&priv->recv_napi);
762 	napi_enable(&priv->send_napi);
763 }
764 
765 static void ipoib_napi_disable(struct net_device *dev)
766 {
767 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
768 
769 	napi_disable(&priv->recv_napi);
770 	napi_disable(&priv->send_napi);
771 }
772 
773 int ipoib_ib_dev_stop_default(struct net_device *dev)
774 {
775 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
776 	struct ib_qp_attr qp_attr;
777 	unsigned long begin;
778 	struct ipoib_tx_buf *tx_req;
779 	int i;
780 
781 	if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
782 		ipoib_napi_disable(dev);
783 
784 	ipoib_cm_dev_stop(dev);
785 
786 	/*
787 	 * Move our QP to the error state and then reinitialize in
788 	 * when all work requests have completed or have been flushed.
789 	 */
790 	qp_attr.qp_state = IB_QPS_ERR;
791 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
792 		check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);
793 
794 	/* Wait for all sends and receives to complete */
795 	begin = jiffies;
796 
797 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
798 		if (time_after(jiffies, begin + 5 * HZ)) {
799 			ipoib_warn(priv,
800 				   "timing out; %d sends %d receives not completed\n",
801 				   priv->tx_head - priv->tx_tail,
802 				   recvs_pending(dev));
803 
804 			/*
805 			 * assume the HW is wedged and just free up
806 			 * all our pending work requests.
807 			 */
808 			while ((int)priv->tx_tail - (int)priv->tx_head < 0) {
809 				tx_req = &priv->tx_ring[priv->tx_tail &
810 							(ipoib_sendq_size - 1)];
811 				ipoib_dma_unmap_tx(priv, tx_req);
812 				dev_kfree_skb_any(tx_req->skb);
813 				++priv->tx_tail;
814 				++priv->global_tx_tail;
815 			}
816 
817 			for (i = 0; i < ipoib_recvq_size; ++i) {
818 				struct ipoib_rx_buf *rx_req;
819 
820 				rx_req = &priv->rx_ring[i];
821 				if (!rx_req->skb)
822 					continue;
823 				ipoib_ud_dma_unmap_rx(priv,
824 						      priv->rx_ring[i].mapping);
825 				dev_kfree_skb_any(rx_req->skb);
826 				rx_req->skb = NULL;
827 			}
828 
829 			goto timeout;
830 		}
831 
832 		ipoib_drain_cq(dev);
833 
834 		usleep_range(1000, 2000);
835 	}
836 
837 	ipoib_dbg(priv, "All sends and receives done.\n");
838 
839 timeout:
840 	qp_attr.qp_state = IB_QPS_RESET;
841 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
842 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
843 
844 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
845 
846 	return 0;
847 }
848 
849 int ipoib_ib_dev_stop(struct net_device *dev)
850 {
851 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
852 
853 	priv->rn_ops->ndo_stop(dev);
854 
855 	clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
856 	ipoib_flush_ah(dev);
857 
858 	return 0;
859 }
860 
861 int ipoib_ib_dev_open_default(struct net_device *dev)
862 {
863 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
864 	int ret;
865 
866 	ret = ipoib_init_qp(dev);
867 	if (ret) {
868 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
869 		return -1;
870 	}
871 
872 	ret = ipoib_ib_post_receives(dev);
873 	if (ret) {
874 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
875 		goto out;
876 	}
877 
878 	ret = ipoib_cm_dev_open(dev);
879 	if (ret) {
880 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
881 		goto out;
882 	}
883 
884 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
885 		ipoib_napi_enable(dev);
886 
887 	return 0;
888 out:
889 	return -1;
890 }
891 
892 int ipoib_ib_dev_open(struct net_device *dev)
893 {
894 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
895 
896 	ipoib_pkey_dev_check_presence(dev);
897 
898 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
899 		ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
900 			   (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
901 		return -1;
902 	}
903 
904 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
905 	queue_delayed_work(priv->wq, &priv->ah_reap_task,
906 			   round_jiffies_relative(HZ));
907 
908 	if (priv->rn_ops->ndo_open(dev)) {
909 		pr_warn("%s: Failed to open dev\n", dev->name);
910 		goto dev_stop;
911 	}
912 
913 	set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
914 
915 	return 0;
916 
917 dev_stop:
918 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
919 	cancel_delayed_work(&priv->ah_reap_task);
920 	set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
921 	ipoib_ib_dev_stop(dev);
922 	return -1;
923 }
924 
925 void ipoib_pkey_dev_check_presence(struct net_device *dev)
926 {
927 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
928 	struct rdma_netdev *rn = netdev_priv(dev);
929 
930 	if (!(priv->pkey & 0x7fff) ||
931 	    ib_find_pkey(priv->ca, priv->port, priv->pkey,
932 			 &priv->pkey_index)) {
933 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
934 	} else {
935 		if (rn->set_id)
936 			rn->set_id(dev, priv->pkey_index);
937 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
938 	}
939 }
940 
941 void ipoib_ib_dev_up(struct net_device *dev)
942 {
943 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
944 
945 	ipoib_pkey_dev_check_presence(dev);
946 
947 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
948 		ipoib_dbg(priv, "PKEY is not assigned.\n");
949 		return;
950 	}
951 
952 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
953 
954 	ipoib_mcast_start_thread(dev);
955 }
956 
957 void ipoib_ib_dev_down(struct net_device *dev)
958 {
959 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
960 
961 	ipoib_dbg(priv, "downing ib_dev\n");
962 
963 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
964 	netif_carrier_off(dev);
965 
966 	ipoib_mcast_stop_thread(dev);
967 	ipoib_mcast_dev_flush(dev);
968 
969 	ipoib_flush_paths(dev);
970 }
971 
972 void ipoib_drain_cq(struct net_device *dev)
973 {
974 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
975 	int i, n;
976 
977 	/*
978 	 * We call completion handling routines that expect to be
979 	 * called from the BH-disabled NAPI poll context, so disable
980 	 * BHs here too.
981 	 */
982 	local_bh_disable();
983 
984 	do {
985 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
986 		for (i = 0; i < n; ++i) {
987 			/*
988 			 * Convert any successful completions to flush
989 			 * errors to avoid passing packets up the
990 			 * stack after bringing the device down.
991 			 */
992 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
993 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
994 
995 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
996 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
997 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
998 				else
999 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
1000 			} else {
1001 				pr_warn("%s: Got unexpected wqe id\n", __func__);
1002 			}
1003 		}
1004 	} while (n == IPOIB_NUM_WC);
1005 
1006 	while (poll_tx(priv))
1007 		; /* nothing */
1008 
1009 	local_bh_enable();
1010 }
1011 
1012 /*
1013  * Takes whatever value which is in pkey index 0 and updates priv->pkey
1014  * returns 0 if the pkey value was changed.
1015  */
1016 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
1017 {
1018 	int result;
1019 	u16 prev_pkey;
1020 
1021 	prev_pkey = priv->pkey;
1022 	result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1023 	if (result) {
1024 		ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
1025 			   priv->port, result);
1026 		return result;
1027 	}
1028 
1029 	priv->pkey |= 0x8000;
1030 
1031 	if (prev_pkey != priv->pkey) {
1032 		ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
1033 			  prev_pkey, priv->pkey);
1034 		/*
1035 		 * Update the pkey in the broadcast address, while making sure to set
1036 		 * the full membership bit, so that we join the right broadcast group.
1037 		 */
1038 		priv->dev->broadcast[8] = priv->pkey >> 8;
1039 		priv->dev->broadcast[9] = priv->pkey & 0xff;
1040 		return 0;
1041 	}
1042 
1043 	return 1;
1044 }
1045 /*
1046  * returns 0 if pkey value was found in a different slot.
1047  */
1048 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
1049 {
1050 	u16 old_index = priv->pkey_index;
1051 
1052 	priv->pkey_index = 0;
1053 	ipoib_pkey_dev_check_presence(priv->dev);
1054 
1055 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
1056 	    (old_index == priv->pkey_index))
1057 		return 1;
1058 	return 0;
1059 }
1060 
1061 /*
1062  * returns true if the device address of the ipoib interface has changed and the
1063  * new address is a valid one (i.e in the gid table), return false otherwise.
1064  */
1065 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)
1066 {
1067 	union ib_gid search_gid;
1068 	union ib_gid gid0;
1069 	union ib_gid *netdev_gid;
1070 	int err;
1071 	u16 index;
1072 	u8 port;
1073 	bool ret = false;
1074 
1075 	netdev_gid = (union ib_gid *)(priv->dev->dev_addr + 4);
1076 	if (rdma_query_gid(priv->ca, priv->port, 0, &gid0))
1077 		return false;
1078 
1079 	netif_addr_lock_bh(priv->dev);
1080 
1081 	/* The subnet prefix may have changed, update it now so we won't have
1082 	 * to do it later
1083 	 */
1084 	priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1085 	netdev_gid->global.subnet_prefix = gid0.global.subnet_prefix;
1086 	search_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1087 
1088 	search_gid.global.interface_id = priv->local_gid.global.interface_id;
1089 
1090 	netif_addr_unlock_bh(priv->dev);
1091 
1092 	err = ib_find_gid(priv->ca, &search_gid, &port, &index);
1093 
1094 	netif_addr_lock_bh(priv->dev);
1095 
1096 	if (search_gid.global.interface_id !=
1097 	    priv->local_gid.global.interface_id)
1098 		/* There was a change while we were looking up the gid, bail
1099 		 * here and let the next work sort this out
1100 		 */
1101 		goto out;
1102 
1103 	/* The next section of code needs some background:
1104 	 * Per IB spec the port GUID can't change if the HCA is powered on.
1105 	 * port GUID is the basis for GID at index 0 which is the basis for
1106 	 * the default device address of a ipoib interface.
1107 	 *
1108 	 * so it seems the flow should be:
1109 	 * if user_changed_dev_addr && gid in gid tbl
1110 	 *	set bit dev_addr_set
1111 	 *	return true
1112 	 * else
1113 	 *	return false
1114 	 *
1115 	 * The issue is that there are devices that don't follow the spec,
1116 	 * they change the port GUID when the HCA is powered, so in order
1117 	 * not to break userspace applications, We need to check if the
1118 	 * user wanted to control the device address and we assume that
1119 	 * if he sets the device address back to be based on GID index 0,
1120 	 * he no longer wishs to control it.
1121 	 *
1122 	 * If the user doesn't control the the device address,
1123 	 * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1124 	 * the port GUID has changed and GID at index 0 has changed
1125 	 * so we need to change priv->local_gid and priv->dev->dev_addr
1126 	 * to reflect the new GID.
1127 	 */
1128 	if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
1129 		if (!err && port == priv->port) {
1130 			set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1131 			if (index == 0)
1132 				clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,
1133 					  &priv->flags);
1134 			else
1135 				set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);
1136 			ret = true;
1137 		} else {
1138 			ret = false;
1139 		}
1140 	} else {
1141 		if (!err && port == priv->port) {
1142 			ret = true;
1143 		} else {
1144 			if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {
1145 				memcpy(&priv->local_gid, &gid0,
1146 				       sizeof(priv->local_gid));
1147 				memcpy(priv->dev->dev_addr + 4, &gid0,
1148 				       sizeof(priv->local_gid));
1149 				ret = true;
1150 			}
1151 		}
1152 	}
1153 
1154 out:
1155 	netif_addr_unlock_bh(priv->dev);
1156 
1157 	return ret;
1158 }
1159 
1160 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1161 				enum ipoib_flush_level level,
1162 				int nesting)
1163 {
1164 	struct ipoib_dev_priv *cpriv;
1165 	struct net_device *dev = priv->dev;
1166 	int result;
1167 
1168 	down_read_nested(&priv->vlan_rwsem, nesting);
1169 
1170 	/*
1171 	 * Flush any child interfaces too -- they might be up even if
1172 	 * the parent is down.
1173 	 */
1174 	list_for_each_entry(cpriv, &priv->child_intfs, list)
1175 		__ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1176 
1177 	up_read(&priv->vlan_rwsem);
1178 
1179 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1180 	    level != IPOIB_FLUSH_HEAVY) {
1181 		/* Make sure the dev_addr is set even if not flushing */
1182 		if (level == IPOIB_FLUSH_LIGHT)
1183 			ipoib_dev_addr_changed_valid(priv);
1184 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1185 		return;
1186 	}
1187 
1188 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1189 		/* interface is down. update pkey and leave. */
1190 		if (level == IPOIB_FLUSH_HEAVY) {
1191 			if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1192 				update_parent_pkey(priv);
1193 			else
1194 				update_child_pkey(priv);
1195 		} else if (level == IPOIB_FLUSH_LIGHT)
1196 			ipoib_dev_addr_changed_valid(priv);
1197 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1198 		return;
1199 	}
1200 
1201 	if (level == IPOIB_FLUSH_HEAVY) {
1202 		/* child devices chase their origin pkey value, while non-child
1203 		 * (parent) devices should always takes what present in pkey index 0
1204 		 */
1205 		if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1206 			result = update_child_pkey(priv);
1207 			if (result) {
1208 				/* restart QP only if P_Key index is changed */
1209 				ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1210 				return;
1211 			}
1212 
1213 		} else {
1214 			result = update_parent_pkey(priv);
1215 			/* restart QP only if P_Key value changed */
1216 			if (result) {
1217 				ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1218 				return;
1219 			}
1220 		}
1221 	}
1222 
1223 	if (level == IPOIB_FLUSH_LIGHT) {
1224 		int oper_up;
1225 		ipoib_mark_paths_invalid(dev);
1226 		/* Set IPoIB operation as down to prevent races between:
1227 		 * the flush flow which leaves MCG and on the fly joins
1228 		 * which can happen during that time. mcast restart task
1229 		 * should deal with join requests we missed.
1230 		 */
1231 		oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1232 		ipoib_mcast_dev_flush(dev);
1233 		if (oper_up)
1234 			set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1235 		ipoib_flush_ah(dev);
1236 	}
1237 
1238 	if (level >= IPOIB_FLUSH_NORMAL)
1239 		ipoib_ib_dev_down(dev);
1240 
1241 	if (level == IPOIB_FLUSH_HEAVY) {
1242 		if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1243 			ipoib_ib_dev_stop(dev);
1244 
1245 		if (ipoib_ib_dev_open(dev))
1246 			return;
1247 
1248 		if (netif_queue_stopped(dev))
1249 			netif_start_queue(dev);
1250 	}
1251 
1252 	/*
1253 	 * The device could have been brought down between the start and when
1254 	 * we get here, don't bring it back up if it's not configured up
1255 	 */
1256 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1257 		if (level >= IPOIB_FLUSH_NORMAL)
1258 			ipoib_ib_dev_up(dev);
1259 		if (ipoib_dev_addr_changed_valid(priv))
1260 			ipoib_mcast_restart_task(&priv->restart_task);
1261 	}
1262 }
1263 
1264 void ipoib_ib_dev_flush_light(struct work_struct *work)
1265 {
1266 	struct ipoib_dev_priv *priv =
1267 		container_of(work, struct ipoib_dev_priv, flush_light);
1268 
1269 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1270 }
1271 
1272 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1273 {
1274 	struct ipoib_dev_priv *priv =
1275 		container_of(work, struct ipoib_dev_priv, flush_normal);
1276 
1277 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1278 }
1279 
1280 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1281 {
1282 	struct ipoib_dev_priv *priv =
1283 		container_of(work, struct ipoib_dev_priv, flush_heavy);
1284 
1285 	rtnl_lock();
1286 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1287 	rtnl_unlock();
1288 }
1289 
1290 void ipoib_ib_dev_cleanup(struct net_device *dev)
1291 {
1292 	struct ipoib_dev_priv *priv = ipoib_priv(dev);
1293 
1294 	ipoib_dbg(priv, "cleaning up ib_dev\n");
1295 	/*
1296 	 * We must make sure there are no more (path) completions
1297 	 * that may wish to touch priv fields that are no longer valid
1298 	 */
1299 	ipoib_flush_paths(dev);
1300 
1301 	ipoib_mcast_stop_thread(dev);
1302 	ipoib_mcast_dev_flush(dev);
1303 
1304 	/*
1305 	 * All of our ah references aren't free until after
1306 	 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1307 	 * the neighbor garbage collection is stopped and reaped.
1308 	 * That should all be done now, so make a final ah flush.
1309 	 */
1310 	ipoib_stop_ah(dev);
1311 
1312 	clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
1313 
1314 	priv->rn_ops->ndo_uninit(dev);
1315 
1316 	if (priv->pd) {
1317 		ib_dealloc_pd(priv->pd);
1318 		priv->pd = NULL;
1319 	}
1320 }
1321 
1322