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 
44 #include "ipoib.h"
45 
46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
47 static int data_debug_level;
48 
49 module_param(data_debug_level, int, 0644);
50 MODULE_PARM_DESC(data_debug_level,
51 		 "Enable data path debug tracing if > 0");
52 #endif
53 
54 static DEFINE_MUTEX(pkey_mutex);
55 
56 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
57 				 struct ib_pd *pd, struct ib_ah_attr *attr)
58 {
59 	struct ipoib_ah *ah;
60 	struct ib_ah *vah;
61 
62 	ah = kmalloc(sizeof *ah, GFP_KERNEL);
63 	if (!ah)
64 		return ERR_PTR(-ENOMEM);
65 
66 	ah->dev       = dev;
67 	ah->last_send = 0;
68 	kref_init(&ah->ref);
69 
70 	vah = ib_create_ah(pd, attr);
71 	if (IS_ERR(vah)) {
72 		kfree(ah);
73 		ah = (struct ipoib_ah *)vah;
74 	} else {
75 		ah->ah = vah;
76 		ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
77 	}
78 
79 	return ah;
80 }
81 
82 void ipoib_free_ah(struct kref *kref)
83 {
84 	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
85 	struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
86 
87 	unsigned long flags;
88 
89 	spin_lock_irqsave(&priv->lock, flags);
90 	list_add_tail(&ah->list, &priv->dead_ahs);
91 	spin_unlock_irqrestore(&priv->lock, flags);
92 }
93 
94 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
95 				  u64 mapping[IPOIB_UD_RX_SG])
96 {
97 	ib_dma_unmap_single(priv->ca, mapping[0],
98 			    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
99 			    DMA_FROM_DEVICE);
100 }
101 
102 static int ipoib_ib_post_receive(struct net_device *dev, int id)
103 {
104 	struct ipoib_dev_priv *priv = netdev_priv(dev);
105 	struct ib_recv_wr *bad_wr;
106 	int ret;
107 
108 	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
109 	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
110 	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
111 
112 
113 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
114 	if (unlikely(ret)) {
115 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
116 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
117 		dev_kfree_skb_any(priv->rx_ring[id].skb);
118 		priv->rx_ring[id].skb = NULL;
119 	}
120 
121 	return ret;
122 }
123 
124 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
125 {
126 	struct ipoib_dev_priv *priv = netdev_priv(dev);
127 	struct sk_buff *skb;
128 	int buf_size;
129 	u64 *mapping;
130 
131 	buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
132 
133 	skb = dev_alloc_skb(buf_size + IPOIB_ENCAP_LEN);
134 	if (unlikely(!skb))
135 		return NULL;
136 
137 	/*
138 	 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
139 	 * header.  So we need 4 more bytes to get to 48 and align the
140 	 * IP header to a multiple of 16.
141 	 */
142 	skb_reserve(skb, 4);
143 
144 	mapping = priv->rx_ring[id].mapping;
145 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
146 				       DMA_FROM_DEVICE);
147 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
148 		goto error;
149 
150 	priv->rx_ring[id].skb = skb;
151 	return skb;
152 error:
153 	dev_kfree_skb_any(skb);
154 	return NULL;
155 }
156 
157 static int ipoib_ib_post_receives(struct net_device *dev)
158 {
159 	struct ipoib_dev_priv *priv = netdev_priv(dev);
160 	int i;
161 
162 	for (i = 0; i < ipoib_recvq_size; ++i) {
163 		if (!ipoib_alloc_rx_skb(dev, i)) {
164 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
165 			return -ENOMEM;
166 		}
167 		if (ipoib_ib_post_receive(dev, i)) {
168 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
169 			return -EIO;
170 		}
171 	}
172 
173 	return 0;
174 }
175 
176 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
177 {
178 	struct ipoib_dev_priv *priv = netdev_priv(dev);
179 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
180 	struct sk_buff *skb;
181 	u64 mapping[IPOIB_UD_RX_SG];
182 	union ib_gid *dgid;
183 
184 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
185 		       wr_id, wc->status);
186 
187 	if (unlikely(wr_id >= ipoib_recvq_size)) {
188 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
189 			   wr_id, ipoib_recvq_size);
190 		return;
191 	}
192 
193 	skb  = priv->rx_ring[wr_id].skb;
194 
195 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
196 		if (wc->status != IB_WC_WR_FLUSH_ERR)
197 			ipoib_warn(priv, "failed recv event "
198 				   "(status=%d, wrid=%d vend_err %x)\n",
199 				   wc->status, wr_id, wc->vendor_err);
200 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
201 		dev_kfree_skb_any(skb);
202 		priv->rx_ring[wr_id].skb = NULL;
203 		return;
204 	}
205 
206 	/*
207 	 * Drop packets that this interface sent, ie multicast packets
208 	 * that the HCA has replicated.
209 	 */
210 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
211 		goto repost;
212 
213 	memcpy(mapping, priv->rx_ring[wr_id].mapping,
214 	       IPOIB_UD_RX_SG * sizeof *mapping);
215 
216 	/*
217 	 * If we can't allocate a new RX buffer, dump
218 	 * this packet and reuse the old buffer.
219 	 */
220 	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
221 		++dev->stats.rx_dropped;
222 		goto repost;
223 	}
224 
225 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
226 		       wc->byte_len, wc->slid);
227 
228 	ipoib_ud_dma_unmap_rx(priv, mapping);
229 
230 	skb_put(skb, wc->byte_len);
231 
232 	/* First byte of dgid signals multicast when 0xff */
233 	dgid = &((struct ib_grh *)skb->data)->dgid;
234 
235 	if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
236 		skb->pkt_type = PACKET_HOST;
237 	else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
238 		skb->pkt_type = PACKET_BROADCAST;
239 	else
240 		skb->pkt_type = PACKET_MULTICAST;
241 
242 	skb_pull(skb, IB_GRH_BYTES);
243 
244 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
245 	skb_reset_mac_header(skb);
246 	skb_pull(skb, IPOIB_ENCAP_LEN);
247 
248 	++dev->stats.rx_packets;
249 	dev->stats.rx_bytes += skb->len;
250 
251 	skb->dev = dev;
252 	if ((dev->features & NETIF_F_RXCSUM) &&
253 			likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
254 		skb->ip_summed = CHECKSUM_UNNECESSARY;
255 
256 	napi_gro_receive(&priv->napi, skb);
257 
258 repost:
259 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
260 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
261 			   "for buf %d\n", wr_id);
262 }
263 
264 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
265 {
266 	struct sk_buff *skb = tx_req->skb;
267 	u64 *mapping = tx_req->mapping;
268 	int i;
269 	int off;
270 
271 	if (skb_headlen(skb)) {
272 		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
273 					       DMA_TO_DEVICE);
274 		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
275 			return -EIO;
276 
277 		off = 1;
278 	} else
279 		off = 0;
280 
281 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
282 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
283 		mapping[i + off] = ib_dma_map_page(ca,
284 						 skb_frag_page(frag),
285 						 frag->page_offset, skb_frag_size(frag),
286 						 DMA_TO_DEVICE);
287 		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
288 			goto partial_error;
289 	}
290 	return 0;
291 
292 partial_error:
293 	for (; i > 0; --i) {
294 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
295 
296 		ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
297 	}
298 
299 	if (off)
300 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
301 
302 	return -EIO;
303 }
304 
305 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
306 			struct ipoib_tx_buf *tx_req)
307 {
308 	struct sk_buff *skb = tx_req->skb;
309 	u64 *mapping = tx_req->mapping;
310 	int i;
311 	int off;
312 
313 	if (skb_headlen(skb)) {
314 		ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
315 				    DMA_TO_DEVICE);
316 		off = 1;
317 	} else
318 		off = 0;
319 
320 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
321 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
322 
323 		ib_dma_unmap_page(priv->ca, mapping[i + off],
324 				  skb_frag_size(frag), DMA_TO_DEVICE);
325 	}
326 }
327 
328 /*
329  * As the result of a completion error the QP Can be transferred to SQE states.
330  * The function checks if the (send)QP is in SQE state and
331  * moves it back to RTS state, that in order to have it functional again.
332  */
333 static void ipoib_qp_state_validate_work(struct work_struct *work)
334 {
335 	struct ipoib_qp_state_validate *qp_work =
336 		container_of(work, struct ipoib_qp_state_validate, work);
337 
338 	struct ipoib_dev_priv *priv = qp_work->priv;
339 	struct ib_qp_attr qp_attr;
340 	struct ib_qp_init_attr query_init_attr;
341 	int ret;
342 
343 	ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
344 	if (ret) {
345 		ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
346 			   __func__, ret);
347 		goto free_res;
348 	}
349 	pr_info("%s: QP: 0x%x is in state: %d\n",
350 		__func__, priv->qp->qp_num, qp_attr.qp_state);
351 
352 	/* currently support only in SQE->RTS transition*/
353 	if (qp_attr.qp_state == IB_QPS_SQE) {
354 		qp_attr.qp_state = IB_QPS_RTS;
355 
356 		ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
357 		if (ret) {
358 			pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
359 				ret, priv->qp->qp_num);
360 			goto free_res;
361 		}
362 		pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
363 			__func__, priv->qp->qp_num);
364 	} else {
365 		pr_warn("QP (%d) will stay in state: %d\n",
366 			priv->qp->qp_num, qp_attr.qp_state);
367 	}
368 
369 free_res:
370 	kfree(qp_work);
371 }
372 
373 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
374 {
375 	struct ipoib_dev_priv *priv = netdev_priv(dev);
376 	unsigned int wr_id = wc->wr_id;
377 	struct ipoib_tx_buf *tx_req;
378 
379 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
380 		       wr_id, wc->status);
381 
382 	if (unlikely(wr_id >= ipoib_sendq_size)) {
383 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
384 			   wr_id, ipoib_sendq_size);
385 		return;
386 	}
387 
388 	tx_req = &priv->tx_ring[wr_id];
389 
390 	ipoib_dma_unmap_tx(priv, tx_req);
391 
392 	++dev->stats.tx_packets;
393 	dev->stats.tx_bytes += tx_req->skb->len;
394 
395 	dev_kfree_skb_any(tx_req->skb);
396 
397 	++priv->tx_tail;
398 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
399 	    netif_queue_stopped(dev) &&
400 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
401 		netif_wake_queue(dev);
402 
403 	if (wc->status != IB_WC_SUCCESS &&
404 	    wc->status != IB_WC_WR_FLUSH_ERR) {
405 		struct ipoib_qp_state_validate *qp_work;
406 		ipoib_warn(priv, "failed send event "
407 			   "(status=%d, wrid=%d vend_err %x)\n",
408 			   wc->status, wr_id, wc->vendor_err);
409 		qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
410 		if (!qp_work) {
411 			ipoib_warn(priv, "%s Failed alloc ipoib_qp_state_validate for qp: 0x%x\n",
412 				   __func__, priv->qp->qp_num);
413 			return;
414 		}
415 
416 		INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
417 		qp_work->priv = priv;
418 		queue_work(priv->wq, &qp_work->work);
419 	}
420 }
421 
422 static int poll_tx(struct ipoib_dev_priv *priv)
423 {
424 	int n, i;
425 
426 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
427 	for (i = 0; i < n; ++i)
428 		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
429 
430 	return n == MAX_SEND_CQE;
431 }
432 
433 int ipoib_poll(struct napi_struct *napi, int budget)
434 {
435 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
436 	struct net_device *dev = priv->dev;
437 	int done;
438 	int t;
439 	int n, i;
440 
441 	done  = 0;
442 
443 poll_more:
444 	while (done < budget) {
445 		int max = (budget - done);
446 
447 		t = min(IPOIB_NUM_WC, max);
448 		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
449 
450 		for (i = 0; i < n; i++) {
451 			struct ib_wc *wc = priv->ibwc + i;
452 
453 			if (wc->wr_id & IPOIB_OP_RECV) {
454 				++done;
455 				if (wc->wr_id & IPOIB_OP_CM)
456 					ipoib_cm_handle_rx_wc(dev, wc);
457 				else
458 					ipoib_ib_handle_rx_wc(dev, wc);
459 			} else
460 				ipoib_cm_handle_tx_wc(priv->dev, wc);
461 		}
462 
463 		if (n != t)
464 			break;
465 	}
466 
467 	if (done < budget) {
468 		napi_complete(napi);
469 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
470 					      IB_CQ_NEXT_COMP |
471 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
472 		    napi_reschedule(napi))
473 			goto poll_more;
474 	}
475 
476 	return done;
477 }
478 
479 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
480 {
481 	struct net_device *dev = dev_ptr;
482 	struct ipoib_dev_priv *priv = netdev_priv(dev);
483 
484 	napi_schedule(&priv->napi);
485 }
486 
487 static void drain_tx_cq(struct net_device *dev)
488 {
489 	struct ipoib_dev_priv *priv = netdev_priv(dev);
490 
491 	netif_tx_lock(dev);
492 	while (poll_tx(priv))
493 		; /* nothing */
494 
495 	if (netif_queue_stopped(dev))
496 		mod_timer(&priv->poll_timer, jiffies + 1);
497 
498 	netif_tx_unlock(dev);
499 }
500 
501 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
502 {
503 	struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
504 
505 	mod_timer(&priv->poll_timer, jiffies);
506 }
507 
508 static inline int post_send(struct ipoib_dev_priv *priv,
509 			    unsigned int wr_id,
510 			    struct ib_ah *address, u32 qpn,
511 			    struct ipoib_tx_buf *tx_req,
512 			    void *head, int hlen)
513 {
514 	struct ib_send_wr *bad_wr;
515 	struct sk_buff *skb = tx_req->skb;
516 
517 	ipoib_build_sge(priv, tx_req);
518 
519 	priv->tx_wr.wr.wr_id	= wr_id;
520 	priv->tx_wr.remote_qpn	= qpn;
521 	priv->tx_wr.ah		= address;
522 
523 	if (head) {
524 		priv->tx_wr.mss		= skb_shinfo(skb)->gso_size;
525 		priv->tx_wr.header	= head;
526 		priv->tx_wr.hlen	= hlen;
527 		priv->tx_wr.wr.opcode	= IB_WR_LSO;
528 	} else
529 		priv->tx_wr.wr.opcode	= IB_WR_SEND;
530 
531 	return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
532 }
533 
534 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
535 		struct ipoib_ah *address, u32 qpn)
536 {
537 	struct ipoib_dev_priv *priv = netdev_priv(dev);
538 	struct ipoib_tx_buf *tx_req;
539 	int hlen, rc;
540 	void *phead;
541 
542 	if (skb_is_gso(skb)) {
543 		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
544 		phead = skb->data;
545 		if (unlikely(!skb_pull(skb, hlen))) {
546 			ipoib_warn(priv, "linear data too small\n");
547 			++dev->stats.tx_dropped;
548 			++dev->stats.tx_errors;
549 			dev_kfree_skb_any(skb);
550 			return;
551 		}
552 	} else {
553 		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
554 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
555 				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
556 			++dev->stats.tx_dropped;
557 			++dev->stats.tx_errors;
558 			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
559 			return;
560 		}
561 		phead = NULL;
562 		hlen  = 0;
563 	}
564 
565 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
566 		       skb->len, address, qpn);
567 
568 	/*
569 	 * We put the skb into the tx_ring _before_ we call post_send()
570 	 * because it's entirely possible that the completion handler will
571 	 * run before we execute anything after the post_send().  That
572 	 * means we have to make sure everything is properly recorded and
573 	 * our state is consistent before we call post_send().
574 	 */
575 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
576 	tx_req->skb = skb;
577 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
578 		++dev->stats.tx_errors;
579 		dev_kfree_skb_any(skb);
580 		return;
581 	}
582 
583 	if (skb->ip_summed == CHECKSUM_PARTIAL)
584 		priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
585 	else
586 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
587 
588 	if (++priv->tx_outstanding == ipoib_sendq_size) {
589 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
590 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
591 			ipoib_warn(priv, "request notify on send CQ failed\n");
592 		netif_stop_queue(dev);
593 	}
594 
595 	skb_orphan(skb);
596 	skb_dst_drop(skb);
597 
598 	rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
599 		       address->ah, qpn, tx_req, phead, hlen);
600 	if (unlikely(rc)) {
601 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
602 		++dev->stats.tx_errors;
603 		--priv->tx_outstanding;
604 		ipoib_dma_unmap_tx(priv, tx_req);
605 		dev_kfree_skb_any(skb);
606 		if (netif_queue_stopped(dev))
607 			netif_wake_queue(dev);
608 	} else {
609 		dev->trans_start = jiffies;
610 
611 		address->last_send = priv->tx_head;
612 		++priv->tx_head;
613 	}
614 
615 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
616 		while (poll_tx(priv))
617 			; /* nothing */
618 }
619 
620 static void __ipoib_reap_ah(struct net_device *dev)
621 {
622 	struct ipoib_dev_priv *priv = netdev_priv(dev);
623 	struct ipoib_ah *ah, *tah;
624 	LIST_HEAD(remove_list);
625 	unsigned long flags;
626 
627 	netif_tx_lock_bh(dev);
628 	spin_lock_irqsave(&priv->lock, flags);
629 
630 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
631 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
632 			list_del(&ah->list);
633 			ib_destroy_ah(ah->ah);
634 			kfree(ah);
635 		}
636 
637 	spin_unlock_irqrestore(&priv->lock, flags);
638 	netif_tx_unlock_bh(dev);
639 }
640 
641 void ipoib_reap_ah(struct work_struct *work)
642 {
643 	struct ipoib_dev_priv *priv =
644 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
645 	struct net_device *dev = priv->dev;
646 
647 	__ipoib_reap_ah(dev);
648 
649 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
650 		queue_delayed_work(priv->wq, &priv->ah_reap_task,
651 				   round_jiffies_relative(HZ));
652 }
653 
654 static void ipoib_flush_ah(struct net_device *dev)
655 {
656 	struct ipoib_dev_priv *priv = netdev_priv(dev);
657 
658 	cancel_delayed_work(&priv->ah_reap_task);
659 	flush_workqueue(priv->wq);
660 	ipoib_reap_ah(&priv->ah_reap_task.work);
661 }
662 
663 static void ipoib_stop_ah(struct net_device *dev)
664 {
665 	struct ipoib_dev_priv *priv = netdev_priv(dev);
666 
667 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
668 	ipoib_flush_ah(dev);
669 }
670 
671 static void ipoib_ib_tx_timer_func(unsigned long ctx)
672 {
673 	drain_tx_cq((struct net_device *)ctx);
674 }
675 
676 int ipoib_ib_dev_open(struct net_device *dev)
677 {
678 	struct ipoib_dev_priv *priv = netdev_priv(dev);
679 	int ret;
680 
681 	ipoib_pkey_dev_check_presence(dev);
682 
683 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
684 		ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
685 			   (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
686 		return -1;
687 	}
688 
689 	ret = ipoib_init_qp(dev);
690 	if (ret) {
691 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
692 		return -1;
693 	}
694 
695 	ret = ipoib_ib_post_receives(dev);
696 	if (ret) {
697 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
698 		goto dev_stop;
699 	}
700 
701 	ret = ipoib_cm_dev_open(dev);
702 	if (ret) {
703 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
704 		goto dev_stop;
705 	}
706 
707 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
708 	queue_delayed_work(priv->wq, &priv->ah_reap_task,
709 			   round_jiffies_relative(HZ));
710 
711 	if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
712 		napi_enable(&priv->napi);
713 
714 	return 0;
715 dev_stop:
716 	if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
717 		napi_enable(&priv->napi);
718 	ipoib_ib_dev_stop(dev);
719 	return -1;
720 }
721 
722 void ipoib_pkey_dev_check_presence(struct net_device *dev)
723 {
724 	struct ipoib_dev_priv *priv = netdev_priv(dev);
725 
726 	if (!(priv->pkey & 0x7fff) ||
727 	    ib_find_pkey(priv->ca, priv->port, priv->pkey,
728 			 &priv->pkey_index))
729 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
730 	else
731 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
732 }
733 
734 int ipoib_ib_dev_up(struct net_device *dev)
735 {
736 	struct ipoib_dev_priv *priv = netdev_priv(dev);
737 
738 	ipoib_pkey_dev_check_presence(dev);
739 
740 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
741 		ipoib_dbg(priv, "PKEY is not assigned.\n");
742 		return 0;
743 	}
744 
745 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
746 
747 	return ipoib_mcast_start_thread(dev);
748 }
749 
750 int ipoib_ib_dev_down(struct net_device *dev)
751 {
752 	struct ipoib_dev_priv *priv = netdev_priv(dev);
753 
754 	ipoib_dbg(priv, "downing ib_dev\n");
755 
756 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
757 	netif_carrier_off(dev);
758 
759 	ipoib_mcast_stop_thread(dev);
760 	ipoib_mcast_dev_flush(dev);
761 
762 	ipoib_flush_paths(dev);
763 
764 	return 0;
765 }
766 
767 static int recvs_pending(struct net_device *dev)
768 {
769 	struct ipoib_dev_priv *priv = netdev_priv(dev);
770 	int pending = 0;
771 	int i;
772 
773 	for (i = 0; i < ipoib_recvq_size; ++i)
774 		if (priv->rx_ring[i].skb)
775 			++pending;
776 
777 	return pending;
778 }
779 
780 void ipoib_drain_cq(struct net_device *dev)
781 {
782 	struct ipoib_dev_priv *priv = netdev_priv(dev);
783 	int i, n;
784 
785 	/*
786 	 * We call completion handling routines that expect to be
787 	 * called from the BH-disabled NAPI poll context, so disable
788 	 * BHs here too.
789 	 */
790 	local_bh_disable();
791 
792 	do {
793 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
794 		for (i = 0; i < n; ++i) {
795 			/*
796 			 * Convert any successful completions to flush
797 			 * errors to avoid passing packets up the
798 			 * stack after bringing the device down.
799 			 */
800 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
801 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
802 
803 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
804 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
805 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
806 				else
807 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
808 			} else
809 				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
810 		}
811 	} while (n == IPOIB_NUM_WC);
812 
813 	while (poll_tx(priv))
814 		; /* nothing */
815 
816 	local_bh_enable();
817 }
818 
819 int ipoib_ib_dev_stop(struct net_device *dev)
820 {
821 	struct ipoib_dev_priv *priv = netdev_priv(dev);
822 	struct ib_qp_attr qp_attr;
823 	unsigned long begin;
824 	struct ipoib_tx_buf *tx_req;
825 	int i;
826 
827 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
828 		napi_disable(&priv->napi);
829 
830 	ipoib_cm_dev_stop(dev);
831 
832 	/*
833 	 * Move our QP to the error state and then reinitialize in
834 	 * when all work requests have completed or have been flushed.
835 	 */
836 	qp_attr.qp_state = IB_QPS_ERR;
837 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
838 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
839 
840 	/* Wait for all sends and receives to complete */
841 	begin = jiffies;
842 
843 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
844 		if (time_after(jiffies, begin + 5 * HZ)) {
845 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
846 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
847 
848 			/*
849 			 * assume the HW is wedged and just free up
850 			 * all our pending work requests.
851 			 */
852 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
853 				tx_req = &priv->tx_ring[priv->tx_tail &
854 							(ipoib_sendq_size - 1)];
855 				ipoib_dma_unmap_tx(priv, tx_req);
856 				dev_kfree_skb_any(tx_req->skb);
857 				++priv->tx_tail;
858 				--priv->tx_outstanding;
859 			}
860 
861 			for (i = 0; i < ipoib_recvq_size; ++i) {
862 				struct ipoib_rx_buf *rx_req;
863 
864 				rx_req = &priv->rx_ring[i];
865 				if (!rx_req->skb)
866 					continue;
867 				ipoib_ud_dma_unmap_rx(priv,
868 						      priv->rx_ring[i].mapping);
869 				dev_kfree_skb_any(rx_req->skb);
870 				rx_req->skb = NULL;
871 			}
872 
873 			goto timeout;
874 		}
875 
876 		ipoib_drain_cq(dev);
877 
878 		msleep(1);
879 	}
880 
881 	ipoib_dbg(priv, "All sends and receives done.\n");
882 
883 timeout:
884 	del_timer_sync(&priv->poll_timer);
885 	qp_attr.qp_state = IB_QPS_RESET;
886 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
887 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
888 
889 	ipoib_flush_ah(dev);
890 
891 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
892 
893 	return 0;
894 }
895 
896 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
897 {
898 	struct ipoib_dev_priv *priv = netdev_priv(dev);
899 
900 	priv->ca = ca;
901 	priv->port = port;
902 	priv->qp = NULL;
903 
904 	if (ipoib_transport_dev_init(dev, ca)) {
905 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
906 		return -ENODEV;
907 	}
908 
909 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
910 		    (unsigned long) dev);
911 
912 	if (dev->flags & IFF_UP) {
913 		if (ipoib_ib_dev_open(dev)) {
914 			ipoib_transport_dev_cleanup(dev);
915 			return -ENODEV;
916 		}
917 	}
918 
919 	return 0;
920 }
921 
922 /*
923  * Takes whatever value which is in pkey index 0 and updates priv->pkey
924  * returns 0 if the pkey value was changed.
925  */
926 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
927 {
928 	int result;
929 	u16 prev_pkey;
930 
931 	prev_pkey = priv->pkey;
932 	result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
933 	if (result) {
934 		ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
935 			   priv->port, result);
936 		return result;
937 	}
938 
939 	priv->pkey |= 0x8000;
940 
941 	if (prev_pkey != priv->pkey) {
942 		ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
943 			  prev_pkey, priv->pkey);
944 		/*
945 		 * Update the pkey in the broadcast address, while making sure to set
946 		 * the full membership bit, so that we join the right broadcast group.
947 		 */
948 		priv->dev->broadcast[8] = priv->pkey >> 8;
949 		priv->dev->broadcast[9] = priv->pkey & 0xff;
950 		return 0;
951 	}
952 
953 	return 1;
954 }
955 /*
956  * returns 0 if pkey value was found in a different slot.
957  */
958 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
959 {
960 	u16 old_index = priv->pkey_index;
961 
962 	priv->pkey_index = 0;
963 	ipoib_pkey_dev_check_presence(priv->dev);
964 
965 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
966 	    (old_index == priv->pkey_index))
967 		return 1;
968 	return 0;
969 }
970 
971 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
972 				enum ipoib_flush_level level,
973 				int nesting)
974 {
975 	struct ipoib_dev_priv *cpriv;
976 	struct net_device *dev = priv->dev;
977 	int result;
978 
979 	down_read_nested(&priv->vlan_rwsem, nesting);
980 
981 	/*
982 	 * Flush any child interfaces too -- they might be up even if
983 	 * the parent is down.
984 	 */
985 	list_for_each_entry(cpriv, &priv->child_intfs, list)
986 		__ipoib_ib_dev_flush(cpriv, level, nesting + 1);
987 
988 	up_read(&priv->vlan_rwsem);
989 
990 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
991 	    level != IPOIB_FLUSH_HEAVY) {
992 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
993 		return;
994 	}
995 
996 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
997 		/* interface is down. update pkey and leave. */
998 		if (level == IPOIB_FLUSH_HEAVY) {
999 			if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1000 				update_parent_pkey(priv);
1001 			else
1002 				update_child_pkey(priv);
1003 		}
1004 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1005 		return;
1006 	}
1007 
1008 	if (level == IPOIB_FLUSH_HEAVY) {
1009 		/* child devices chase their origin pkey value, while non-child
1010 		 * (parent) devices should always takes what present in pkey index 0
1011 		 */
1012 		if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1013 			result = update_child_pkey(priv);
1014 			if (result) {
1015 				/* restart QP only if P_Key index is changed */
1016 				ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1017 				return;
1018 			}
1019 
1020 		} else {
1021 			result = update_parent_pkey(priv);
1022 			/* restart QP only if P_Key value changed */
1023 			if (result) {
1024 				ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1025 				return;
1026 			}
1027 		}
1028 	}
1029 
1030 	if (level == IPOIB_FLUSH_LIGHT) {
1031 		ipoib_mark_paths_invalid(dev);
1032 		ipoib_mcast_dev_flush(dev);
1033 		ipoib_flush_ah(dev);
1034 	}
1035 
1036 	if (level >= IPOIB_FLUSH_NORMAL)
1037 		ipoib_ib_dev_down(dev);
1038 
1039 	if (level == IPOIB_FLUSH_HEAVY) {
1040 		if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1041 			ipoib_ib_dev_stop(dev);
1042 		if (ipoib_ib_dev_open(dev) != 0)
1043 			return;
1044 		if (netif_queue_stopped(dev))
1045 			netif_start_queue(dev);
1046 	}
1047 
1048 	/*
1049 	 * The device could have been brought down between the start and when
1050 	 * we get here, don't bring it back up if it's not configured up
1051 	 */
1052 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1053 		if (level >= IPOIB_FLUSH_NORMAL)
1054 			ipoib_ib_dev_up(dev);
1055 		ipoib_mcast_restart_task(&priv->restart_task);
1056 	}
1057 }
1058 
1059 void ipoib_ib_dev_flush_light(struct work_struct *work)
1060 {
1061 	struct ipoib_dev_priv *priv =
1062 		container_of(work, struct ipoib_dev_priv, flush_light);
1063 
1064 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1065 }
1066 
1067 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1068 {
1069 	struct ipoib_dev_priv *priv =
1070 		container_of(work, struct ipoib_dev_priv, flush_normal);
1071 
1072 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1073 }
1074 
1075 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1076 {
1077 	struct ipoib_dev_priv *priv =
1078 		container_of(work, struct ipoib_dev_priv, flush_heavy);
1079 
1080 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1081 }
1082 
1083 void ipoib_ib_dev_cleanup(struct net_device *dev)
1084 {
1085 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1086 
1087 	ipoib_dbg(priv, "cleaning up ib_dev\n");
1088 	/*
1089 	 * We must make sure there are no more (path) completions
1090 	 * that may wish to touch priv fields that are no longer valid
1091 	 */
1092 	ipoib_flush_paths(dev);
1093 
1094 	ipoib_mcast_stop_thread(dev);
1095 	ipoib_mcast_dev_flush(dev);
1096 
1097 	/*
1098 	 * All of our ah references aren't free until after
1099 	 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1100 	 * the neighbor garbage collection is stopped and reaped.
1101 	 * That should all be done now, so make a final ah flush.
1102 	 */
1103 	ipoib_stop_ah(dev);
1104 
1105 	ipoib_transport_dev_cleanup(dev);
1106 }
1107 
1108 
1109