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/dma-mapping.h>
38 
39 #include <linux/ip.h>
40 #include <linux/tcp.h>
41 
42 #include "ipoib.h"
43 
44 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
45 static int data_debug_level;
46 
47 module_param(data_debug_level, int, 0644);
48 MODULE_PARM_DESC(data_debug_level,
49 		 "Enable data path debug tracing if > 0");
50 #endif
51 
52 static DEFINE_MUTEX(pkey_mutex);
53 
54 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
55 				 struct ib_pd *pd, struct ib_ah_attr *attr)
56 {
57 	struct ipoib_ah *ah;
58 
59 	ah = kmalloc(sizeof *ah, GFP_KERNEL);
60 	if (!ah)
61 		return NULL;
62 
63 	ah->dev       = dev;
64 	ah->last_send = 0;
65 	kref_init(&ah->ref);
66 
67 	ah->ah = ib_create_ah(pd, attr);
68 	if (IS_ERR(ah->ah)) {
69 		kfree(ah);
70 		ah = NULL;
71 	} else
72 		ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
73 
74 	return ah;
75 }
76 
77 void ipoib_free_ah(struct kref *kref)
78 {
79 	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
80 	struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
81 
82 	unsigned long flags;
83 
84 	spin_lock_irqsave(&priv->lock, flags);
85 	list_add_tail(&ah->list, &priv->dead_ahs);
86 	spin_unlock_irqrestore(&priv->lock, flags);
87 }
88 
89 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
90 				  u64 mapping[IPOIB_UD_RX_SG])
91 {
92 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
93 		ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
94 				    DMA_FROM_DEVICE);
95 		ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
96 				  DMA_FROM_DEVICE);
97 	} else
98 		ib_dma_unmap_single(priv->ca, mapping[0],
99 				    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
100 				    DMA_FROM_DEVICE);
101 }
102 
103 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
104 				   struct sk_buff *skb,
105 				   unsigned int length)
106 {
107 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
108 		skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
109 		unsigned int size;
110 		/*
111 		 * There is only two buffers needed for max_payload = 4K,
112 		 * first buf size is IPOIB_UD_HEAD_SIZE
113 		 */
114 		skb->tail += IPOIB_UD_HEAD_SIZE;
115 		skb->len  += length;
116 
117 		size = length - IPOIB_UD_HEAD_SIZE;
118 
119 		frag->size     = size;
120 		skb->data_len += size;
121 		skb->truesize += size;
122 	} else
123 		skb_put(skb, length);
124 
125 }
126 
127 static int ipoib_ib_post_receive(struct net_device *dev, int id)
128 {
129 	struct ipoib_dev_priv *priv = netdev_priv(dev);
130 	struct ib_recv_wr *bad_wr;
131 	int ret;
132 
133 	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
134 	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
135 	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
136 
137 
138 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
139 	if (unlikely(ret)) {
140 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
141 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
142 		dev_kfree_skb_any(priv->rx_ring[id].skb);
143 		priv->rx_ring[id].skb = NULL;
144 	}
145 
146 	return ret;
147 }
148 
149 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
150 {
151 	struct ipoib_dev_priv *priv = netdev_priv(dev);
152 	struct sk_buff *skb;
153 	int buf_size;
154 	u64 *mapping;
155 
156 	if (ipoib_ud_need_sg(priv->max_ib_mtu))
157 		buf_size = IPOIB_UD_HEAD_SIZE;
158 	else
159 		buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
160 
161 	skb = dev_alloc_skb(buf_size + 4);
162 	if (unlikely(!skb))
163 		return NULL;
164 
165 	/*
166 	 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
167 	 * header.  So we need 4 more bytes to get to 48 and align the
168 	 * IP header to a multiple of 16.
169 	 */
170 	skb_reserve(skb, 4);
171 
172 	mapping = priv->rx_ring[id].mapping;
173 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
174 				       DMA_FROM_DEVICE);
175 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
176 		goto error;
177 
178 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
179 		struct page *page = alloc_page(GFP_ATOMIC);
180 		if (!page)
181 			goto partial_error;
182 		skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
183 		mapping[1] =
184 			ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[0].page,
185 					0, PAGE_SIZE, DMA_FROM_DEVICE);
186 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
187 			goto partial_error;
188 	}
189 
190 	priv->rx_ring[id].skb = skb;
191 	return skb;
192 
193 partial_error:
194 	ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
195 error:
196 	dev_kfree_skb_any(skb);
197 	return NULL;
198 }
199 
200 static int ipoib_ib_post_receives(struct net_device *dev)
201 {
202 	struct ipoib_dev_priv *priv = netdev_priv(dev);
203 	int i;
204 
205 	for (i = 0; i < ipoib_recvq_size; ++i) {
206 		if (!ipoib_alloc_rx_skb(dev, i)) {
207 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
208 			return -ENOMEM;
209 		}
210 		if (ipoib_ib_post_receive(dev, i)) {
211 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
212 			return -EIO;
213 		}
214 	}
215 
216 	return 0;
217 }
218 
219 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
220 {
221 	struct ipoib_dev_priv *priv = netdev_priv(dev);
222 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
223 	struct sk_buff *skb;
224 	u64 mapping[IPOIB_UD_RX_SG];
225 
226 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
227 		       wr_id, wc->status);
228 
229 	if (unlikely(wr_id >= ipoib_recvq_size)) {
230 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
231 			   wr_id, ipoib_recvq_size);
232 		return;
233 	}
234 
235 	skb  = priv->rx_ring[wr_id].skb;
236 
237 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
238 		if (wc->status != IB_WC_WR_FLUSH_ERR)
239 			ipoib_warn(priv, "failed recv event "
240 				   "(status=%d, wrid=%d vend_err %x)\n",
241 				   wc->status, wr_id, wc->vendor_err);
242 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
243 		dev_kfree_skb_any(skb);
244 		priv->rx_ring[wr_id].skb = NULL;
245 		return;
246 	}
247 
248 	/*
249 	 * Drop packets that this interface sent, ie multicast packets
250 	 * that the HCA has replicated.
251 	 */
252 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
253 		goto repost;
254 
255 	memcpy(mapping, priv->rx_ring[wr_id].mapping,
256 	       IPOIB_UD_RX_SG * sizeof *mapping);
257 
258 	/*
259 	 * If we can't allocate a new RX buffer, dump
260 	 * this packet and reuse the old buffer.
261 	 */
262 	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
263 		++dev->stats.rx_dropped;
264 		goto repost;
265 	}
266 
267 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
268 		       wc->byte_len, wc->slid);
269 
270 	ipoib_ud_dma_unmap_rx(priv, mapping);
271 	ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
272 
273 	skb_pull(skb, IB_GRH_BYTES);
274 
275 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
276 	skb_reset_mac_header(skb);
277 	skb_pull(skb, IPOIB_ENCAP_LEN);
278 
279 	++dev->stats.rx_packets;
280 	dev->stats.rx_bytes += skb->len;
281 
282 	skb->dev = dev;
283 	/* XXX get correct PACKET_ type here */
284 	skb->pkt_type = PACKET_HOST;
285 
286 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
287 		skb->ip_summed = CHECKSUM_UNNECESSARY;
288 
289 	if (dev->features & NETIF_F_LRO)
290 		lro_receive_skb(&priv->lro.lro_mgr, skb, NULL);
291 	else
292 		netif_receive_skb(skb);
293 
294 repost:
295 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
296 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
297 			   "for buf %d\n", wr_id);
298 }
299 
300 static int ipoib_dma_map_tx(struct ib_device *ca,
301 			    struct ipoib_tx_buf *tx_req)
302 {
303 	struct sk_buff *skb = tx_req->skb;
304 	u64 *mapping = tx_req->mapping;
305 	int i;
306 	int off;
307 
308 	if (skb_headlen(skb)) {
309 		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
310 					       DMA_TO_DEVICE);
311 		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
312 			return -EIO;
313 
314 		off = 1;
315 	} else
316 		off = 0;
317 
318 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
319 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
320 		mapping[i + off] = ib_dma_map_page(ca, frag->page,
321 						 frag->page_offset, frag->size,
322 						 DMA_TO_DEVICE);
323 		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
324 			goto partial_error;
325 	}
326 	return 0;
327 
328 partial_error:
329 	for (; i > 0; --i) {
330 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
331 		ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
332 	}
333 
334 	if (off)
335 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
336 
337 	return -EIO;
338 }
339 
340 static void ipoib_dma_unmap_tx(struct ib_device *ca,
341 			       struct ipoib_tx_buf *tx_req)
342 {
343 	struct sk_buff *skb = tx_req->skb;
344 	u64 *mapping = tx_req->mapping;
345 	int i;
346 	int off;
347 
348 	if (skb_headlen(skb)) {
349 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
350 		off = 1;
351 	} else
352 		off = 0;
353 
354 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
355 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
356 		ib_dma_unmap_page(ca, mapping[i + off], frag->size,
357 				  DMA_TO_DEVICE);
358 	}
359 }
360 
361 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
362 {
363 	struct ipoib_dev_priv *priv = netdev_priv(dev);
364 	unsigned int wr_id = wc->wr_id;
365 	struct ipoib_tx_buf *tx_req;
366 
367 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
368 		       wr_id, wc->status);
369 
370 	if (unlikely(wr_id >= ipoib_sendq_size)) {
371 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
372 			   wr_id, ipoib_sendq_size);
373 		return;
374 	}
375 
376 	tx_req = &priv->tx_ring[wr_id];
377 
378 	ipoib_dma_unmap_tx(priv->ca, tx_req);
379 
380 	++dev->stats.tx_packets;
381 	dev->stats.tx_bytes += tx_req->skb->len;
382 
383 	dev_kfree_skb_any(tx_req->skb);
384 
385 	++priv->tx_tail;
386 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
387 	    netif_queue_stopped(dev) &&
388 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
389 		netif_wake_queue(dev);
390 
391 	if (wc->status != IB_WC_SUCCESS &&
392 	    wc->status != IB_WC_WR_FLUSH_ERR)
393 		ipoib_warn(priv, "failed send event "
394 			   "(status=%d, wrid=%d vend_err %x)\n",
395 			   wc->status, wr_id, wc->vendor_err);
396 }
397 
398 static int poll_tx(struct ipoib_dev_priv *priv)
399 {
400 	int n, i;
401 
402 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
403 	for (i = 0; i < n; ++i)
404 		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
405 
406 	return n == MAX_SEND_CQE;
407 }
408 
409 int ipoib_poll(struct napi_struct *napi, int budget)
410 {
411 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
412 	struct net_device *dev = priv->dev;
413 	int done;
414 	int t;
415 	int n, i;
416 
417 	done  = 0;
418 
419 poll_more:
420 	while (done < budget) {
421 		int max = (budget - done);
422 
423 		t = min(IPOIB_NUM_WC, max);
424 		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
425 
426 		for (i = 0; i < n; i++) {
427 			struct ib_wc *wc = priv->ibwc + i;
428 
429 			if (wc->wr_id & IPOIB_OP_RECV) {
430 				++done;
431 				if (wc->wr_id & IPOIB_OP_CM)
432 					ipoib_cm_handle_rx_wc(dev, wc);
433 				else
434 					ipoib_ib_handle_rx_wc(dev, wc);
435 			} else
436 				ipoib_cm_handle_tx_wc(priv->dev, wc);
437 		}
438 
439 		if (n != t)
440 			break;
441 	}
442 
443 	if (done < budget) {
444 		if (dev->features & NETIF_F_LRO)
445 			lro_flush_all(&priv->lro.lro_mgr);
446 
447 		napi_complete(napi);
448 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
449 					      IB_CQ_NEXT_COMP |
450 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
451 		    napi_reschedule(napi))
452 			goto poll_more;
453 	}
454 
455 	return done;
456 }
457 
458 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
459 {
460 	struct net_device *dev = dev_ptr;
461 	struct ipoib_dev_priv *priv = netdev_priv(dev);
462 
463 	napi_schedule(&priv->napi);
464 }
465 
466 static void drain_tx_cq(struct net_device *dev)
467 {
468 	struct ipoib_dev_priv *priv = netdev_priv(dev);
469 
470 	netif_tx_lock(dev);
471 	while (poll_tx(priv))
472 		; /* nothing */
473 
474 	if (netif_queue_stopped(dev))
475 		mod_timer(&priv->poll_timer, jiffies + 1);
476 
477 	netif_tx_unlock(dev);
478 }
479 
480 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
481 {
482 	struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
483 
484 	mod_timer(&priv->poll_timer, jiffies);
485 }
486 
487 static inline int post_send(struct ipoib_dev_priv *priv,
488 			    unsigned int wr_id,
489 			    struct ib_ah *address, u32 qpn,
490 			    struct ipoib_tx_buf *tx_req,
491 			    void *head, int hlen)
492 {
493 	struct ib_send_wr *bad_wr;
494 	int i, off;
495 	struct sk_buff *skb = tx_req->skb;
496 	skb_frag_t *frags = skb_shinfo(skb)->frags;
497 	int nr_frags = skb_shinfo(skb)->nr_frags;
498 	u64 *mapping = tx_req->mapping;
499 
500 	if (skb_headlen(skb)) {
501 		priv->tx_sge[0].addr         = mapping[0];
502 		priv->tx_sge[0].length       = skb_headlen(skb);
503 		off = 1;
504 	} else
505 		off = 0;
506 
507 	for (i = 0; i < nr_frags; ++i) {
508 		priv->tx_sge[i + off].addr = mapping[i + off];
509 		priv->tx_sge[i + off].length = frags[i].size;
510 	}
511 	priv->tx_wr.num_sge	     = nr_frags + off;
512 	priv->tx_wr.wr_id 	     = wr_id;
513 	priv->tx_wr.wr.ud.remote_qpn = qpn;
514 	priv->tx_wr.wr.ud.ah 	     = address;
515 
516 	if (head) {
517 		priv->tx_wr.wr.ud.mss	 = skb_shinfo(skb)->gso_size;
518 		priv->tx_wr.wr.ud.header = head;
519 		priv->tx_wr.wr.ud.hlen	 = hlen;
520 		priv->tx_wr.opcode	 = IB_WR_LSO;
521 	} else
522 		priv->tx_wr.opcode	 = IB_WR_SEND;
523 
524 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
525 }
526 
527 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
528 		struct ipoib_ah *address, u32 qpn)
529 {
530 	struct ipoib_dev_priv *priv = netdev_priv(dev);
531 	struct ipoib_tx_buf *tx_req;
532 	int hlen;
533 	void *phead;
534 
535 	if (skb_is_gso(skb)) {
536 		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
537 		phead = skb->data;
538 		if (unlikely(!skb_pull(skb, hlen))) {
539 			ipoib_warn(priv, "linear data too small\n");
540 			++dev->stats.tx_dropped;
541 			++dev->stats.tx_errors;
542 			dev_kfree_skb_any(skb);
543 			return;
544 		}
545 	} else {
546 		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
547 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
548 				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
549 			++dev->stats.tx_dropped;
550 			++dev->stats.tx_errors;
551 			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
552 			return;
553 		}
554 		phead = NULL;
555 		hlen  = 0;
556 	}
557 
558 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
559 		       skb->len, address, qpn);
560 
561 	/*
562 	 * We put the skb into the tx_ring _before_ we call post_send()
563 	 * because it's entirely possible that the completion handler will
564 	 * run before we execute anything after the post_send().  That
565 	 * means we have to make sure everything is properly recorded and
566 	 * our state is consistent before we call post_send().
567 	 */
568 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
569 	tx_req->skb = skb;
570 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
571 		++dev->stats.tx_errors;
572 		dev_kfree_skb_any(skb);
573 		return;
574 	}
575 
576 	if (skb->ip_summed == CHECKSUM_PARTIAL)
577 		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
578 	else
579 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
580 
581 	if (++priv->tx_outstanding == ipoib_sendq_size) {
582 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
583 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
584 			ipoib_warn(priv, "request notify on send CQ failed\n");
585 		netif_stop_queue(dev);
586 	}
587 
588 	if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
589 			       address->ah, qpn, tx_req, phead, hlen))) {
590 		ipoib_warn(priv, "post_send failed\n");
591 		++dev->stats.tx_errors;
592 		--priv->tx_outstanding;
593 		ipoib_dma_unmap_tx(priv->ca, tx_req);
594 		dev_kfree_skb_any(skb);
595 		if (netif_queue_stopped(dev))
596 			netif_wake_queue(dev);
597 	} else {
598 		dev->trans_start = jiffies;
599 
600 		address->last_send = priv->tx_head;
601 		++priv->tx_head;
602 		skb_orphan(skb);
603 
604 	}
605 
606 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
607 		while (poll_tx(priv))
608 			; /* nothing */
609 }
610 
611 static void __ipoib_reap_ah(struct net_device *dev)
612 {
613 	struct ipoib_dev_priv *priv = netdev_priv(dev);
614 	struct ipoib_ah *ah, *tah;
615 	LIST_HEAD(remove_list);
616 	unsigned long flags;
617 
618 	netif_tx_lock_bh(dev);
619 	spin_lock_irqsave(&priv->lock, flags);
620 
621 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
622 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
623 			list_del(&ah->list);
624 			ib_destroy_ah(ah->ah);
625 			kfree(ah);
626 		}
627 
628 	spin_unlock_irqrestore(&priv->lock, flags);
629 	netif_tx_unlock_bh(dev);
630 }
631 
632 void ipoib_reap_ah(struct work_struct *work)
633 {
634 	struct ipoib_dev_priv *priv =
635 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
636 	struct net_device *dev = priv->dev;
637 
638 	__ipoib_reap_ah(dev);
639 
640 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
641 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
642 				   round_jiffies_relative(HZ));
643 }
644 
645 static void ipoib_ib_tx_timer_func(unsigned long ctx)
646 {
647 	drain_tx_cq((struct net_device *)ctx);
648 }
649 
650 int ipoib_ib_dev_open(struct net_device *dev)
651 {
652 	struct ipoib_dev_priv *priv = netdev_priv(dev);
653 	int ret;
654 
655 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
656 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
657 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
658 		return -1;
659 	}
660 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
661 
662 	ret = ipoib_init_qp(dev);
663 	if (ret) {
664 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
665 		return -1;
666 	}
667 
668 	ret = ipoib_ib_post_receives(dev);
669 	if (ret) {
670 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
671 		ipoib_ib_dev_stop(dev, 1);
672 		return -1;
673 	}
674 
675 	ret = ipoib_cm_dev_open(dev);
676 	if (ret) {
677 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
678 		ipoib_ib_dev_stop(dev, 1);
679 		return -1;
680 	}
681 
682 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
683 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
684 			   round_jiffies_relative(HZ));
685 
686 	if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
687 		napi_enable(&priv->napi);
688 
689 	return 0;
690 }
691 
692 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
693 {
694 	struct ipoib_dev_priv *priv = netdev_priv(dev);
695 	u16 pkey_index = 0;
696 
697 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
698 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
699 	else
700 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
701 }
702 
703 int ipoib_ib_dev_up(struct net_device *dev)
704 {
705 	struct ipoib_dev_priv *priv = netdev_priv(dev);
706 
707 	ipoib_pkey_dev_check_presence(dev);
708 
709 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
710 		ipoib_dbg(priv, "PKEY is not assigned.\n");
711 		return 0;
712 	}
713 
714 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
715 
716 	return ipoib_mcast_start_thread(dev);
717 }
718 
719 int ipoib_ib_dev_down(struct net_device *dev, int flush)
720 {
721 	struct ipoib_dev_priv *priv = netdev_priv(dev);
722 
723 	ipoib_dbg(priv, "downing ib_dev\n");
724 
725 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
726 	netif_carrier_off(dev);
727 
728 	/* Shutdown the P_Key thread if still active */
729 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
730 		mutex_lock(&pkey_mutex);
731 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
732 		cancel_delayed_work(&priv->pkey_poll_task);
733 		mutex_unlock(&pkey_mutex);
734 		if (flush)
735 			flush_workqueue(ipoib_workqueue);
736 	}
737 
738 	ipoib_mcast_stop_thread(dev, flush);
739 	ipoib_mcast_dev_flush(dev);
740 
741 	ipoib_flush_paths(dev);
742 
743 	return 0;
744 }
745 
746 static int recvs_pending(struct net_device *dev)
747 {
748 	struct ipoib_dev_priv *priv = netdev_priv(dev);
749 	int pending = 0;
750 	int i;
751 
752 	for (i = 0; i < ipoib_recvq_size; ++i)
753 		if (priv->rx_ring[i].skb)
754 			++pending;
755 
756 	return pending;
757 }
758 
759 void ipoib_drain_cq(struct net_device *dev)
760 {
761 	struct ipoib_dev_priv *priv = netdev_priv(dev);
762 	int i, n;
763 
764 	/*
765 	 * We call completion handling routines that expect to be
766 	 * called from the BH-disabled NAPI poll context, so disable
767 	 * BHs here too.
768 	 */
769 	local_bh_disable();
770 
771 	do {
772 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
773 		for (i = 0; i < n; ++i) {
774 			/*
775 			 * Convert any successful completions to flush
776 			 * errors to avoid passing packets up the
777 			 * stack after bringing the device down.
778 			 */
779 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
780 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
781 
782 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
783 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
784 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
785 				else
786 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
787 			} else
788 				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
789 		}
790 	} while (n == IPOIB_NUM_WC);
791 
792 	while (poll_tx(priv))
793 		; /* nothing */
794 
795 	local_bh_enable();
796 }
797 
798 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
799 {
800 	struct ipoib_dev_priv *priv = netdev_priv(dev);
801 	struct ib_qp_attr qp_attr;
802 	unsigned long begin;
803 	struct ipoib_tx_buf *tx_req;
804 	int i;
805 
806 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
807 		napi_disable(&priv->napi);
808 
809 	ipoib_cm_dev_stop(dev);
810 
811 	/*
812 	 * Move our QP to the error state and then reinitialize in
813 	 * when all work requests have completed or have been flushed.
814 	 */
815 	qp_attr.qp_state = IB_QPS_ERR;
816 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
817 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
818 
819 	/* Wait for all sends and receives to complete */
820 	begin = jiffies;
821 
822 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
823 		if (time_after(jiffies, begin + 5 * HZ)) {
824 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
825 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
826 
827 			/*
828 			 * assume the HW is wedged and just free up
829 			 * all our pending work requests.
830 			 */
831 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
832 				tx_req = &priv->tx_ring[priv->tx_tail &
833 							(ipoib_sendq_size - 1)];
834 				ipoib_dma_unmap_tx(priv->ca, tx_req);
835 				dev_kfree_skb_any(tx_req->skb);
836 				++priv->tx_tail;
837 				--priv->tx_outstanding;
838 			}
839 
840 			for (i = 0; i < ipoib_recvq_size; ++i) {
841 				struct ipoib_rx_buf *rx_req;
842 
843 				rx_req = &priv->rx_ring[i];
844 				if (!rx_req->skb)
845 					continue;
846 				ipoib_ud_dma_unmap_rx(priv,
847 						      priv->rx_ring[i].mapping);
848 				dev_kfree_skb_any(rx_req->skb);
849 				rx_req->skb = NULL;
850 			}
851 
852 			goto timeout;
853 		}
854 
855 		ipoib_drain_cq(dev);
856 
857 		msleep(1);
858 	}
859 
860 	ipoib_dbg(priv, "All sends and receives done.\n");
861 
862 timeout:
863 	del_timer_sync(&priv->poll_timer);
864 	qp_attr.qp_state = IB_QPS_RESET;
865 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
866 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
867 
868 	/* Wait for all AHs to be reaped */
869 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
870 	cancel_delayed_work(&priv->ah_reap_task);
871 	if (flush)
872 		flush_workqueue(ipoib_workqueue);
873 
874 	begin = jiffies;
875 
876 	while (!list_empty(&priv->dead_ahs)) {
877 		__ipoib_reap_ah(dev);
878 
879 		if (time_after(jiffies, begin + HZ)) {
880 			ipoib_warn(priv, "timing out; will leak address handles\n");
881 			break;
882 		}
883 
884 		msleep(1);
885 	}
886 
887 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
888 
889 	return 0;
890 }
891 
892 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
893 {
894 	struct ipoib_dev_priv *priv = netdev_priv(dev);
895 
896 	priv->ca = ca;
897 	priv->port = port;
898 	priv->qp = NULL;
899 
900 	if (ipoib_transport_dev_init(dev, ca)) {
901 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
902 		return -ENODEV;
903 	}
904 
905 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
906 		    (unsigned long) dev);
907 
908 	if (dev->flags & IFF_UP) {
909 		if (ipoib_ib_dev_open(dev)) {
910 			ipoib_transport_dev_cleanup(dev);
911 			return -ENODEV;
912 		}
913 	}
914 
915 	return 0;
916 }
917 
918 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
919 				enum ipoib_flush_level level)
920 {
921 	struct ipoib_dev_priv *cpriv;
922 	struct net_device *dev = priv->dev;
923 	u16 new_index;
924 
925 	mutex_lock(&priv->vlan_mutex);
926 
927 	/*
928 	 * Flush any child interfaces too -- they might be up even if
929 	 * the parent is down.
930 	 */
931 	list_for_each_entry(cpriv, &priv->child_intfs, list)
932 		__ipoib_ib_dev_flush(cpriv, level);
933 
934 	mutex_unlock(&priv->vlan_mutex);
935 
936 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
937 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
938 		return;
939 	}
940 
941 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
942 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
943 		return;
944 	}
945 
946 	if (level == IPOIB_FLUSH_HEAVY) {
947 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
948 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
949 			ipoib_ib_dev_down(dev, 0);
950 			ipoib_ib_dev_stop(dev, 0);
951 			if (ipoib_pkey_dev_delay_open(dev))
952 				return;
953 		}
954 
955 		/* restart QP only if P_Key index is changed */
956 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
957 		    new_index == priv->pkey_index) {
958 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
959 			return;
960 		}
961 		priv->pkey_index = new_index;
962 	}
963 
964 	if (level == IPOIB_FLUSH_LIGHT) {
965 		ipoib_mark_paths_invalid(dev);
966 		ipoib_mcast_dev_flush(dev);
967 	}
968 
969 	if (level >= IPOIB_FLUSH_NORMAL)
970 		ipoib_ib_dev_down(dev, 0);
971 
972 	if (level == IPOIB_FLUSH_HEAVY) {
973 		ipoib_ib_dev_stop(dev, 0);
974 		ipoib_ib_dev_open(dev);
975 	}
976 
977 	/*
978 	 * The device could have been brought down between the start and when
979 	 * we get here, don't bring it back up if it's not configured up
980 	 */
981 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
982 		if (level >= IPOIB_FLUSH_NORMAL)
983 			ipoib_ib_dev_up(dev);
984 		ipoib_mcast_restart_task(&priv->restart_task);
985 	}
986 }
987 
988 void ipoib_ib_dev_flush_light(struct work_struct *work)
989 {
990 	struct ipoib_dev_priv *priv =
991 		container_of(work, struct ipoib_dev_priv, flush_light);
992 
993 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
994 }
995 
996 void ipoib_ib_dev_flush_normal(struct work_struct *work)
997 {
998 	struct ipoib_dev_priv *priv =
999 		container_of(work, struct ipoib_dev_priv, flush_normal);
1000 
1001 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1002 }
1003 
1004 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1005 {
1006 	struct ipoib_dev_priv *priv =
1007 		container_of(work, struct ipoib_dev_priv, flush_heavy);
1008 
1009 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1010 }
1011 
1012 void ipoib_ib_dev_cleanup(struct net_device *dev)
1013 {
1014 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1015 
1016 	ipoib_dbg(priv, "cleaning up ib_dev\n");
1017 
1018 	ipoib_mcast_stop_thread(dev, 1);
1019 	ipoib_mcast_dev_flush(dev);
1020 
1021 	ipoib_transport_dev_cleanup(dev);
1022 }
1023 
1024 /*
1025  * Delayed P_Key Assigment Interim Support
1026  *
1027  * The following is initial implementation of delayed P_Key assigment
1028  * mechanism. It is using the same approach implemented for the multicast
1029  * group join. The single goal of this implementation is to quickly address
1030  * Bug #2507. This implementation will probably be removed when the P_Key
1031  * change async notification is available.
1032  */
1033 
1034 void ipoib_pkey_poll(struct work_struct *work)
1035 {
1036 	struct ipoib_dev_priv *priv =
1037 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1038 	struct net_device *dev = priv->dev;
1039 
1040 	ipoib_pkey_dev_check_presence(dev);
1041 
1042 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1043 		ipoib_open(dev);
1044 	else {
1045 		mutex_lock(&pkey_mutex);
1046 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1047 			queue_delayed_work(ipoib_workqueue,
1048 					   &priv->pkey_poll_task,
1049 					   HZ);
1050 		mutex_unlock(&pkey_mutex);
1051 	}
1052 }
1053 
1054 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1055 {
1056 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1057 
1058 	/* Look for the interface pkey value in the IB Port P_Key table and */
1059 	/* set the interface pkey assigment flag                            */
1060 	ipoib_pkey_dev_check_presence(dev);
1061 
1062 	/* P_Key value not assigned yet - start polling */
1063 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1064 		mutex_lock(&pkey_mutex);
1065 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1066 		queue_delayed_work(ipoib_workqueue,
1067 				   &priv->pkey_poll_task,
1068 				   HZ);
1069 		mutex_unlock(&pkey_mutex);
1070 		return 1;
1071 	}
1072 
1073 	return 0;
1074 }
1075