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 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
98 		ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
99 				    DMA_FROM_DEVICE);
100 		ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
101 				  DMA_FROM_DEVICE);
102 	} else
103 		ib_dma_unmap_single(priv->ca, mapping[0],
104 				    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
105 				    DMA_FROM_DEVICE);
106 }
107 
108 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
109 				   struct sk_buff *skb,
110 				   unsigned int length)
111 {
112 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
113 		skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
114 		unsigned int size;
115 		/*
116 		 * There is only two buffers needed for max_payload = 4K,
117 		 * first buf size is IPOIB_UD_HEAD_SIZE
118 		 */
119 		skb->tail += IPOIB_UD_HEAD_SIZE;
120 		skb->len  += length;
121 
122 		size = length - IPOIB_UD_HEAD_SIZE;
123 
124 		skb_frag_size_set(frag, size);
125 		skb->data_len += size;
126 		skb->truesize += size;
127 	} else
128 		skb_put(skb, length);
129 
130 }
131 
132 static int ipoib_ib_post_receive(struct net_device *dev, int id)
133 {
134 	struct ipoib_dev_priv *priv = netdev_priv(dev);
135 	struct ib_recv_wr *bad_wr;
136 	int ret;
137 
138 	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
139 	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
140 	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
141 
142 
143 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
144 	if (unlikely(ret)) {
145 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
146 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
147 		dev_kfree_skb_any(priv->rx_ring[id].skb);
148 		priv->rx_ring[id].skb = NULL;
149 	}
150 
151 	return ret;
152 }
153 
154 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
155 {
156 	struct ipoib_dev_priv *priv = netdev_priv(dev);
157 	struct sk_buff *skb;
158 	int buf_size;
159 	u64 *mapping;
160 
161 	if (ipoib_ud_need_sg(priv->max_ib_mtu))
162 		buf_size = IPOIB_UD_HEAD_SIZE;
163 	else
164 		buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
165 
166 	skb = dev_alloc_skb(buf_size + 4);
167 	if (unlikely(!skb))
168 		return NULL;
169 
170 	/*
171 	 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
172 	 * header.  So we need 4 more bytes to get to 48 and align the
173 	 * IP header to a multiple of 16.
174 	 */
175 	skb_reserve(skb, 4);
176 
177 	mapping = priv->rx_ring[id].mapping;
178 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
179 				       DMA_FROM_DEVICE);
180 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
181 		goto error;
182 
183 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
184 		struct page *page = alloc_page(GFP_ATOMIC);
185 		if (!page)
186 			goto partial_error;
187 		skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
188 		mapping[1] =
189 			ib_dma_map_page(priv->ca, page,
190 					0, PAGE_SIZE, DMA_FROM_DEVICE);
191 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
192 			goto partial_error;
193 	}
194 
195 	priv->rx_ring[id].skb = skb;
196 	return skb;
197 
198 partial_error:
199 	ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
200 error:
201 	dev_kfree_skb_any(skb);
202 	return NULL;
203 }
204 
205 static int ipoib_ib_post_receives(struct net_device *dev)
206 {
207 	struct ipoib_dev_priv *priv = netdev_priv(dev);
208 	int i;
209 
210 	for (i = 0; i < ipoib_recvq_size; ++i) {
211 		if (!ipoib_alloc_rx_skb(dev, i)) {
212 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
213 			return -ENOMEM;
214 		}
215 		if (ipoib_ib_post_receive(dev, i)) {
216 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
217 			return -EIO;
218 		}
219 	}
220 
221 	return 0;
222 }
223 
224 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
225 {
226 	struct ipoib_dev_priv *priv = netdev_priv(dev);
227 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
228 	struct sk_buff *skb;
229 	u64 mapping[IPOIB_UD_RX_SG];
230 	union ib_gid *dgid;
231 
232 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
233 		       wr_id, wc->status);
234 
235 	if (unlikely(wr_id >= ipoib_recvq_size)) {
236 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
237 			   wr_id, ipoib_recvq_size);
238 		return;
239 	}
240 
241 	skb  = priv->rx_ring[wr_id].skb;
242 
243 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
244 		if (wc->status != IB_WC_WR_FLUSH_ERR)
245 			ipoib_warn(priv, "failed recv event "
246 				   "(status=%d, wrid=%d vend_err %x)\n",
247 				   wc->status, wr_id, wc->vendor_err);
248 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
249 		dev_kfree_skb_any(skb);
250 		priv->rx_ring[wr_id].skb = NULL;
251 		return;
252 	}
253 
254 	/*
255 	 * Drop packets that this interface sent, ie multicast packets
256 	 * that the HCA has replicated.
257 	 */
258 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
259 		goto repost;
260 
261 	memcpy(mapping, priv->rx_ring[wr_id].mapping,
262 	       IPOIB_UD_RX_SG * sizeof *mapping);
263 
264 	/*
265 	 * If we can't allocate a new RX buffer, dump
266 	 * this packet and reuse the old buffer.
267 	 */
268 	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
269 		++dev->stats.rx_dropped;
270 		goto repost;
271 	}
272 
273 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
274 		       wc->byte_len, wc->slid);
275 
276 	ipoib_ud_dma_unmap_rx(priv, mapping);
277 	ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
278 
279 	/* First byte of dgid signals multicast when 0xff */
280 	dgid = &((struct ib_grh *)skb->data)->dgid;
281 
282 	if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
283 		skb->pkt_type = PACKET_HOST;
284 	else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
285 		skb->pkt_type = PACKET_BROADCAST;
286 	else
287 		skb->pkt_type = PACKET_MULTICAST;
288 
289 	skb_pull(skb, IB_GRH_BYTES);
290 
291 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
292 	skb_reset_mac_header(skb);
293 	skb_pull(skb, IPOIB_ENCAP_LEN);
294 
295 	++dev->stats.rx_packets;
296 	dev->stats.rx_bytes += skb->len;
297 
298 	skb->dev = dev;
299 	if ((dev->features & NETIF_F_RXCSUM) &&
300 			likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
301 		skb->ip_summed = CHECKSUM_UNNECESSARY;
302 
303 	napi_gro_receive(&priv->napi, skb);
304 
305 repost:
306 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
307 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
308 			   "for buf %d\n", wr_id);
309 }
310 
311 static int ipoib_dma_map_tx(struct ib_device *ca,
312 			    struct ipoib_tx_buf *tx_req)
313 {
314 	struct sk_buff *skb = tx_req->skb;
315 	u64 *mapping = tx_req->mapping;
316 	int i;
317 	int off;
318 
319 	if (skb_headlen(skb)) {
320 		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
321 					       DMA_TO_DEVICE);
322 		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
323 			return -EIO;
324 
325 		off = 1;
326 	} else
327 		off = 0;
328 
329 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
330 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
331 		mapping[i + off] = ib_dma_map_page(ca,
332 						 skb_frag_page(frag),
333 						 frag->page_offset, skb_frag_size(frag),
334 						 DMA_TO_DEVICE);
335 		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
336 			goto partial_error;
337 	}
338 	return 0;
339 
340 partial_error:
341 	for (; i > 0; --i) {
342 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
343 
344 		ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
345 	}
346 
347 	if (off)
348 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
349 
350 	return -EIO;
351 }
352 
353 static void ipoib_dma_unmap_tx(struct ib_device *ca,
354 			       struct ipoib_tx_buf *tx_req)
355 {
356 	struct sk_buff *skb = tx_req->skb;
357 	u64 *mapping = tx_req->mapping;
358 	int i;
359 	int off;
360 
361 	if (skb_headlen(skb)) {
362 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
363 		off = 1;
364 	} else
365 		off = 0;
366 
367 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
368 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
369 
370 		ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
371 				  DMA_TO_DEVICE);
372 	}
373 }
374 
375 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
376 {
377 	struct ipoib_dev_priv *priv = netdev_priv(dev);
378 	unsigned int wr_id = wc->wr_id;
379 	struct ipoib_tx_buf *tx_req;
380 
381 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
382 		       wr_id, wc->status);
383 
384 	if (unlikely(wr_id >= ipoib_sendq_size)) {
385 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
386 			   wr_id, ipoib_sendq_size);
387 		return;
388 	}
389 
390 	tx_req = &priv->tx_ring[wr_id];
391 
392 	ipoib_dma_unmap_tx(priv->ca, tx_req);
393 
394 	++dev->stats.tx_packets;
395 	dev->stats.tx_bytes += tx_req->skb->len;
396 
397 	dev_kfree_skb_any(tx_req->skb);
398 
399 	++priv->tx_tail;
400 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
401 	    netif_queue_stopped(dev) &&
402 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
403 		netif_wake_queue(dev);
404 
405 	if (wc->status != IB_WC_SUCCESS &&
406 	    wc->status != IB_WC_WR_FLUSH_ERR)
407 		ipoib_warn(priv, "failed send event "
408 			   "(status=%d, wrid=%d vend_err %x)\n",
409 			   wc->status, wr_id, wc->vendor_err);
410 }
411 
412 static int poll_tx(struct ipoib_dev_priv *priv)
413 {
414 	int n, i;
415 
416 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
417 	for (i = 0; i < n; ++i)
418 		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
419 
420 	return n == MAX_SEND_CQE;
421 }
422 
423 int ipoib_poll(struct napi_struct *napi, int budget)
424 {
425 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
426 	struct net_device *dev = priv->dev;
427 	int done;
428 	int t;
429 	int n, i;
430 
431 	done  = 0;
432 
433 poll_more:
434 	while (done < budget) {
435 		int max = (budget - done);
436 
437 		t = min(IPOIB_NUM_WC, max);
438 		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
439 
440 		for (i = 0; i < n; i++) {
441 			struct ib_wc *wc = priv->ibwc + i;
442 
443 			if (wc->wr_id & IPOIB_OP_RECV) {
444 				++done;
445 				if (wc->wr_id & IPOIB_OP_CM)
446 					ipoib_cm_handle_rx_wc(dev, wc);
447 				else
448 					ipoib_ib_handle_rx_wc(dev, wc);
449 			} else
450 				ipoib_cm_handle_tx_wc(priv->dev, wc);
451 		}
452 
453 		if (n != t)
454 			break;
455 	}
456 
457 	if (done < budget) {
458 		napi_complete(napi);
459 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
460 					      IB_CQ_NEXT_COMP |
461 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
462 		    napi_reschedule(napi))
463 			goto poll_more;
464 	}
465 
466 	return done;
467 }
468 
469 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
470 {
471 	struct net_device *dev = dev_ptr;
472 	struct ipoib_dev_priv *priv = netdev_priv(dev);
473 
474 	napi_schedule(&priv->napi);
475 }
476 
477 static void drain_tx_cq(struct net_device *dev)
478 {
479 	struct ipoib_dev_priv *priv = netdev_priv(dev);
480 
481 	netif_tx_lock(dev);
482 	while (poll_tx(priv))
483 		; /* nothing */
484 
485 	if (netif_queue_stopped(dev))
486 		mod_timer(&priv->poll_timer, jiffies + 1);
487 
488 	netif_tx_unlock(dev);
489 }
490 
491 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
492 {
493 	struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
494 
495 	mod_timer(&priv->poll_timer, jiffies);
496 }
497 
498 static inline int post_send(struct ipoib_dev_priv *priv,
499 			    unsigned int wr_id,
500 			    struct ib_ah *address, u32 qpn,
501 			    struct ipoib_tx_buf *tx_req,
502 			    void *head, int hlen)
503 {
504 	struct ib_send_wr *bad_wr;
505 	int i, off;
506 	struct sk_buff *skb = tx_req->skb;
507 	skb_frag_t *frags = skb_shinfo(skb)->frags;
508 	int nr_frags = skb_shinfo(skb)->nr_frags;
509 	u64 *mapping = tx_req->mapping;
510 
511 	if (skb_headlen(skb)) {
512 		priv->tx_sge[0].addr         = mapping[0];
513 		priv->tx_sge[0].length       = skb_headlen(skb);
514 		off = 1;
515 	} else
516 		off = 0;
517 
518 	for (i = 0; i < nr_frags; ++i) {
519 		priv->tx_sge[i + off].addr = mapping[i + off];
520 		priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
521 	}
522 	priv->tx_wr.num_sge	     = nr_frags + off;
523 	priv->tx_wr.wr_id 	     = wr_id;
524 	priv->tx_wr.wr.ud.remote_qpn = qpn;
525 	priv->tx_wr.wr.ud.ah 	     = address;
526 
527 	if (head) {
528 		priv->tx_wr.wr.ud.mss	 = skb_shinfo(skb)->gso_size;
529 		priv->tx_wr.wr.ud.header = head;
530 		priv->tx_wr.wr.ud.hlen	 = hlen;
531 		priv->tx_wr.opcode	 = IB_WR_LSO;
532 	} else
533 		priv->tx_wr.opcode	 = IB_WR_SEND;
534 
535 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
536 }
537 
538 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
539 		struct ipoib_ah *address, u32 qpn)
540 {
541 	struct ipoib_dev_priv *priv = netdev_priv(dev);
542 	struct ipoib_tx_buf *tx_req;
543 	int hlen, rc;
544 	void *phead;
545 
546 	if (skb_is_gso(skb)) {
547 		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
548 		phead = skb->data;
549 		if (unlikely(!skb_pull(skb, hlen))) {
550 			ipoib_warn(priv, "linear data too small\n");
551 			++dev->stats.tx_dropped;
552 			++dev->stats.tx_errors;
553 			dev_kfree_skb_any(skb);
554 			return;
555 		}
556 	} else {
557 		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
558 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
559 				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
560 			++dev->stats.tx_dropped;
561 			++dev->stats.tx_errors;
562 			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
563 			return;
564 		}
565 		phead = NULL;
566 		hlen  = 0;
567 	}
568 
569 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
570 		       skb->len, address, qpn);
571 
572 	/*
573 	 * We put the skb into the tx_ring _before_ we call post_send()
574 	 * because it's entirely possible that the completion handler will
575 	 * run before we execute anything after the post_send().  That
576 	 * means we have to make sure everything is properly recorded and
577 	 * our state is consistent before we call post_send().
578 	 */
579 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
580 	tx_req->skb = skb;
581 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
582 		++dev->stats.tx_errors;
583 		dev_kfree_skb_any(skb);
584 		return;
585 	}
586 
587 	if (skb->ip_summed == CHECKSUM_PARTIAL)
588 		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
589 	else
590 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
591 
592 	if (++priv->tx_outstanding == ipoib_sendq_size) {
593 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
594 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
595 			ipoib_warn(priv, "request notify on send CQ failed\n");
596 		netif_stop_queue(dev);
597 	}
598 
599 	rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
600 		       address->ah, qpn, tx_req, phead, hlen);
601 	if (unlikely(rc)) {
602 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
603 		++dev->stats.tx_errors;
604 		--priv->tx_outstanding;
605 		ipoib_dma_unmap_tx(priv->ca, tx_req);
606 		dev_kfree_skb_any(skb);
607 		if (netif_queue_stopped(dev))
608 			netif_wake_queue(dev);
609 	} else {
610 		dev->trans_start = jiffies;
611 
612 		address->last_send = priv->tx_head;
613 		++priv->tx_head;
614 		skb_orphan(skb);
615 
616 	}
617 
618 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
619 		while (poll_tx(priv))
620 			; /* nothing */
621 }
622 
623 static void __ipoib_reap_ah(struct net_device *dev)
624 {
625 	struct ipoib_dev_priv *priv = netdev_priv(dev);
626 	struct ipoib_ah *ah, *tah;
627 	LIST_HEAD(remove_list);
628 	unsigned long flags;
629 
630 	netif_tx_lock_bh(dev);
631 	spin_lock_irqsave(&priv->lock, flags);
632 
633 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
634 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
635 			list_del(&ah->list);
636 			ib_destroy_ah(ah->ah);
637 			kfree(ah);
638 		}
639 
640 	spin_unlock_irqrestore(&priv->lock, flags);
641 	netif_tx_unlock_bh(dev);
642 }
643 
644 void ipoib_reap_ah(struct work_struct *work)
645 {
646 	struct ipoib_dev_priv *priv =
647 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
648 	struct net_device *dev = priv->dev;
649 
650 	__ipoib_reap_ah(dev);
651 
652 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
653 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
654 				   round_jiffies_relative(HZ));
655 }
656 
657 static void ipoib_ib_tx_timer_func(unsigned long ctx)
658 {
659 	drain_tx_cq((struct net_device *)ctx);
660 }
661 
662 int ipoib_ib_dev_open(struct net_device *dev)
663 {
664 	struct ipoib_dev_priv *priv = netdev_priv(dev);
665 	int ret;
666 
667 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
668 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
669 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
670 		return -1;
671 	}
672 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
673 
674 	ret = ipoib_init_qp(dev);
675 	if (ret) {
676 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
677 		return -1;
678 	}
679 
680 	ret = ipoib_ib_post_receives(dev);
681 	if (ret) {
682 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
683 		ipoib_ib_dev_stop(dev, 1);
684 		return -1;
685 	}
686 
687 	ret = ipoib_cm_dev_open(dev);
688 	if (ret) {
689 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
690 		ipoib_ib_dev_stop(dev, 1);
691 		return -1;
692 	}
693 
694 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
695 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
696 			   round_jiffies_relative(HZ));
697 
698 	if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
699 		napi_enable(&priv->napi);
700 
701 	return 0;
702 }
703 
704 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
705 {
706 	struct ipoib_dev_priv *priv = netdev_priv(dev);
707 	u16 pkey_index = 0;
708 
709 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
710 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
711 	else
712 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
713 }
714 
715 int ipoib_ib_dev_up(struct net_device *dev)
716 {
717 	struct ipoib_dev_priv *priv = netdev_priv(dev);
718 
719 	ipoib_pkey_dev_check_presence(dev);
720 
721 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
722 		ipoib_dbg(priv, "PKEY is not assigned.\n");
723 		return 0;
724 	}
725 
726 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
727 
728 	return ipoib_mcast_start_thread(dev);
729 }
730 
731 int ipoib_ib_dev_down(struct net_device *dev, int flush)
732 {
733 	struct ipoib_dev_priv *priv = netdev_priv(dev);
734 
735 	ipoib_dbg(priv, "downing ib_dev\n");
736 
737 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
738 	netif_carrier_off(dev);
739 
740 	/* Shutdown the P_Key thread if still active */
741 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
742 		mutex_lock(&pkey_mutex);
743 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
744 		cancel_delayed_work(&priv->pkey_poll_task);
745 		mutex_unlock(&pkey_mutex);
746 		if (flush)
747 			flush_workqueue(ipoib_workqueue);
748 	}
749 
750 	ipoib_mcast_stop_thread(dev, flush);
751 	ipoib_mcast_dev_flush(dev);
752 
753 	ipoib_flush_paths(dev);
754 
755 	return 0;
756 }
757 
758 static int recvs_pending(struct net_device *dev)
759 {
760 	struct ipoib_dev_priv *priv = netdev_priv(dev);
761 	int pending = 0;
762 	int i;
763 
764 	for (i = 0; i < ipoib_recvq_size; ++i)
765 		if (priv->rx_ring[i].skb)
766 			++pending;
767 
768 	return pending;
769 }
770 
771 void ipoib_drain_cq(struct net_device *dev)
772 {
773 	struct ipoib_dev_priv *priv = netdev_priv(dev);
774 	int i, n;
775 
776 	/*
777 	 * We call completion handling routines that expect to be
778 	 * called from the BH-disabled NAPI poll context, so disable
779 	 * BHs here too.
780 	 */
781 	local_bh_disable();
782 
783 	do {
784 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
785 		for (i = 0; i < n; ++i) {
786 			/*
787 			 * Convert any successful completions to flush
788 			 * errors to avoid passing packets up the
789 			 * stack after bringing the device down.
790 			 */
791 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
792 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
793 
794 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
795 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
796 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
797 				else
798 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
799 			} else
800 				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
801 		}
802 	} while (n == IPOIB_NUM_WC);
803 
804 	while (poll_tx(priv))
805 		; /* nothing */
806 
807 	local_bh_enable();
808 }
809 
810 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
811 {
812 	struct ipoib_dev_priv *priv = netdev_priv(dev);
813 	struct ib_qp_attr qp_attr;
814 	unsigned long begin;
815 	struct ipoib_tx_buf *tx_req;
816 	int i;
817 
818 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
819 		napi_disable(&priv->napi);
820 
821 	ipoib_cm_dev_stop(dev);
822 
823 	/*
824 	 * Move our QP to the error state and then reinitialize in
825 	 * when all work requests have completed or have been flushed.
826 	 */
827 	qp_attr.qp_state = IB_QPS_ERR;
828 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
829 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
830 
831 	/* Wait for all sends and receives to complete */
832 	begin = jiffies;
833 
834 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
835 		if (time_after(jiffies, begin + 5 * HZ)) {
836 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
837 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
838 
839 			/*
840 			 * assume the HW is wedged and just free up
841 			 * all our pending work requests.
842 			 */
843 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
844 				tx_req = &priv->tx_ring[priv->tx_tail &
845 							(ipoib_sendq_size - 1)];
846 				ipoib_dma_unmap_tx(priv->ca, tx_req);
847 				dev_kfree_skb_any(tx_req->skb);
848 				++priv->tx_tail;
849 				--priv->tx_outstanding;
850 			}
851 
852 			for (i = 0; i < ipoib_recvq_size; ++i) {
853 				struct ipoib_rx_buf *rx_req;
854 
855 				rx_req = &priv->rx_ring[i];
856 				if (!rx_req->skb)
857 					continue;
858 				ipoib_ud_dma_unmap_rx(priv,
859 						      priv->rx_ring[i].mapping);
860 				dev_kfree_skb_any(rx_req->skb);
861 				rx_req->skb = NULL;
862 			}
863 
864 			goto timeout;
865 		}
866 
867 		ipoib_drain_cq(dev);
868 
869 		msleep(1);
870 	}
871 
872 	ipoib_dbg(priv, "All sends and receives done.\n");
873 
874 timeout:
875 	del_timer_sync(&priv->poll_timer);
876 	qp_attr.qp_state = IB_QPS_RESET;
877 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
878 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
879 
880 	/* Wait for all AHs to be reaped */
881 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
882 	cancel_delayed_work(&priv->ah_reap_task);
883 	if (flush)
884 		flush_workqueue(ipoib_workqueue);
885 
886 	begin = jiffies;
887 
888 	while (!list_empty(&priv->dead_ahs)) {
889 		__ipoib_reap_ah(dev);
890 
891 		if (time_after(jiffies, begin + HZ)) {
892 			ipoib_warn(priv, "timing out; will leak address handles\n");
893 			break;
894 		}
895 
896 		msleep(1);
897 	}
898 
899 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
900 
901 	return 0;
902 }
903 
904 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
905 {
906 	struct ipoib_dev_priv *priv = netdev_priv(dev);
907 
908 	priv->ca = ca;
909 	priv->port = port;
910 	priv->qp = NULL;
911 
912 	if (ipoib_transport_dev_init(dev, ca)) {
913 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
914 		return -ENODEV;
915 	}
916 
917 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
918 		    (unsigned long) dev);
919 
920 	if (dev->flags & IFF_UP) {
921 		if (ipoib_ib_dev_open(dev)) {
922 			ipoib_transport_dev_cleanup(dev);
923 			return -ENODEV;
924 		}
925 	}
926 
927 	return 0;
928 }
929 
930 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
931 				enum ipoib_flush_level level)
932 {
933 	struct ipoib_dev_priv *cpriv;
934 	struct net_device *dev = priv->dev;
935 	u16 new_index;
936 
937 	mutex_lock(&priv->vlan_mutex);
938 
939 	/*
940 	 * Flush any child interfaces too -- they might be up even if
941 	 * the parent is down.
942 	 */
943 	list_for_each_entry(cpriv, &priv->child_intfs, list)
944 		__ipoib_ib_dev_flush(cpriv, level);
945 
946 	mutex_unlock(&priv->vlan_mutex);
947 
948 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
949 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
950 		return;
951 	}
952 
953 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
954 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
955 		return;
956 	}
957 
958 	if (level == IPOIB_FLUSH_HEAVY) {
959 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
960 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
961 			ipoib_ib_dev_down(dev, 0);
962 			ipoib_ib_dev_stop(dev, 0);
963 			if (ipoib_pkey_dev_delay_open(dev))
964 				return;
965 		}
966 
967 		/* restart QP only if P_Key index is changed */
968 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
969 		    new_index == priv->pkey_index) {
970 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
971 			return;
972 		}
973 		priv->pkey_index = new_index;
974 	}
975 
976 	if (level == IPOIB_FLUSH_LIGHT) {
977 		ipoib_mark_paths_invalid(dev);
978 		ipoib_mcast_dev_flush(dev);
979 	}
980 
981 	if (level >= IPOIB_FLUSH_NORMAL)
982 		ipoib_ib_dev_down(dev, 0);
983 
984 	if (level == IPOIB_FLUSH_HEAVY) {
985 		ipoib_ib_dev_stop(dev, 0);
986 		ipoib_ib_dev_open(dev);
987 	}
988 
989 	/*
990 	 * The device could have been brought down between the start and when
991 	 * we get here, don't bring it back up if it's not configured up
992 	 */
993 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
994 		if (level >= IPOIB_FLUSH_NORMAL)
995 			ipoib_ib_dev_up(dev);
996 		ipoib_mcast_restart_task(&priv->restart_task);
997 	}
998 }
999 
1000 void ipoib_ib_dev_flush_light(struct work_struct *work)
1001 {
1002 	struct ipoib_dev_priv *priv =
1003 		container_of(work, struct ipoib_dev_priv, flush_light);
1004 
1005 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1006 }
1007 
1008 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1009 {
1010 	struct ipoib_dev_priv *priv =
1011 		container_of(work, struct ipoib_dev_priv, flush_normal);
1012 
1013 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1014 }
1015 
1016 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1017 {
1018 	struct ipoib_dev_priv *priv =
1019 		container_of(work, struct ipoib_dev_priv, flush_heavy);
1020 
1021 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1022 }
1023 
1024 void ipoib_ib_dev_cleanup(struct net_device *dev)
1025 {
1026 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1027 
1028 	ipoib_dbg(priv, "cleaning up ib_dev\n");
1029 
1030 	ipoib_mcast_stop_thread(dev, 1);
1031 	ipoib_mcast_dev_flush(dev);
1032 
1033 	ipoib_transport_dev_cleanup(dev);
1034 }
1035 
1036 /*
1037  * Delayed P_Key Assigment Interim Support
1038  *
1039  * The following is initial implementation of delayed P_Key assigment
1040  * mechanism. It is using the same approach implemented for the multicast
1041  * group join. The single goal of this implementation is to quickly address
1042  * Bug #2507. This implementation will probably be removed when the P_Key
1043  * change async notification is available.
1044  */
1045 
1046 void ipoib_pkey_poll(struct work_struct *work)
1047 {
1048 	struct ipoib_dev_priv *priv =
1049 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1050 	struct net_device *dev = priv->dev;
1051 
1052 	ipoib_pkey_dev_check_presence(dev);
1053 
1054 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1055 		ipoib_open(dev);
1056 	else {
1057 		mutex_lock(&pkey_mutex);
1058 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1059 			queue_delayed_work(ipoib_workqueue,
1060 					   &priv->pkey_poll_task,
1061 					   HZ);
1062 		mutex_unlock(&pkey_mutex);
1063 	}
1064 }
1065 
1066 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1067 {
1068 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1069 
1070 	/* Look for the interface pkey value in the IB Port P_Key table and */
1071 	/* set the interface pkey assigment flag                            */
1072 	ipoib_pkey_dev_check_presence(dev);
1073 
1074 	/* P_Key value not assigned yet - start polling */
1075 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1076 		mutex_lock(&pkey_mutex);
1077 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1078 		queue_delayed_work(ipoib_workqueue,
1079 				   &priv->pkey_poll_task,
1080 				   HZ);
1081 		mutex_unlock(&pkey_mutex);
1082 		return 1;
1083 	}
1084 
1085 	return 0;
1086 }
1087