xref: /openbmc/linux/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision baa7eb025ab14f3cba2e35c0a8648f9c9f01d24f)
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 #include <linux/slab.h>
39 
40 #include <linux/ip.h>
41 #include <linux/tcp.h>
42 
43 #include "ipoib.h"
44 
45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
46 static int data_debug_level;
47 
48 module_param(data_debug_level, int, 0644);
49 MODULE_PARM_DESC(data_debug_level,
50 		 "Enable data path debug tracing if > 0");
51 #endif
52 
53 static DEFINE_MUTEX(pkey_mutex);
54 
55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
56 				 struct ib_pd *pd, struct ib_ah_attr *attr)
57 {
58 	struct ipoib_ah *ah;
59 
60 	ah = kmalloc(sizeof *ah, GFP_KERNEL);
61 	if (!ah)
62 		return NULL;
63 
64 	ah->dev       = dev;
65 	ah->last_send = 0;
66 	kref_init(&ah->ref);
67 
68 	ah->ah = ib_create_ah(pd, attr);
69 	if (IS_ERR(ah->ah)) {
70 		kfree(ah);
71 		ah = NULL;
72 	} else
73 		ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
74 
75 	return ah;
76 }
77 
78 void ipoib_free_ah(struct kref *kref)
79 {
80 	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
81 	struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
82 
83 	unsigned long flags;
84 
85 	spin_lock_irqsave(&priv->lock, flags);
86 	list_add_tail(&ah->list, &priv->dead_ahs);
87 	spin_unlock_irqrestore(&priv->lock, flags);
88 }
89 
90 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
91 				  u64 mapping[IPOIB_UD_RX_SG])
92 {
93 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
94 		ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
95 				    DMA_FROM_DEVICE);
96 		ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
97 				  DMA_FROM_DEVICE);
98 	} else
99 		ib_dma_unmap_single(priv->ca, mapping[0],
100 				    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
101 				    DMA_FROM_DEVICE);
102 }
103 
104 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
105 				   struct sk_buff *skb,
106 				   unsigned int length)
107 {
108 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
109 		skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
110 		unsigned int size;
111 		/*
112 		 * There is only two buffers needed for max_payload = 4K,
113 		 * first buf size is IPOIB_UD_HEAD_SIZE
114 		 */
115 		skb->tail += IPOIB_UD_HEAD_SIZE;
116 		skb->len  += length;
117 
118 		size = length - IPOIB_UD_HEAD_SIZE;
119 
120 		frag->size     = size;
121 		skb->data_len += size;
122 		skb->truesize += size;
123 	} else
124 		skb_put(skb, length);
125 
126 }
127 
128 static int ipoib_ib_post_receive(struct net_device *dev, int id)
129 {
130 	struct ipoib_dev_priv *priv = netdev_priv(dev);
131 	struct ib_recv_wr *bad_wr;
132 	int ret;
133 
134 	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
135 	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
136 	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
137 
138 
139 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
140 	if (unlikely(ret)) {
141 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
142 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
143 		dev_kfree_skb_any(priv->rx_ring[id].skb);
144 		priv->rx_ring[id].skb = NULL;
145 	}
146 
147 	return ret;
148 }
149 
150 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
151 {
152 	struct ipoib_dev_priv *priv = netdev_priv(dev);
153 	struct sk_buff *skb;
154 	int buf_size;
155 	u64 *mapping;
156 
157 	if (ipoib_ud_need_sg(priv->max_ib_mtu))
158 		buf_size = IPOIB_UD_HEAD_SIZE;
159 	else
160 		buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
161 
162 	skb = dev_alloc_skb(buf_size + 4);
163 	if (unlikely(!skb))
164 		return NULL;
165 
166 	/*
167 	 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
168 	 * header.  So we need 4 more bytes to get to 48 and align the
169 	 * IP header to a multiple of 16.
170 	 */
171 	skb_reserve(skb, 4);
172 
173 	mapping = priv->rx_ring[id].mapping;
174 	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
175 				       DMA_FROM_DEVICE);
176 	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
177 		goto error;
178 
179 	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
180 		struct page *page = alloc_page(GFP_ATOMIC);
181 		if (!page)
182 			goto partial_error;
183 		skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
184 		mapping[1] =
185 			ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[0].page,
186 					0, PAGE_SIZE, DMA_FROM_DEVICE);
187 		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
188 			goto partial_error;
189 	}
190 
191 	priv->rx_ring[id].skb = skb;
192 	return skb;
193 
194 partial_error:
195 	ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
196 error:
197 	dev_kfree_skb_any(skb);
198 	return NULL;
199 }
200 
201 static int ipoib_ib_post_receives(struct net_device *dev)
202 {
203 	struct ipoib_dev_priv *priv = netdev_priv(dev);
204 	int i;
205 
206 	for (i = 0; i < ipoib_recvq_size; ++i) {
207 		if (!ipoib_alloc_rx_skb(dev, i)) {
208 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
209 			return -ENOMEM;
210 		}
211 		if (ipoib_ib_post_receive(dev, i)) {
212 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
213 			return -EIO;
214 		}
215 	}
216 
217 	return 0;
218 }
219 
220 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
221 {
222 	struct ipoib_dev_priv *priv = netdev_priv(dev);
223 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
224 	struct sk_buff *skb;
225 	u64 mapping[IPOIB_UD_RX_SG];
226 	union ib_gid *dgid;
227 
228 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
229 		       wr_id, wc->status);
230 
231 	if (unlikely(wr_id >= ipoib_recvq_size)) {
232 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
233 			   wr_id, ipoib_recvq_size);
234 		return;
235 	}
236 
237 	skb  = priv->rx_ring[wr_id].skb;
238 
239 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
240 		if (wc->status != IB_WC_WR_FLUSH_ERR)
241 			ipoib_warn(priv, "failed recv event "
242 				   "(status=%d, wrid=%d vend_err %x)\n",
243 				   wc->status, wr_id, wc->vendor_err);
244 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
245 		dev_kfree_skb_any(skb);
246 		priv->rx_ring[wr_id].skb = NULL;
247 		return;
248 	}
249 
250 	/*
251 	 * Drop packets that this interface sent, ie multicast packets
252 	 * that the HCA has replicated.
253 	 */
254 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
255 		goto repost;
256 
257 	memcpy(mapping, priv->rx_ring[wr_id].mapping,
258 	       IPOIB_UD_RX_SG * sizeof *mapping);
259 
260 	/*
261 	 * If we can't allocate a new RX buffer, dump
262 	 * this packet and reuse the old buffer.
263 	 */
264 	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
265 		++dev->stats.rx_dropped;
266 		goto repost;
267 	}
268 
269 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
270 		       wc->byte_len, wc->slid);
271 
272 	ipoib_ud_dma_unmap_rx(priv, mapping);
273 	ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
274 
275 	/* First byte of dgid signals multicast when 0xff */
276 	dgid = &((struct ib_grh *)skb->data)->dgid;
277 
278 	if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
279 		skb->pkt_type = PACKET_HOST;
280 	else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
281 		skb->pkt_type = PACKET_BROADCAST;
282 	else
283 		skb->pkt_type = PACKET_MULTICAST;
284 
285 	skb_pull(skb, IB_GRH_BYTES);
286 
287 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
288 	skb_reset_mac_header(skb);
289 	skb_pull(skb, IPOIB_ENCAP_LEN);
290 
291 	++dev->stats.rx_packets;
292 	dev->stats.rx_bytes += skb->len;
293 
294 	skb->dev = dev;
295 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
296 		skb->ip_summed = CHECKSUM_UNNECESSARY;
297 
298 	if (dev->features & NETIF_F_LRO)
299 		lro_receive_skb(&priv->lro.lro_mgr, skb, NULL);
300 	else
301 		netif_receive_skb(skb);
302 
303 repost:
304 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
305 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
306 			   "for buf %d\n", wr_id);
307 }
308 
309 static int ipoib_dma_map_tx(struct ib_device *ca,
310 			    struct ipoib_tx_buf *tx_req)
311 {
312 	struct sk_buff *skb = tx_req->skb;
313 	u64 *mapping = tx_req->mapping;
314 	int i;
315 	int off;
316 
317 	if (skb_headlen(skb)) {
318 		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
319 					       DMA_TO_DEVICE);
320 		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
321 			return -EIO;
322 
323 		off = 1;
324 	} else
325 		off = 0;
326 
327 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
328 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
329 		mapping[i + off] = ib_dma_map_page(ca, frag->page,
330 						 frag->page_offset, frag->size,
331 						 DMA_TO_DEVICE);
332 		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
333 			goto partial_error;
334 	}
335 	return 0;
336 
337 partial_error:
338 	for (; i > 0; --i) {
339 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
340 		ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
341 	}
342 
343 	if (off)
344 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
345 
346 	return -EIO;
347 }
348 
349 static void ipoib_dma_unmap_tx(struct ib_device *ca,
350 			       struct ipoib_tx_buf *tx_req)
351 {
352 	struct sk_buff *skb = tx_req->skb;
353 	u64 *mapping = tx_req->mapping;
354 	int i;
355 	int off;
356 
357 	if (skb_headlen(skb)) {
358 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
359 		off = 1;
360 	} else
361 		off = 0;
362 
363 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
364 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
365 		ib_dma_unmap_page(ca, mapping[i + off], frag->size,
366 				  DMA_TO_DEVICE);
367 	}
368 }
369 
370 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
371 {
372 	struct ipoib_dev_priv *priv = netdev_priv(dev);
373 	unsigned int wr_id = wc->wr_id;
374 	struct ipoib_tx_buf *tx_req;
375 
376 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
377 		       wr_id, wc->status);
378 
379 	if (unlikely(wr_id >= ipoib_sendq_size)) {
380 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
381 			   wr_id, ipoib_sendq_size);
382 		return;
383 	}
384 
385 	tx_req = &priv->tx_ring[wr_id];
386 
387 	ipoib_dma_unmap_tx(priv->ca, tx_req);
388 
389 	++dev->stats.tx_packets;
390 	dev->stats.tx_bytes += tx_req->skb->len;
391 
392 	dev_kfree_skb_any(tx_req->skb);
393 
394 	++priv->tx_tail;
395 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
396 	    netif_queue_stopped(dev) &&
397 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
398 		netif_wake_queue(dev);
399 
400 	if (wc->status != IB_WC_SUCCESS &&
401 	    wc->status != IB_WC_WR_FLUSH_ERR)
402 		ipoib_warn(priv, "failed send event "
403 			   "(status=%d, wrid=%d vend_err %x)\n",
404 			   wc->status, wr_id, wc->vendor_err);
405 }
406 
407 static int poll_tx(struct ipoib_dev_priv *priv)
408 {
409 	int n, i;
410 
411 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
412 	for (i = 0; i < n; ++i)
413 		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
414 
415 	return n == MAX_SEND_CQE;
416 }
417 
418 int ipoib_poll(struct napi_struct *napi, int budget)
419 {
420 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
421 	struct net_device *dev = priv->dev;
422 	int done;
423 	int t;
424 	int n, i;
425 
426 	done  = 0;
427 
428 poll_more:
429 	while (done < budget) {
430 		int max = (budget - done);
431 
432 		t = min(IPOIB_NUM_WC, max);
433 		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
434 
435 		for (i = 0; i < n; i++) {
436 			struct ib_wc *wc = priv->ibwc + i;
437 
438 			if (wc->wr_id & IPOIB_OP_RECV) {
439 				++done;
440 				if (wc->wr_id & IPOIB_OP_CM)
441 					ipoib_cm_handle_rx_wc(dev, wc);
442 				else
443 					ipoib_ib_handle_rx_wc(dev, wc);
444 			} else
445 				ipoib_cm_handle_tx_wc(priv->dev, wc);
446 		}
447 
448 		if (n != t)
449 			break;
450 	}
451 
452 	if (done < budget) {
453 		if (dev->features & NETIF_F_LRO)
454 			lro_flush_all(&priv->lro.lro_mgr);
455 
456 		napi_complete(napi);
457 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
458 					      IB_CQ_NEXT_COMP |
459 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
460 		    napi_reschedule(napi))
461 			goto poll_more;
462 	}
463 
464 	return done;
465 }
466 
467 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
468 {
469 	struct net_device *dev = dev_ptr;
470 	struct ipoib_dev_priv *priv = netdev_priv(dev);
471 
472 	napi_schedule(&priv->napi);
473 }
474 
475 static void drain_tx_cq(struct net_device *dev)
476 {
477 	struct ipoib_dev_priv *priv = netdev_priv(dev);
478 
479 	netif_tx_lock(dev);
480 	while (poll_tx(priv))
481 		; /* nothing */
482 
483 	if (netif_queue_stopped(dev))
484 		mod_timer(&priv->poll_timer, jiffies + 1);
485 
486 	netif_tx_unlock(dev);
487 }
488 
489 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
490 {
491 	struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
492 
493 	mod_timer(&priv->poll_timer, jiffies);
494 }
495 
496 static inline int post_send(struct ipoib_dev_priv *priv,
497 			    unsigned int wr_id,
498 			    struct ib_ah *address, u32 qpn,
499 			    struct ipoib_tx_buf *tx_req,
500 			    void *head, int hlen)
501 {
502 	struct ib_send_wr *bad_wr;
503 	int i, off;
504 	struct sk_buff *skb = tx_req->skb;
505 	skb_frag_t *frags = skb_shinfo(skb)->frags;
506 	int nr_frags = skb_shinfo(skb)->nr_frags;
507 	u64 *mapping = tx_req->mapping;
508 
509 	if (skb_headlen(skb)) {
510 		priv->tx_sge[0].addr         = mapping[0];
511 		priv->tx_sge[0].length       = skb_headlen(skb);
512 		off = 1;
513 	} else
514 		off = 0;
515 
516 	for (i = 0; i < nr_frags; ++i) {
517 		priv->tx_sge[i + off].addr = mapping[i + off];
518 		priv->tx_sge[i + off].length = frags[i].size;
519 	}
520 	priv->tx_wr.num_sge	     = nr_frags + off;
521 	priv->tx_wr.wr_id 	     = wr_id;
522 	priv->tx_wr.wr.ud.remote_qpn = qpn;
523 	priv->tx_wr.wr.ud.ah 	     = address;
524 
525 	if (head) {
526 		priv->tx_wr.wr.ud.mss	 = skb_shinfo(skb)->gso_size;
527 		priv->tx_wr.wr.ud.header = head;
528 		priv->tx_wr.wr.ud.hlen	 = hlen;
529 		priv->tx_wr.opcode	 = IB_WR_LSO;
530 	} else
531 		priv->tx_wr.opcode	 = IB_WR_SEND;
532 
533 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
534 }
535 
536 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
537 		struct ipoib_ah *address, u32 qpn)
538 {
539 	struct ipoib_dev_priv *priv = netdev_priv(dev);
540 	struct ipoib_tx_buf *tx_req;
541 	int hlen, rc;
542 	void *phead;
543 
544 	if (skb_is_gso(skb)) {
545 		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
546 		phead = skb->data;
547 		if (unlikely(!skb_pull(skb, hlen))) {
548 			ipoib_warn(priv, "linear data too small\n");
549 			++dev->stats.tx_dropped;
550 			++dev->stats.tx_errors;
551 			dev_kfree_skb_any(skb);
552 			return;
553 		}
554 	} else {
555 		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
556 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
557 				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
558 			++dev->stats.tx_dropped;
559 			++dev->stats.tx_errors;
560 			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
561 			return;
562 		}
563 		phead = NULL;
564 		hlen  = 0;
565 	}
566 
567 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
568 		       skb->len, address, qpn);
569 
570 	/*
571 	 * We put the skb into the tx_ring _before_ we call post_send()
572 	 * because it's entirely possible that the completion handler will
573 	 * run before we execute anything after the post_send().  That
574 	 * means we have to make sure everything is properly recorded and
575 	 * our state is consistent before we call post_send().
576 	 */
577 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
578 	tx_req->skb = skb;
579 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
580 		++dev->stats.tx_errors;
581 		dev_kfree_skb_any(skb);
582 		return;
583 	}
584 
585 	if (skb->ip_summed == CHECKSUM_PARTIAL)
586 		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
587 	else
588 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
589 
590 	if (++priv->tx_outstanding == ipoib_sendq_size) {
591 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
592 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
593 			ipoib_warn(priv, "request notify on send CQ failed\n");
594 		netif_stop_queue(dev);
595 	}
596 
597 	rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
598 		       address->ah, qpn, tx_req, phead, hlen);
599 	if (unlikely(rc)) {
600 		ipoib_warn(priv, "post_send failed, error %d\n", rc);
601 		++dev->stats.tx_errors;
602 		--priv->tx_outstanding;
603 		ipoib_dma_unmap_tx(priv->ca, tx_req);
604 		dev_kfree_skb_any(skb);
605 		if (netif_queue_stopped(dev))
606 			netif_wake_queue(dev);
607 	} else {
608 		dev->trans_start = jiffies;
609 
610 		address->last_send = priv->tx_head;
611 		++priv->tx_head;
612 		skb_orphan(skb);
613 
614 	}
615 
616 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
617 		while (poll_tx(priv))
618 			; /* nothing */
619 }
620 
621 static void __ipoib_reap_ah(struct net_device *dev)
622 {
623 	struct ipoib_dev_priv *priv = netdev_priv(dev);
624 	struct ipoib_ah *ah, *tah;
625 	LIST_HEAD(remove_list);
626 	unsigned long flags;
627 
628 	netif_tx_lock_bh(dev);
629 	spin_lock_irqsave(&priv->lock, flags);
630 
631 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
632 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
633 			list_del(&ah->list);
634 			ib_destroy_ah(ah->ah);
635 			kfree(ah);
636 		}
637 
638 	spin_unlock_irqrestore(&priv->lock, flags);
639 	netif_tx_unlock_bh(dev);
640 }
641 
642 void ipoib_reap_ah(struct work_struct *work)
643 {
644 	struct ipoib_dev_priv *priv =
645 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
646 	struct net_device *dev = priv->dev;
647 
648 	__ipoib_reap_ah(dev);
649 
650 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
651 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
652 				   round_jiffies_relative(HZ));
653 }
654 
655 static void ipoib_ib_tx_timer_func(unsigned long ctx)
656 {
657 	drain_tx_cq((struct net_device *)ctx);
658 }
659 
660 int ipoib_ib_dev_open(struct net_device *dev)
661 {
662 	struct ipoib_dev_priv *priv = netdev_priv(dev);
663 	int ret;
664 
665 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
666 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
667 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
668 		return -1;
669 	}
670 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
671 
672 	ret = ipoib_init_qp(dev);
673 	if (ret) {
674 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
675 		return -1;
676 	}
677 
678 	ret = ipoib_ib_post_receives(dev);
679 	if (ret) {
680 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
681 		ipoib_ib_dev_stop(dev, 1);
682 		return -1;
683 	}
684 
685 	ret = ipoib_cm_dev_open(dev);
686 	if (ret) {
687 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
688 		ipoib_ib_dev_stop(dev, 1);
689 		return -1;
690 	}
691 
692 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
693 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
694 			   round_jiffies_relative(HZ));
695 
696 	if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
697 		napi_enable(&priv->napi);
698 
699 	return 0;
700 }
701 
702 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
703 {
704 	struct ipoib_dev_priv *priv = netdev_priv(dev);
705 	u16 pkey_index = 0;
706 
707 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
708 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
709 	else
710 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
711 }
712 
713 int ipoib_ib_dev_up(struct net_device *dev)
714 {
715 	struct ipoib_dev_priv *priv = netdev_priv(dev);
716 
717 	ipoib_pkey_dev_check_presence(dev);
718 
719 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
720 		ipoib_dbg(priv, "PKEY is not assigned.\n");
721 		return 0;
722 	}
723 
724 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
725 
726 	return ipoib_mcast_start_thread(dev);
727 }
728 
729 int ipoib_ib_dev_down(struct net_device *dev, int flush)
730 {
731 	struct ipoib_dev_priv *priv = netdev_priv(dev);
732 
733 	ipoib_dbg(priv, "downing ib_dev\n");
734 
735 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
736 	netif_carrier_off(dev);
737 
738 	/* Shutdown the P_Key thread if still active */
739 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
740 		mutex_lock(&pkey_mutex);
741 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
742 		cancel_delayed_work(&priv->pkey_poll_task);
743 		mutex_unlock(&pkey_mutex);
744 		if (flush)
745 			flush_workqueue(ipoib_workqueue);
746 	}
747 
748 	ipoib_mcast_stop_thread(dev, flush);
749 	ipoib_mcast_dev_flush(dev);
750 
751 	ipoib_flush_paths(dev);
752 
753 	return 0;
754 }
755 
756 static int recvs_pending(struct net_device *dev)
757 {
758 	struct ipoib_dev_priv *priv = netdev_priv(dev);
759 	int pending = 0;
760 	int i;
761 
762 	for (i = 0; i < ipoib_recvq_size; ++i)
763 		if (priv->rx_ring[i].skb)
764 			++pending;
765 
766 	return pending;
767 }
768 
769 void ipoib_drain_cq(struct net_device *dev)
770 {
771 	struct ipoib_dev_priv *priv = netdev_priv(dev);
772 	int i, n;
773 
774 	/*
775 	 * We call completion handling routines that expect to be
776 	 * called from the BH-disabled NAPI poll context, so disable
777 	 * BHs here too.
778 	 */
779 	local_bh_disable();
780 
781 	do {
782 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
783 		for (i = 0; i < n; ++i) {
784 			/*
785 			 * Convert any successful completions to flush
786 			 * errors to avoid passing packets up the
787 			 * stack after bringing the device down.
788 			 */
789 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
790 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
791 
792 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
793 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
794 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
795 				else
796 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
797 			} else
798 				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
799 		}
800 	} while (n == IPOIB_NUM_WC);
801 
802 	while (poll_tx(priv))
803 		; /* nothing */
804 
805 	local_bh_enable();
806 }
807 
808 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
809 {
810 	struct ipoib_dev_priv *priv = netdev_priv(dev);
811 	struct ib_qp_attr qp_attr;
812 	unsigned long begin;
813 	struct ipoib_tx_buf *tx_req;
814 	int i;
815 
816 	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
817 		napi_disable(&priv->napi);
818 
819 	ipoib_cm_dev_stop(dev);
820 
821 	/*
822 	 * Move our QP to the error state and then reinitialize in
823 	 * when all work requests have completed or have been flushed.
824 	 */
825 	qp_attr.qp_state = IB_QPS_ERR;
826 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
827 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
828 
829 	/* Wait for all sends and receives to complete */
830 	begin = jiffies;
831 
832 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
833 		if (time_after(jiffies, begin + 5 * HZ)) {
834 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
835 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
836 
837 			/*
838 			 * assume the HW is wedged and just free up
839 			 * all our pending work requests.
840 			 */
841 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
842 				tx_req = &priv->tx_ring[priv->tx_tail &
843 							(ipoib_sendq_size - 1)];
844 				ipoib_dma_unmap_tx(priv->ca, tx_req);
845 				dev_kfree_skb_any(tx_req->skb);
846 				++priv->tx_tail;
847 				--priv->tx_outstanding;
848 			}
849 
850 			for (i = 0; i < ipoib_recvq_size; ++i) {
851 				struct ipoib_rx_buf *rx_req;
852 
853 				rx_req = &priv->rx_ring[i];
854 				if (!rx_req->skb)
855 					continue;
856 				ipoib_ud_dma_unmap_rx(priv,
857 						      priv->rx_ring[i].mapping);
858 				dev_kfree_skb_any(rx_req->skb);
859 				rx_req->skb = NULL;
860 			}
861 
862 			goto timeout;
863 		}
864 
865 		ipoib_drain_cq(dev);
866 
867 		msleep(1);
868 	}
869 
870 	ipoib_dbg(priv, "All sends and receives done.\n");
871 
872 timeout:
873 	del_timer_sync(&priv->poll_timer);
874 	qp_attr.qp_state = IB_QPS_RESET;
875 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
876 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
877 
878 	/* Wait for all AHs to be reaped */
879 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
880 	cancel_delayed_work(&priv->ah_reap_task);
881 	if (flush)
882 		flush_workqueue(ipoib_workqueue);
883 
884 	begin = jiffies;
885 
886 	while (!list_empty(&priv->dead_ahs)) {
887 		__ipoib_reap_ah(dev);
888 
889 		if (time_after(jiffies, begin + HZ)) {
890 			ipoib_warn(priv, "timing out; will leak address handles\n");
891 			break;
892 		}
893 
894 		msleep(1);
895 	}
896 
897 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
898 
899 	return 0;
900 }
901 
902 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
903 {
904 	struct ipoib_dev_priv *priv = netdev_priv(dev);
905 
906 	priv->ca = ca;
907 	priv->port = port;
908 	priv->qp = NULL;
909 
910 	if (ipoib_transport_dev_init(dev, ca)) {
911 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
912 		return -ENODEV;
913 	}
914 
915 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
916 		    (unsigned long) dev);
917 
918 	if (dev->flags & IFF_UP) {
919 		if (ipoib_ib_dev_open(dev)) {
920 			ipoib_transport_dev_cleanup(dev);
921 			return -ENODEV;
922 		}
923 	}
924 
925 	return 0;
926 }
927 
928 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
929 				enum ipoib_flush_level level)
930 {
931 	struct ipoib_dev_priv *cpriv;
932 	struct net_device *dev = priv->dev;
933 	u16 new_index;
934 
935 	mutex_lock(&priv->vlan_mutex);
936 
937 	/*
938 	 * Flush any child interfaces too -- they might be up even if
939 	 * the parent is down.
940 	 */
941 	list_for_each_entry(cpriv, &priv->child_intfs, list)
942 		__ipoib_ib_dev_flush(cpriv, level);
943 
944 	mutex_unlock(&priv->vlan_mutex);
945 
946 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
947 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
948 		return;
949 	}
950 
951 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
952 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
953 		return;
954 	}
955 
956 	if (level == IPOIB_FLUSH_HEAVY) {
957 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
958 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
959 			ipoib_ib_dev_down(dev, 0);
960 			ipoib_ib_dev_stop(dev, 0);
961 			if (ipoib_pkey_dev_delay_open(dev))
962 				return;
963 		}
964 
965 		/* restart QP only if P_Key index is changed */
966 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
967 		    new_index == priv->pkey_index) {
968 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
969 			return;
970 		}
971 		priv->pkey_index = new_index;
972 	}
973 
974 	if (level == IPOIB_FLUSH_LIGHT) {
975 		ipoib_mark_paths_invalid(dev);
976 		ipoib_mcast_dev_flush(dev);
977 	}
978 
979 	if (level >= IPOIB_FLUSH_NORMAL)
980 		ipoib_ib_dev_down(dev, 0);
981 
982 	if (level == IPOIB_FLUSH_HEAVY) {
983 		ipoib_ib_dev_stop(dev, 0);
984 		ipoib_ib_dev_open(dev);
985 	}
986 
987 	/*
988 	 * The device could have been brought down between the start and when
989 	 * we get here, don't bring it back up if it's not configured up
990 	 */
991 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
992 		if (level >= IPOIB_FLUSH_NORMAL)
993 			ipoib_ib_dev_up(dev);
994 		ipoib_mcast_restart_task(&priv->restart_task);
995 	}
996 }
997 
998 void ipoib_ib_dev_flush_light(struct work_struct *work)
999 {
1000 	struct ipoib_dev_priv *priv =
1001 		container_of(work, struct ipoib_dev_priv, flush_light);
1002 
1003 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1004 }
1005 
1006 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1007 {
1008 	struct ipoib_dev_priv *priv =
1009 		container_of(work, struct ipoib_dev_priv, flush_normal);
1010 
1011 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1012 }
1013 
1014 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1015 {
1016 	struct ipoib_dev_priv *priv =
1017 		container_of(work, struct ipoib_dev_priv, flush_heavy);
1018 
1019 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1020 }
1021 
1022 void ipoib_ib_dev_cleanup(struct net_device *dev)
1023 {
1024 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1025 
1026 	ipoib_dbg(priv, "cleaning up ib_dev\n");
1027 
1028 	ipoib_mcast_stop_thread(dev, 1);
1029 	ipoib_mcast_dev_flush(dev);
1030 
1031 	ipoib_transport_dev_cleanup(dev);
1032 }
1033 
1034 /*
1035  * Delayed P_Key Assigment Interim Support
1036  *
1037  * The following is initial implementation of delayed P_Key assigment
1038  * mechanism. It is using the same approach implemented for the multicast
1039  * group join. The single goal of this implementation is to quickly address
1040  * Bug #2507. This implementation will probably be removed when the P_Key
1041  * change async notification is available.
1042  */
1043 
1044 void ipoib_pkey_poll(struct work_struct *work)
1045 {
1046 	struct ipoib_dev_priv *priv =
1047 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1048 	struct net_device *dev = priv->dev;
1049 
1050 	ipoib_pkey_dev_check_presence(dev);
1051 
1052 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1053 		ipoib_open(dev);
1054 	else {
1055 		mutex_lock(&pkey_mutex);
1056 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1057 			queue_delayed_work(ipoib_workqueue,
1058 					   &priv->pkey_poll_task,
1059 					   HZ);
1060 		mutex_unlock(&pkey_mutex);
1061 	}
1062 }
1063 
1064 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1065 {
1066 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1067 
1068 	/* Look for the interface pkey value in the IB Port P_Key table and */
1069 	/* set the interface pkey assigment flag                            */
1070 	ipoib_pkey_dev_check_presence(dev);
1071 
1072 	/* P_Key value not assigned yet - start polling */
1073 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1074 		mutex_lock(&pkey_mutex);
1075 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1076 		queue_delayed_work(ipoib_workqueue,
1077 				   &priv->pkey_poll_task,
1078 				   HZ);
1079 		mutex_unlock(&pkey_mutex);
1080 		return 1;
1081 	}
1082 
1083 	return 0;
1084 }
1085