xref: /openbmc/linux/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision f15cbe6f1a4b4d9df59142fc8e4abb973302cf44)
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 <rdma/ib_cache.h>
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 
227 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
228 		       wr_id, wc->status);
229 
230 	if (unlikely(wr_id >= ipoib_recvq_size)) {
231 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
232 			   wr_id, ipoib_recvq_size);
233 		return;
234 	}
235 
236 	skb  = priv->rx_ring[wr_id].skb;
237 
238 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
239 		if (wc->status != IB_WC_WR_FLUSH_ERR)
240 			ipoib_warn(priv, "failed recv event "
241 				   "(status=%d, wrid=%d vend_err %x)\n",
242 				   wc->status, wr_id, wc->vendor_err);
243 		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
244 		dev_kfree_skb_any(skb);
245 		priv->rx_ring[wr_id].skb = NULL;
246 		return;
247 	}
248 
249 	/*
250 	 * Drop packets that this interface sent, ie multicast packets
251 	 * that the HCA has replicated.
252 	 */
253 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
254 		goto repost;
255 
256 	memcpy(mapping, priv->rx_ring[wr_id].mapping,
257 	       IPOIB_UD_RX_SG * sizeof *mapping);
258 
259 	/*
260 	 * If we can't allocate a new RX buffer, dump
261 	 * this packet and reuse the old buffer.
262 	 */
263 	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
264 		++dev->stats.rx_dropped;
265 		goto repost;
266 	}
267 
268 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
269 		       wc->byte_len, wc->slid);
270 
271 	ipoib_ud_dma_unmap_rx(priv, mapping);
272 	ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
273 
274 	skb_pull(skb, IB_GRH_BYTES);
275 
276 	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
277 	skb_reset_mac_header(skb);
278 	skb_pull(skb, IPOIB_ENCAP_LEN);
279 
280 	dev->last_rx = jiffies;
281 	++dev->stats.rx_packets;
282 	dev->stats.rx_bytes += skb->len;
283 
284 	skb->dev = dev;
285 	/* XXX get correct PACKET_ type here */
286 	skb->pkt_type = PACKET_HOST;
287 
288 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
289 		skb->ip_summed = CHECKSUM_UNNECESSARY;
290 
291 	if (dev->features & NETIF_F_LRO)
292 		lro_receive_skb(&priv->lro.lro_mgr, skb, NULL);
293 	else
294 		netif_receive_skb(skb);
295 
296 repost:
297 	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
298 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
299 			   "for buf %d\n", wr_id);
300 }
301 
302 static int ipoib_dma_map_tx(struct ib_device *ca,
303 			    struct ipoib_tx_buf *tx_req)
304 {
305 	struct sk_buff *skb = tx_req->skb;
306 	u64 *mapping = tx_req->mapping;
307 	int i;
308 	int off;
309 
310 	if (skb_headlen(skb)) {
311 		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
312 					       DMA_TO_DEVICE);
313 		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
314 			return -EIO;
315 
316 		off = 1;
317 	} else
318 		off = 0;
319 
320 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
321 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
322 		mapping[i + off] = ib_dma_map_page(ca, frag->page,
323 						 frag->page_offset, frag->size,
324 						 DMA_TO_DEVICE);
325 		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
326 			goto partial_error;
327 	}
328 	return 0;
329 
330 partial_error:
331 	for (; i > 0; --i) {
332 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
333 		ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
334 	}
335 
336 	if (off)
337 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
338 
339 	return -EIO;
340 }
341 
342 static void ipoib_dma_unmap_tx(struct ib_device *ca,
343 			       struct ipoib_tx_buf *tx_req)
344 {
345 	struct sk_buff *skb = tx_req->skb;
346 	u64 *mapping = tx_req->mapping;
347 	int i;
348 	int off;
349 
350 	if (skb_headlen(skb)) {
351 		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
352 		off = 1;
353 	} else
354 		off = 0;
355 
356 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
357 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
358 		ib_dma_unmap_page(ca, mapping[i + off], frag->size,
359 				  DMA_TO_DEVICE);
360 	}
361 }
362 
363 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
364 {
365 	struct ipoib_dev_priv *priv = netdev_priv(dev);
366 	unsigned int wr_id = wc->wr_id;
367 	struct ipoib_tx_buf *tx_req;
368 
369 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
370 		       wr_id, wc->status);
371 
372 	if (unlikely(wr_id >= ipoib_sendq_size)) {
373 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
374 			   wr_id, ipoib_sendq_size);
375 		return;
376 	}
377 
378 	tx_req = &priv->tx_ring[wr_id];
379 
380 	ipoib_dma_unmap_tx(priv->ca, tx_req);
381 
382 	++dev->stats.tx_packets;
383 	dev->stats.tx_bytes += tx_req->skb->len;
384 
385 	dev_kfree_skb_any(tx_req->skb);
386 
387 	++priv->tx_tail;
388 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
389 	    netif_queue_stopped(dev) &&
390 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
391 		netif_wake_queue(dev);
392 
393 	if (wc->status != IB_WC_SUCCESS &&
394 	    wc->status != IB_WC_WR_FLUSH_ERR)
395 		ipoib_warn(priv, "failed send event "
396 			   "(status=%d, wrid=%d vend_err %x)\n",
397 			   wc->status, wr_id, wc->vendor_err);
398 }
399 
400 static int poll_tx(struct ipoib_dev_priv *priv)
401 {
402 	int n, i;
403 
404 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
405 	for (i = 0; i < n; ++i)
406 		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
407 
408 	return n == MAX_SEND_CQE;
409 }
410 
411 int ipoib_poll(struct napi_struct *napi, int budget)
412 {
413 	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
414 	struct net_device *dev = priv->dev;
415 	int done;
416 	int t;
417 	int n, i;
418 
419 	done  = 0;
420 
421 poll_more:
422 	while (done < budget) {
423 		int max = (budget - done);
424 
425 		t = min(IPOIB_NUM_WC, max);
426 		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
427 
428 		for (i = 0; i < n; i++) {
429 			struct ib_wc *wc = priv->ibwc + i;
430 
431 			if (wc->wr_id & IPOIB_OP_RECV) {
432 				++done;
433 				if (wc->wr_id & IPOIB_OP_CM)
434 					ipoib_cm_handle_rx_wc(dev, wc);
435 				else
436 					ipoib_ib_handle_rx_wc(dev, wc);
437 			} else
438 				ipoib_cm_handle_tx_wc(priv->dev, wc);
439 		}
440 
441 		if (n != t)
442 			break;
443 	}
444 
445 	if (done < budget) {
446 		if (dev->features & NETIF_F_LRO)
447 			lro_flush_all(&priv->lro.lro_mgr);
448 
449 		netif_rx_complete(dev, napi);
450 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
451 					      IB_CQ_NEXT_COMP |
452 					      IB_CQ_REPORT_MISSED_EVENTS)) &&
453 		    netif_rx_reschedule(dev, napi))
454 			goto poll_more;
455 	}
456 
457 	return done;
458 }
459 
460 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
461 {
462 	struct net_device *dev = dev_ptr;
463 	struct ipoib_dev_priv *priv = netdev_priv(dev);
464 
465 	netif_rx_schedule(dev, &priv->napi);
466 }
467 
468 static void drain_tx_cq(struct net_device *dev)
469 {
470 	struct ipoib_dev_priv *priv = netdev_priv(dev);
471 	unsigned long flags;
472 
473 	spin_lock_irqsave(&priv->tx_lock, flags);
474 	while (poll_tx(priv))
475 		; /* nothing */
476 
477 	if (netif_queue_stopped(dev))
478 		mod_timer(&priv->poll_timer, jiffies + 1);
479 
480 	spin_unlock_irqrestore(&priv->tx_lock, flags);
481 }
482 
483 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
484 {
485 	drain_tx_cq((struct net_device *)dev_ptr);
486 }
487 
488 static inline int post_send(struct ipoib_dev_priv *priv,
489 			    unsigned int wr_id,
490 			    struct ib_ah *address, u32 qpn,
491 			    struct ipoib_tx_buf *tx_req,
492 			    void *head, int hlen)
493 {
494 	struct ib_send_wr *bad_wr;
495 	int i, off;
496 	struct sk_buff *skb = tx_req->skb;
497 	skb_frag_t *frags = skb_shinfo(skb)->frags;
498 	int nr_frags = skb_shinfo(skb)->nr_frags;
499 	u64 *mapping = tx_req->mapping;
500 
501 	if (skb_headlen(skb)) {
502 		priv->tx_sge[0].addr         = mapping[0];
503 		priv->tx_sge[0].length       = skb_headlen(skb);
504 		off = 1;
505 	} else
506 		off = 0;
507 
508 	for (i = 0; i < nr_frags; ++i) {
509 		priv->tx_sge[i + off].addr = mapping[i + off];
510 		priv->tx_sge[i + off].length = frags[i].size;
511 	}
512 	priv->tx_wr.num_sge	     = nr_frags + off;
513 	priv->tx_wr.wr_id 	     = wr_id;
514 	priv->tx_wr.wr.ud.remote_qpn = qpn;
515 	priv->tx_wr.wr.ud.ah 	     = address;
516 
517 	if (head) {
518 		priv->tx_wr.wr.ud.mss	 = skb_shinfo(skb)->gso_size;
519 		priv->tx_wr.wr.ud.header = head;
520 		priv->tx_wr.wr.ud.hlen	 = hlen;
521 		priv->tx_wr.opcode	 = IB_WR_LSO;
522 	} else
523 		priv->tx_wr.opcode	 = IB_WR_SEND;
524 
525 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
526 }
527 
528 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
529 		struct ipoib_ah *address, u32 qpn)
530 {
531 	struct ipoib_dev_priv *priv = netdev_priv(dev);
532 	struct ipoib_tx_buf *tx_req;
533 	int hlen;
534 	void *phead;
535 
536 	if (skb_is_gso(skb)) {
537 		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
538 		phead = skb->data;
539 		if (unlikely(!skb_pull(skb, hlen))) {
540 			ipoib_warn(priv, "linear data too small\n");
541 			++dev->stats.tx_dropped;
542 			++dev->stats.tx_errors;
543 			dev_kfree_skb_any(skb);
544 			return;
545 		}
546 	} else {
547 		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
548 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
549 				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
550 			++dev->stats.tx_dropped;
551 			++dev->stats.tx_errors;
552 			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
553 			return;
554 		}
555 		phead = NULL;
556 		hlen  = 0;
557 	}
558 
559 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
560 		       skb->len, address, qpn);
561 
562 	/*
563 	 * We put the skb into the tx_ring _before_ we call post_send()
564 	 * because it's entirely possible that the completion handler will
565 	 * run before we execute anything after the post_send().  That
566 	 * means we have to make sure everything is properly recorded and
567 	 * our state is consistent before we call post_send().
568 	 */
569 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
570 	tx_req->skb = skb;
571 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
572 		++dev->stats.tx_errors;
573 		dev_kfree_skb_any(skb);
574 		return;
575 	}
576 
577 	if (skb->ip_summed == CHECKSUM_PARTIAL)
578 		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
579 	else
580 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
581 
582 	if (++priv->tx_outstanding == ipoib_sendq_size) {
583 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
584 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
585 			ipoib_warn(priv, "request notify on send CQ failed\n");
586 		netif_stop_queue(dev);
587 	}
588 
589 	if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
590 			       address->ah, qpn, tx_req, phead, hlen))) {
591 		ipoib_warn(priv, "post_send failed\n");
592 		++dev->stats.tx_errors;
593 		--priv->tx_outstanding;
594 		ipoib_dma_unmap_tx(priv->ca, tx_req);
595 		dev_kfree_skb_any(skb);
596 		if (netif_queue_stopped(dev))
597 			netif_wake_queue(dev);
598 	} else {
599 		dev->trans_start = jiffies;
600 
601 		address->last_send = priv->tx_head;
602 		++priv->tx_head;
603 		skb_orphan(skb);
604 
605 	}
606 
607 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
608 		while (poll_tx(priv))
609 			; /* nothing */
610 }
611 
612 static void __ipoib_reap_ah(struct net_device *dev)
613 {
614 	struct ipoib_dev_priv *priv = netdev_priv(dev);
615 	struct ipoib_ah *ah, *tah;
616 	LIST_HEAD(remove_list);
617 
618 	spin_lock_irq(&priv->tx_lock);
619 	spin_lock(&priv->lock);
620 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
621 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
622 			list_del(&ah->list);
623 			ib_destroy_ah(ah->ah);
624 			kfree(ah);
625 		}
626 	spin_unlock(&priv->lock);
627 	spin_unlock_irq(&priv->tx_lock);
628 }
629 
630 void ipoib_reap_ah(struct work_struct *work)
631 {
632 	struct ipoib_dev_priv *priv =
633 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
634 	struct net_device *dev = priv->dev;
635 
636 	__ipoib_reap_ah(dev);
637 
638 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
639 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
640 				   round_jiffies_relative(HZ));
641 }
642 
643 static void ipoib_ib_tx_timer_func(unsigned long ctx)
644 {
645 	drain_tx_cq((struct net_device *)ctx);
646 }
647 
648 int ipoib_ib_dev_open(struct net_device *dev)
649 {
650 	struct ipoib_dev_priv *priv = netdev_priv(dev);
651 	int ret;
652 
653 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
654 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
655 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
656 		return -1;
657 	}
658 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
659 
660 	ret = ipoib_init_qp(dev);
661 	if (ret) {
662 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
663 		return -1;
664 	}
665 
666 	ret = ipoib_ib_post_receives(dev);
667 	if (ret) {
668 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
669 		ipoib_ib_dev_stop(dev, 1);
670 		return -1;
671 	}
672 
673 	ret = ipoib_cm_dev_open(dev);
674 	if (ret) {
675 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
676 		ipoib_ib_dev_stop(dev, 1);
677 		return -1;
678 	}
679 
680 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
681 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
682 			   round_jiffies_relative(HZ));
683 
684 	init_timer(&priv->poll_timer);
685 	priv->poll_timer.function = ipoib_ib_tx_timer_func;
686 	priv->poll_timer.data = (unsigned long)dev;
687 
688 	set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
689 
690 	return 0;
691 }
692 
693 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
694 {
695 	struct ipoib_dev_priv *priv = netdev_priv(dev);
696 	u16 pkey_index = 0;
697 
698 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
699 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
700 	else
701 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
702 }
703 
704 int ipoib_ib_dev_up(struct net_device *dev)
705 {
706 	struct ipoib_dev_priv *priv = netdev_priv(dev);
707 
708 	ipoib_pkey_dev_check_presence(dev);
709 
710 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
711 		ipoib_dbg(priv, "PKEY is not assigned.\n");
712 		return 0;
713 	}
714 
715 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
716 
717 	return ipoib_mcast_start_thread(dev);
718 }
719 
720 int ipoib_ib_dev_down(struct net_device *dev, int flush)
721 {
722 	struct ipoib_dev_priv *priv = netdev_priv(dev);
723 
724 	ipoib_dbg(priv, "downing ib_dev\n");
725 
726 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
727 	netif_carrier_off(dev);
728 
729 	/* Shutdown the P_Key thread if still active */
730 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
731 		mutex_lock(&pkey_mutex);
732 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
733 		cancel_delayed_work(&priv->pkey_poll_task);
734 		mutex_unlock(&pkey_mutex);
735 		if (flush)
736 			flush_workqueue(ipoib_workqueue);
737 	}
738 
739 	ipoib_mcast_stop_thread(dev, flush);
740 	ipoib_mcast_dev_flush(dev);
741 
742 	ipoib_flush_paths(dev);
743 
744 	return 0;
745 }
746 
747 static int recvs_pending(struct net_device *dev)
748 {
749 	struct ipoib_dev_priv *priv = netdev_priv(dev);
750 	int pending = 0;
751 	int i;
752 
753 	for (i = 0; i < ipoib_recvq_size; ++i)
754 		if (priv->rx_ring[i].skb)
755 			++pending;
756 
757 	return pending;
758 }
759 
760 void ipoib_drain_cq(struct net_device *dev)
761 {
762 	struct ipoib_dev_priv *priv = netdev_priv(dev);
763 	int i, n;
764 	do {
765 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
766 		for (i = 0; i < n; ++i) {
767 			/*
768 			 * Convert any successful completions to flush
769 			 * errors to avoid passing packets up the
770 			 * stack after bringing the device down.
771 			 */
772 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
773 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
774 
775 			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
776 				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
777 					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
778 				else
779 					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
780 			} else
781 				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
782 		}
783 	} while (n == IPOIB_NUM_WC);
784 
785 	while (poll_tx(priv))
786 		; /* nothing */
787 }
788 
789 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
790 {
791 	struct ipoib_dev_priv *priv = netdev_priv(dev);
792 	struct ib_qp_attr qp_attr;
793 	unsigned long begin;
794 	struct ipoib_tx_buf *tx_req;
795 	int i;
796 
797 	clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
798 
799 	ipoib_cm_dev_stop(dev);
800 
801 	/*
802 	 * Move our QP to the error state and then reinitialize in
803 	 * when all work requests have completed or have been flushed.
804 	 */
805 	qp_attr.qp_state = IB_QPS_ERR;
806 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
807 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
808 
809 	/* Wait for all sends and receives to complete */
810 	begin = jiffies;
811 
812 	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
813 		if (time_after(jiffies, begin + 5 * HZ)) {
814 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
815 				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
816 
817 			/*
818 			 * assume the HW is wedged and just free up
819 			 * all our pending work requests.
820 			 */
821 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
822 				tx_req = &priv->tx_ring[priv->tx_tail &
823 							(ipoib_sendq_size - 1)];
824 				ipoib_dma_unmap_tx(priv->ca, tx_req);
825 				dev_kfree_skb_any(tx_req->skb);
826 				++priv->tx_tail;
827 				--priv->tx_outstanding;
828 			}
829 
830 			for (i = 0; i < ipoib_recvq_size; ++i) {
831 				struct ipoib_rx_buf *rx_req;
832 
833 				rx_req = &priv->rx_ring[i];
834 				if (!rx_req->skb)
835 					continue;
836 				ipoib_ud_dma_unmap_rx(priv,
837 						      priv->rx_ring[i].mapping);
838 				dev_kfree_skb_any(rx_req->skb);
839 				rx_req->skb = NULL;
840 			}
841 
842 			goto timeout;
843 		}
844 
845 		ipoib_drain_cq(dev);
846 
847 		msleep(1);
848 	}
849 
850 	ipoib_dbg(priv, "All sends and receives done.\n");
851 
852 timeout:
853 	del_timer_sync(&priv->poll_timer);
854 	qp_attr.qp_state = IB_QPS_RESET;
855 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
856 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
857 
858 	/* Wait for all AHs to be reaped */
859 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
860 	cancel_delayed_work(&priv->ah_reap_task);
861 	if (flush)
862 		flush_workqueue(ipoib_workqueue);
863 
864 	begin = jiffies;
865 
866 	while (!list_empty(&priv->dead_ahs)) {
867 		__ipoib_reap_ah(dev);
868 
869 		if (time_after(jiffies, begin + HZ)) {
870 			ipoib_warn(priv, "timing out; will leak address handles\n");
871 			break;
872 		}
873 
874 		msleep(1);
875 	}
876 
877 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
878 
879 	return 0;
880 }
881 
882 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
883 {
884 	struct ipoib_dev_priv *priv = netdev_priv(dev);
885 
886 	priv->ca = ca;
887 	priv->port = port;
888 	priv->qp = NULL;
889 
890 	if (ipoib_transport_dev_init(dev, ca)) {
891 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
892 		return -ENODEV;
893 	}
894 
895 	if (dev->flags & IFF_UP) {
896 		if (ipoib_ib_dev_open(dev)) {
897 			ipoib_transport_dev_cleanup(dev);
898 			return -ENODEV;
899 		}
900 	}
901 
902 	return 0;
903 }
904 
905 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
906 				enum ipoib_flush_level level)
907 {
908 	struct ipoib_dev_priv *cpriv;
909 	struct net_device *dev = priv->dev;
910 	u16 new_index;
911 
912 	mutex_lock(&priv->vlan_mutex);
913 
914 	/*
915 	 * Flush any child interfaces too -- they might be up even if
916 	 * the parent is down.
917 	 */
918 	list_for_each_entry(cpriv, &priv->child_intfs, list)
919 		__ipoib_ib_dev_flush(cpriv, level);
920 
921 	mutex_unlock(&priv->vlan_mutex);
922 
923 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
924 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
925 		return;
926 	}
927 
928 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
929 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
930 		return;
931 	}
932 
933 	if (level == IPOIB_FLUSH_HEAVY) {
934 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
935 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
936 			ipoib_ib_dev_down(dev, 0);
937 			ipoib_ib_dev_stop(dev, 0);
938 			if (ipoib_pkey_dev_delay_open(dev))
939 				return;
940 		}
941 
942 		/* restart QP only if P_Key index is changed */
943 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
944 		    new_index == priv->pkey_index) {
945 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
946 			return;
947 		}
948 		priv->pkey_index = new_index;
949 	}
950 
951 	if (level == IPOIB_FLUSH_LIGHT) {
952 		ipoib_mark_paths_invalid(dev);
953 		ipoib_mcast_dev_flush(dev);
954 	}
955 
956 	if (level >= IPOIB_FLUSH_NORMAL)
957 		ipoib_ib_dev_down(dev, 0);
958 
959 	if (level == IPOIB_FLUSH_HEAVY) {
960 		ipoib_ib_dev_stop(dev, 0);
961 		ipoib_ib_dev_open(dev);
962 	}
963 
964 	/*
965 	 * The device could have been brought down between the start and when
966 	 * we get here, don't bring it back up if it's not configured up
967 	 */
968 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
969 		if (level >= IPOIB_FLUSH_NORMAL)
970 			ipoib_ib_dev_up(dev);
971 		ipoib_mcast_restart_task(&priv->restart_task);
972 	}
973 }
974 
975 void ipoib_ib_dev_flush_light(struct work_struct *work)
976 {
977 	struct ipoib_dev_priv *priv =
978 		container_of(work, struct ipoib_dev_priv, flush_light);
979 
980 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
981 }
982 
983 void ipoib_ib_dev_flush_normal(struct work_struct *work)
984 {
985 	struct ipoib_dev_priv *priv =
986 		container_of(work, struct ipoib_dev_priv, flush_normal);
987 
988 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
989 }
990 
991 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
992 {
993 	struct ipoib_dev_priv *priv =
994 		container_of(work, struct ipoib_dev_priv, flush_heavy);
995 
996 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
997 }
998 
999 void ipoib_ib_dev_cleanup(struct net_device *dev)
1000 {
1001 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1002 
1003 	ipoib_dbg(priv, "cleaning up ib_dev\n");
1004 
1005 	ipoib_mcast_stop_thread(dev, 1);
1006 	ipoib_mcast_dev_flush(dev);
1007 
1008 	ipoib_transport_dev_cleanup(dev);
1009 }
1010 
1011 /*
1012  * Delayed P_Key Assigment Interim Support
1013  *
1014  * The following is initial implementation of delayed P_Key assigment
1015  * mechanism. It is using the same approach implemented for the multicast
1016  * group join. The single goal of this implementation is to quickly address
1017  * Bug #2507. This implementation will probably be removed when the P_Key
1018  * change async notification is available.
1019  */
1020 
1021 void ipoib_pkey_poll(struct work_struct *work)
1022 {
1023 	struct ipoib_dev_priv *priv =
1024 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1025 	struct net_device *dev = priv->dev;
1026 
1027 	ipoib_pkey_dev_check_presence(dev);
1028 
1029 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1030 		ipoib_open(dev);
1031 	else {
1032 		mutex_lock(&pkey_mutex);
1033 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1034 			queue_delayed_work(ipoib_workqueue,
1035 					   &priv->pkey_poll_task,
1036 					   HZ);
1037 		mutex_unlock(&pkey_mutex);
1038 	}
1039 }
1040 
1041 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1042 {
1043 	struct ipoib_dev_priv *priv = netdev_priv(dev);
1044 
1045 	/* Look for the interface pkey value in the IB Port P_Key table and */
1046 	/* set the interface pkey assigment flag                            */
1047 	ipoib_pkey_dev_check_presence(dev);
1048 
1049 	/* P_Key value not assigned yet - start polling */
1050 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1051 		mutex_lock(&pkey_mutex);
1052 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1053 		queue_delayed_work(ipoib_workqueue,
1054 				   &priv->pkey_poll_task,
1055 				   HZ);
1056 		mutex_unlock(&pkey_mutex);
1057 		return 1;
1058 	}
1059 
1060 	return 0;
1061 }
1062