xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision 641a6cfb86023499caafe26a4d821a0b885cf00b)
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 "ipoib.h"
37 
38 #include <rdma/ib_cache.h>
39 
40 #include <security/mac/mac_framework.h>
41 
42 #include <linux/delay.h>
43 #include <linux/dma-mapping.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 ipoib_dev_priv *priv,
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->priv      = priv;
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(priv, "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 = ah->priv;
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 void
91 ipoib_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req)
92 {
93 	struct mbuf *m;
94 	int i;
95 
96 	for (i = 0, m = rx_req->mb; m != NULL; m = m->m_next, i++)
97 		ib_dma_unmap_single(priv->ca, rx_req->mapping[i], m->m_len,
98 		    DMA_FROM_DEVICE);
99 }
100 
101 void
102 ipoib_dma_mb(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int length)
103 {
104 
105 	m_adj(mb, -(mb->m_pkthdr.len - length));
106 }
107 
108 struct mbuf *
109 ipoib_alloc_map_mb(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req,
110     int size)
111 {
112 	struct mbuf *mb, *m;
113 	int i, j;
114 
115 	rx_req->mb = NULL;
116 	mb = m_getm2(NULL, size, M_NOWAIT, MT_DATA, M_PKTHDR);
117 	if (mb == NULL)
118 		return (NULL);
119 	for (i = 0, m = mb; m != NULL; m = m->m_next, i++) {
120 		m->m_len = (m->m_flags & M_EXT) ? m->m_ext.ext_size :
121 		    ((m->m_flags & M_PKTHDR) ? MHLEN : MLEN);
122 		mb->m_pkthdr.len += m->m_len;
123 		rx_req->mapping[i] = ib_dma_map_single(priv->ca,
124 		    mtod(m, void *), m->m_len, DMA_FROM_DEVICE);
125 		if (unlikely(ib_dma_mapping_error(priv->ca,
126 		    rx_req->mapping[i])))
127 			goto error;
128 
129 	}
130 	rx_req->mb = mb;
131 	return (mb);
132 error:
133 	for (j = 0, m = mb; j < i; m = m->m_next, j++)
134 		ib_dma_unmap_single(priv->ca, rx_req->mapping[j], m->m_len,
135 		    DMA_FROM_DEVICE);
136 	m_freem(mb);
137 	return (NULL);
138 
139 }
140 
141 static int ipoib_ib_post_receive(struct ipoib_dev_priv *priv, int id)
142 {
143 	struct ipoib_rx_buf *rx_req;
144 	struct ib_recv_wr *bad_wr;
145 	struct mbuf *m;
146 	int ret;
147 	int i;
148 
149 	rx_req = &priv->rx_ring[id];
150 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
151 		priv->rx_sge[i].addr = rx_req->mapping[i];
152 		priv->rx_sge[i].length = m->m_len;
153 	}
154 	priv->rx_wr.num_sge = i;
155 	priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
156 
157 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
158 	if (unlikely(ret)) {
159 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
160 		ipoib_dma_unmap_rx(priv, &priv->rx_ring[id]);
161 		m_freem(priv->rx_ring[id].mb);
162 		priv->rx_ring[id].mb = NULL;
163 	}
164 
165 	return ret;
166 }
167 
168 static struct mbuf *
169 ipoib_alloc_rx_mb(struct ipoib_dev_priv *priv, int id)
170 {
171 
172 	return ipoib_alloc_map_mb(priv, &priv->rx_ring[id],
173 	    priv->max_ib_mtu + IB_GRH_BYTES);
174 }
175 
176 static int ipoib_ib_post_receives(struct ipoib_dev_priv *priv)
177 {
178 	int i;
179 
180 	for (i = 0; i < ipoib_recvq_size; ++i) {
181 		if (!ipoib_alloc_rx_mb(priv, i)) {
182 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
183 			return -ENOMEM;
184 		}
185 		if (ipoib_ib_post_receive(priv, i)) {
186 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
187 			return -EIO;
188 		}
189 	}
190 
191 	return 0;
192 }
193 
194 static void
195 ipoib_ib_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
196 {
197 	struct ipoib_rx_buf saverx;
198 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
199 	struct ifnet *dev = priv->dev;
200 	struct ipoib_header *eh;
201 	struct mbuf *mb;
202 
203 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
204 		       wr_id, wc->status);
205 
206 	if (unlikely(wr_id >= ipoib_recvq_size)) {
207 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
208 			   wr_id, ipoib_recvq_size);
209 		return;
210 	}
211 
212 	mb  = priv->rx_ring[wr_id].mb;
213 
214 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
215 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
216 			ipoib_warn(priv, "failed recv event "
217 				   "(status=%d, wrid=%d vend_err %x)\n",
218 				   wc->status, wr_id, wc->vendor_err);
219 			goto repost;
220 		}
221 		if (mb) {
222 			ipoib_dma_unmap_rx(priv, &priv->rx_ring[wr_id]);
223 			m_freem(mb);
224 			priv->rx_ring[wr_id].mb = NULL;
225 		}
226 		return;
227 	}
228 
229 	/*
230 	 * Drop packets that this interface sent, ie multicast packets
231 	 * that the HCA has replicated.
232 	 */
233 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
234 		goto repost;
235 
236 	memcpy(&saverx, &priv->rx_ring[wr_id], sizeof(saverx));
237 	/*
238 	 * If we can't allocate a new RX buffer, dump
239 	 * this packet and reuse the old buffer.
240 	 */
241 	if (unlikely(!ipoib_alloc_rx_mb(priv, wr_id))) {
242 		memcpy(&priv->rx_ring[wr_id], &saverx, sizeof(saverx));
243 		dev->if_iqdrops++;
244 		goto repost;
245 	}
246 
247 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
248 		       wc->byte_len, wc->slid);
249 
250 	ipoib_dma_unmap_rx(priv, &saverx);
251 	ipoib_dma_mb(priv, mb, wc->byte_len);
252 
253 	++dev->if_ipackets;
254 	dev->if_ibytes += mb->m_pkthdr.len;
255 	mb->m_pkthdr.rcvif = dev;
256 	m_adj(mb, sizeof(struct ib_grh) - INFINIBAND_ALEN);
257 	eh = mtod(mb, struct ipoib_header *);
258 	bzero(eh->hwaddr, 4);	/* Zero the queue pair, only dgid is in grh */
259 
260 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
261 		mb->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID;
262 
263 	dev->if_input(dev, mb);
264 
265 repost:
266 	if (unlikely(ipoib_ib_post_receive(priv, wr_id)))
267 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
268 			   "for buf %d\n", wr_id);
269 }
270 
271 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req, int max)
272 {
273 	struct mbuf *mb = tx_req->mb;
274 	u64 *mapping = tx_req->mapping;
275 	struct mbuf *m, *p;
276 	int error;
277 	int i;
278 
279 	for (m = mb, p = NULL, i = 0; m != NULL; p = m, m = m->m_next, i++) {
280 		if (m->m_len != 0)
281 			continue;
282 		if (p == NULL)
283 			panic("ipoib_dma_map_tx: First mbuf empty\n");
284 		p->m_next = m_free(m);
285 		m = p;
286 		i--;
287 	}
288 	i--;
289 	if (i >= max) {
290 		tx_req->mb = mb = m_defrag(mb, M_DONTWAIT);
291 		if (mb == NULL)
292 			return -EIO;
293 		for (m = mb, i = 0; m != NULL; m = m->m_next, i++);
294 		if (i >= max)
295 			return -EIO;
296 	}
297 	error = 0;
298 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
299 		mapping[i] = ib_dma_map_single(ca, mtod(m, void *),
300 					       m->m_len, DMA_TO_DEVICE);
301 		if (unlikely(ib_dma_mapping_error(ca, mapping[i]))) {
302 			error = -EIO;
303 			break;
304 		}
305 	}
306 	if (error) {
307 		int end;
308 
309 		end = i;
310 		for (m = mb, i = 0; i < end; m = m->m_next, i++)
311 			ib_dma_unmap_single(ca, mapping[i], m->m_len,
312 					    DMA_TO_DEVICE);
313 	}
314 	return error;
315 }
316 
317 void ipoib_dma_unmap_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
318 {
319 	struct mbuf *mb = tx_req->mb;
320 	u64 *mapping = tx_req->mapping;
321 	struct mbuf *m;
322 	int i;
323 
324 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++)
325 		ib_dma_unmap_single(ca, mapping[i], m->m_len, DMA_TO_DEVICE);
326 }
327 
328 static void ipoib_ib_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
329 {
330 	struct ifnet *dev = priv->dev;
331 	unsigned int wr_id = wc->wr_id;
332 	struct ipoib_tx_buf *tx_req;
333 
334 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
335 		       wr_id, wc->status);
336 
337 	if (unlikely(wr_id >= ipoib_sendq_size)) {
338 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
339 			   wr_id, ipoib_sendq_size);
340 		return;
341 	}
342 
343 	tx_req = &priv->tx_ring[wr_id];
344 
345 	ipoib_dma_unmap_tx(priv->ca, tx_req);
346 
347 	++dev->if_opackets;
348 
349 	m_freem(tx_req->mb);
350 
351 	++priv->tx_tail;
352 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
353 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) &&
354 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
355 		dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
356 
357 	if (wc->status != IB_WC_SUCCESS &&
358 	    wc->status != IB_WC_WR_FLUSH_ERR)
359 		ipoib_warn(priv, "failed send event "
360 			   "(status=%d, wrid=%d vend_err %x)\n",
361 			   wc->status, wr_id, wc->vendor_err);
362 }
363 
364 int
365 ipoib_poll_tx(struct ipoib_dev_priv *priv)
366 {
367 	int n, i;
368 
369 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
370 	for (i = 0; i < n; ++i) {
371 		struct ib_wc *wc = priv->send_wc + i;
372 		if (wc->wr_id & IPOIB_OP_CM)
373 			ipoib_cm_handle_tx_wc(priv, wc);
374 		else
375 			ipoib_ib_handle_tx_wc(priv, wc);
376 	}
377 
378 	return n == MAX_SEND_CQE;
379 }
380 
381 static void
382 ipoib_poll(struct ipoib_dev_priv *priv)
383 {
384 	int n, i;
385 
386 poll_more:
387 	for (;;) {
388 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
389 
390 		for (i = 0; i < n; i++) {
391 			struct ib_wc *wc = priv->ibwc + i;
392 
393 			if ((wc->wr_id & IPOIB_OP_RECV) == 0)
394 				panic("ipoib_poll: Bad wr_id 0x%jX\n",
395 				    (intmax_t)wc->wr_id);
396 			if (wc->wr_id & IPOIB_OP_CM)
397 				ipoib_cm_handle_rx_wc(priv, wc);
398 			else
399 				ipoib_ib_handle_rx_wc(priv, wc);
400 		}
401 
402 		if (n != IPOIB_NUM_WC)
403 			break;
404 	}
405 
406 	if (ib_req_notify_cq(priv->recv_cq,
407 	    IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS))
408 		goto poll_more;
409 }
410 
411 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
412 {
413 	struct ipoib_dev_priv *priv = dev_ptr;
414 
415 	ipoib_poll(priv);
416 }
417 
418 static void drain_tx_cq(struct ipoib_dev_priv *priv)
419 {
420 	struct ifnet *dev = priv->dev;
421 
422 	spin_lock(&priv->lock);
423 	while (ipoib_poll_tx(priv))
424 		; /* nothing */
425 
426 	if (dev->if_drv_flags & IFF_DRV_OACTIVE)
427 		mod_timer(&priv->poll_timer, jiffies + 1);
428 
429 	spin_unlock(&priv->lock);
430 }
431 
432 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
433 {
434 	struct ipoib_dev_priv *priv = dev_ptr;
435 
436 	mod_timer(&priv->poll_timer, jiffies);
437 }
438 
439 static inline int
440 post_send(struct ipoib_dev_priv *priv, unsigned int wr_id,
441     struct ib_ah *address, u32 qpn, struct ipoib_tx_buf *tx_req, void *head,
442     int hlen)
443 {
444 	struct ib_send_wr *bad_wr;
445 	struct mbuf *mb = tx_req->mb;
446 	u64 *mapping = tx_req->mapping;
447 	struct mbuf *m;
448 	int i;
449 
450 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
451 		priv->tx_sge[i].addr         = mapping[i];
452 		priv->tx_sge[i].length       = m->m_len;
453 	}
454 	priv->tx_wr.num_sge	     = i;
455 	priv->tx_wr.wr_id 	     = wr_id;
456 	priv->tx_wr.wr.ud.remote_qpn = qpn;
457 	priv->tx_wr.wr.ud.ah 	     = address;
458 
459 
460 	if (head) {
461 		priv->tx_wr.wr.ud.mss	 = 0; /* XXX mb_shinfo(mb)->gso_size; */
462 		priv->tx_wr.wr.ud.header = head;
463 		priv->tx_wr.wr.ud.hlen	 = hlen;
464 		priv->tx_wr.opcode	 = IB_WR_LSO;
465 	} else
466 		priv->tx_wr.opcode	 = IB_WR_SEND;
467 
468 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
469 }
470 
471 void
472 ipoib_send(struct ipoib_dev_priv *priv, struct mbuf *mb,
473     struct ipoib_ah *address, u32 qpn)
474 {
475 	struct ifnet *dev = priv->dev;
476 	struct ipoib_tx_buf *tx_req;
477 	int hlen;
478 	void *phead;
479 
480 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
481 		while (ipoib_poll_tx(priv))
482 			; /* nothing */
483 
484 	m_adj(mb, sizeof (struct ipoib_pseudoheader));
485 	if (0 /* XXX segment offload mb_is_gso(mb) */) {
486 		/* XXX hlen = mb_transport_offset(mb) + tcp_hdrlen(mb); */
487 		phead = mtod(mb, void *);
488 		if (mb->m_len < hlen) {
489 			ipoib_warn(priv, "linear data too small\n");
490 			++dev->if_oerrors;
491 			m_freem(mb);
492 			return;
493 		}
494 		m_adj(mb, hlen);
495 	} else {
496 		if (unlikely(mb->m_pkthdr.len - IPOIB_ENCAP_LEN > priv->mcast_mtu)) {
497 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
498 				   mb->m_pkthdr.len, priv->mcast_mtu);
499 			++dev->if_oerrors;
500 			ipoib_cm_mb_too_long(priv, mb, priv->mcast_mtu);
501 			return;
502 		}
503 		phead = NULL;
504 		hlen  = 0;
505 	}
506 
507 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
508 		       mb->m_pkthdr.len, address, qpn);
509 
510 	/*
511 	 * We put the mb into the tx_ring _before_ we call post_send()
512 	 * because it's entirely possible that the completion handler will
513 	 * run before we execute anything after the post_send().  That
514 	 * means we have to make sure everything is properly recorded and
515 	 * our state is consistent before we call post_send().
516 	 */
517 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
518 	tx_req->mb = mb;
519 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req, IPOIB_UD_TX_SG))) {
520 		++dev->if_oerrors;
521 		if (tx_req->mb)
522 			m_freem(tx_req->mb);
523 		return;
524 	}
525 
526 	if (mb->m_pkthdr.csum_flags & (CSUM_IP|CSUM_TCP|CSUM_UDP))
527 		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
528 	else
529 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
530 
531 	if (++priv->tx_outstanding == ipoib_sendq_size) {
532 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
533 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
534 			ipoib_warn(priv, "request notify on send CQ failed\n");
535 		dev->if_drv_flags |= IFF_DRV_OACTIVE;
536 	}
537 
538 	if (unlikely(post_send(priv,
539 	    priv->tx_head & (ipoib_sendq_size - 1), address->ah, qpn,
540 	    tx_req, phead, hlen))) {
541 		ipoib_warn(priv, "post_send failed\n");
542 		++dev->if_oerrors;
543 		--priv->tx_outstanding;
544 		ipoib_dma_unmap_tx(priv->ca, tx_req);
545 		m_freem(mb);
546 		if (dev->if_drv_flags & IFF_DRV_OACTIVE)
547 			dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
548 	} else {
549 		address->last_send = priv->tx_head;
550 		++priv->tx_head;
551 	}
552 }
553 
554 static void __ipoib_reap_ah(struct ipoib_dev_priv *priv)
555 {
556 	struct ipoib_ah *ah, *tah;
557 	LIST_HEAD(remove_list);
558 	unsigned long flags;
559 
560 	spin_lock_irqsave(&priv->lock, flags);
561 
562 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
563 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
564 			list_del(&ah->list);
565 			ib_destroy_ah(ah->ah);
566 			kfree(ah);
567 		}
568 
569 	spin_unlock_irqrestore(&priv->lock, flags);
570 }
571 
572 void ipoib_reap_ah(struct work_struct *work)
573 {
574 	struct ipoib_dev_priv *priv =
575 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
576 
577 	__ipoib_reap_ah(priv);
578 
579 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
580 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
581 				   HZ);
582 }
583 
584 static void ipoib_ah_dev_cleanup(struct ipoib_dev_priv *priv)
585 {
586 	unsigned long begin;
587 
588 	begin = jiffies;
589 
590 	while (!list_empty(&priv->dead_ahs)) {
591 		__ipoib_reap_ah(priv);
592 
593 		if (time_after(jiffies, begin + HZ)) {
594 			ipoib_warn(priv, "timing out; will leak address handles\n");
595 			break;
596 		}
597 
598 		msleep(1);
599 	}
600 }
601 
602 static void ipoib_ib_tx_timer_func(unsigned long ctx)
603 {
604 	drain_tx_cq((struct ipoib_dev_priv *)ctx);
605 }
606 
607 int ipoib_ib_dev_open(struct ipoib_dev_priv *priv)
608 {
609 	int ret;
610 
611 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
612 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
613 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
614 		return -1;
615 	}
616 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
617 
618 	ret = ipoib_init_qp(priv);
619 	if (ret) {
620 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
621 		return -1;
622 	}
623 
624 	ret = ipoib_ib_post_receives(priv);
625 	if (ret) {
626 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
627 		ipoib_ib_dev_stop(priv, 1);
628 		return -1;
629 	}
630 
631 	ret = ipoib_cm_dev_open(priv);
632 	if (ret) {
633 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
634 		ipoib_ib_dev_stop(priv, 1);
635 		return -1;
636 	}
637 
638 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
639 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
640 
641 	return 0;
642 }
643 
644 static void ipoib_pkey_dev_check_presence(struct ipoib_dev_priv *priv)
645 {
646 	u16 pkey_index = 0;
647 
648 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
649 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
650 	else
651 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
652 }
653 
654 int ipoib_ib_dev_up(struct ipoib_dev_priv *priv)
655 {
656 
657 	ipoib_pkey_dev_check_presence(priv);
658 
659 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
660 		ipoib_dbg(priv, "PKEY is not assigned.\n");
661 		return 0;
662 	}
663 
664 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
665 
666 	return ipoib_mcast_start_thread(priv);
667 }
668 
669 int ipoib_ib_dev_down(struct ipoib_dev_priv *priv, int flush)
670 {
671 
672 	ipoib_dbg(priv, "downing ib_dev\n");
673 
674 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
675 	if_link_state_change(priv->dev, LINK_STATE_DOWN);
676 
677 	/* Shutdown the P_Key thread if still active */
678 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
679 		mutex_lock(&pkey_mutex);
680 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
681 		cancel_delayed_work(&priv->pkey_poll_task);
682 		mutex_unlock(&pkey_mutex);
683 		if (flush)
684 			flush_workqueue(ipoib_workqueue);
685 	}
686 
687 	ipoib_mcast_stop_thread(priv, flush);
688 	ipoib_mcast_dev_flush(priv);
689 
690 	ipoib_flush_paths(priv);
691 
692 	return 0;
693 }
694 
695 static int recvs_pending(struct ipoib_dev_priv *priv)
696 {
697 	int pending = 0;
698 	int i;
699 
700 	for (i = 0; i < ipoib_recvq_size; ++i)
701 		if (priv->rx_ring[i].mb)
702 			++pending;
703 
704 	return pending;
705 }
706 
707 void ipoib_drain_cq(struct ipoib_dev_priv *priv)
708 {
709 	int i, n;
710 
711 	do {
712 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
713 		for (i = 0; i < n; ++i) {
714 			/*
715 			 * Convert any successful completions to flush
716 			 * errors to avoid passing packets up the
717 			 * stack after bringing the device down.
718 			 */
719 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
720 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
721 
722 			if ((priv->ibwc[i].wr_id & IPOIB_OP_RECV) == 0)
723 				panic("ipoib_drain_cq:  Bad wrid 0x%jX\n",
724 				    (intmax_t)priv->ibwc[i].wr_id);
725 			if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
726 				ipoib_cm_handle_rx_wc(priv, priv->ibwc + i);
727 			else
728 				ipoib_ib_handle_rx_wc(priv, priv->ibwc + i);
729 		}
730 	} while (n == IPOIB_NUM_WC);
731 
732 	spin_lock(&priv->lock);
733 	while (ipoib_poll_tx(priv))
734 		; /* nothing */
735 
736 	spin_unlock(&priv->lock);
737 }
738 
739 int ipoib_ib_dev_stop(struct ipoib_dev_priv *priv, int flush)
740 {
741 	struct ib_qp_attr qp_attr;
742 	unsigned long begin;
743 	struct ipoib_tx_buf *tx_req;
744 	int i;
745 
746 	ipoib_cm_dev_stop(priv);
747 
748 	/*
749 	 * Move our QP to the error state and then reinitialize in
750 	 * when all work requests have completed or have been flushed.
751 	 */
752 	qp_attr.qp_state = IB_QPS_ERR;
753 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
754 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
755 
756 	/* Wait for all sends and receives to complete */
757 	begin = jiffies;
758 
759 	while (priv->tx_head != priv->tx_tail || recvs_pending(priv)) {
760 		if (time_after(jiffies, begin + 5 * HZ)) {
761 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
762 				   priv->tx_head - priv->tx_tail, recvs_pending(priv));
763 
764 			/*
765 			 * assume the HW is wedged and just free up
766 			 * all our pending work requests.
767 			 */
768 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
769 				tx_req = &priv->tx_ring[priv->tx_tail &
770 							(ipoib_sendq_size - 1)];
771 				ipoib_dma_unmap_tx(priv->ca, tx_req);
772 				m_freem(tx_req->mb);
773 				++priv->tx_tail;
774 				--priv->tx_outstanding;
775 			}
776 
777 			for (i = 0; i < ipoib_recvq_size; ++i) {
778 				struct ipoib_rx_buf *rx_req;
779 
780 				rx_req = &priv->rx_ring[i];
781 				if (!rx_req->mb)
782 					continue;
783 				ipoib_dma_unmap_rx(priv, &priv->rx_ring[i]);
784 				m_freem(rx_req->mb);
785 				rx_req->mb = NULL;
786 			}
787 
788 			goto timeout;
789 		}
790 
791 		ipoib_drain_cq(priv);
792 
793 		msleep(1);
794 	}
795 
796 	ipoib_dbg(priv, "All sends and receives done.\n");
797 
798 timeout:
799 	del_timer_sync(&priv->poll_timer);
800 	qp_attr.qp_state = IB_QPS_RESET;
801 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
802 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
803 
804 	/* Wait for all AHs to be reaped */
805 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
806 	cancel_delayed_work(&priv->ah_reap_task);
807 	if (flush)
808 		flush_workqueue(ipoib_workqueue);
809 
810 	ipoib_ah_dev_cleanup(priv);
811 
812 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
813 
814 	return 0;
815 }
816 
817 int ipoib_ib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
818 {
819 	struct ifnet *dev = priv->dev;
820 
821 	priv->ca = ca;
822 	priv->port = port;
823 	priv->qp = NULL;
824 
825 	if (ipoib_transport_dev_init(priv, ca)) {
826 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
827 		return -ENODEV;
828 	}
829 
830 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
831 		    (unsigned long) priv);
832 
833 	if (dev->if_flags & IFF_UP) {
834 		if (ipoib_ib_dev_open(priv)) {
835 			ipoib_transport_dev_cleanup(priv);
836 			return -ENODEV;
837 		}
838 	}
839 
840 	return 0;
841 }
842 
843 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
844 				enum ipoib_flush_level level)
845 {
846 	struct ipoib_dev_priv *cpriv;
847 	u16 new_index;
848 
849 	mutex_lock(&priv->vlan_mutex);
850 
851 	/*
852 	 * Flush any child interfaces too -- they might be up even if
853 	 * the parent is down.
854 	 */
855 	list_for_each_entry(cpriv, &priv->child_intfs, list)
856 		__ipoib_ib_dev_flush(cpriv, level);
857 
858 	mutex_unlock(&priv->vlan_mutex);
859 
860 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
861 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
862 		return;
863 	}
864 
865 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
866 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
867 		return;
868 	}
869 
870 	if (level == IPOIB_FLUSH_HEAVY) {
871 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
872 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
873 			ipoib_ib_dev_down(priv, 0);
874 			ipoib_ib_dev_stop(priv, 0);
875 			if (ipoib_pkey_dev_delay_open(priv))
876 				return;
877 		}
878 
879 		/* restart QP only if P_Key index is changed */
880 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
881 		    new_index == priv->pkey_index) {
882 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
883 			return;
884 		}
885 		priv->pkey_index = new_index;
886 	}
887 
888 	if (level == IPOIB_FLUSH_LIGHT) {
889 		ipoib_mark_paths_invalid(priv);
890 		ipoib_mcast_dev_flush(priv);
891 	}
892 
893 	if (level >= IPOIB_FLUSH_NORMAL)
894 		ipoib_ib_dev_down(priv, 0);
895 
896 	if (level == IPOIB_FLUSH_HEAVY) {
897 		ipoib_ib_dev_stop(priv, 0);
898 		ipoib_ib_dev_open(priv);
899 	}
900 
901 	/*
902 	 * The device could have been brought down between the start and when
903 	 * we get here, don't bring it back up if it's not configured up
904 	 */
905 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
906 		if (level >= IPOIB_FLUSH_NORMAL)
907 			ipoib_ib_dev_up(priv);
908 		ipoib_mcast_restart_task(&priv->restart_task);
909 	}
910 }
911 
912 void ipoib_ib_dev_flush_light(struct work_struct *work)
913 {
914 	struct ipoib_dev_priv *priv =
915 		container_of(work, struct ipoib_dev_priv, flush_light);
916 
917 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
918 }
919 
920 void ipoib_ib_dev_flush_normal(struct work_struct *work)
921 {
922 	struct ipoib_dev_priv *priv =
923 		container_of(work, struct ipoib_dev_priv, flush_normal);
924 
925 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
926 }
927 
928 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
929 {
930 	struct ipoib_dev_priv *priv =
931 		container_of(work, struct ipoib_dev_priv, flush_heavy);
932 
933 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
934 }
935 
936 void ipoib_ib_dev_cleanup(struct ipoib_dev_priv *priv)
937 {
938 
939 	ipoib_dbg(priv, "cleaning up ib_dev\n");
940 
941 	ipoib_mcast_stop_thread(priv, 1);
942 	ipoib_mcast_dev_flush(priv);
943 
944 	ipoib_ah_dev_cleanup(priv);
945 	ipoib_transport_dev_cleanup(priv);
946 }
947 
948 /*
949  * Delayed P_Key Assigment Interim Support
950  *
951  * The following is initial implementation of delayed P_Key assigment
952  * mechanism. It is using the same approach implemented for the multicast
953  * group join. The single goal of this implementation is to quickly address
954  * Bug #2507. This implementation will probably be removed when the P_Key
955  * change async notification is available.
956  */
957 
958 void ipoib_pkey_poll(struct work_struct *work)
959 {
960 	struct ipoib_dev_priv *priv =
961 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
962 
963 	ipoib_pkey_dev_check_presence(priv);
964 
965 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
966 		ipoib_open(priv);
967 	else {
968 		mutex_lock(&pkey_mutex);
969 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
970 			queue_delayed_work(ipoib_workqueue,
971 					   &priv->pkey_poll_task,
972 					   HZ);
973 		mutex_unlock(&pkey_mutex);
974 	}
975 }
976 
977 int ipoib_pkey_dev_delay_open(struct ipoib_dev_priv *priv)
978 {
979 
980 	/* Look for the interface pkey value in the IB Port P_Key table and */
981 	/* set the interface pkey assigment flag                            */
982 	ipoib_pkey_dev_check_presence(priv);
983 
984 	/* P_Key value not assigned yet - start polling */
985 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
986 		mutex_lock(&pkey_mutex);
987 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
988 		queue_delayed_work(ipoib_workqueue,
989 				   &priv->pkey_poll_task,
990 				   HZ);
991 		mutex_unlock(&pkey_mutex);
992 		return 1;
993 	}
994 
995 	return 0;
996 }
997