xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision 4436b51dff5736e74da464946049ea6899a88938)
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_SIZE(m);
121 		mb->m_pkthdr.len += m->m_len;
122 		rx_req->mapping[i] = ib_dma_map_single(priv->ca,
123 		    mtod(m, void *), m->m_len, DMA_FROM_DEVICE);
124 		if (unlikely(ib_dma_mapping_error(priv->ca,
125 		    rx_req->mapping[i])))
126 			goto error;
127 
128 	}
129 	rx_req->mb = mb;
130 	return (mb);
131 error:
132 	for (j = 0, m = mb; j < i; m = m->m_next, j++)
133 		ib_dma_unmap_single(priv->ca, rx_req->mapping[j], m->m_len,
134 		    DMA_FROM_DEVICE);
135 	m_freem(mb);
136 	return (NULL);
137 
138 }
139 
140 static int ipoib_ib_post_receive(struct ipoib_dev_priv *priv, int id)
141 {
142 	struct ipoib_rx_buf *rx_req;
143 	struct ib_recv_wr *bad_wr;
144 	struct mbuf *m;
145 	int ret;
146 	int i;
147 
148 	rx_req = &priv->rx_ring[id];
149 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
150 		priv->rx_sge[i].addr = rx_req->mapping[i];
151 		priv->rx_sge[i].length = m->m_len;
152 	}
153 	priv->rx_wr.num_sge = i;
154 	priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
155 
156 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
157 	if (unlikely(ret)) {
158 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
159 		ipoib_dma_unmap_rx(priv, &priv->rx_ring[id]);
160 		m_freem(priv->rx_ring[id].mb);
161 		priv->rx_ring[id].mb = NULL;
162 	}
163 
164 	return ret;
165 }
166 
167 static struct mbuf *
168 ipoib_alloc_rx_mb(struct ipoib_dev_priv *priv, int id)
169 {
170 
171 	return ipoib_alloc_map_mb(priv, &priv->rx_ring[id],
172 	    priv->max_ib_mtu + IB_GRH_BYTES);
173 }
174 
175 static int ipoib_ib_post_receives(struct ipoib_dev_priv *priv)
176 {
177 	int i;
178 
179 	for (i = 0; i < ipoib_recvq_size; ++i) {
180 		if (!ipoib_alloc_rx_mb(priv, i)) {
181 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
182 			return -ENOMEM;
183 		}
184 		if (ipoib_ib_post_receive(priv, i)) {
185 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
186 			return -EIO;
187 		}
188 	}
189 
190 	return 0;
191 }
192 
193 static void
194 ipoib_ib_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
195 {
196 	struct ipoib_rx_buf saverx;
197 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
198 	struct ifnet *dev = priv->dev;
199 	struct ipoib_header *eh;
200 	struct mbuf *mb;
201 
202 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
203 		       wr_id, wc->status);
204 
205 	if (unlikely(wr_id >= ipoib_recvq_size)) {
206 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
207 			   wr_id, ipoib_recvq_size);
208 		return;
209 	}
210 
211 	mb  = priv->rx_ring[wr_id].mb;
212 
213 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
214 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
215 			ipoib_warn(priv, "failed recv event "
216 				   "(status=%d, wrid=%d vend_err %x)\n",
217 				   wc->status, wr_id, wc->vendor_err);
218 			goto repost;
219 		}
220 		if (mb) {
221 			ipoib_dma_unmap_rx(priv, &priv->rx_ring[wr_id]);
222 			m_freem(mb);
223 			priv->rx_ring[wr_id].mb = NULL;
224 		}
225 		return;
226 	}
227 
228 	/*
229 	 * Drop packets that this interface sent, ie multicast packets
230 	 * that the HCA has replicated.
231 	 */
232 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
233 		goto repost;
234 
235 	memcpy(&saverx, &priv->rx_ring[wr_id], sizeof(saverx));
236 	/*
237 	 * If we can't allocate a new RX buffer, dump
238 	 * this packet and reuse the old buffer.
239 	 */
240 	if (unlikely(!ipoib_alloc_rx_mb(priv, wr_id))) {
241 		memcpy(&priv->rx_ring[wr_id], &saverx, sizeof(saverx));
242 		if_inc_counter(dev, IFCOUNTER_IQDROPS, 1);
243 		goto repost;
244 	}
245 
246 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
247 		       wc->byte_len, wc->slid);
248 
249 	ipoib_dma_unmap_rx(priv, &saverx);
250 	ipoib_dma_mb(priv, mb, wc->byte_len);
251 
252 	if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
253 	if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
254 	mb->m_pkthdr.rcvif = dev;
255 	m_adj(mb, sizeof(struct ib_grh) - INFINIBAND_ALEN);
256 	eh = mtod(mb, struct ipoib_header *);
257 	bzero(eh->hwaddr, 4);	/* Zero the queue pair, only dgid is in grh */
258 
259 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
260 		mb->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID;
261 
262 	dev->if_input(dev, mb);
263 
264 repost:
265 	if (unlikely(ipoib_ib_post_receive(priv, wr_id)))
266 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
267 			   "for buf %d\n", wr_id);
268 }
269 
270 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req, int max)
271 {
272 	struct mbuf *mb = tx_req->mb;
273 	u64 *mapping = tx_req->mapping;
274 	struct mbuf *m, *p;
275 	int error;
276 	int i;
277 
278 	for (m = mb, p = NULL, i = 0; m != NULL; p = m, m = m->m_next, i++) {
279 		if (m->m_len != 0)
280 			continue;
281 		if (p == NULL)
282 			panic("ipoib_dma_map_tx: First mbuf empty\n");
283 		p->m_next = m_free(m);
284 		m = p;
285 		i--;
286 	}
287 	i--;
288 	if (i >= max) {
289 		tx_req->mb = mb = m_defrag(mb, M_NOWAIT);
290 		if (mb == NULL)
291 			return -EIO;
292 		for (m = mb, i = 0; m != NULL; m = m->m_next, i++);
293 		if (i >= max)
294 			return -EIO;
295 	}
296 	error = 0;
297 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
298 		mapping[i] = ib_dma_map_single(ca, mtod(m, void *),
299 					       m->m_len, DMA_TO_DEVICE);
300 		if (unlikely(ib_dma_mapping_error(ca, mapping[i]))) {
301 			error = -EIO;
302 			break;
303 		}
304 	}
305 	if (error) {
306 		int end;
307 
308 		end = i;
309 		for (m = mb, i = 0; i < end; m = m->m_next, i++)
310 			ib_dma_unmap_single(ca, mapping[i], m->m_len,
311 					    DMA_TO_DEVICE);
312 	}
313 	return error;
314 }
315 
316 void ipoib_dma_unmap_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
317 {
318 	struct mbuf *mb = tx_req->mb;
319 	u64 *mapping = tx_req->mapping;
320 	struct mbuf *m;
321 	int i;
322 
323 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++)
324 		ib_dma_unmap_single(ca, mapping[i], m->m_len, DMA_TO_DEVICE);
325 }
326 
327 static void ipoib_ib_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
328 {
329 	struct ifnet *dev = priv->dev;
330 	unsigned int wr_id = wc->wr_id;
331 	struct ipoib_tx_buf *tx_req;
332 
333 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
334 		       wr_id, wc->status);
335 
336 	if (unlikely(wr_id >= ipoib_sendq_size)) {
337 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
338 			   wr_id, ipoib_sendq_size);
339 		return;
340 	}
341 
342 	tx_req = &priv->tx_ring[wr_id];
343 
344 	ipoib_dma_unmap_tx(priv->ca, tx_req);
345 
346 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
347 
348 	m_freem(tx_req->mb);
349 
350 	++priv->tx_tail;
351 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
352 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) &&
353 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
354 		dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
355 
356 	if (wc->status != IB_WC_SUCCESS &&
357 	    wc->status != IB_WC_WR_FLUSH_ERR)
358 		ipoib_warn(priv, "failed send event "
359 			   "(status=%d, wrid=%d vend_err %x)\n",
360 			   wc->status, wr_id, wc->vendor_err);
361 }
362 
363 int
364 ipoib_poll_tx(struct ipoib_dev_priv *priv)
365 {
366 	int n, i;
367 
368 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
369 	for (i = 0; i < n; ++i) {
370 		struct ib_wc *wc = priv->send_wc + i;
371 		if (wc->wr_id & IPOIB_OP_CM)
372 			ipoib_cm_handle_tx_wc(priv, wc);
373 		else
374 			ipoib_ib_handle_tx_wc(priv, wc);
375 	}
376 
377 	return n == MAX_SEND_CQE;
378 }
379 
380 static void
381 ipoib_poll(struct ipoib_dev_priv *priv)
382 {
383 	int n, i;
384 
385 poll_more:
386 	spin_lock(&priv->drain_lock);
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 	spin_unlock(&priv->drain_lock);
406 
407 	if (ib_req_notify_cq(priv->recv_cq,
408 	    IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS))
409 		goto poll_more;
410 }
411 
412 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
413 {
414 	struct ipoib_dev_priv *priv = dev_ptr;
415 
416 	ipoib_poll(priv);
417 }
418 
419 static void drain_tx_cq(struct ipoib_dev_priv *priv)
420 {
421 	struct ifnet *dev = priv->dev;
422 
423 	spin_lock(&priv->lock);
424 	while (ipoib_poll_tx(priv))
425 		; /* nothing */
426 
427 	if (dev->if_drv_flags & IFF_DRV_OACTIVE)
428 		mod_timer(&priv->poll_timer, jiffies + 1);
429 
430 	spin_unlock(&priv->lock);
431 }
432 
433 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
434 {
435 	struct ipoib_dev_priv *priv = dev_ptr;
436 
437 	mod_timer(&priv->poll_timer, jiffies);
438 }
439 
440 static inline int
441 post_send(struct ipoib_dev_priv *priv, unsigned int wr_id,
442     struct ib_ah *address, u32 qpn, struct ipoib_tx_buf *tx_req, void *head,
443     int hlen)
444 {
445 	struct ib_send_wr *bad_wr;
446 	struct mbuf *mb = tx_req->mb;
447 	u64 *mapping = tx_req->mapping;
448 	struct mbuf *m;
449 	int i;
450 
451 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
452 		priv->tx_sge[i].addr         = mapping[i];
453 		priv->tx_sge[i].length       = m->m_len;
454 	}
455 	priv->tx_wr.num_sge	     = i;
456 	priv->tx_wr.wr_id 	     = wr_id;
457 	priv->tx_wr.wr.ud.remote_qpn = qpn;
458 	priv->tx_wr.wr.ud.ah 	     = address;
459 
460 
461 	if (head) {
462 		priv->tx_wr.wr.ud.mss	 = 0; /* XXX mb_shinfo(mb)->gso_size; */
463 		priv->tx_wr.wr.ud.header = head;
464 		priv->tx_wr.wr.ud.hlen	 = hlen;
465 		priv->tx_wr.opcode	 = IB_WR_LSO;
466 	} else
467 		priv->tx_wr.opcode	 = IB_WR_SEND;
468 
469 	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
470 }
471 
472 void
473 ipoib_send(struct ipoib_dev_priv *priv, struct mbuf *mb,
474     struct ipoib_ah *address, u32 qpn)
475 {
476 	struct ifnet *dev = priv->dev;
477 	struct ipoib_tx_buf *tx_req;
478 	int hlen;
479 	void *phead;
480 
481 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
482 		while (ipoib_poll_tx(priv))
483 			; /* nothing */
484 
485 	m_adj(mb, sizeof (struct ipoib_pseudoheader));
486 	if (0 /* XXX segment offload mb_is_gso(mb) */) {
487 		/* XXX hlen = mb_transport_offset(mb) + tcp_hdrlen(mb); */
488 		phead = mtod(mb, void *);
489 		if (mb->m_len < hlen) {
490 			ipoib_warn(priv, "linear data too small\n");
491 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
492 			m_freem(mb);
493 			return;
494 		}
495 		m_adj(mb, hlen);
496 	} else {
497 		if (unlikely(mb->m_pkthdr.len - IPOIB_ENCAP_LEN > priv->mcast_mtu)) {
498 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
499 				   mb->m_pkthdr.len, priv->mcast_mtu);
500 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
501 			ipoib_cm_mb_too_long(priv, mb, priv->mcast_mtu);
502 			return;
503 		}
504 		phead = NULL;
505 		hlen  = 0;
506 	}
507 
508 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
509 		       mb->m_pkthdr.len, address, qpn);
510 
511 	/*
512 	 * We put the mb into the tx_ring _before_ we call post_send()
513 	 * because it's entirely possible that the completion handler will
514 	 * run before we execute anything after the post_send().  That
515 	 * means we have to make sure everything is properly recorded and
516 	 * our state is consistent before we call post_send().
517 	 */
518 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
519 	tx_req->mb = mb;
520 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req, IPOIB_UD_TX_SG))) {
521 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
522 		if (tx_req->mb)
523 			m_freem(tx_req->mb);
524 		return;
525 	}
526 
527 	if (mb->m_pkthdr.csum_flags & (CSUM_IP|CSUM_TCP|CSUM_UDP))
528 		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
529 	else
530 		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
531 
532 	if (++priv->tx_outstanding == ipoib_sendq_size) {
533 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
534 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
535 			ipoib_warn(priv, "request notify on send CQ failed\n");
536 		dev->if_drv_flags |= IFF_DRV_OACTIVE;
537 	}
538 
539 	if (unlikely(post_send(priv,
540 	    priv->tx_head & (ipoib_sendq_size - 1), address->ah, qpn,
541 	    tx_req, phead, hlen))) {
542 		ipoib_warn(priv, "post_send failed\n");
543 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
544 		--priv->tx_outstanding;
545 		ipoib_dma_unmap_tx(priv->ca, tx_req);
546 		m_freem(mb);
547 		if (dev->if_drv_flags & IFF_DRV_OACTIVE)
548 			dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
549 	} else {
550 		address->last_send = priv->tx_head;
551 		++priv->tx_head;
552 	}
553 }
554 
555 static void __ipoib_reap_ah(struct ipoib_dev_priv *priv)
556 {
557 	struct ipoib_ah *ah, *tah;
558 	LIST_HEAD(remove_list);
559 	unsigned long flags;
560 
561 	spin_lock_irqsave(&priv->lock, flags);
562 
563 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
564 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
565 			list_del(&ah->list);
566 			ib_destroy_ah(ah->ah);
567 			kfree(ah);
568 		}
569 
570 	spin_unlock_irqrestore(&priv->lock, flags);
571 }
572 
573 void ipoib_reap_ah(struct work_struct *work)
574 {
575 	struct ipoib_dev_priv *priv =
576 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
577 
578 	__ipoib_reap_ah(priv);
579 
580 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
581 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
582 				   HZ);
583 }
584 
585 static void ipoib_ah_dev_cleanup(struct ipoib_dev_priv *priv)
586 {
587 	unsigned long begin;
588 
589 	begin = jiffies;
590 
591 	while (!list_empty(&priv->dead_ahs)) {
592 		__ipoib_reap_ah(priv);
593 
594 		if (time_after(jiffies, begin + HZ)) {
595 			ipoib_warn(priv, "timing out; will leak address handles\n");
596 			break;
597 		}
598 
599 		msleep(1);
600 	}
601 }
602 
603 static void ipoib_ib_tx_timer_func(unsigned long ctx)
604 {
605 	drain_tx_cq((struct ipoib_dev_priv *)ctx);
606 }
607 
608 int ipoib_ib_dev_open(struct ipoib_dev_priv *priv)
609 {
610 	int ret;
611 
612 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
613 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
614 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
615 		return -1;
616 	}
617 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
618 
619 	ret = ipoib_init_qp(priv);
620 	if (ret) {
621 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
622 		return -1;
623 	}
624 
625 	ret = ipoib_ib_post_receives(priv);
626 	if (ret) {
627 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
628 		ipoib_ib_dev_stop(priv, 1);
629 		return -1;
630 	}
631 
632 	ret = ipoib_cm_dev_open(priv);
633 	if (ret) {
634 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
635 		ipoib_ib_dev_stop(priv, 1);
636 		return -1;
637 	}
638 
639 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
640 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
641 
642 	return 0;
643 }
644 
645 static void ipoib_pkey_dev_check_presence(struct ipoib_dev_priv *priv)
646 {
647 	u16 pkey_index = 0;
648 
649 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
650 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
651 	else
652 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
653 }
654 
655 int ipoib_ib_dev_up(struct ipoib_dev_priv *priv)
656 {
657 
658 	ipoib_pkey_dev_check_presence(priv);
659 
660 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
661 		ipoib_dbg(priv, "PKEY is not assigned.\n");
662 		return 0;
663 	}
664 
665 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
666 
667 	return ipoib_mcast_start_thread(priv);
668 }
669 
670 int ipoib_ib_dev_down(struct ipoib_dev_priv *priv, int flush)
671 {
672 
673 	ipoib_dbg(priv, "downing ib_dev\n");
674 
675 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
676 	if_link_state_change(priv->dev, LINK_STATE_DOWN);
677 
678 	/* Shutdown the P_Key thread if still active */
679 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
680 		mutex_lock(&pkey_mutex);
681 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
682 		cancel_delayed_work(&priv->pkey_poll_task);
683 		mutex_unlock(&pkey_mutex);
684 		if (flush)
685 			flush_workqueue(ipoib_workqueue);
686 	}
687 
688 	ipoib_mcast_stop_thread(priv, flush);
689 	ipoib_mcast_dev_flush(priv);
690 
691 	ipoib_flush_paths(priv);
692 
693 	return 0;
694 }
695 
696 static int recvs_pending(struct ipoib_dev_priv *priv)
697 {
698 	int pending = 0;
699 	int i;
700 
701 	for (i = 0; i < ipoib_recvq_size; ++i)
702 		if (priv->rx_ring[i].mb)
703 			++pending;
704 
705 	return pending;
706 }
707 
708 void ipoib_drain_cq(struct ipoib_dev_priv *priv)
709 {
710 	int i, n;
711 
712 	spin_lock(&priv->drain_lock);
713 	do {
714 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
715 		for (i = 0; i < n; ++i) {
716 			/*
717 			 * Convert any successful completions to flush
718 			 * errors to avoid passing packets up the
719 			 * stack after bringing the device down.
720 			 */
721 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
722 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
723 
724 			if ((priv->ibwc[i].wr_id & IPOIB_OP_RECV) == 0)
725 				panic("ipoib_drain_cq:  Bad wrid 0x%jX\n",
726 				    (intmax_t)priv->ibwc[i].wr_id);
727 			if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
728 				ipoib_cm_handle_rx_wc(priv, priv->ibwc + i);
729 			else
730 				ipoib_ib_handle_rx_wc(priv, priv->ibwc + i);
731 		}
732 	} while (n == IPOIB_NUM_WC);
733 	spin_unlock(&priv->drain_lock);
734 
735 	spin_lock(&priv->lock);
736 	while (ipoib_poll_tx(priv))
737 		; /* nothing */
738 
739 	spin_unlock(&priv->lock);
740 }
741 
742 int ipoib_ib_dev_stop(struct ipoib_dev_priv *priv, int flush)
743 {
744 	struct ib_qp_attr qp_attr;
745 	unsigned long begin;
746 	struct ipoib_tx_buf *tx_req;
747 	int i;
748 
749 	ipoib_cm_dev_stop(priv);
750 
751 	/*
752 	 * Move our QP to the error state and then reinitialize in
753 	 * when all work requests have completed or have been flushed.
754 	 */
755 	qp_attr.qp_state = IB_QPS_ERR;
756 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
757 		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
758 
759 	/* Wait for all sends and receives to complete */
760 	begin = jiffies;
761 
762 	while (priv->tx_head != priv->tx_tail || recvs_pending(priv)) {
763 		if (time_after(jiffies, begin + 5 * HZ)) {
764 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
765 				   priv->tx_head - priv->tx_tail, recvs_pending(priv));
766 
767 			/*
768 			 * assume the HW is wedged and just free up
769 			 * all our pending work requests.
770 			 */
771 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
772 				tx_req = &priv->tx_ring[priv->tx_tail &
773 							(ipoib_sendq_size - 1)];
774 				ipoib_dma_unmap_tx(priv->ca, tx_req);
775 				m_freem(tx_req->mb);
776 				++priv->tx_tail;
777 				--priv->tx_outstanding;
778 			}
779 
780 			for (i = 0; i < ipoib_recvq_size; ++i) {
781 				struct ipoib_rx_buf *rx_req;
782 
783 				rx_req = &priv->rx_ring[i];
784 				if (!rx_req->mb)
785 					continue;
786 				ipoib_dma_unmap_rx(priv, &priv->rx_ring[i]);
787 				m_freem(rx_req->mb);
788 				rx_req->mb = NULL;
789 			}
790 
791 			goto timeout;
792 		}
793 
794 		ipoib_drain_cq(priv);
795 
796 		msleep(1);
797 	}
798 
799 	ipoib_dbg(priv, "All sends and receives done.\n");
800 
801 timeout:
802 	del_timer_sync(&priv->poll_timer);
803 	qp_attr.qp_state = IB_QPS_RESET;
804 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
805 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
806 
807 	/* Wait for all AHs to be reaped */
808 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
809 	cancel_delayed_work(&priv->ah_reap_task);
810 	if (flush)
811 		flush_workqueue(ipoib_workqueue);
812 
813 	ipoib_ah_dev_cleanup(priv);
814 
815 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
816 
817 	return 0;
818 }
819 
820 int ipoib_ib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
821 {
822 	struct ifnet *dev = priv->dev;
823 
824 	priv->ca = ca;
825 	priv->port = port;
826 	priv->qp = NULL;
827 
828 	if (ipoib_transport_dev_init(priv, ca)) {
829 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
830 		return -ENODEV;
831 	}
832 
833 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
834 		    (unsigned long) priv);
835 
836 	if (dev->if_flags & IFF_UP) {
837 		if (ipoib_ib_dev_open(priv)) {
838 			ipoib_transport_dev_cleanup(priv);
839 			return -ENODEV;
840 		}
841 	}
842 
843 	return 0;
844 }
845 
846 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
847 				enum ipoib_flush_level level)
848 {
849 	struct ipoib_dev_priv *cpriv;
850 	u16 new_index;
851 
852 	mutex_lock(&priv->vlan_mutex);
853 
854 	/*
855 	 * Flush any child interfaces too -- they might be up even if
856 	 * the parent is down.
857 	 */
858 	list_for_each_entry(cpriv, &priv->child_intfs, list)
859 		__ipoib_ib_dev_flush(cpriv, level);
860 
861 	mutex_unlock(&priv->vlan_mutex);
862 
863 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
864 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
865 		return;
866 	}
867 
868 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
869 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
870 		return;
871 	}
872 
873 	if (level == IPOIB_FLUSH_HEAVY) {
874 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
875 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
876 			ipoib_ib_dev_down(priv, 0);
877 			ipoib_ib_dev_stop(priv, 0);
878 			if (ipoib_pkey_dev_delay_open(priv))
879 				return;
880 		}
881 
882 		/* restart QP only if P_Key index is changed */
883 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
884 		    new_index == priv->pkey_index) {
885 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
886 			return;
887 		}
888 		priv->pkey_index = new_index;
889 	}
890 
891 	if (level == IPOIB_FLUSH_LIGHT) {
892 		ipoib_mark_paths_invalid(priv);
893 		ipoib_mcast_dev_flush(priv);
894 	}
895 
896 	if (level >= IPOIB_FLUSH_NORMAL)
897 		ipoib_ib_dev_down(priv, 0);
898 
899 	if (level == IPOIB_FLUSH_HEAVY) {
900 		ipoib_ib_dev_stop(priv, 0);
901 		ipoib_ib_dev_open(priv);
902 	}
903 
904 	/*
905 	 * The device could have been brought down between the start and when
906 	 * we get here, don't bring it back up if it's not configured up
907 	 */
908 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
909 		if (level >= IPOIB_FLUSH_NORMAL)
910 			ipoib_ib_dev_up(priv);
911 		ipoib_mcast_restart_task(&priv->restart_task);
912 	}
913 }
914 
915 void ipoib_ib_dev_flush_light(struct work_struct *work)
916 {
917 	struct ipoib_dev_priv *priv =
918 		container_of(work, struct ipoib_dev_priv, flush_light);
919 
920 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
921 }
922 
923 void ipoib_ib_dev_flush_normal(struct work_struct *work)
924 {
925 	struct ipoib_dev_priv *priv =
926 		container_of(work, struct ipoib_dev_priv, flush_normal);
927 
928 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
929 }
930 
931 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
932 {
933 	struct ipoib_dev_priv *priv =
934 		container_of(work, struct ipoib_dev_priv, flush_heavy);
935 
936 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
937 }
938 
939 void ipoib_ib_dev_cleanup(struct ipoib_dev_priv *priv)
940 {
941 
942 	ipoib_dbg(priv, "cleaning up ib_dev\n");
943 
944 	ipoib_mcast_stop_thread(priv, 1);
945 	ipoib_mcast_dev_flush(priv);
946 
947 	ipoib_ah_dev_cleanup(priv);
948 	ipoib_transport_dev_cleanup(priv);
949 }
950 
951 /*
952  * Delayed P_Key Assigment Interim Support
953  *
954  * The following is initial implementation of delayed P_Key assigment
955  * mechanism. It is using the same approach implemented for the multicast
956  * group join. The single goal of this implementation is to quickly address
957  * Bug #2507. This implementation will probably be removed when the P_Key
958  * change async notification is available.
959  */
960 
961 void ipoib_pkey_poll(struct work_struct *work)
962 {
963 	struct ipoib_dev_priv *priv =
964 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
965 
966 	ipoib_pkey_dev_check_presence(priv);
967 
968 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
969 		ipoib_open(priv);
970 	else {
971 		mutex_lock(&pkey_mutex);
972 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
973 			queue_delayed_work(ipoib_workqueue,
974 					   &priv->pkey_poll_task,
975 					   HZ);
976 		mutex_unlock(&pkey_mutex);
977 	}
978 }
979 
980 int ipoib_pkey_dev_delay_open(struct ipoib_dev_priv *priv)
981 {
982 
983 	/* Look for the interface pkey value in the IB Port P_Key table and */
984 	/* set the interface pkey assigment flag                            */
985 	ipoib_pkey_dev_check_presence(priv);
986 
987 	/* P_Key value not assigned yet - start polling */
988 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
989 		mutex_lock(&pkey_mutex);
990 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
991 		queue_delayed_work(ipoib_workqueue,
992 				   &priv->pkey_poll_task,
993 				   HZ);
994 		mutex_unlock(&pkey_mutex);
995 		return 1;
996 	}
997 
998 	return 0;
999 }
1000