xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision 7648bc9fee8dec6cb3c4941e0165a930fbe8dcb0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
3  *
4  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
6  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
7  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
8  *
9  * This software is available to you under a choice of one of two
10  * licenses.  You may choose to be licensed under the terms of the GNU
11  * General Public License (GPL) Version 2, available from the file
12  * COPYING in the main directory of this source tree, or the
13  * OpenIB.org BSD license below:
14  *
15  *     Redistribution and use in source and binary forms, with or
16  *     without modification, are permitted provided that the following
17  *     conditions are met:
18  *
19  *      - Redistributions of source code must retain the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer.
22  *
23  *      - Redistributions in binary form must reproduce the above
24  *        copyright notice, this list of conditions and the following
25  *        disclaimer in the documentation and/or other materials
26  *        provided with the distribution.
27  *
28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
29  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
30  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
31  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
32  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
33  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
34  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35  * SOFTWARE.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "ipoib.h"
42 
43 #include <rdma/ib_cache.h>
44 
45 #include <security/mac/mac_framework.h>
46 
47 #include <linux/delay.h>
48 #include <linux/dma-mapping.h>
49 
50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
51 static int data_debug_level;
52 
53 module_param(data_debug_level, int, 0644);
54 MODULE_PARM_DESC(data_debug_level,
55 		 "Enable data path debug tracing if > 0");
56 #endif
57 
58 static DEFINE_MUTEX(pkey_mutex);
59 
60 struct ipoib_ah *ipoib_create_ah(struct ipoib_dev_priv *priv,
61 				 struct ib_pd *pd, struct ib_ah_attr *attr)
62 {
63 	struct ipoib_ah *ah;
64 
65 	ah = kmalloc(sizeof *ah, GFP_KERNEL);
66 	if (!ah)
67 		return NULL;
68 
69 	ah->priv      = priv;
70 	ah->last_send = 0;
71 	kref_init(&ah->ref);
72 
73 	ah->ah = ib_create_ah(pd, attr);
74 	if (IS_ERR(ah->ah)) {
75 		kfree(ah);
76 		ah = NULL;
77 	} else
78 		ipoib_dbg(priv, "Created ah %p\n", ah->ah);
79 
80 	return ah;
81 }
82 
83 void ipoib_free_ah(struct kref *kref)
84 {
85 	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
86 	struct ipoib_dev_priv *priv = ah->priv;
87 
88 	unsigned long flags;
89 
90 	spin_lock_irqsave(&priv->lock, flags);
91 	list_add_tail(&ah->list, &priv->dead_ahs);
92 	spin_unlock_irqrestore(&priv->lock, flags);
93 }
94 
95 void
96 ipoib_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req)
97 {
98 	struct mbuf *m;
99 	int i;
100 
101 	for (i = 0, m = rx_req->mb; m != NULL; m = m->m_next, i++)
102 		ib_dma_unmap_single(priv->ca, rx_req->mapping[i], m->m_len,
103 		    DMA_FROM_DEVICE);
104 }
105 
106 void
107 ipoib_dma_mb(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int length)
108 {
109 
110 	m_adj(mb, -(mb->m_pkthdr.len - length));
111 }
112 
113 struct mbuf *
114 ipoib_alloc_map_mb(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req,
115     int size)
116 {
117 	struct mbuf *mb, *m;
118 	int i, j;
119 
120 	rx_req->mb = NULL;
121 	mb = m_getm2(NULL, size, M_NOWAIT, MT_DATA, M_PKTHDR);
122 	if (mb == NULL)
123 		return (NULL);
124 	for (i = 0, m = mb; m != NULL; m = m->m_next, i++) {
125 		m->m_len = M_SIZE(m);
126 		mb->m_pkthdr.len += m->m_len;
127 		rx_req->mapping[i] = ib_dma_map_single(priv->ca,
128 		    mtod(m, void *), m->m_len, DMA_FROM_DEVICE);
129 		if (unlikely(ib_dma_mapping_error(priv->ca,
130 		    rx_req->mapping[i])))
131 			goto error;
132 
133 	}
134 	rx_req->mb = mb;
135 	return (mb);
136 error:
137 	for (j = 0, m = mb; j < i; m = m->m_next, j++)
138 		ib_dma_unmap_single(priv->ca, rx_req->mapping[j], m->m_len,
139 		    DMA_FROM_DEVICE);
140 	m_freem(mb);
141 	return (NULL);
142 
143 }
144 
145 static int ipoib_ib_post_receive(struct ipoib_dev_priv *priv, int id)
146 {
147 	struct ipoib_rx_buf *rx_req;
148 	struct ib_recv_wr *bad_wr;
149 	struct mbuf *m;
150 	int ret;
151 	int i;
152 
153 	rx_req = &priv->rx_ring[id];
154 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
155 		priv->rx_sge[i].addr = rx_req->mapping[i];
156 		priv->rx_sge[i].length = m->m_len;
157 	}
158 	priv->rx_wr.num_sge = i;
159 	priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
160 
161 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
162 	if (unlikely(ret)) {
163 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
164 		ipoib_dma_unmap_rx(priv, &priv->rx_ring[id]);
165 		m_freem(priv->rx_ring[id].mb);
166 		priv->rx_ring[id].mb = NULL;
167 	}
168 
169 	return ret;
170 }
171 
172 static struct mbuf *
173 ipoib_alloc_rx_mb(struct ipoib_dev_priv *priv, int id)
174 {
175 
176 	return ipoib_alloc_map_mb(priv, &priv->rx_ring[id],
177 	    priv->max_ib_mtu + IB_GRH_BYTES);
178 }
179 
180 static int ipoib_ib_post_receives(struct ipoib_dev_priv *priv)
181 {
182 	int i;
183 
184 	for (i = 0; i < ipoib_recvq_size; ++i) {
185 		if (!ipoib_alloc_rx_mb(priv, i)) {
186 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
187 			return -ENOMEM;
188 		}
189 		if (ipoib_ib_post_receive(priv, i)) {
190 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
191 			return -EIO;
192 		}
193 	}
194 
195 	return 0;
196 }
197 
198 static void
199 ipoib_ib_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
200 {
201 	struct ipoib_rx_buf saverx;
202 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
203 	struct ifnet *dev = priv->dev;
204 	struct ipoib_header *eh;
205 	struct mbuf *mb;
206 
207 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
208 		       wr_id, wc->status);
209 
210 	if (unlikely(wr_id >= ipoib_recvq_size)) {
211 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
212 			   wr_id, ipoib_recvq_size);
213 		return;
214 	}
215 
216 	mb  = priv->rx_ring[wr_id].mb;
217 
218 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
219 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
220 			ipoib_warn(priv, "failed recv event "
221 				   "(status=%d, wrid=%d vend_err %x)\n",
222 				   wc->status, wr_id, wc->vendor_err);
223 			goto repost;
224 		}
225 		if (mb) {
226 			ipoib_dma_unmap_rx(priv, &priv->rx_ring[wr_id]);
227 			m_freem(mb);
228 			priv->rx_ring[wr_id].mb = NULL;
229 		}
230 		return;
231 	}
232 
233 	/*
234 	 * Drop packets that this interface sent, ie multicast packets
235 	 * that the HCA has replicated.
236 	 */
237 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
238 		goto repost;
239 
240 	memcpy(&saverx, &priv->rx_ring[wr_id], sizeof(saverx));
241 	/*
242 	 * If we can't allocate a new RX buffer, dump
243 	 * this packet and reuse the old buffer.
244 	 */
245 	if (unlikely(!ipoib_alloc_rx_mb(priv, wr_id))) {
246 		memcpy(&priv->rx_ring[wr_id], &saverx, sizeof(saverx));
247 		if_inc_counter(dev, IFCOUNTER_IQDROPS, 1);
248 		goto repost;
249 	}
250 
251 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
252 		       wc->byte_len, wc->slid);
253 
254 	ipoib_dma_unmap_rx(priv, &saverx);
255 	ipoib_dma_mb(priv, mb, wc->byte_len);
256 
257 	if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
258 	if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
259 	mb->m_pkthdr.rcvif = dev;
260 	m_adj(mb, sizeof(struct ib_grh) - INFINIBAND_ALEN);
261 	eh = mtod(mb, struct ipoib_header *);
262 	bzero(eh->hwaddr, 4);	/* Zero the queue pair, only dgid is in grh */
263 
264 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
265 		mb->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID;
266 
267 	dev->if_input(dev, mb);
268 
269 repost:
270 	if (unlikely(ipoib_ib_post_receive(priv, wr_id)))
271 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
272 			   "for buf %d\n", wr_id);
273 }
274 
275 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req, int max)
276 {
277 	struct mbuf *mb = tx_req->mb;
278 	u64 *mapping = tx_req->mapping;
279 	struct mbuf *m, *p;
280 	int error;
281 	int i;
282 
283 	for (m = mb, p = NULL, i = 0; m != NULL; p = m, m = m->m_next, i++) {
284 		if (m->m_len != 0)
285 			continue;
286 		if (p == NULL)
287 			panic("ipoib_dma_map_tx: First mbuf empty\n");
288 		p->m_next = m_free(m);
289 		m = p;
290 		i--;
291 	}
292 	i--;
293 	if (i >= max) {
294 		tx_req->mb = mb = m_defrag(mb, M_NOWAIT);
295 		if (mb == NULL)
296 			return -EIO;
297 		for (m = mb, i = 0; m != NULL; m = m->m_next, i++);
298 		if (i >= max)
299 			return -EIO;
300 	}
301 	error = 0;
302 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
303 		mapping[i] = ib_dma_map_single(ca, mtod(m, void *),
304 					       m->m_len, DMA_TO_DEVICE);
305 		if (unlikely(ib_dma_mapping_error(ca, mapping[i]))) {
306 			error = -EIO;
307 			break;
308 		}
309 	}
310 	if (error) {
311 		int end;
312 
313 		end = i;
314 		for (m = mb, i = 0; i < end; m = m->m_next, i++)
315 			ib_dma_unmap_single(ca, mapping[i], m->m_len,
316 					    DMA_TO_DEVICE);
317 	}
318 	return error;
319 }
320 
321 void ipoib_dma_unmap_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
322 {
323 	struct mbuf *mb = tx_req->mb;
324 	u64 *mapping = tx_req->mapping;
325 	struct mbuf *m;
326 	int i;
327 
328 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++)
329 		ib_dma_unmap_single(ca, mapping[i], m->m_len, DMA_TO_DEVICE);
330 }
331 
332 static void ipoib_ib_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
333 {
334 	struct ifnet *dev = priv->dev;
335 	unsigned int wr_id = wc->wr_id;
336 	struct ipoib_tx_buf *tx_req;
337 
338 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
339 		       wr_id, wc->status);
340 
341 	if (unlikely(wr_id >= ipoib_sendq_size)) {
342 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
343 			   wr_id, ipoib_sendq_size);
344 		return;
345 	}
346 
347 	tx_req = &priv->tx_ring[wr_id];
348 
349 	ipoib_dma_unmap_tx(priv->ca, tx_req);
350 
351 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
352 
353 	m_freem(tx_req->mb);
354 
355 	++priv->tx_tail;
356 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
357 	    (dev->if_drv_flags & IFF_DRV_OACTIVE) &&
358 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
359 		dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
360 
361 	if (wc->status != IB_WC_SUCCESS &&
362 	    wc->status != IB_WC_WR_FLUSH_ERR)
363 		ipoib_warn(priv, "failed send event "
364 			   "(status=%d, wrid=%d vend_err %x)\n",
365 			   wc->status, wr_id, wc->vendor_err);
366 }
367 
368 int
369 ipoib_poll_tx(struct ipoib_dev_priv *priv)
370 {
371 	int n, i;
372 
373 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
374 	for (i = 0; i < n; ++i) {
375 		struct ib_wc *wc = priv->send_wc + i;
376 		if (wc->wr_id & IPOIB_OP_CM)
377 			ipoib_cm_handle_tx_wc(priv, wc);
378 		else
379 			ipoib_ib_handle_tx_wc(priv, wc);
380 	}
381 
382 	return n == MAX_SEND_CQE;
383 }
384 
385 static void
386 ipoib_poll(struct ipoib_dev_priv *priv)
387 {
388 	int n, i;
389 
390 poll_more:
391 	spin_lock(&priv->drain_lock);
392 	for (;;) {
393 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
394 		for (i = 0; i < n; i++) {
395 			struct ib_wc *wc = priv->ibwc + i;
396 
397 			if ((wc->wr_id & IPOIB_OP_RECV) == 0)
398 				panic("ipoib_poll: Bad wr_id 0x%jX\n",
399 				    (intmax_t)wc->wr_id);
400 			if (wc->wr_id & IPOIB_OP_CM)
401 				ipoib_cm_handle_rx_wc(priv, wc);
402 			else
403 				ipoib_ib_handle_rx_wc(priv, wc);
404 		}
405 
406 		if (n != IPOIB_NUM_WC)
407 			break;
408 	}
409 	spin_unlock(&priv->drain_lock);
410 
411 	if (ib_req_notify_cq(priv->recv_cq,
412 	    IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS) > 0)
413 		goto poll_more;
414 }
415 
416 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
417 {
418 	struct ipoib_dev_priv *priv = dev_ptr;
419 
420 	ipoib_poll(priv);
421 }
422 
423 static void drain_tx_cq(struct ipoib_dev_priv *priv)
424 {
425 	struct ifnet *dev = priv->dev;
426 
427 	spin_lock(&priv->lock);
428 	while (ipoib_poll_tx(priv))
429 		; /* nothing */
430 
431 	if (dev->if_drv_flags & IFF_DRV_OACTIVE)
432 		mod_timer(&priv->poll_timer, jiffies + 1);
433 
434 	spin_unlock(&priv->lock);
435 }
436 
437 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
438 {
439 	struct ipoib_dev_priv *priv = dev_ptr;
440 
441 	mod_timer(&priv->poll_timer, jiffies);
442 }
443 
444 static inline int
445 post_send(struct ipoib_dev_priv *priv, unsigned int wr_id,
446     struct ib_ah *address, u32 qpn, struct ipoib_tx_buf *tx_req, void *head,
447     int hlen)
448 {
449 	struct ib_send_wr *bad_wr;
450 	struct mbuf *mb = tx_req->mb;
451 	u64 *mapping = tx_req->mapping;
452 	struct mbuf *m;
453 	int i;
454 
455 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
456 		priv->tx_sge[i].addr         = mapping[i];
457 		priv->tx_sge[i].length       = m->m_len;
458 	}
459 	priv->tx_wr.wr.num_sge	= i;
460 	priv->tx_wr.wr.wr_id	= wr_id;
461 	priv->tx_wr.remote_qpn	= qpn;
462 	priv->tx_wr.ah		= address;
463 
464 	if (head) {
465 		priv->tx_wr.mss		= 0; /* XXX mb_shinfo(mb)->gso_size; */
466 		priv->tx_wr.header	= head;
467 		priv->tx_wr.hlen	= hlen;
468 		priv->tx_wr.wr.opcode	= IB_WR_LSO;
469 	} else
470 		priv->tx_wr.wr.opcode	= IB_WR_SEND;
471 
472 	return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
473 }
474 
475 void
476 ipoib_send(struct ipoib_dev_priv *priv, struct mbuf *mb,
477     struct ipoib_ah *address, u32 qpn)
478 {
479 	struct ifnet *dev = priv->dev;
480 	struct ipoib_tx_buf *tx_req;
481 	int hlen;
482 	void *phead;
483 
484 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
485 		while (ipoib_poll_tx(priv))
486 			; /* nothing */
487 
488 	m_adj(mb, sizeof (struct ipoib_pseudoheader));
489 	if (0 /* XXX segment offload mb_is_gso(mb) */) {
490 		/* XXX hlen = mb_transport_offset(mb) + tcp_hdrlen(mb); */
491 		phead = mtod(mb, void *);
492 		if (mb->m_len < hlen) {
493 			ipoib_warn(priv, "linear data too small\n");
494 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
495 			m_freem(mb);
496 			return;
497 		}
498 		m_adj(mb, hlen);
499 	} else {
500 		if (unlikely(mb->m_pkthdr.len - IPOIB_ENCAP_LEN > priv->mcast_mtu)) {
501 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
502 				   mb->m_pkthdr.len, priv->mcast_mtu);
503 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
504 			ipoib_cm_mb_too_long(priv, mb, priv->mcast_mtu);
505 			return;
506 		}
507 		phead = NULL;
508 		hlen  = 0;
509 	}
510 
511 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
512 		       mb->m_pkthdr.len, address, qpn);
513 
514 	/*
515 	 * We put the mb into the tx_ring _before_ we call post_send()
516 	 * because it's entirely possible that the completion handler will
517 	 * run before we execute anything after the post_send().  That
518 	 * means we have to make sure everything is properly recorded and
519 	 * our state is consistent before we call post_send().
520 	 */
521 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
522 	tx_req->mb = mb;
523 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req, IPOIB_UD_TX_SG))) {
524 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
525 		if (tx_req->mb)
526 			m_freem(tx_req->mb);
527 		return;
528 	}
529 
530 	if (mb->m_pkthdr.csum_flags & (CSUM_IP|CSUM_TCP|CSUM_UDP))
531 		priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
532 	else
533 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
534 
535 	if (++priv->tx_outstanding == ipoib_sendq_size) {
536 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
537 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
538 			ipoib_warn(priv, "request notify on send CQ failed\n");
539 		dev->if_drv_flags |= IFF_DRV_OACTIVE;
540 	}
541 
542 	if (unlikely(post_send(priv,
543 	    priv->tx_head & (ipoib_sendq_size - 1), address->ah, qpn,
544 	    tx_req, phead, hlen))) {
545 		ipoib_warn(priv, "post_send failed\n");
546 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
547 		--priv->tx_outstanding;
548 		ipoib_dma_unmap_tx(priv->ca, tx_req);
549 		m_freem(mb);
550 		if (dev->if_drv_flags & IFF_DRV_OACTIVE)
551 			dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
552 	} else {
553 		address->last_send = priv->tx_head;
554 		++priv->tx_head;
555 	}
556 }
557 
558 static void __ipoib_reap_ah(struct ipoib_dev_priv *priv)
559 {
560 	struct ipoib_ah *ah, *tah;
561 	LIST_HEAD(remove_list);
562 	unsigned long flags;
563 
564 	spin_lock_irqsave(&priv->lock, flags);
565 
566 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
567 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
568 			list_del(&ah->list);
569 			ib_destroy_ah(ah->ah);
570 			kfree(ah);
571 		}
572 
573 	spin_unlock_irqrestore(&priv->lock, flags);
574 }
575 
576 void ipoib_reap_ah(struct work_struct *work)
577 {
578 	struct ipoib_dev_priv *priv =
579 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
580 
581 	__ipoib_reap_ah(priv);
582 
583 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
584 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
585 				   HZ);
586 }
587 
588 static void ipoib_ah_dev_cleanup(struct ipoib_dev_priv *priv)
589 {
590 	unsigned long begin;
591 
592 	begin = jiffies;
593 
594 	while (!list_empty(&priv->dead_ahs)) {
595 		__ipoib_reap_ah(priv);
596 
597 		if (time_after(jiffies, begin + HZ)) {
598 			ipoib_warn(priv, "timing out; will leak address handles\n");
599 			break;
600 		}
601 
602 		msleep(1);
603 	}
604 }
605 
606 static void ipoib_ib_tx_timer_func(unsigned long ctx)
607 {
608 	drain_tx_cq((struct ipoib_dev_priv *)ctx);
609 }
610 
611 int ipoib_ib_dev_open(struct ipoib_dev_priv *priv)
612 {
613 	int ret;
614 
615 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
616 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
617 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
618 		return -1;
619 	}
620 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
621 
622 	ret = ipoib_init_qp(priv);
623 	if (ret) {
624 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
625 		return -1;
626 	}
627 
628 	ret = ipoib_ib_post_receives(priv);
629 	if (ret) {
630 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
631 		ipoib_ib_dev_stop(priv, 1);
632 		return -1;
633 	}
634 
635 	ret = ipoib_cm_dev_open(priv);
636 	if (ret) {
637 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
638 		ipoib_ib_dev_stop(priv, 1);
639 		return -1;
640 	}
641 
642 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
643 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
644 
645 	set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
646 
647 	return 0;
648 }
649 
650 static void ipoib_pkey_dev_check_presence(struct ipoib_dev_priv *priv)
651 {
652 	u16 pkey_index = 0;
653 
654 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
655 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
656 	else
657 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
658 }
659 
660 int ipoib_ib_dev_up(struct ipoib_dev_priv *priv)
661 {
662 
663 	ipoib_pkey_dev_check_presence(priv);
664 
665 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
666 		ipoib_dbg(priv, "PKEY is not assigned.\n");
667 		return 0;
668 	}
669 
670 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
671 
672 	return ipoib_mcast_start_thread(priv);
673 }
674 
675 int ipoib_ib_dev_down(struct ipoib_dev_priv *priv, int flush)
676 {
677 
678 	ipoib_dbg(priv, "downing ib_dev\n");
679 
680 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
681 	if_link_state_change(priv->dev, LINK_STATE_DOWN);
682 
683 	/* Shutdown the P_Key thread if still active */
684 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
685 		mutex_lock(&pkey_mutex);
686 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
687 		cancel_delayed_work(&priv->pkey_poll_task);
688 		mutex_unlock(&pkey_mutex);
689 		if (flush)
690 			flush_workqueue(ipoib_workqueue);
691 	}
692 
693 	ipoib_mcast_stop_thread(priv, flush);
694 	ipoib_mcast_dev_flush(priv);
695 
696 	ipoib_flush_paths(priv);
697 
698 	return 0;
699 }
700 
701 static int recvs_pending(struct ipoib_dev_priv *priv)
702 {
703 	int pending = 0;
704 	int i;
705 
706 	for (i = 0; i < ipoib_recvq_size; ++i)
707 		if (priv->rx_ring[i].mb)
708 			++pending;
709 
710 	return pending;
711 }
712 
713 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,
714 					struct ib_qp *qp,
715 					enum ib_qp_state new_state)
716 {
717 	struct ib_qp_attr qp_attr;
718 	struct ib_qp_init_attr query_init_attr;
719 	int ret;
720 
721 	ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);
722 	if (ret) {
723 		ipoib_warn(priv, "%s: Failed to query QP (%d)\n", __func__, ret);
724 		return;
725 	}
726 
727 	/* print according to the new-state and the previous state */
728 	if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET) {
729 		ipoib_dbg(priv, "Failed to modify QP %d->%d, acceptable\n",
730 			  qp_attr.qp_state, new_state);
731 	} else {
732 		ipoib_warn(priv, "Failed to modify QP %d->%d\n",
733 			   qp_attr.qp_state, new_state);
734 	}
735 }
736 
737 void ipoib_drain_cq(struct ipoib_dev_priv *priv)
738 {
739 	int i, n;
740 
741 	spin_lock(&priv->drain_lock);
742 	do {
743 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
744 		for (i = 0; i < n; ++i) {
745 			/*
746 			 * Convert any successful completions to flush
747 			 * errors to avoid passing packets up the
748 			 * stack after bringing the device down.
749 			 */
750 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
751 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
752 
753 			if ((priv->ibwc[i].wr_id & IPOIB_OP_RECV) == 0)
754 				panic("ipoib_drain_cq:  Bad wrid 0x%jX\n",
755 				    (intmax_t)priv->ibwc[i].wr_id);
756 			if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
757 				ipoib_cm_handle_rx_wc(priv, priv->ibwc + i);
758 			else
759 				ipoib_ib_handle_rx_wc(priv, priv->ibwc + i);
760 		}
761 	} while (n == IPOIB_NUM_WC);
762 	spin_unlock(&priv->drain_lock);
763 
764 	spin_lock(&priv->lock);
765 	while (ipoib_poll_tx(priv))
766 		; /* nothing */
767 
768 	spin_unlock(&priv->lock);
769 }
770 
771 int ipoib_ib_dev_stop(struct ipoib_dev_priv *priv, int flush)
772 {
773 	struct ib_qp_attr qp_attr;
774 	unsigned long begin;
775 	struct ipoib_tx_buf *tx_req;
776 	int i;
777 
778 	clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
779 
780 	ipoib_cm_dev_stop(priv);
781 
782 	/*
783 	 * Move our QP to the error state and then reinitialize in
784 	 * when all work requests have completed or have been flushed.
785 	 */
786 	qp_attr.qp_state = IB_QPS_ERR;
787 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
788 		check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);
789 
790 	/* Wait for all sends and receives to complete */
791 	begin = jiffies;
792 
793 	while (priv->tx_head != priv->tx_tail || recvs_pending(priv)) {
794 		if (time_after(jiffies, begin + 5 * HZ)) {
795 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
796 				   priv->tx_head - priv->tx_tail, recvs_pending(priv));
797 
798 			/*
799 			 * assume the HW is wedged and just free up
800 			 * all our pending work requests.
801 			 */
802 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
803 				tx_req = &priv->tx_ring[priv->tx_tail &
804 							(ipoib_sendq_size - 1)];
805 				ipoib_dma_unmap_tx(priv->ca, tx_req);
806 				m_freem(tx_req->mb);
807 				++priv->tx_tail;
808 				--priv->tx_outstanding;
809 			}
810 
811 			for (i = 0; i < ipoib_recvq_size; ++i) {
812 				struct ipoib_rx_buf *rx_req;
813 
814 				rx_req = &priv->rx_ring[i];
815 				if (!rx_req->mb)
816 					continue;
817 				ipoib_dma_unmap_rx(priv, &priv->rx_ring[i]);
818 				m_freem(rx_req->mb);
819 				rx_req->mb = NULL;
820 			}
821 
822 			goto timeout;
823 		}
824 
825 		ipoib_drain_cq(priv);
826 
827 		msleep(1);
828 	}
829 
830 	ipoib_dbg(priv, "All sends and receives done.\n");
831 
832 timeout:
833 	del_timer_sync(&priv->poll_timer);
834 	qp_attr.qp_state = IB_QPS_RESET;
835 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
836 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
837 
838 	/* Wait for all AHs to be reaped */
839 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
840 	cancel_delayed_work(&priv->ah_reap_task);
841 	if (flush)
842 		flush_workqueue(ipoib_workqueue);
843 
844 	ipoib_ah_dev_cleanup(priv);
845 
846 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
847 
848 	return 0;
849 }
850 
851 int ipoib_ib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
852 {
853 	struct ifnet *dev = priv->dev;
854 
855 	priv->ca = ca;
856 	priv->port = port;
857 	priv->qp = NULL;
858 
859 	if (ipoib_transport_dev_init(priv, ca)) {
860 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
861 		return -ENODEV;
862 	}
863 
864 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
865 		    (unsigned long) priv);
866 
867 	if (dev->if_flags & IFF_UP) {
868 		if (ipoib_ib_dev_open(priv)) {
869 			ipoib_transport_dev_cleanup(priv);
870 			return -ENODEV;
871 		}
872 	}
873 
874 	return 0;
875 }
876 
877 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
878 				enum ipoib_flush_level level)
879 {
880 	struct ipoib_dev_priv *cpriv;
881 	u16 new_index;
882 
883 	mutex_lock(&priv->vlan_mutex);
884 
885 	/*
886 	 * Flush any child interfaces too -- they might be up even if
887 	 * the parent is down.
888 	 */
889 	list_for_each_entry(cpriv, &priv->child_intfs, list)
890 		__ipoib_ib_dev_flush(cpriv, level);
891 
892 	mutex_unlock(&priv->vlan_mutex);
893 
894 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
895 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
896 		return;
897 	}
898 
899 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
900 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
901 		return;
902 	}
903 
904 	if (level == IPOIB_FLUSH_HEAVY) {
905 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
906 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
907 			ipoib_ib_dev_down(priv, 0);
908 			ipoib_ib_dev_stop(priv, 0);
909 			if (ipoib_pkey_dev_delay_open(priv))
910 				return;
911 		}
912 
913 		/* restart QP only if P_Key index is changed */
914 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
915 		    new_index == priv->pkey_index) {
916 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
917 			return;
918 		}
919 		priv->pkey_index = new_index;
920 	}
921 
922 	if (level == IPOIB_FLUSH_LIGHT) {
923 		ipoib_mark_paths_invalid(priv);
924 		ipoib_mcast_dev_flush(priv);
925 	}
926 
927 	if (level >= IPOIB_FLUSH_NORMAL)
928 		ipoib_ib_dev_down(priv, 0);
929 
930 	if (level == IPOIB_FLUSH_HEAVY) {
931 		ipoib_ib_dev_stop(priv, 0);
932 		ipoib_ib_dev_open(priv);
933 	}
934 
935 	/*
936 	 * The device could have been brought down between the start and when
937 	 * we get here, don't bring it back up if it's not configured up
938 	 */
939 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
940 		if (level >= IPOIB_FLUSH_NORMAL)
941 			ipoib_ib_dev_up(priv);
942 		ipoib_mcast_restart_task(&priv->restart_task);
943 	}
944 }
945 
946 void ipoib_ib_dev_flush_light(struct work_struct *work)
947 {
948 	struct ipoib_dev_priv *priv =
949 		container_of(work, struct ipoib_dev_priv, flush_light);
950 
951 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
952 }
953 
954 void ipoib_ib_dev_flush_normal(struct work_struct *work)
955 {
956 	struct ipoib_dev_priv *priv =
957 		container_of(work, struct ipoib_dev_priv, flush_normal);
958 
959 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
960 }
961 
962 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
963 {
964 	struct ipoib_dev_priv *priv =
965 		container_of(work, struct ipoib_dev_priv, flush_heavy);
966 
967 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
968 }
969 
970 void ipoib_ib_dev_cleanup(struct ipoib_dev_priv *priv)
971 {
972 
973 	ipoib_dbg(priv, "cleaning up ib_dev\n");
974 
975 	ipoib_mcast_stop_thread(priv, 1);
976 	ipoib_mcast_dev_flush(priv);
977 
978 	ipoib_ah_dev_cleanup(priv);
979 	ipoib_transport_dev_cleanup(priv);
980 }
981 
982 /*
983  * Delayed P_Key Assigment Interim Support
984  *
985  * The following is initial implementation of delayed P_Key assigment
986  * mechanism. It is using the same approach implemented for the multicast
987  * group join. The single goal of this implementation is to quickly address
988  * Bug #2507. This implementation will probably be removed when the P_Key
989  * change async notification is available.
990  */
991 
992 void ipoib_pkey_poll(struct work_struct *work)
993 {
994 	struct ipoib_dev_priv *priv =
995 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
996 
997 	ipoib_pkey_dev_check_presence(priv);
998 
999 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1000 		ipoib_open(priv);
1001 	else {
1002 		mutex_lock(&pkey_mutex);
1003 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1004 			queue_delayed_work(ipoib_workqueue,
1005 					   &priv->pkey_poll_task,
1006 					   HZ);
1007 		mutex_unlock(&pkey_mutex);
1008 	}
1009 }
1010 
1011 int ipoib_pkey_dev_delay_open(struct ipoib_dev_priv *priv)
1012 {
1013 
1014 	/* Look for the interface pkey value in the IB Port P_Key table and */
1015 	/* set the interface pkey assigment flag                            */
1016 	ipoib_pkey_dev_check_presence(priv);
1017 
1018 	/* P_Key value not assigned yet - start polling */
1019 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1020 		mutex_lock(&pkey_mutex);
1021 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1022 		queue_delayed_work(ipoib_workqueue,
1023 				   &priv->pkey_poll_task,
1024 				   HZ);
1025 		mutex_unlock(&pkey_mutex);
1026 		return 1;
1027 	}
1028 
1029 	return 0;
1030 }
1031