xref: /freebsd/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_ib.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
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 #include "ipoib.h"
40 
41 #include <rdma/ib_cache.h>
42 
43 #include <security/mac/mac_framework.h>
44 
45 #include <linux/delay.h>
46 #include <linux/dma-mapping.h>
47 
48 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
49 static int data_debug_level;
50 
51 module_param(data_debug_level, int, 0644);
52 MODULE_PARM_DESC(data_debug_level,
53 		 "Enable data path debug tracing if > 0");
54 #endif
55 
56 static DEFINE_MUTEX(pkey_mutex);
57 
58 struct ipoib_ah *ipoib_create_ah(struct ipoib_dev_priv *priv,
59 				 struct ib_pd *pd, struct ib_ah_attr *attr)
60 {
61 	struct ipoib_ah *ah;
62 
63 	ah = kmalloc(sizeof *ah, GFP_KERNEL);
64 	if (!ah)
65 		return NULL;
66 
67 	ah->priv      = priv;
68 	ah->last_send = 0;
69 	kref_init(&ah->ref);
70 
71 	ah->ah = ib_create_ah(pd, attr, RDMA_CREATE_AH_SLEEPABLE);
72 	if (IS_ERR(ah->ah)) {
73 		kfree(ah);
74 		ah = NULL;
75 	} else
76 		ipoib_dbg(priv, "Created ah %p\n", ah->ah);
77 
78 	return ah;
79 }
80 
81 void ipoib_free_ah(struct kref *kref)
82 {
83 	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
84 	struct ipoib_dev_priv *priv = ah->priv;
85 
86 	unsigned long flags;
87 
88 	spin_lock_irqsave(&priv->lock, flags);
89 	list_add_tail(&ah->list, &priv->dead_ahs);
90 	spin_unlock_irqrestore(&priv->lock, flags);
91 }
92 
93 void
94 ipoib_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req)
95 {
96 	struct mbuf *m;
97 	int i;
98 
99 	for (i = 0, m = rx_req->mb; m != NULL; m = m->m_next, i++)
100 		ib_dma_unmap_single(priv->ca, rx_req->mapping[i], m->m_len,
101 		    DMA_FROM_DEVICE);
102 }
103 
104 void
105 ipoib_dma_mb(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int length)
106 {
107 
108 	m_adj(mb, -(mb->m_pkthdr.len - length));
109 }
110 
111 struct mbuf *
112 ipoib_alloc_map_mb(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req,
113     int align, int size, int max_frags)
114 {
115 	struct mbuf *mb, *m;
116 	int i, j;
117 
118 	rx_req->mb = NULL;
119 	mb = m_getm2(NULL, align + size, M_NOWAIT, MT_DATA, M_PKTHDR);
120 	if (mb == NULL)
121 		return (NULL);
122 	for (i = 0, m = mb; m != NULL; m = m->m_next, i++) {
123 		MPASS(i < max_frags);
124 
125 		m->m_len = M_SIZE(m) - align;
126 		m->m_data += align;
127 		align = 0;
128 		mb->m_pkthdr.len += m->m_len;
129 		rx_req->mapping[i] = ib_dma_map_single(priv->ca,
130 		    mtod(m, void *), m->m_len, DMA_FROM_DEVICE);
131 		if (unlikely(ib_dma_mapping_error(priv->ca,
132 		    rx_req->mapping[i])))
133 			goto error;
134 
135 	}
136 	rx_req->mb = mb;
137 	return (mb);
138 error:
139 	for (j = 0, m = mb; j < i; m = m->m_next, j++)
140 		ib_dma_unmap_single(priv->ca, rx_req->mapping[j], m->m_len,
141 		    DMA_FROM_DEVICE);
142 	m_freem(mb);
143 	return (NULL);
144 
145 }
146 
147 static int ipoib_ib_post_receive(struct ipoib_dev_priv *priv, int id)
148 {
149 	struct ipoib_rx_buf *rx_req;
150 	const struct ib_recv_wr *bad_wr;
151 	struct mbuf *m;
152 	int ret;
153 	int i;
154 
155 	rx_req = &priv->rx_ring[id];
156 	for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
157 		priv->rx_sge[i].addr = rx_req->mapping[i];
158 		priv->rx_sge[i].length = m->m_len;
159 	}
160 	priv->rx_wr.num_sge = i;
161 	priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
162 
163 	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
164 	if (unlikely(ret)) {
165 		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
166 		ipoib_dma_unmap_rx(priv, &priv->rx_ring[id]);
167 		m_freem(priv->rx_ring[id].mb);
168 		priv->rx_ring[id].mb = NULL;
169 	}
170 
171 	return ret;
172 }
173 
174 static struct mbuf *
175 ipoib_alloc_rx_mb(struct ipoib_dev_priv *priv, int id)
176 {
177 	return ipoib_alloc_map_mb(priv, &priv->rx_ring[id],
178 	    0, priv->max_ib_mtu + IB_GRH_BYTES, IPOIB_UD_RX_SG);
179 }
180 
181 static int ipoib_ib_post_receives(struct ipoib_dev_priv *priv)
182 {
183 	int i;
184 
185 	for (i = 0; i < ipoib_recvq_size; ++i) {
186 		if (!ipoib_alloc_rx_mb(priv, i)) {
187 			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
188 			return -ENOMEM;
189 		}
190 		if (ipoib_ib_post_receive(priv, i)) {
191 			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
192 			return -EIO;
193 		}
194 	}
195 
196 	return 0;
197 }
198 
199 static void
200 ipoib_ib_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
201 {
202 	struct ipoib_rx_buf saverx;
203 	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
204 	if_t dev = priv->dev;
205 	struct ipoib_header *eh;
206 	struct mbuf *mb;
207 
208 	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
209 		       wr_id, wc->status);
210 
211 	if (unlikely(wr_id >= ipoib_recvq_size)) {
212 		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
213 			   wr_id, ipoib_recvq_size);
214 		return;
215 	}
216 
217 	mb  = priv->rx_ring[wr_id].mb;
218 
219 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
220 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
221 			ipoib_warn(priv, "failed recv event "
222 				   "(status=%d, wrid=%d vend_err %x)\n",
223 				   wc->status, wr_id, wc->vendor_err);
224 			goto repost;
225 		}
226 		if (mb) {
227 			ipoib_dma_unmap_rx(priv, &priv->rx_ring[wr_id]);
228 			m_freem(mb);
229 			priv->rx_ring[wr_id].mb = NULL;
230 		}
231 		return;
232 	}
233 
234 	/*
235 	 * Drop packets that this interface sent, ie multicast packets
236 	 * that the HCA has replicated.
237 	 */
238 	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
239 		goto repost;
240 
241 	memcpy(&saverx, &priv->rx_ring[wr_id], sizeof(saverx));
242 	/*
243 	 * If we can't allocate a new RX buffer, dump
244 	 * this packet and reuse the old buffer.
245 	 */
246 	if (unlikely(!ipoib_alloc_rx_mb(priv, wr_id))) {
247 		memcpy(&priv->rx_ring[wr_id], &saverx, sizeof(saverx));
248 		if_inc_counter(dev, IFCOUNTER_IQDROPS, 1);
249 		goto repost;
250 	}
251 
252 	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
253 		       wc->byte_len, wc->slid);
254 
255 	ipoib_dma_unmap_rx(priv, &saverx);
256 	ipoib_dma_mb(priv, mb, wc->byte_len);
257 
258 	if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
259 	if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
260 	mb->m_pkthdr.rcvif = dev;
261 	m_adj(mb, sizeof(struct ib_grh) - INFINIBAND_ALEN);
262 	eh = mtod(mb, struct ipoib_header *);
263 	bzero(eh->hwaddr, 4);	/* Zero the queue pair, only dgid is in grh */
264 
265 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
266 		mb->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID;
267 
268 	if_input(dev, mb);
269 
270 repost:
271 	if (unlikely(ipoib_ib_post_receive(priv, wr_id)))
272 		ipoib_warn(priv, "ipoib_ib_post_receive failed "
273 			   "for buf %d\n", wr_id);
274 }
275 
276 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req, int max)
277 {
278 	struct mbuf *mb = tx_req->mb;
279 	u64 *mapping = tx_req->mapping;
280 	struct mbuf *m, *p;
281 	int error;
282 	int i;
283 
284 	for (m = mb, p = NULL, i = 0; m != NULL; p = m, m = m->m_next, i++) {
285 		if (m->m_len != 0)
286 			continue;
287 		if (p == NULL)
288 			panic("ipoib_dma_map_tx: First mbuf empty\n");
289 		p->m_next = m_free(m);
290 		m = p;
291 		i--;
292 	}
293 	i--;
294 	if (i >= max) {
295 		tx_req->mb = mb = m_defrag(mb, M_NOWAIT);
296 		if (mb == NULL)
297 			return -EIO;
298 		for (m = mb, i = 0; m != NULL; m = m->m_next, i++);
299 		if (i >= max)
300 			return -EIO;
301 	}
302 	error = 0;
303 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
304 		mapping[i] = ib_dma_map_single(ca, mtod(m, void *),
305 					       m->m_len, DMA_TO_DEVICE);
306 		if (unlikely(ib_dma_mapping_error(ca, mapping[i]))) {
307 			error = -EIO;
308 			break;
309 		}
310 	}
311 	if (error) {
312 		int end;
313 
314 		end = i;
315 		for (m = mb, i = 0; i < end; m = m->m_next, i++)
316 			ib_dma_unmap_single(ca, mapping[i], m->m_len,
317 					    DMA_TO_DEVICE);
318 	}
319 	return error;
320 }
321 
322 void ipoib_dma_unmap_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
323 {
324 	struct mbuf *mb = tx_req->mb;
325 	u64 *mapping = tx_req->mapping;
326 	struct mbuf *m;
327 	int i;
328 
329 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++)
330 		ib_dma_unmap_single(ca, mapping[i], m->m_len, DMA_TO_DEVICE);
331 }
332 
333 static void ipoib_ib_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
334 {
335 	if_t dev = priv->dev;
336 	unsigned int wr_id = wc->wr_id;
337 	struct ipoib_tx_buf *tx_req;
338 
339 	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
340 		       wr_id, wc->status);
341 
342 	if (unlikely(wr_id >= ipoib_sendq_size)) {
343 		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
344 			   wr_id, ipoib_sendq_size);
345 		return;
346 	}
347 
348 	tx_req = &priv->tx_ring[wr_id];
349 
350 	ipoib_dma_unmap_tx(priv->ca, tx_req);
351 
352 	if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
353 
354 	m_freem(tx_req->mb);
355 
356 	++priv->tx_tail;
357 	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
358 	    (if_getdrvflags(dev) & IFF_DRV_OACTIVE) &&
359 	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
360 		if_setdrvflagbits(dev, 0, IFF_DRV_OACTIVE);
361 
362 	if (wc->status != IB_WC_SUCCESS &&
363 	    wc->status != IB_WC_WR_FLUSH_ERR)
364 		ipoib_warn(priv, "failed send event "
365 			   "(status=%d, wrid=%d vend_err %x)\n",
366 			   wc->status, wr_id, wc->vendor_err);
367 }
368 
369 int
370 ipoib_poll_tx(struct ipoib_dev_priv *priv, bool do_start)
371 {
372 	int n, i;
373 
374 	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
375 	for (i = 0; i < n; ++i) {
376 		struct ib_wc *wc = priv->send_wc + i;
377 		if (wc->wr_id & IPOIB_OP_CM)
378 			ipoib_cm_handle_tx_wc(priv, wc);
379 		else
380 			ipoib_ib_handle_tx_wc(priv, wc);
381 	}
382 
383 	if (do_start && n != 0)
384 		ipoib_start_locked(priv->dev, priv);
385 
386 	return n == MAX_SEND_CQE;
387 }
388 
389 static void
390 ipoib_poll(struct ipoib_dev_priv *priv)
391 {
392 	int n, i;
393 
394 poll_more:
395 	spin_lock(&priv->drain_lock);
396 	for (;;) {
397 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
398 		for (i = 0; i < n; i++) {
399 			struct ib_wc *wc = priv->ibwc + i;
400 
401 			if ((wc->wr_id & IPOIB_OP_RECV) == 0)
402 				panic("ipoib_poll: Bad wr_id 0x%jX\n",
403 				    (intmax_t)wc->wr_id);
404 			if (wc->wr_id & IPOIB_OP_CM)
405 				ipoib_cm_handle_rx_wc(priv, wc);
406 			else
407 				ipoib_ib_handle_rx_wc(priv, wc);
408 		}
409 
410 		if (n != IPOIB_NUM_WC)
411 			break;
412 	}
413 	spin_unlock(&priv->drain_lock);
414 
415 	if (ib_req_notify_cq(priv->recv_cq,
416 	    IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS) > 0)
417 		goto poll_more;
418 }
419 
420 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
421 {
422 	struct ipoib_dev_priv *priv = dev_ptr;
423 
424 	ipoib_poll(priv);
425 }
426 
427 static void drain_tx_cq(struct ipoib_dev_priv *priv)
428 {
429 	if_t dev = priv->dev;
430 
431 	spin_lock(&priv->lock);
432 	while (ipoib_poll_tx(priv, true))
433 		; /* nothing */
434 
435 	if (if_getdrvflags(dev) & IFF_DRV_OACTIVE)
436 		mod_timer(&priv->poll_timer, jiffies + 1);
437 
438 	spin_unlock(&priv->lock);
439 }
440 
441 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
442 {
443 	struct ipoib_dev_priv *priv = dev_ptr;
444 
445 	mod_timer(&priv->poll_timer, jiffies);
446 }
447 
448 static inline int
449 post_send(struct ipoib_dev_priv *priv, unsigned int wr_id,
450     struct ib_ah *address, u32 qpn, struct ipoib_tx_buf *tx_req, void *head,
451     int hlen)
452 {
453 	const struct ib_send_wr *bad_wr;
454 	struct mbuf *mb = tx_req->mb;
455 	u64 *mapping = tx_req->mapping;
456 	struct mbuf *m;
457 	int i;
458 
459 	for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
460 		priv->tx_sge[i].addr         = mapping[i];
461 		priv->tx_sge[i].length       = m->m_len;
462 	}
463 	priv->tx_wr.wr.num_sge	= i;
464 	priv->tx_wr.wr.wr_id	= wr_id;
465 	priv->tx_wr.remote_qpn	= qpn;
466 	priv->tx_wr.ah		= address;
467 
468 	if (head) {
469 		priv->tx_wr.mss		= 0; /* XXX mb_shinfo(mb)->gso_size; */
470 		priv->tx_wr.header	= head;
471 		priv->tx_wr.hlen	= hlen;
472 		priv->tx_wr.wr.opcode	= IB_WR_LSO;
473 	} else
474 		priv->tx_wr.wr.opcode	= IB_WR_SEND;
475 
476 	return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
477 }
478 
479 void
480 ipoib_send(struct ipoib_dev_priv *priv, struct mbuf *mb,
481     struct ipoib_ah *address, u32 qpn)
482 {
483 	if_t dev = priv->dev;
484 	struct ipoib_tx_buf *tx_req;
485 	int hlen;
486 	void *phead;
487 
488 	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
489 		while (ipoib_poll_tx(priv, false))
490 			; /* nothing */
491 
492 	m_adj(mb, sizeof (struct ipoib_pseudoheader));
493 	if (0 /* XXX segment offload mb_is_gso(mb) */) {
494 		/* XXX hlen = mb_transport_offset(mb) + tcp_hdrlen(mb); */
495 		phead = mtod(mb, void *);
496 		if (mb->m_len < hlen) {
497 			ipoib_warn(priv, "linear data too small\n");
498 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
499 			m_freem(mb);
500 			return;
501 		}
502 		m_adj(mb, hlen);
503 	} else {
504 		if (unlikely(mb->m_pkthdr.len - IPOIB_ENCAP_LEN > priv->mcast_mtu)) {
505 			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
506 				   mb->m_pkthdr.len, priv->mcast_mtu);
507 			if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
508 			ipoib_cm_mb_too_long(priv, mb, priv->mcast_mtu);
509 			return;
510 		}
511 		phead = NULL;
512 		hlen  = 0;
513 	}
514 
515 	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
516 		       mb->m_pkthdr.len, address, qpn);
517 
518 	/*
519 	 * We put the mb into the tx_ring _before_ we call post_send()
520 	 * because it's entirely possible that the completion handler will
521 	 * run before we execute anything after the post_send().  That
522 	 * means we have to make sure everything is properly recorded and
523 	 * our state is consistent before we call post_send().
524 	 */
525 	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
526 	tx_req->mb = mb;
527 	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req, IPOIB_UD_TX_SG))) {
528 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
529 		if (tx_req->mb)
530 			m_freem(tx_req->mb);
531 		return;
532 	}
533 
534 	if (mb->m_pkthdr.csum_flags & (CSUM_IP|CSUM_TCP|CSUM_UDP))
535 		priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
536 	else
537 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
538 
539 	if (++priv->tx_outstanding == ipoib_sendq_size) {
540 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
541 		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
542 			ipoib_warn(priv, "request notify on send CQ failed\n");
543 		if_setdrvflagbits(dev, IFF_DRV_OACTIVE, 0);
544 	}
545 
546 	if (unlikely(post_send(priv,
547 	    priv->tx_head & (ipoib_sendq_size - 1), address->ah, qpn,
548 	    tx_req, phead, hlen))) {
549 		ipoib_warn(priv, "post_send failed\n");
550 		if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
551 		--priv->tx_outstanding;
552 		ipoib_dma_unmap_tx(priv->ca, tx_req);
553 		m_freem(mb);
554 		if (if_getdrvflags(dev) & IFF_DRV_OACTIVE)
555 			if_setdrvflagbits(dev, 0, IFF_DRV_OACTIVE);
556 	} else {
557 		address->last_send = priv->tx_head;
558 		++priv->tx_head;
559 	}
560 }
561 
562 static void __ipoib_reap_ah(struct ipoib_dev_priv *priv)
563 {
564 	struct ipoib_ah *ah, *tah;
565 	LIST_HEAD(remove_list);
566 	unsigned long flags;
567 
568 	spin_lock_irqsave(&priv->lock, flags);
569 
570 	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
571 		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
572 			list_del(&ah->list);
573 			ib_destroy_ah(ah->ah, 0);
574 			kfree(ah);
575 		}
576 
577 	spin_unlock_irqrestore(&priv->lock, flags);
578 }
579 
580 void ipoib_reap_ah(struct work_struct *work)
581 {
582 	struct ipoib_dev_priv *priv =
583 		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
584 
585 	__ipoib_reap_ah(priv);
586 
587 	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
588 		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
589 				   HZ);
590 }
591 
592 static void ipoib_ah_dev_cleanup(struct ipoib_dev_priv *priv)
593 {
594 	unsigned long begin;
595 
596 	begin = jiffies;
597 
598 	while (!list_empty(&priv->dead_ahs)) {
599 		__ipoib_reap_ah(priv);
600 
601 		if (time_after(jiffies, begin + HZ)) {
602 			ipoib_warn(priv, "timing out; will leak address handles\n");
603 			break;
604 		}
605 
606 		msleep(1);
607 	}
608 }
609 
610 static void ipoib_ib_tx_timer_func(unsigned long ctx)
611 {
612 	drain_tx_cq((struct ipoib_dev_priv *)ctx);
613 }
614 
615 int ipoib_ib_dev_open(struct ipoib_dev_priv *priv)
616 {
617 	int ret;
618 
619 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
620 		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
621 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
622 		return -1;
623 	}
624 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
625 
626 	ret = ipoib_init_qp(priv);
627 	if (ret) {
628 		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
629 		return -1;
630 	}
631 
632 	ret = ipoib_ib_post_receives(priv);
633 	if (ret) {
634 		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
635 		ipoib_ib_dev_stop(priv, 1);
636 		return -1;
637 	}
638 
639 	ret = ipoib_cm_dev_open(priv);
640 	if (ret) {
641 		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
642 		ipoib_ib_dev_stop(priv, 1);
643 		return -1;
644 	}
645 
646 	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
647 	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
648 
649 	set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
650 
651 	return 0;
652 }
653 
654 static void ipoib_pkey_dev_check_presence(struct ipoib_dev_priv *priv)
655 {
656 	u16 pkey_index = 0;
657 
658 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
659 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
660 	else
661 		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
662 }
663 
664 int ipoib_ib_dev_up(struct ipoib_dev_priv *priv)
665 {
666 
667 	ipoib_pkey_dev_check_presence(priv);
668 
669 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
670 		ipoib_dbg(priv, "PKEY is not assigned.\n");
671 		return 0;
672 	}
673 
674 	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
675 
676 	return ipoib_mcast_start_thread(priv);
677 }
678 
679 int ipoib_ib_dev_down(struct ipoib_dev_priv *priv, int flush)
680 {
681 
682 	ipoib_dbg(priv, "downing ib_dev\n");
683 
684 	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
685 	if_link_state_change(priv->dev, LINK_STATE_DOWN);
686 
687 	/* Shutdown the P_Key thread if still active */
688 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
689 		mutex_lock(&pkey_mutex);
690 		set_bit(IPOIB_PKEY_STOP, &priv->flags);
691 		cancel_delayed_work(&priv->pkey_poll_task);
692 		mutex_unlock(&pkey_mutex);
693 		if (flush)
694 			flush_workqueue(ipoib_workqueue);
695 	}
696 
697 	ipoib_mcast_stop_thread(priv, flush);
698 	ipoib_mcast_dev_flush(priv);
699 
700 	ipoib_flush_paths(priv);
701 
702 	return 0;
703 }
704 
705 static int recvs_pending(struct ipoib_dev_priv *priv)
706 {
707 	int pending = 0;
708 	int i;
709 
710 	for (i = 0; i < ipoib_recvq_size; ++i)
711 		if (priv->rx_ring[i].mb)
712 			++pending;
713 
714 	return pending;
715 }
716 
717 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,
718 					struct ib_qp *qp,
719 					enum ib_qp_state new_state)
720 {
721 	struct ib_qp_attr qp_attr;
722 	struct ib_qp_init_attr query_init_attr;
723 	int ret;
724 
725 	ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);
726 	if (ret) {
727 		ipoib_warn(priv, "%s: Failed to query QP (%d)\n", __func__, ret);
728 		return;
729 	}
730 
731 	/* print according to the new-state and the previous state */
732 	if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET) {
733 		ipoib_dbg(priv, "Failed to modify QP %d->%d, acceptable\n",
734 			  qp_attr.qp_state, new_state);
735 	} else {
736 		ipoib_warn(priv, "Failed to modify QP %d->%d\n",
737 			   qp_attr.qp_state, new_state);
738 	}
739 }
740 
741 void ipoib_drain_cq(struct ipoib_dev_priv *priv)
742 {
743 	int i, n;
744 
745 	spin_lock(&priv->drain_lock);
746 	do {
747 		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
748 		for (i = 0; i < n; ++i) {
749 			/*
750 			 * Convert any successful completions to flush
751 			 * errors to avoid passing packets up the
752 			 * stack after bringing the device down.
753 			 */
754 			if (priv->ibwc[i].status == IB_WC_SUCCESS)
755 				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
756 
757 			if ((priv->ibwc[i].wr_id & IPOIB_OP_RECV) == 0)
758 				panic("ipoib_drain_cq:  Bad wrid 0x%jX\n",
759 				    (intmax_t)priv->ibwc[i].wr_id);
760 			if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
761 				ipoib_cm_handle_rx_wc(priv, priv->ibwc + i);
762 			else
763 				ipoib_ib_handle_rx_wc(priv, priv->ibwc + i);
764 		}
765 	} while (n == IPOIB_NUM_WC);
766 	spin_unlock(&priv->drain_lock);
767 
768 	spin_lock(&priv->lock);
769 	while (ipoib_poll_tx(priv, true))
770 		; /* nothing */
771 
772 	spin_unlock(&priv->lock);
773 }
774 
775 int ipoib_ib_dev_stop(struct ipoib_dev_priv *priv, int flush)
776 {
777 	struct ib_qp_attr qp_attr;
778 	unsigned long begin;
779 	struct ipoib_tx_buf *tx_req;
780 	int i;
781 
782 	clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
783 
784 	ipoib_cm_dev_stop(priv);
785 
786 	/*
787 	 * Move our QP to the error state and then reinitialize in
788 	 * when all work requests have completed or have been flushed.
789 	 */
790 	qp_attr.qp_state = IB_QPS_ERR;
791 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
792 		check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);
793 
794 	/* Wait for all sends and receives to complete */
795 	begin = jiffies;
796 
797 	while (priv->tx_head != priv->tx_tail || recvs_pending(priv)) {
798 		if (time_after(jiffies, begin + 5 * HZ)) {
799 			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
800 				   priv->tx_head - priv->tx_tail, recvs_pending(priv));
801 
802 			/*
803 			 * assume the HW is wedged and just free up
804 			 * all our pending work requests.
805 			 */
806 			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
807 				tx_req = &priv->tx_ring[priv->tx_tail &
808 							(ipoib_sendq_size - 1)];
809 				ipoib_dma_unmap_tx(priv->ca, tx_req);
810 				m_freem(tx_req->mb);
811 				++priv->tx_tail;
812 				--priv->tx_outstanding;
813 			}
814 
815 			for (i = 0; i < ipoib_recvq_size; ++i) {
816 				struct ipoib_rx_buf *rx_req;
817 
818 				rx_req = &priv->rx_ring[i];
819 				if (!rx_req->mb)
820 					continue;
821 				ipoib_dma_unmap_rx(priv, &priv->rx_ring[i]);
822 				m_freem(rx_req->mb);
823 				rx_req->mb = NULL;
824 			}
825 
826 			goto timeout;
827 		}
828 
829 		ipoib_drain_cq(priv);
830 
831 		msleep(1);
832 	}
833 
834 	ipoib_dbg(priv, "All sends and receives done.\n");
835 
836 timeout:
837 	del_timer_sync(&priv->poll_timer);
838 	qp_attr.qp_state = IB_QPS_RESET;
839 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
840 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
841 
842 	/* Wait for all AHs to be reaped */
843 	set_bit(IPOIB_STOP_REAPER, &priv->flags);
844 	cancel_delayed_work(&priv->ah_reap_task);
845 	if (flush)
846 		flush_workqueue(ipoib_workqueue);
847 
848 	ipoib_ah_dev_cleanup(priv);
849 
850 	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
851 
852 	return 0;
853 }
854 
855 int ipoib_ib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port)
856 {
857 	if_t dev = priv->dev;
858 
859 	priv->ca = ca;
860 	priv->port = port;
861 	priv->qp = NULL;
862 
863 	if (ipoib_transport_dev_init(priv, ca)) {
864 		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
865 		return -ENODEV;
866 	}
867 
868 	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
869 		    (unsigned long) priv);
870 
871 	if (if_getflags(dev) & IFF_UP) {
872 		if (ipoib_ib_dev_open(priv)) {
873 			ipoib_transport_dev_cleanup(priv);
874 			return -ENODEV;
875 		}
876 	}
877 
878 	return 0;
879 }
880 
881 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
882 				enum ipoib_flush_level level)
883 {
884 	struct ipoib_dev_priv *cpriv;
885 	u16 new_index;
886 
887 	mutex_lock(&priv->vlan_mutex);
888 
889 	/*
890 	 * Flush any child interfaces too -- they might be up even if
891 	 * the parent is down.
892 	 */
893 	list_for_each_entry(cpriv, &priv->child_intfs, list)
894 		__ipoib_ib_dev_flush(cpriv, level);
895 
896 	mutex_unlock(&priv->vlan_mutex);
897 
898 	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
899 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
900 		return;
901 	}
902 
903 	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
904 		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
905 		return;
906 	}
907 
908 	if (level == IPOIB_FLUSH_HEAVY) {
909 		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
910 			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
911 			ipoib_ib_dev_down(priv, 0);
912 			ipoib_ib_dev_stop(priv, 0);
913 			if (ipoib_pkey_dev_delay_open(priv))
914 				return;
915 		}
916 
917 		/* restart QP only if P_Key index is changed */
918 		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
919 		    new_index == priv->pkey_index) {
920 			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
921 			return;
922 		}
923 		priv->pkey_index = new_index;
924 	}
925 
926 	if (level == IPOIB_FLUSH_LIGHT) {
927 		ipoib_mark_paths_invalid(priv);
928 		ipoib_mcast_dev_flush(priv);
929 	}
930 
931 	if (level >= IPOIB_FLUSH_NORMAL)
932 		ipoib_ib_dev_down(priv, 0);
933 
934 	if (level == IPOIB_FLUSH_HEAVY) {
935 		ipoib_ib_dev_stop(priv, 0);
936 		ipoib_ib_dev_open(priv);
937 	}
938 
939 	/*
940 	 * The device could have been brought down between the start and when
941 	 * we get here, don't bring it back up if it's not configured up
942 	 */
943 	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
944 		if (level >= IPOIB_FLUSH_NORMAL)
945 			ipoib_ib_dev_up(priv);
946 		ipoib_mcast_restart_task(&priv->restart_task);
947 	}
948 }
949 
950 void ipoib_ib_dev_flush_light(struct work_struct *work)
951 {
952 	struct ipoib_dev_priv *priv =
953 		container_of(work, struct ipoib_dev_priv, flush_light);
954 
955 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
956 }
957 
958 void ipoib_ib_dev_flush_normal(struct work_struct *work)
959 {
960 	struct ipoib_dev_priv *priv =
961 		container_of(work, struct ipoib_dev_priv, flush_normal);
962 
963 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
964 }
965 
966 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
967 {
968 	struct ipoib_dev_priv *priv =
969 		container_of(work, struct ipoib_dev_priv, flush_heavy);
970 
971 	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
972 }
973 
974 void ipoib_ib_dev_cleanup(struct ipoib_dev_priv *priv)
975 {
976 
977 	ipoib_dbg(priv, "cleaning up ib_dev\n");
978 
979 	ipoib_mcast_stop_thread(priv, 1);
980 	ipoib_mcast_dev_flush(priv);
981 
982 	ipoib_ah_dev_cleanup(priv);
983 	ipoib_transport_dev_cleanup(priv);
984 }
985 
986 /*
987  * Delayed P_Key Assigment Interim Support
988  *
989  * The following is initial implementation of delayed P_Key assigment
990  * mechanism. It is using the same approach implemented for the multicast
991  * group join. The single goal of this implementation is to quickly address
992  * Bug #2507. This implementation will probably be removed when the P_Key
993  * change async notification is available.
994  */
995 
996 void ipoib_pkey_poll(struct work_struct *work)
997 {
998 	struct ipoib_dev_priv *priv =
999 		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1000 
1001 	ipoib_pkey_dev_check_presence(priv);
1002 
1003 	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1004 		ipoib_open(priv);
1005 	else {
1006 		mutex_lock(&pkey_mutex);
1007 		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1008 			queue_delayed_work(ipoib_workqueue,
1009 					   &priv->pkey_poll_task,
1010 					   HZ);
1011 		mutex_unlock(&pkey_mutex);
1012 	}
1013 }
1014 
1015 int ipoib_pkey_dev_delay_open(struct ipoib_dev_priv *priv)
1016 {
1017 
1018 	/* Look for the interface pkey value in the IB Port P_Key table and */
1019 	/* set the interface pkey assigment flag                            */
1020 	ipoib_pkey_dev_check_presence(priv);
1021 
1022 	/* P_Key value not assigned yet - start polling */
1023 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1024 		mutex_lock(&pkey_mutex);
1025 		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1026 		queue_delayed_work(ipoib_workqueue,
1027 				   &priv->pkey_poll_task,
1028 				   HZ);
1029 		mutex_unlock(&pkey_mutex);
1030 		return 1;
1031 	}
1032 
1033 	return 0;
1034 }
1035