1/*
2 * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
6 *
7 * This software is available to you under a choice of one of two
8 * licenses.  You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
12 *
13 *     Redistribution and use in source and binary forms, with or
14 *     without modification, are permitted provided that the following
15 *     conditions are met:
16 *
17 *      - Redistributions of source code must retain the above
18 *        copyright notice, this list of conditions and the following
19 *        disclaimer.
20 *
21 *      - Redistributions in binary form must reproduce the above
22 *        copyright notice, this list of conditions and the following
23 *        disclaimer in the documentation and/or other materials
24 *        provided with the distribution.
25 *
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 * SOFTWARE.
34 */
35
36#include <linux/delay.h>
37#include <linux/moduleparam.h>
38#include <linux/dma-mapping.h>
39#include <linux/slab.h>
40
41#include <linux/ip.h>
42#include <linux/tcp.h>
43
44#include "ipoib.h"
45
46#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
47static int data_debug_level;
48
49module_param(data_debug_level, int, 0644);
50MODULE_PARM_DESC(data_debug_level,
51		 "Enable data path debug tracing if > 0");
52#endif
53
54static DEFINE_MUTEX(pkey_mutex);
55
56struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
57				 struct ib_pd *pd, struct ib_ah_attr *attr)
58{
59	struct ipoib_ah *ah;
60	struct ib_ah *vah;
61
62	ah = kmalloc(sizeof *ah, GFP_KERNEL);
63	if (!ah)
64		return ERR_PTR(-ENOMEM);
65
66	ah->dev       = dev;
67	ah->last_send = 0;
68	kref_init(&ah->ref);
69
70	vah = ib_create_ah(pd, attr);
71	if (IS_ERR(vah)) {
72		kfree(ah);
73		ah = (struct ipoib_ah *)vah;
74	} else {
75		ah->ah = vah;
76		ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
77	}
78
79	return ah;
80}
81
82void ipoib_free_ah(struct kref *kref)
83{
84	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
85	struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
86
87	unsigned long flags;
88
89	spin_lock_irqsave(&priv->lock, flags);
90	list_add_tail(&ah->list, &priv->dead_ahs);
91	spin_unlock_irqrestore(&priv->lock, flags);
92}
93
94static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
95				  u64 mapping[IPOIB_UD_RX_SG])
96{
97	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
98		ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
99				    DMA_FROM_DEVICE);
100		ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
101				  DMA_FROM_DEVICE);
102	} else
103		ib_dma_unmap_single(priv->ca, mapping[0],
104				    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
105				    DMA_FROM_DEVICE);
106}
107
108static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
109				   struct sk_buff *skb,
110				   unsigned int length)
111{
112	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
113		skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
114		unsigned int size;
115		/*
116		 * There is only two buffers needed for max_payload = 4K,
117		 * first buf size is IPOIB_UD_HEAD_SIZE
118		 */
119		skb->tail += IPOIB_UD_HEAD_SIZE;
120		skb->len  += length;
121
122		size = length - IPOIB_UD_HEAD_SIZE;
123
124		skb_frag_size_set(frag, size);
125		skb->data_len += size;
126		skb->truesize += size;
127	} else
128		skb_put(skb, length);
129
130}
131
132static int ipoib_ib_post_receive(struct net_device *dev, int id)
133{
134	struct ipoib_dev_priv *priv = netdev_priv(dev);
135	struct ib_recv_wr *bad_wr;
136	int ret;
137
138	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
139	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
140	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
141
142
143	ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
144	if (unlikely(ret)) {
145		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
146		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
147		dev_kfree_skb_any(priv->rx_ring[id].skb);
148		priv->rx_ring[id].skb = NULL;
149	}
150
151	return ret;
152}
153
154static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
155{
156	struct ipoib_dev_priv *priv = netdev_priv(dev);
157	struct sk_buff *skb;
158	int buf_size;
159	u64 *mapping;
160
161	if (ipoib_ud_need_sg(priv->max_ib_mtu))
162		buf_size = IPOIB_UD_HEAD_SIZE;
163	else
164		buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
165
166	skb = dev_alloc_skb(buf_size + 4);
167	if (unlikely(!skb))
168		return NULL;
169
170	/*
171	 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
172	 * header.  So we need 4 more bytes to get to 48 and align the
173	 * IP header to a multiple of 16.
174	 */
175	skb_reserve(skb, 4);
176
177	mapping = priv->rx_ring[id].mapping;
178	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
179				       DMA_FROM_DEVICE);
180	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
181		goto error;
182
183	if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
184		struct page *page = alloc_page(GFP_ATOMIC);
185		if (!page)
186			goto partial_error;
187		skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
188		mapping[1] =
189			ib_dma_map_page(priv->ca, page,
190					0, PAGE_SIZE, DMA_FROM_DEVICE);
191		if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
192			goto partial_error;
193	}
194
195	priv->rx_ring[id].skb = skb;
196	return skb;
197
198partial_error:
199	ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
200error:
201	dev_kfree_skb_any(skb);
202	return NULL;
203}
204
205static int ipoib_ib_post_receives(struct net_device *dev)
206{
207	struct ipoib_dev_priv *priv = netdev_priv(dev);
208	int i;
209
210	for (i = 0; i < ipoib_recvq_size; ++i) {
211		if (!ipoib_alloc_rx_skb(dev, i)) {
212			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
213			return -ENOMEM;
214		}
215		if (ipoib_ib_post_receive(dev, i)) {
216			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
217			return -EIO;
218		}
219	}
220
221	return 0;
222}
223
224static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
225{
226	struct ipoib_dev_priv *priv = netdev_priv(dev);
227	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
228	struct sk_buff *skb;
229	u64 mapping[IPOIB_UD_RX_SG];
230	union ib_gid *dgid;
231
232	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
233		       wr_id, wc->status);
234
235	if (unlikely(wr_id >= ipoib_recvq_size)) {
236		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
237			   wr_id, ipoib_recvq_size);
238		return;
239	}
240
241	skb  = priv->rx_ring[wr_id].skb;
242
243	if (unlikely(wc->status != IB_WC_SUCCESS)) {
244		if (wc->status != IB_WC_WR_FLUSH_ERR)
245			ipoib_warn(priv, "failed recv event "
246				   "(status=%d, wrid=%d vend_err %x)\n",
247				   wc->status, wr_id, wc->vendor_err);
248		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
249		dev_kfree_skb_any(skb);
250		priv->rx_ring[wr_id].skb = NULL;
251		return;
252	}
253
254	/*
255	 * Drop packets that this interface sent, ie multicast packets
256	 * that the HCA has replicated.
257	 */
258	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
259		goto repost;
260
261	memcpy(mapping, priv->rx_ring[wr_id].mapping,
262	       IPOIB_UD_RX_SG * sizeof *mapping);
263
264	/*
265	 * If we can't allocate a new RX buffer, dump
266	 * this packet and reuse the old buffer.
267	 */
268	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
269		++dev->stats.rx_dropped;
270		goto repost;
271	}
272
273	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
274		       wc->byte_len, wc->slid);
275
276	ipoib_ud_dma_unmap_rx(priv, mapping);
277	ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
278
279	/* First byte of dgid signals multicast when 0xff */
280	dgid = &((struct ib_grh *)skb->data)->dgid;
281
282	if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
283		skb->pkt_type = PACKET_HOST;
284	else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
285		skb->pkt_type = PACKET_BROADCAST;
286	else
287		skb->pkt_type = PACKET_MULTICAST;
288
289	skb_pull(skb, IB_GRH_BYTES);
290
291	skb->protocol = ((struct ipoib_header *) skb->data)->proto;
292	skb_reset_mac_header(skb);
293	skb_pull(skb, IPOIB_ENCAP_LEN);
294
295	++dev->stats.rx_packets;
296	dev->stats.rx_bytes += skb->len;
297
298	skb->dev = dev;
299	if ((dev->features & NETIF_F_RXCSUM) && likely(wc->csum_ok))
300		skb->ip_summed = CHECKSUM_UNNECESSARY;
301
302	napi_gro_receive(&priv->napi, skb);
303
304repost:
305	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
306		ipoib_warn(priv, "ipoib_ib_post_receive failed "
307			   "for buf %d\n", wr_id);
308}
309
310static int ipoib_dma_map_tx(struct ib_device *ca,
311			    struct ipoib_tx_buf *tx_req)
312{
313	struct sk_buff *skb = tx_req->skb;
314	u64 *mapping = tx_req->mapping;
315	int i;
316	int off;
317
318	if (skb_headlen(skb)) {
319		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
320					       DMA_TO_DEVICE);
321		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
322			return -EIO;
323
324		off = 1;
325	} else
326		off = 0;
327
328	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
329		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
330		mapping[i + off] = ib_dma_map_page(ca,
331						 skb_frag_page(frag),
332						 frag->page_offset, skb_frag_size(frag),
333						 DMA_TO_DEVICE);
334		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
335			goto partial_error;
336	}
337	return 0;
338
339partial_error:
340	for (; i > 0; --i) {
341		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
342
343		ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
344	}
345
346	if (off)
347		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
348
349	return -EIO;
350}
351
352static void ipoib_dma_unmap_tx(struct ib_device *ca,
353			       struct ipoib_tx_buf *tx_req)
354{
355	struct sk_buff *skb = tx_req->skb;
356	u64 *mapping = tx_req->mapping;
357	int i;
358	int off;
359
360	if (skb_headlen(skb)) {
361		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
362		off = 1;
363	} else
364		off = 0;
365
366	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
367		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
368
369		ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
370				  DMA_TO_DEVICE);
371	}
372}
373
374static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
375{
376	struct ipoib_dev_priv *priv = netdev_priv(dev);
377	unsigned int wr_id = wc->wr_id;
378	struct ipoib_tx_buf *tx_req;
379
380	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
381		       wr_id, wc->status);
382
383	if (unlikely(wr_id >= ipoib_sendq_size)) {
384		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
385			   wr_id, ipoib_sendq_size);
386		return;
387	}
388
389	tx_req = &priv->tx_ring[wr_id];
390
391	ipoib_dma_unmap_tx(priv->ca, tx_req);
392
393	++dev->stats.tx_packets;
394	dev->stats.tx_bytes += tx_req->skb->len;
395
396	dev_kfree_skb_any(tx_req->skb);
397
398	++priv->tx_tail;
399	if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
400	    netif_queue_stopped(dev) &&
401	    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
402		netif_wake_queue(dev);
403
404	if (wc->status != IB_WC_SUCCESS &&
405	    wc->status != IB_WC_WR_FLUSH_ERR)
406		ipoib_warn(priv, "failed send event "
407			   "(status=%d, wrid=%d vend_err %x)\n",
408			   wc->status, wr_id, wc->vendor_err);
409}
410
411static int poll_tx(struct ipoib_dev_priv *priv)
412{
413	int n, i;
414
415	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
416	for (i = 0; i < n; ++i)
417		ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
418
419	return n == MAX_SEND_CQE;
420}
421
422int ipoib_poll(struct napi_struct *napi, int budget)
423{
424	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
425	struct net_device *dev = priv->dev;
426	int done;
427	int t;
428	int n, i;
429
430	done  = 0;
431
432poll_more:
433	while (done < budget) {
434		int max = (budget - done);
435
436		t = min(IPOIB_NUM_WC, max);
437		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
438
439		for (i = 0; i < n; i++) {
440			struct ib_wc *wc = priv->ibwc + i;
441
442			if (wc->wr_id & IPOIB_OP_RECV) {
443				++done;
444				if (wc->wr_id & IPOIB_OP_CM)
445					ipoib_cm_handle_rx_wc(dev, wc);
446				else
447					ipoib_ib_handle_rx_wc(dev, wc);
448			} else
449				ipoib_cm_handle_tx_wc(priv->dev, wc);
450		}
451
452		if (n != t)
453			break;
454	}
455
456	if (done < budget) {
457		napi_complete(napi);
458		if (unlikely(ib_req_notify_cq(priv->recv_cq,
459					      IB_CQ_NEXT_COMP |
460					      IB_CQ_REPORT_MISSED_EVENTS)) &&
461		    napi_reschedule(napi))
462			goto poll_more;
463	}
464
465	return done;
466}
467
468void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
469{
470	struct net_device *dev = dev_ptr;
471	struct ipoib_dev_priv *priv = netdev_priv(dev);
472
473	napi_schedule(&priv->napi);
474}
475
476static void drain_tx_cq(struct net_device *dev)
477{
478	struct ipoib_dev_priv *priv = netdev_priv(dev);
479
480	netif_tx_lock(dev);
481	while (poll_tx(priv))
482		; /* nothing */
483
484	if (netif_queue_stopped(dev))
485		mod_timer(&priv->poll_timer, jiffies + 1);
486
487	netif_tx_unlock(dev);
488}
489
490void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
491{
492	struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
493
494	mod_timer(&priv->poll_timer, jiffies);
495}
496
497static inline int post_send(struct ipoib_dev_priv *priv,
498			    unsigned int wr_id,
499			    struct ib_ah *address, u32 qpn,
500			    struct ipoib_tx_buf *tx_req,
501			    void *head, int hlen)
502{
503	struct ib_send_wr *bad_wr;
504	int i, off;
505	struct sk_buff *skb = tx_req->skb;
506	skb_frag_t *frags = skb_shinfo(skb)->frags;
507	int nr_frags = skb_shinfo(skb)->nr_frags;
508	u64 *mapping = tx_req->mapping;
509
510	if (skb_headlen(skb)) {
511		priv->tx_sge[0].addr         = mapping[0];
512		priv->tx_sge[0].length       = skb_headlen(skb);
513		off = 1;
514	} else
515		off = 0;
516
517	for (i = 0; i < nr_frags; ++i) {
518		priv->tx_sge[i + off].addr = mapping[i + off];
519		priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
520	}
521	priv->tx_wr.num_sge	     = nr_frags + off;
522	priv->tx_wr.wr_id 	     = wr_id;
523	priv->tx_wr.wr.ud.remote_qpn = qpn;
524	priv->tx_wr.wr.ud.ah 	     = address;
525
526	if (head) {
527		priv->tx_wr.wr.ud.mss	 = skb_shinfo(skb)->gso_size;
528		priv->tx_wr.wr.ud.header = head;
529		priv->tx_wr.wr.ud.hlen	 = hlen;
530		priv->tx_wr.opcode	 = IB_WR_LSO;
531	} else
532		priv->tx_wr.opcode	 = IB_WR_SEND;
533
534	return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
535}
536
537void ipoib_send(struct net_device *dev, struct sk_buff *skb,
538		struct ipoib_ah *address, u32 qpn)
539{
540	struct ipoib_dev_priv *priv = netdev_priv(dev);
541	struct ipoib_tx_buf *tx_req;
542	int hlen, rc;
543	void *phead;
544
545	if (skb_is_gso(skb)) {
546		hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
547		phead = skb->data;
548		if (unlikely(!skb_pull(skb, hlen))) {
549			ipoib_warn(priv, "linear data too small\n");
550			++dev->stats.tx_dropped;
551			++dev->stats.tx_errors;
552			dev_kfree_skb_any(skb);
553			return;
554		}
555	} else {
556		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
557			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
558				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
559			++dev->stats.tx_dropped;
560			++dev->stats.tx_errors;
561			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
562			return;
563		}
564		phead = NULL;
565		hlen  = 0;
566	}
567
568	ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
569		       skb->len, address, qpn);
570
571	/*
572	 * We put the skb into the tx_ring _before_ we call post_send()
573	 * because it's entirely possible that the completion handler will
574	 * run before we execute anything after the post_send().  That
575	 * means we have to make sure everything is properly recorded and
576	 * our state is consistent before we call post_send().
577	 */
578	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
579	tx_req->skb = skb;
580	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
581		++dev->stats.tx_errors;
582		dev_kfree_skb_any(skb);
583		return;
584	}
585
586	if (skb->ip_summed == CHECKSUM_PARTIAL)
587		priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
588	else
589		priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
590
591	if (++priv->tx_outstanding == ipoib_sendq_size) {
592		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
593		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
594			ipoib_warn(priv, "request notify on send CQ failed\n");
595		netif_stop_queue(dev);
596	}
597
598	rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
599		       address->ah, qpn, tx_req, phead, hlen);
600	if (unlikely(rc)) {
601		ipoib_warn(priv, "post_send failed, error %d\n", rc);
602		++dev->stats.tx_errors;
603		--priv->tx_outstanding;
604		ipoib_dma_unmap_tx(priv->ca, tx_req);
605		dev_kfree_skb_any(skb);
606		if (netif_queue_stopped(dev))
607			netif_wake_queue(dev);
608	} else {
609		dev->trans_start = jiffies;
610
611		address->last_send = priv->tx_head;
612		++priv->tx_head;
613		skb_orphan(skb);
614
615	}
616
617	if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
618		while (poll_tx(priv))
619			; /* nothing */
620}
621
622static void __ipoib_reap_ah(struct net_device *dev)
623{
624	struct ipoib_dev_priv *priv = netdev_priv(dev);
625	struct ipoib_ah *ah, *tah;
626	LIST_HEAD(remove_list);
627	unsigned long flags;
628
629	netif_tx_lock_bh(dev);
630	spin_lock_irqsave(&priv->lock, flags);
631
632	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
633		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
634			list_del(&ah->list);
635			ib_destroy_ah(ah->ah);
636			kfree(ah);
637		}
638
639	spin_unlock_irqrestore(&priv->lock, flags);
640	netif_tx_unlock_bh(dev);
641}
642
643void ipoib_reap_ah(struct work_struct *work)
644{
645	struct ipoib_dev_priv *priv =
646		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
647	struct net_device *dev = priv->dev;
648
649	__ipoib_reap_ah(dev);
650
651	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
652		queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
653				   round_jiffies_relative(HZ));
654}
655
656static void ipoib_ib_tx_timer_func(unsigned long ctx)
657{
658	drain_tx_cq((struct net_device *)ctx);
659}
660
661int ipoib_ib_dev_open(struct net_device *dev)
662{
663	struct ipoib_dev_priv *priv = netdev_priv(dev);
664	int ret;
665
666	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
667		ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
668		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
669		return -1;
670	}
671	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
672
673	ret = ipoib_init_qp(dev);
674	if (ret) {
675		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
676		return -1;
677	}
678
679	ret = ipoib_ib_post_receives(dev);
680	if (ret) {
681		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
682		ipoib_ib_dev_stop(dev, 1);
683		return -1;
684	}
685
686	ret = ipoib_cm_dev_open(dev);
687	if (ret) {
688		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
689		ipoib_ib_dev_stop(dev, 1);
690		return -1;
691	}
692
693	clear_bit(IPOIB_STOP_REAPER, &priv->flags);
694	queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
695			   round_jiffies_relative(HZ));
696
697	if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
698		napi_enable(&priv->napi);
699
700	return 0;
701}
702
703static void ipoib_pkey_dev_check_presence(struct net_device *dev)
704{
705	struct ipoib_dev_priv *priv = netdev_priv(dev);
706	u16 pkey_index = 0;
707
708	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
709		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
710	else
711		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
712}
713
714int ipoib_ib_dev_up(struct net_device *dev)
715{
716	struct ipoib_dev_priv *priv = netdev_priv(dev);
717
718	ipoib_pkey_dev_check_presence(dev);
719
720	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
721		ipoib_dbg(priv, "PKEY is not assigned.\n");
722		return 0;
723	}
724
725	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
726
727	return ipoib_mcast_start_thread(dev);
728}
729
730int ipoib_ib_dev_down(struct net_device *dev, int flush)
731{
732	struct ipoib_dev_priv *priv = netdev_priv(dev);
733
734	ipoib_dbg(priv, "downing ib_dev\n");
735
736	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
737	netif_carrier_off(dev);
738
739	/* Shutdown the P_Key thread if still active */
740	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
741		mutex_lock(&pkey_mutex);
742		set_bit(IPOIB_PKEY_STOP, &priv->flags);
743		cancel_delayed_work(&priv->pkey_poll_task);
744		mutex_unlock(&pkey_mutex);
745		if (flush)
746			flush_workqueue(ipoib_workqueue);
747	}
748
749	ipoib_mcast_stop_thread(dev, flush);
750	ipoib_mcast_dev_flush(dev);
751
752	ipoib_flush_paths(dev);
753
754	return 0;
755}
756
757static int recvs_pending(struct net_device *dev)
758{
759	struct ipoib_dev_priv *priv = netdev_priv(dev);
760	int pending = 0;
761	int i;
762
763	for (i = 0; i < ipoib_recvq_size; ++i)
764		if (priv->rx_ring[i].skb)
765			++pending;
766
767	return pending;
768}
769
770void ipoib_drain_cq(struct net_device *dev)
771{
772	struct ipoib_dev_priv *priv = netdev_priv(dev);
773	int i, n;
774
775	/*
776	 * We call completion handling routines that expect to be
777	 * called from the BH-disabled NAPI poll context, so disable
778	 * BHs here too.
779	 */
780	local_bh_disable();
781
782	do {
783		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
784		for (i = 0; i < n; ++i) {
785			/*
786			 * Convert any successful completions to flush
787			 * errors to avoid passing packets up the
788			 * stack after bringing the device down.
789			 */
790			if (priv->ibwc[i].status == IB_WC_SUCCESS)
791				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
792
793			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
794				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
795					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
796				else
797					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
798			} else
799				ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
800		}
801	} while (n == IPOIB_NUM_WC);
802
803	while (poll_tx(priv))
804		; /* nothing */
805
806	local_bh_enable();
807}
808
809int ipoib_ib_dev_stop(struct net_device *dev, int flush)
810{
811	struct ipoib_dev_priv *priv = netdev_priv(dev);
812	struct ib_qp_attr qp_attr;
813	unsigned long begin;
814	struct ipoib_tx_buf *tx_req;
815	int i;
816
817	if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
818		napi_disable(&priv->napi);
819
820	ipoib_cm_dev_stop(dev);
821
822	/*
823	 * Move our QP to the error state and then reinitialize in
824	 * when all work requests have completed or have been flushed.
825	 */
826	qp_attr.qp_state = IB_QPS_ERR;
827	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
828		ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
829
830	/* Wait for all sends and receives to complete */
831	begin = jiffies;
832
833	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
834		if (time_after(jiffies, begin + 5 * HZ)) {
835			ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
836				   priv->tx_head - priv->tx_tail, recvs_pending(dev));
837
838			/*
839			 * assume the HW is wedged and just free up
840			 * all our pending work requests.
841			 */
842			while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
843				tx_req = &priv->tx_ring[priv->tx_tail &
844							(ipoib_sendq_size - 1)];
845				ipoib_dma_unmap_tx(priv->ca, tx_req);
846				dev_kfree_skb_any(tx_req->skb);
847				++priv->tx_tail;
848				--priv->tx_outstanding;
849			}
850
851			for (i = 0; i < ipoib_recvq_size; ++i) {
852				struct ipoib_rx_buf *rx_req;
853
854				rx_req = &priv->rx_ring[i];
855				if (!rx_req->skb)
856					continue;
857				ipoib_ud_dma_unmap_rx(priv,
858						      priv->rx_ring[i].mapping);
859				dev_kfree_skb_any(rx_req->skb);
860				rx_req->skb = NULL;
861			}
862
863			goto timeout;
864		}
865
866		ipoib_drain_cq(dev);
867
868		msleep(1);
869	}
870
871	ipoib_dbg(priv, "All sends and receives done.\n");
872
873timeout:
874	del_timer_sync(&priv->poll_timer);
875	qp_attr.qp_state = IB_QPS_RESET;
876	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
877		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
878
879	/* Wait for all AHs to be reaped */
880	set_bit(IPOIB_STOP_REAPER, &priv->flags);
881	cancel_delayed_work(&priv->ah_reap_task);
882	if (flush)
883		flush_workqueue(ipoib_workqueue);
884
885	begin = jiffies;
886
887	while (!list_empty(&priv->dead_ahs)) {
888		__ipoib_reap_ah(dev);
889
890		if (time_after(jiffies, begin + HZ)) {
891			ipoib_warn(priv, "timing out; will leak address handles\n");
892			break;
893		}
894
895		msleep(1);
896	}
897
898	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
899
900	return 0;
901}
902
903int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
904{
905	struct ipoib_dev_priv *priv = netdev_priv(dev);
906
907	priv->ca = ca;
908	priv->port = port;
909	priv->qp = NULL;
910
911	if (ipoib_transport_dev_init(dev, ca)) {
912		printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
913		return -ENODEV;
914	}
915
916	setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
917		    (unsigned long) dev);
918
919	if (dev->flags & IFF_UP) {
920		if (ipoib_ib_dev_open(dev)) {
921			ipoib_transport_dev_cleanup(dev);
922			return -ENODEV;
923		}
924	}
925
926	return 0;
927}
928
929static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
930				enum ipoib_flush_level level)
931{
932	struct ipoib_dev_priv *cpriv;
933	struct net_device *dev = priv->dev;
934	u16 new_index;
935
936	mutex_lock(&priv->vlan_mutex);
937
938	/*
939	 * Flush any child interfaces too -- they might be up even if
940	 * the parent is down.
941	 */
942	list_for_each_entry(cpriv, &priv->child_intfs, list)
943		__ipoib_ib_dev_flush(cpriv, level);
944
945	mutex_unlock(&priv->vlan_mutex);
946
947	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
948		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
949		return;
950	}
951
952	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
953		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
954		return;
955	}
956
957	if (level == IPOIB_FLUSH_HEAVY) {
958		if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
959			clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
960			ipoib_ib_dev_down(dev, 0);
961			ipoib_ib_dev_stop(dev, 0);
962			if (ipoib_pkey_dev_delay_open(dev))
963				return;
964		}
965
966		/* restart QP only if P_Key index is changed */
967		if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
968		    new_index == priv->pkey_index) {
969			ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
970			return;
971		}
972		priv->pkey_index = new_index;
973	}
974
975	if (level == IPOIB_FLUSH_LIGHT) {
976		ipoib_mark_paths_invalid(dev);
977		ipoib_mcast_dev_flush(dev);
978	}
979
980	if (level >= IPOIB_FLUSH_NORMAL)
981		ipoib_ib_dev_down(dev, 0);
982
983	if (level == IPOIB_FLUSH_HEAVY) {
984		ipoib_ib_dev_stop(dev, 0);
985		ipoib_ib_dev_open(dev);
986	}
987
988	/*
989	 * The device could have been brought down between the start and when
990	 * we get here, don't bring it back up if it's not configured up
991	 */
992	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
993		if (level >= IPOIB_FLUSH_NORMAL)
994			ipoib_ib_dev_up(dev);
995		ipoib_mcast_restart_task(&priv->restart_task);
996	}
997}
998
999void ipoib_ib_dev_flush_light(struct work_struct *work)
1000{
1001	struct ipoib_dev_priv *priv =
1002		container_of(work, struct ipoib_dev_priv, flush_light);
1003
1004	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1005}
1006
1007void ipoib_ib_dev_flush_normal(struct work_struct *work)
1008{
1009	struct ipoib_dev_priv *priv =
1010		container_of(work, struct ipoib_dev_priv, flush_normal);
1011
1012	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1013}
1014
1015void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1016{
1017	struct ipoib_dev_priv *priv =
1018		container_of(work, struct ipoib_dev_priv, flush_heavy);
1019
1020	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1021}
1022
1023void ipoib_ib_dev_cleanup(struct net_device *dev)
1024{
1025	struct ipoib_dev_priv *priv = netdev_priv(dev);
1026
1027	ipoib_dbg(priv, "cleaning up ib_dev\n");
1028
1029	ipoib_mcast_stop_thread(dev, 1);
1030	ipoib_mcast_dev_flush(dev);
1031
1032	ipoib_transport_dev_cleanup(dev);
1033}
1034
1035/*
1036 * Delayed P_Key Assigment Interim Support
1037 *
1038 * The following is initial implementation of delayed P_Key assigment
1039 * mechanism. It is using the same approach implemented for the multicast
1040 * group join. The single goal of this implementation is to quickly address
1041 * Bug #2507. This implementation will probably be removed when the P_Key
1042 * change async notification is available.
1043 */
1044
1045void ipoib_pkey_poll(struct work_struct *work)
1046{
1047	struct ipoib_dev_priv *priv =
1048		container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1049	struct net_device *dev = priv->dev;
1050
1051	ipoib_pkey_dev_check_presence(dev);
1052
1053	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1054		ipoib_open(dev);
1055	else {
1056		mutex_lock(&pkey_mutex);
1057		if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1058			queue_delayed_work(ipoib_workqueue,
1059					   &priv->pkey_poll_task,
1060					   HZ);
1061		mutex_unlock(&pkey_mutex);
1062	}
1063}
1064
1065int ipoib_pkey_dev_delay_open(struct net_device *dev)
1066{
1067	struct ipoib_dev_priv *priv = netdev_priv(dev);
1068
1069	/* Look for the interface pkey value in the IB Port P_Key table and */
1070	/* set the interface pkey assigment flag                            */
1071	ipoib_pkey_dev_check_presence(dev);
1072
1073	/* P_Key value not assigned yet - start polling */
1074	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1075		mutex_lock(&pkey_mutex);
1076		clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1077		queue_delayed_work(ipoib_workqueue,
1078				   &priv->pkey_poll_task,
1079				   HZ);
1080		mutex_unlock(&pkey_mutex);
1081		return 1;
1082	}
1083
1084	return 0;
1085}
1086