Merge branch 'sa1100'
[linux-2.6] / drivers / infiniband / ulp / ipoib / ipoib_ib.c
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  * $Id: ipoib_ib.c 1386 2004-12-27 16:23:17Z roland $
36  */
37
38 #include <linux/delay.h>
39 #include <linux/dma-mapping.h>
40
41 #include <rdma/ib_cache.h>
42 #include <linux/ip.h>
43 #include <linux/tcp.h>
44
45 #include "ipoib.h"
46
47 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
48 static int data_debug_level;
49
50 module_param(data_debug_level, int, 0644);
51 MODULE_PARM_DESC(data_debug_level,
52                  "Enable data path debug tracing if > 0");
53 #endif
54
55 static DEFINE_MUTEX(pkey_mutex);
56
57 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
58                                  struct ib_pd *pd, struct ib_ah_attr *attr)
59 {
60         struct ipoib_ah *ah;
61
62         ah = kmalloc(sizeof *ah, GFP_KERNEL);
63         if (!ah)
64                 return NULL;
65
66         ah->dev       = dev;
67         ah->last_send = 0;
68         kref_init(&ah->ref);
69
70         ah->ah = ib_create_ah(pd, attr);
71         if (IS_ERR(ah->ah)) {
72                 kfree(ah);
73                 ah = NULL;
74         } else
75                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
76
77         return ah;
78 }
79
80 void ipoib_free_ah(struct kref *kref)
81 {
82         struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
83         struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
84
85         unsigned long flags;
86
87         spin_lock_irqsave(&priv->lock, flags);
88         list_add_tail(&ah->list, &priv->dead_ahs);
89         spin_unlock_irqrestore(&priv->lock, flags);
90 }
91
92 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
93                                   u64 mapping[IPOIB_UD_RX_SG])
94 {
95         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
96                 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
97                                     DMA_FROM_DEVICE);
98                 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
99                                   DMA_FROM_DEVICE);
100         } else
101                 ib_dma_unmap_single(priv->ca, mapping[0],
102                                     IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
103                                     DMA_FROM_DEVICE);
104 }
105
106 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
107                                    struct sk_buff *skb,
108                                    unsigned int length)
109 {
110         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
111                 skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
112                 unsigned int size;
113                 /*
114                  * There is only two buffers needed for max_payload = 4K,
115                  * first buf size is IPOIB_UD_HEAD_SIZE
116                  */
117                 skb->tail += IPOIB_UD_HEAD_SIZE;
118                 skb->len  += length;
119
120                 size = length - IPOIB_UD_HEAD_SIZE;
121
122                 frag->size     = size;
123                 skb->data_len += size;
124                 skb->truesize += size;
125         } else
126                 skb_put(skb, length);
127
128 }
129
130 static int ipoib_ib_post_receive(struct net_device *dev, int id)
131 {
132         struct ipoib_dev_priv *priv = netdev_priv(dev);
133         struct ib_recv_wr *bad_wr;
134         int ret;
135
136         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
137         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
138         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
139
140
141         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
142         if (unlikely(ret)) {
143                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
144                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
145                 dev_kfree_skb_any(priv->rx_ring[id].skb);
146                 priv->rx_ring[id].skb = NULL;
147         }
148
149         return ret;
150 }
151
152 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
153 {
154         struct ipoib_dev_priv *priv = netdev_priv(dev);
155         struct sk_buff *skb;
156         int buf_size;
157         u64 *mapping;
158
159         if (ipoib_ud_need_sg(priv->max_ib_mtu))
160                 buf_size = IPOIB_UD_HEAD_SIZE;
161         else
162                 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
163
164         skb = dev_alloc_skb(buf_size + 4);
165         if (unlikely(!skb))
166                 return NULL;
167
168         /*
169          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
170          * header.  So we need 4 more bytes to get to 48 and align the
171          * IP header to a multiple of 16.
172          */
173         skb_reserve(skb, 4);
174
175         mapping = priv->rx_ring[id].mapping;
176         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
177                                        DMA_FROM_DEVICE);
178         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
179                 goto error;
180
181         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
182                 struct page *page = alloc_page(GFP_ATOMIC);
183                 if (!page)
184                         goto partial_error;
185                 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
186                 mapping[1] =
187                         ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[0].page,
188                                         0, PAGE_SIZE, DMA_FROM_DEVICE);
189                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
190                         goto partial_error;
191         }
192
193         priv->rx_ring[id].skb = skb;
194         return skb;
195
196 partial_error:
197         ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
198 error:
199         dev_kfree_skb_any(skb);
200         return NULL;
201 }
202
203 static int ipoib_ib_post_receives(struct net_device *dev)
204 {
205         struct ipoib_dev_priv *priv = netdev_priv(dev);
206         int i;
207
208         for (i = 0; i < ipoib_recvq_size; ++i) {
209                 if (!ipoib_alloc_rx_skb(dev, i)) {
210                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
211                         return -ENOMEM;
212                 }
213                 if (ipoib_ib_post_receive(dev, i)) {
214                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
215                         return -EIO;
216                 }
217         }
218
219         return 0;
220 }
221
222 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
223 {
224         struct ipoib_dev_priv *priv = netdev_priv(dev);
225         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
226         struct sk_buff *skb;
227         u64 mapping[IPOIB_UD_RX_SG];
228
229         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
230                        wr_id, wc->status);
231
232         if (unlikely(wr_id >= ipoib_recvq_size)) {
233                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
234                            wr_id, ipoib_recvq_size);
235                 return;
236         }
237
238         skb  = priv->rx_ring[wr_id].skb;
239
240         if (unlikely(wc->status != IB_WC_SUCCESS)) {
241                 if (wc->status != IB_WC_WR_FLUSH_ERR)
242                         ipoib_warn(priv, "failed recv event "
243                                    "(status=%d, wrid=%d vend_err %x)\n",
244                                    wc->status, wr_id, wc->vendor_err);
245                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
246                 dev_kfree_skb_any(skb);
247                 priv->rx_ring[wr_id].skb = NULL;
248                 return;
249         }
250
251         /*
252          * Drop packets that this interface sent, ie multicast packets
253          * that the HCA has replicated.
254          */
255         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
256                 goto repost;
257
258         memcpy(mapping, priv->rx_ring[wr_id].mapping,
259                IPOIB_UD_RX_SG * sizeof *mapping);
260
261         /*
262          * If we can't allocate a new RX buffer, dump
263          * this packet and reuse the old buffer.
264          */
265         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
266                 ++dev->stats.rx_dropped;
267                 goto repost;
268         }
269
270         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
271                        wc->byte_len, wc->slid);
272
273         ipoib_ud_dma_unmap_rx(priv, mapping);
274         ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
275
276         skb_pull(skb, IB_GRH_BYTES);
277
278         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
279         skb_reset_mac_header(skb);
280         skb_pull(skb, IPOIB_ENCAP_LEN);
281
282         dev->last_rx = jiffies;
283         ++dev->stats.rx_packets;
284         dev->stats.rx_bytes += skb->len;
285
286         skb->dev = dev;
287         /* XXX get correct PACKET_ type here */
288         skb->pkt_type = PACKET_HOST;
289
290         if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
291                 skb->ip_summed = CHECKSUM_UNNECESSARY;
292
293         netif_receive_skb(skb);
294
295 repost:
296         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
297                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
298                            "for buf %d\n", wr_id);
299 }
300
301 static int ipoib_dma_map_tx(struct ib_device *ca,
302                             struct ipoib_tx_buf *tx_req)
303 {
304         struct sk_buff *skb = tx_req->skb;
305         u64 *mapping = tx_req->mapping;
306         int i;
307         int off;
308
309         if (skb_headlen(skb)) {
310                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
311                                                DMA_TO_DEVICE);
312                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
313                         return -EIO;
314
315                 off = 1;
316         } else
317                 off = 0;
318
319         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
320                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
321                 mapping[i + off] = ib_dma_map_page(ca, frag->page,
322                                                  frag->page_offset, frag->size,
323                                                  DMA_TO_DEVICE);
324                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
325                         goto partial_error;
326         }
327         return 0;
328
329 partial_error:
330         for (; i > 0; --i) {
331                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
332                 ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
333         }
334
335         if (off)
336                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
337
338         return -EIO;
339 }
340
341 static void ipoib_dma_unmap_tx(struct ib_device *ca,
342                                struct ipoib_tx_buf *tx_req)
343 {
344         struct sk_buff *skb = tx_req->skb;
345         u64 *mapping = tx_req->mapping;
346         int i;
347         int off;
348
349         if (skb_headlen(skb)) {
350                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
351                 off = 1;
352         } else
353                 off = 0;
354
355         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
356                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
357                 ib_dma_unmap_page(ca, mapping[i + off], frag->size,
358                                   DMA_TO_DEVICE);
359         }
360 }
361
362 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
363 {
364         struct ipoib_dev_priv *priv = netdev_priv(dev);
365         unsigned int wr_id = wc->wr_id;
366         struct ipoib_tx_buf *tx_req;
367
368         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
369                        wr_id, wc->status);
370
371         if (unlikely(wr_id >= ipoib_sendq_size)) {
372                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
373                            wr_id, ipoib_sendq_size);
374                 return;
375         }
376
377         tx_req = &priv->tx_ring[wr_id];
378
379         ipoib_dma_unmap_tx(priv->ca, tx_req);
380
381         ++dev->stats.tx_packets;
382         dev->stats.tx_bytes += tx_req->skb->len;
383
384         dev_kfree_skb_any(tx_req->skb);
385
386         ++priv->tx_tail;
387         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
388             netif_queue_stopped(dev) &&
389             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
390                 netif_wake_queue(dev);
391
392         if (wc->status != IB_WC_SUCCESS &&
393             wc->status != IB_WC_WR_FLUSH_ERR)
394                 ipoib_warn(priv, "failed send event "
395                            "(status=%d, wrid=%d vend_err %x)\n",
396                            wc->status, wr_id, wc->vendor_err);
397 }
398
399 static int poll_tx(struct ipoib_dev_priv *priv)
400 {
401         int n, i;
402
403         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
404         for (i = 0; i < n; ++i)
405                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
406
407         return n == MAX_SEND_CQE;
408 }
409
410 int ipoib_poll(struct napi_struct *napi, int budget)
411 {
412         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
413         struct net_device *dev = priv->dev;
414         int done;
415         int t;
416         int n, i;
417
418         done  = 0;
419
420 poll_more:
421         while (done < budget) {
422                 int max = (budget - done);
423
424                 t = min(IPOIB_NUM_WC, max);
425                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
426
427                 for (i = 0; i < n; i++) {
428                         struct ib_wc *wc = priv->ibwc + i;
429
430                         if (wc->wr_id & IPOIB_OP_RECV) {
431                                 ++done;
432                                 if (wc->wr_id & IPOIB_OP_CM)
433                                         ipoib_cm_handle_rx_wc(dev, wc);
434                                 else
435                                         ipoib_ib_handle_rx_wc(dev, wc);
436                         } else
437                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
438                 }
439
440                 if (n != t)
441                         break;
442         }
443
444         if (done < budget) {
445                 netif_rx_complete(dev, napi);
446                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
447                                               IB_CQ_NEXT_COMP |
448                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
449                     netif_rx_reschedule(dev, napi))
450                         goto poll_more;
451         }
452
453         return done;
454 }
455
456 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
457 {
458         struct net_device *dev = dev_ptr;
459         struct ipoib_dev_priv *priv = netdev_priv(dev);
460
461         netif_rx_schedule(dev, &priv->napi);
462 }
463
464 static void drain_tx_cq(struct net_device *dev)
465 {
466         struct ipoib_dev_priv *priv = netdev_priv(dev);
467         unsigned long flags;
468
469         spin_lock_irqsave(&priv->tx_lock, flags);
470         while (poll_tx(priv))
471                 ; /* nothing */
472
473         if (netif_queue_stopped(dev))
474                 mod_timer(&priv->poll_timer, jiffies + 1);
475
476         spin_unlock_irqrestore(&priv->tx_lock, flags);
477 }
478
479 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
480 {
481         drain_tx_cq((struct net_device *)dev_ptr);
482 }
483
484 static inline int post_send(struct ipoib_dev_priv *priv,
485                             unsigned int wr_id,
486                             struct ib_ah *address, u32 qpn,
487                             struct ipoib_tx_buf *tx_req,
488                             void *head, int hlen)
489 {
490         struct ib_send_wr *bad_wr;
491         int i, off;
492         struct sk_buff *skb = tx_req->skb;
493         skb_frag_t *frags = skb_shinfo(skb)->frags;
494         int nr_frags = skb_shinfo(skb)->nr_frags;
495         u64 *mapping = tx_req->mapping;
496
497         if (skb_headlen(skb)) {
498                 priv->tx_sge[0].addr         = mapping[0];
499                 priv->tx_sge[0].length       = skb_headlen(skb);
500                 off = 1;
501         } else
502                 off = 0;
503
504         for (i = 0; i < nr_frags; ++i) {
505                 priv->tx_sge[i + off].addr = mapping[i + off];
506                 priv->tx_sge[i + off].length = frags[i].size;
507         }
508         priv->tx_wr.num_sge          = nr_frags + off;
509         priv->tx_wr.wr_id            = wr_id;
510         priv->tx_wr.wr.ud.remote_qpn = qpn;
511         priv->tx_wr.wr.ud.ah         = address;
512
513         if (head) {
514                 priv->tx_wr.wr.ud.mss    = skb_shinfo(skb)->gso_size;
515                 priv->tx_wr.wr.ud.header = head;
516                 priv->tx_wr.wr.ud.hlen   = hlen;
517                 priv->tx_wr.opcode       = IB_WR_LSO;
518         } else
519                 priv->tx_wr.opcode       = IB_WR_SEND;
520
521         return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
522 }
523
524 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
525                 struct ipoib_ah *address, u32 qpn)
526 {
527         struct ipoib_dev_priv *priv = netdev_priv(dev);
528         struct ipoib_tx_buf *tx_req;
529         int hlen;
530         void *phead;
531
532         if (skb_is_gso(skb)) {
533                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
534                 phead = skb->data;
535                 if (unlikely(!skb_pull(skb, hlen))) {
536                         ipoib_warn(priv, "linear data too small\n");
537                         ++dev->stats.tx_dropped;
538                         ++dev->stats.tx_errors;
539                         dev_kfree_skb_any(skb);
540                         return;
541                 }
542         } else {
543                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
544                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
545                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
546                         ++dev->stats.tx_dropped;
547                         ++dev->stats.tx_errors;
548                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
549                         return;
550                 }
551                 phead = NULL;
552                 hlen  = 0;
553         }
554
555         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
556                        skb->len, address, qpn);
557
558         /*
559          * We put the skb into the tx_ring _before_ we call post_send()
560          * because it's entirely possible that the completion handler will
561          * run before we execute anything after the post_send().  That
562          * means we have to make sure everything is properly recorded and
563          * our state is consistent before we call post_send().
564          */
565         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
566         tx_req->skb = skb;
567         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
568                 ++dev->stats.tx_errors;
569                 dev_kfree_skb_any(skb);
570                 return;
571         }
572
573         if (skb->ip_summed == CHECKSUM_PARTIAL)
574                 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
575         else
576                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
577
578         if (++priv->tx_outstanding == ipoib_sendq_size) {
579                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
580                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
581                         ipoib_warn(priv, "request notify on send CQ failed\n");
582                 netif_stop_queue(dev);
583         }
584
585         if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
586                                address->ah, qpn, tx_req, phead, hlen))) {
587                 ipoib_warn(priv, "post_send failed\n");
588                 ++dev->stats.tx_errors;
589                 --priv->tx_outstanding;
590                 ipoib_dma_unmap_tx(priv->ca, tx_req);
591                 dev_kfree_skb_any(skb);
592                 if (netif_queue_stopped(dev))
593                         netif_wake_queue(dev);
594         } else {
595                 dev->trans_start = jiffies;
596
597                 address->last_send = priv->tx_head;
598                 ++priv->tx_head;
599                 skb_orphan(skb);
600
601         }
602
603         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
604                 while (poll_tx(priv))
605                         ; /* nothing */
606 }
607
608 static void __ipoib_reap_ah(struct net_device *dev)
609 {
610         struct ipoib_dev_priv *priv = netdev_priv(dev);
611         struct ipoib_ah *ah, *tah;
612         LIST_HEAD(remove_list);
613
614         spin_lock_irq(&priv->tx_lock);
615         spin_lock(&priv->lock);
616         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
617                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
618                         list_del(&ah->list);
619                         ib_destroy_ah(ah->ah);
620                         kfree(ah);
621                 }
622         spin_unlock(&priv->lock);
623         spin_unlock_irq(&priv->tx_lock);
624 }
625
626 void ipoib_reap_ah(struct work_struct *work)
627 {
628         struct ipoib_dev_priv *priv =
629                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
630         struct net_device *dev = priv->dev;
631
632         __ipoib_reap_ah(dev);
633
634         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
635                 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
636                                    round_jiffies_relative(HZ));
637 }
638
639 static void ipoib_ib_tx_timer_func(unsigned long ctx)
640 {
641         drain_tx_cq((struct net_device *)ctx);
642 }
643
644 int ipoib_ib_dev_open(struct net_device *dev)
645 {
646         struct ipoib_dev_priv *priv = netdev_priv(dev);
647         int ret;
648
649         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
650                 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
651                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
652                 return -1;
653         }
654         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
655
656         ret = ipoib_init_qp(dev);
657         if (ret) {
658                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
659                 return -1;
660         }
661
662         ret = ipoib_ib_post_receives(dev);
663         if (ret) {
664                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
665                 ipoib_ib_dev_stop(dev, 1);
666                 return -1;
667         }
668
669         ret = ipoib_cm_dev_open(dev);
670         if (ret) {
671                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
672                 ipoib_ib_dev_stop(dev, 1);
673                 return -1;
674         }
675
676         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
677         queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
678                            round_jiffies_relative(HZ));
679
680         init_timer(&priv->poll_timer);
681         priv->poll_timer.function = ipoib_ib_tx_timer_func;
682         priv->poll_timer.data = (unsigned long)dev;
683
684         set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
685
686         return 0;
687 }
688
689 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
690 {
691         struct ipoib_dev_priv *priv = netdev_priv(dev);
692         u16 pkey_index = 0;
693
694         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
695                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
696         else
697                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
698 }
699
700 int ipoib_ib_dev_up(struct net_device *dev)
701 {
702         struct ipoib_dev_priv *priv = netdev_priv(dev);
703
704         ipoib_pkey_dev_check_presence(dev);
705
706         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
707                 ipoib_dbg(priv, "PKEY is not assigned.\n");
708                 return 0;
709         }
710
711         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
712
713         return ipoib_mcast_start_thread(dev);
714 }
715
716 int ipoib_ib_dev_down(struct net_device *dev, int flush)
717 {
718         struct ipoib_dev_priv *priv = netdev_priv(dev);
719
720         ipoib_dbg(priv, "downing ib_dev\n");
721
722         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
723         netif_carrier_off(dev);
724
725         /* Shutdown the P_Key thread if still active */
726         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
727                 mutex_lock(&pkey_mutex);
728                 set_bit(IPOIB_PKEY_STOP, &priv->flags);
729                 cancel_delayed_work(&priv->pkey_poll_task);
730                 mutex_unlock(&pkey_mutex);
731                 if (flush)
732                         flush_workqueue(ipoib_workqueue);
733         }
734
735         ipoib_mcast_stop_thread(dev, flush);
736         ipoib_mcast_dev_flush(dev);
737
738         ipoib_flush_paths(dev);
739
740         return 0;
741 }
742
743 static int recvs_pending(struct net_device *dev)
744 {
745         struct ipoib_dev_priv *priv = netdev_priv(dev);
746         int pending = 0;
747         int i;
748
749         for (i = 0; i < ipoib_recvq_size; ++i)
750                 if (priv->rx_ring[i].skb)
751                         ++pending;
752
753         return pending;
754 }
755
756 void ipoib_drain_cq(struct net_device *dev)
757 {
758         struct ipoib_dev_priv *priv = netdev_priv(dev);
759         int i, n;
760         do {
761                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
762                 for (i = 0; i < n; ++i) {
763                         /*
764                          * Convert any successful completions to flush
765                          * errors to avoid passing packets up the
766                          * stack after bringing the device down.
767                          */
768                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
769                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
770
771                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
772                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
773                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
774                                 else
775                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
776                         } else
777                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
778                 }
779         } while (n == IPOIB_NUM_WC);
780
781         while (poll_tx(priv))
782                 ; /* nothing */
783 }
784
785 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
786 {
787         struct ipoib_dev_priv *priv = netdev_priv(dev);
788         struct ib_qp_attr qp_attr;
789         unsigned long begin;
790         struct ipoib_tx_buf *tx_req;
791         int i;
792
793         clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
794
795         ipoib_cm_dev_stop(dev);
796
797         /*
798          * Move our QP to the error state and then reinitialize in
799          * when all work requests have completed or have been flushed.
800          */
801         qp_attr.qp_state = IB_QPS_ERR;
802         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
803                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
804
805         /* Wait for all sends and receives to complete */
806         begin = jiffies;
807
808         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
809                 if (time_after(jiffies, begin + 5 * HZ)) {
810                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
811                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
812
813                         /*
814                          * assume the HW is wedged and just free up
815                          * all our pending work requests.
816                          */
817                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
818                                 tx_req = &priv->tx_ring[priv->tx_tail &
819                                                         (ipoib_sendq_size - 1)];
820                                 ipoib_dma_unmap_tx(priv->ca, tx_req);
821                                 dev_kfree_skb_any(tx_req->skb);
822                                 ++priv->tx_tail;
823                                 --priv->tx_outstanding;
824                         }
825
826                         for (i = 0; i < ipoib_recvq_size; ++i) {
827                                 struct ipoib_rx_buf *rx_req;
828
829                                 rx_req = &priv->rx_ring[i];
830                                 if (!rx_req->skb)
831                                         continue;
832                                 ipoib_ud_dma_unmap_rx(priv,
833                                                       priv->rx_ring[i].mapping);
834                                 dev_kfree_skb_any(rx_req->skb);
835                                 rx_req->skb = NULL;
836                         }
837
838                         goto timeout;
839                 }
840
841                 ipoib_drain_cq(dev);
842
843                 msleep(1);
844         }
845
846         ipoib_dbg(priv, "All sends and receives done.\n");
847
848 timeout:
849         del_timer_sync(&priv->poll_timer);
850         qp_attr.qp_state = IB_QPS_RESET;
851         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
852                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
853
854         /* Wait for all AHs to be reaped */
855         set_bit(IPOIB_STOP_REAPER, &priv->flags);
856         cancel_delayed_work(&priv->ah_reap_task);
857         if (flush)
858                 flush_workqueue(ipoib_workqueue);
859
860         begin = jiffies;
861
862         while (!list_empty(&priv->dead_ahs)) {
863                 __ipoib_reap_ah(dev);
864
865                 if (time_after(jiffies, begin + HZ)) {
866                         ipoib_warn(priv, "timing out; will leak address handles\n");
867                         break;
868                 }
869
870                 msleep(1);
871         }
872
873         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
874
875         return 0;
876 }
877
878 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
879 {
880         struct ipoib_dev_priv *priv = netdev_priv(dev);
881
882         priv->ca = ca;
883         priv->port = port;
884         priv->qp = NULL;
885
886         if (ipoib_transport_dev_init(dev, ca)) {
887                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
888                 return -ENODEV;
889         }
890
891         if (dev->flags & IFF_UP) {
892                 if (ipoib_ib_dev_open(dev)) {
893                         ipoib_transport_dev_cleanup(dev);
894                         return -ENODEV;
895                 }
896         }
897
898         return 0;
899 }
900
901 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, int pkey_event)
902 {
903         struct ipoib_dev_priv *cpriv;
904         struct net_device *dev = priv->dev;
905         u16 new_index;
906
907         mutex_lock(&priv->vlan_mutex);
908
909         /*
910          * Flush any child interfaces too -- they might be up even if
911          * the parent is down.
912          */
913         list_for_each_entry(cpriv, &priv->child_intfs, list)
914                 __ipoib_ib_dev_flush(cpriv, pkey_event);
915
916         mutex_unlock(&priv->vlan_mutex);
917
918         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
919                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
920                 return;
921         }
922
923         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
924                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
925                 return;
926         }
927
928         if (pkey_event) {
929                 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
930                         clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
931                         ipoib_ib_dev_down(dev, 0);
932                         ipoib_ib_dev_stop(dev, 0);
933                         if (ipoib_pkey_dev_delay_open(dev))
934                                 return;
935                 }
936
937                 /* restart QP only if P_Key index is changed */
938                 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
939                     new_index == priv->pkey_index) {
940                         ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
941                         return;
942                 }
943                 priv->pkey_index = new_index;
944         }
945
946         ipoib_dbg(priv, "flushing\n");
947
948         ipoib_ib_dev_down(dev, 0);
949
950         if (pkey_event) {
951                 ipoib_ib_dev_stop(dev, 0);
952                 ipoib_ib_dev_open(dev);
953         }
954
955         /*
956          * The device could have been brought down between the start and when
957          * we get here, don't bring it back up if it's not configured up
958          */
959         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
960                 ipoib_ib_dev_up(dev);
961                 ipoib_mcast_restart_task(&priv->restart_task);
962         }
963 }
964
965 void ipoib_ib_dev_flush(struct work_struct *work)
966 {
967         struct ipoib_dev_priv *priv =
968                 container_of(work, struct ipoib_dev_priv, flush_task);
969
970         ipoib_dbg(priv, "Flushing %s\n", priv->dev->name);
971         __ipoib_ib_dev_flush(priv, 0);
972 }
973
974 void ipoib_pkey_event(struct work_struct *work)
975 {
976         struct ipoib_dev_priv *priv =
977                 container_of(work, struct ipoib_dev_priv, pkey_event_task);
978
979         ipoib_dbg(priv, "Flushing %s and restarting its QP\n", priv->dev->name);
980         __ipoib_ib_dev_flush(priv, 1);
981 }
982
983 void ipoib_ib_dev_cleanup(struct net_device *dev)
984 {
985         struct ipoib_dev_priv *priv = netdev_priv(dev);
986
987         ipoib_dbg(priv, "cleaning up ib_dev\n");
988
989         ipoib_mcast_stop_thread(dev, 1);
990         ipoib_mcast_dev_flush(dev);
991
992         ipoib_transport_dev_cleanup(dev);
993 }
994
995 /*
996  * Delayed P_Key Assigment Interim Support
997  *
998  * The following is initial implementation of delayed P_Key assigment
999  * mechanism. It is using the same approach implemented for the multicast
1000  * group join. The single goal of this implementation is to quickly address
1001  * Bug #2507. This implementation will probably be removed when the P_Key
1002  * change async notification is available.
1003  */
1004
1005 void ipoib_pkey_poll(struct work_struct *work)
1006 {
1007         struct ipoib_dev_priv *priv =
1008                 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1009         struct net_device *dev = priv->dev;
1010
1011         ipoib_pkey_dev_check_presence(dev);
1012
1013         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1014                 ipoib_open(dev);
1015         else {
1016                 mutex_lock(&pkey_mutex);
1017                 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1018                         queue_delayed_work(ipoib_workqueue,
1019                                            &priv->pkey_poll_task,
1020                                            HZ);
1021                 mutex_unlock(&pkey_mutex);
1022         }
1023 }
1024
1025 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1026 {
1027         struct ipoib_dev_priv *priv = netdev_priv(dev);
1028
1029         /* Look for the interface pkey value in the IB Port P_Key table and */
1030         /* set the interface pkey assigment flag                            */
1031         ipoib_pkey_dev_check_presence(dev);
1032
1033         /* P_Key value not assigned yet - start polling */
1034         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1035                 mutex_lock(&pkey_mutex);
1036                 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1037                 queue_delayed_work(ipoib_workqueue,
1038                                    &priv->pkey_poll_task,
1039                                    HZ);
1040                 mutex_unlock(&pkey_mutex);
1041                 return 1;
1042         }
1043
1044         return 0;
1045 }