45c40a17d6a6ffc60cac0cab92bc24c689f75c83
[platform/kernel/linux-rpi.git] / 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
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
47 static int data_debug_level;
48
49 module_param(data_debug_level, int, 0644);
50 MODULE_PARM_DESC(data_debug_level,
51                  "Enable data path debug tracing if > 0");
52 #endif
53
54 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
55                                  struct ib_pd *pd, struct ib_ah_attr *attr)
56 {
57         struct ipoib_ah *ah;
58         struct ib_ah *vah;
59
60         ah = kmalloc(sizeof *ah, GFP_KERNEL);
61         if (!ah)
62                 return ERR_PTR(-ENOMEM);
63
64         ah->dev       = dev;
65         ah->last_send = 0;
66         kref_init(&ah->ref);
67
68         vah = ib_create_ah(pd, attr);
69         if (IS_ERR(vah)) {
70                 kfree(ah);
71                 ah = (struct ipoib_ah *)vah;
72         } else {
73                 ah->ah = vah;
74                 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
75         }
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         ib_dma_unmap_single(priv->ca, mapping[0],
96                             IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
97                             DMA_FROM_DEVICE);
98 }
99
100 static int ipoib_ib_post_receive(struct net_device *dev, int id)
101 {
102         struct ipoib_dev_priv *priv = netdev_priv(dev);
103         struct ib_recv_wr *bad_wr;
104         int ret;
105
106         priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;
107         priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
108         priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
109
110
111         ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
112         if (unlikely(ret)) {
113                 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
114                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
115                 dev_kfree_skb_any(priv->rx_ring[id].skb);
116                 priv->rx_ring[id].skb = NULL;
117         }
118
119         return ret;
120 }
121
122 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
123 {
124         struct ipoib_dev_priv *priv = netdev_priv(dev);
125         struct sk_buff *skb;
126         int buf_size;
127         u64 *mapping;
128
129         buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
130
131         skb = dev_alloc_skb(buf_size + IPOIB_ENCAP_LEN);
132         if (unlikely(!skb))
133                 return NULL;
134
135         /*
136          * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
137          * header.  So we need 4 more bytes to get to 48 and align the
138          * IP header to a multiple of 16.
139          */
140         skb_reserve(skb, 4);
141
142         mapping = priv->rx_ring[id].mapping;
143         mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
144                                        DMA_FROM_DEVICE);
145         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
146                 goto error;
147
148         priv->rx_ring[id].skb = skb;
149         return skb;
150 error:
151         dev_kfree_skb_any(skb);
152         return NULL;
153 }
154
155 static int ipoib_ib_post_receives(struct net_device *dev)
156 {
157         struct ipoib_dev_priv *priv = netdev_priv(dev);
158         int i;
159
160         for (i = 0; i < ipoib_recvq_size; ++i) {
161                 if (!ipoib_alloc_rx_skb(dev, i)) {
162                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
163                         return -ENOMEM;
164                 }
165                 if (ipoib_ib_post_receive(dev, i)) {
166                         ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
167                         return -EIO;
168                 }
169         }
170
171         return 0;
172 }
173
174 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
175 {
176         struct ipoib_dev_priv *priv = netdev_priv(dev);
177         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
178         struct sk_buff *skb;
179         u64 mapping[IPOIB_UD_RX_SG];
180         union ib_gid *dgid;
181         union ib_gid *sgid;
182
183         ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
184                        wr_id, wc->status);
185
186         if (unlikely(wr_id >= ipoib_recvq_size)) {
187                 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
188                            wr_id, ipoib_recvq_size);
189                 return;
190         }
191
192         skb  = priv->rx_ring[wr_id].skb;
193
194         if (unlikely(wc->status != IB_WC_SUCCESS)) {
195                 if (wc->status != IB_WC_WR_FLUSH_ERR)
196                         ipoib_warn(priv, "failed recv event "
197                                    "(status=%d, wrid=%d vend_err %x)\n",
198                                    wc->status, wr_id, wc->vendor_err);
199                 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
200                 dev_kfree_skb_any(skb);
201                 priv->rx_ring[wr_id].skb = NULL;
202                 return;
203         }
204
205         memcpy(mapping, priv->rx_ring[wr_id].mapping,
206                IPOIB_UD_RX_SG * sizeof *mapping);
207
208         /*
209          * If we can't allocate a new RX buffer, dump
210          * this packet and reuse the old buffer.
211          */
212         if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
213                 ++dev->stats.rx_dropped;
214                 goto repost;
215         }
216
217         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
218                        wc->byte_len, wc->slid);
219
220         ipoib_ud_dma_unmap_rx(priv, mapping);
221
222         skb_put(skb, wc->byte_len);
223
224         /* First byte of dgid signals multicast when 0xff */
225         dgid = &((struct ib_grh *)skb->data)->dgid;
226
227         if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
228                 skb->pkt_type = PACKET_HOST;
229         else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
230                 skb->pkt_type = PACKET_BROADCAST;
231         else
232                 skb->pkt_type = PACKET_MULTICAST;
233
234         sgid = &((struct ib_grh *)skb->data)->sgid;
235
236         /*
237          * Drop packets that this interface sent, ie multicast packets
238          * that the HCA has replicated.
239          */
240         if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) {
241                 int need_repost = 1;
242
243                 if ((wc->wc_flags & IB_WC_GRH) &&
244                     sgid->global.interface_id != priv->local_gid.global.interface_id)
245                         need_repost = 0;
246
247                 if (need_repost) {
248                         dev_kfree_skb_any(skb);
249                         goto repost;
250                 }
251         }
252
253         skb_pull(skb, IB_GRH_BYTES);
254
255         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
256         skb_reset_mac_header(skb);
257         skb_pull(skb, IPOIB_ENCAP_LEN);
258
259         ++dev->stats.rx_packets;
260         dev->stats.rx_bytes += skb->len;
261
262         skb->dev = dev;
263         if ((dev->features & NETIF_F_RXCSUM) &&
264                         likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
265                 skb->ip_summed = CHECKSUM_UNNECESSARY;
266
267         napi_gro_receive(&priv->napi, skb);
268
269 repost:
270         if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
271                 ipoib_warn(priv, "ipoib_ib_post_receive failed "
272                            "for buf %d\n", wr_id);
273 }
274
275 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
276 {
277         struct sk_buff *skb = tx_req->skb;
278         u64 *mapping = tx_req->mapping;
279         int i;
280         int off;
281
282         if (skb_headlen(skb)) {
283                 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
284                                                DMA_TO_DEVICE);
285                 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
286                         return -EIO;
287
288                 off = 1;
289         } else
290                 off = 0;
291
292         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
293                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
294                 mapping[i + off] = ib_dma_map_page(ca,
295                                                  skb_frag_page(frag),
296                                                  frag->page_offset, skb_frag_size(frag),
297                                                  DMA_TO_DEVICE);
298                 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
299                         goto partial_error;
300         }
301         return 0;
302
303 partial_error:
304         for (; i > 0; --i) {
305                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
306
307                 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
308         }
309
310         if (off)
311                 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
312
313         return -EIO;
314 }
315
316 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
317                         struct ipoib_tx_buf *tx_req)
318 {
319         struct sk_buff *skb = tx_req->skb;
320         u64 *mapping = tx_req->mapping;
321         int i;
322         int off;
323
324         if (skb_headlen(skb)) {
325                 ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
326                                     DMA_TO_DEVICE);
327                 off = 1;
328         } else
329                 off = 0;
330
331         for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
332                 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
333
334                 ib_dma_unmap_page(priv->ca, mapping[i + off],
335                                   skb_frag_size(frag), DMA_TO_DEVICE);
336         }
337 }
338
339 /*
340  * As the result of a completion error the QP Can be transferred to SQE states.
341  * The function checks if the (send)QP is in SQE state and
342  * moves it back to RTS state, that in order to have it functional again.
343  */
344 static void ipoib_qp_state_validate_work(struct work_struct *work)
345 {
346         struct ipoib_qp_state_validate *qp_work =
347                 container_of(work, struct ipoib_qp_state_validate, work);
348
349         struct ipoib_dev_priv *priv = qp_work->priv;
350         struct ib_qp_attr qp_attr;
351         struct ib_qp_init_attr query_init_attr;
352         int ret;
353
354         ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
355         if (ret) {
356                 ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
357                            __func__, ret);
358                 goto free_res;
359         }
360         pr_info("%s: QP: 0x%x is in state: %d\n",
361                 __func__, priv->qp->qp_num, qp_attr.qp_state);
362
363         /* currently support only in SQE->RTS transition*/
364         if (qp_attr.qp_state == IB_QPS_SQE) {
365                 qp_attr.qp_state = IB_QPS_RTS;
366
367                 ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
368                 if (ret) {
369                         pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
370                                 ret, priv->qp->qp_num);
371                         goto free_res;
372                 }
373                 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
374                         __func__, priv->qp->qp_num);
375         } else {
376                 pr_warn("QP (%d) will stay in state: %d\n",
377                         priv->qp->qp_num, qp_attr.qp_state);
378         }
379
380 free_res:
381         kfree(qp_work);
382 }
383
384 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
385 {
386         struct ipoib_dev_priv *priv = netdev_priv(dev);
387         unsigned int wr_id = wc->wr_id;
388         struct ipoib_tx_buf *tx_req;
389
390         ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
391                        wr_id, wc->status);
392
393         if (unlikely(wr_id >= ipoib_sendq_size)) {
394                 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
395                            wr_id, ipoib_sendq_size);
396                 return;
397         }
398
399         tx_req = &priv->tx_ring[wr_id];
400
401         ipoib_dma_unmap_tx(priv, tx_req);
402
403         ++dev->stats.tx_packets;
404         dev->stats.tx_bytes += tx_req->skb->len;
405
406         dev_kfree_skb_any(tx_req->skb);
407
408         ++priv->tx_tail;
409         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
410             netif_queue_stopped(dev) &&
411             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
412                 netif_wake_queue(dev);
413
414         if (wc->status != IB_WC_SUCCESS &&
415             wc->status != IB_WC_WR_FLUSH_ERR) {
416                 struct ipoib_qp_state_validate *qp_work;
417                 ipoib_warn(priv, "failed send event "
418                            "(status=%d, wrid=%d vend_err %x)\n",
419                            wc->status, wr_id, wc->vendor_err);
420                 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
421                 if (!qp_work) {
422                         ipoib_warn(priv, "%s Failed alloc ipoib_qp_state_validate for qp: 0x%x\n",
423                                    __func__, priv->qp->qp_num);
424                         return;
425                 }
426
427                 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
428                 qp_work->priv = priv;
429                 queue_work(priv->wq, &qp_work->work);
430         }
431 }
432
433 static int poll_tx(struct ipoib_dev_priv *priv)
434 {
435         int n, i;
436
437         n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
438         for (i = 0; i < n; ++i)
439                 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
440
441         return n == MAX_SEND_CQE;
442 }
443
444 int ipoib_poll(struct napi_struct *napi, int budget)
445 {
446         struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
447         struct net_device *dev = priv->dev;
448         int done;
449         int t;
450         int n, i;
451
452         done  = 0;
453
454 poll_more:
455         while (done < budget) {
456                 int max = (budget - done);
457
458                 t = min(IPOIB_NUM_WC, max);
459                 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
460
461                 for (i = 0; i < n; i++) {
462                         struct ib_wc *wc = priv->ibwc + i;
463
464                         if (wc->wr_id & IPOIB_OP_RECV) {
465                                 ++done;
466                                 if (wc->wr_id & IPOIB_OP_CM)
467                                         ipoib_cm_handle_rx_wc(dev, wc);
468                                 else
469                                         ipoib_ib_handle_rx_wc(dev, wc);
470                         } else
471                                 ipoib_cm_handle_tx_wc(priv->dev, wc);
472                 }
473
474                 if (n != t)
475                         break;
476         }
477
478         if (done < budget) {
479                 napi_complete(napi);
480                 if (unlikely(ib_req_notify_cq(priv->recv_cq,
481                                               IB_CQ_NEXT_COMP |
482                                               IB_CQ_REPORT_MISSED_EVENTS)) &&
483                     napi_reschedule(napi))
484                         goto poll_more;
485         }
486
487         return done;
488 }
489
490 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
491 {
492         struct net_device *dev = dev_ptr;
493         struct ipoib_dev_priv *priv = netdev_priv(dev);
494
495         napi_schedule(&priv->napi);
496 }
497
498 static void drain_tx_cq(struct net_device *dev)
499 {
500         struct ipoib_dev_priv *priv = netdev_priv(dev);
501
502         netif_tx_lock(dev);
503         while (poll_tx(priv))
504                 ; /* nothing */
505
506         if (netif_queue_stopped(dev))
507                 mod_timer(&priv->poll_timer, jiffies + 1);
508
509         netif_tx_unlock(dev);
510 }
511
512 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
513 {
514         struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
515
516         mod_timer(&priv->poll_timer, jiffies);
517 }
518
519 static inline int post_send(struct ipoib_dev_priv *priv,
520                             unsigned int wr_id,
521                             struct ib_ah *address, u32 qpn,
522                             struct ipoib_tx_buf *tx_req,
523                             void *head, int hlen)
524 {
525         struct ib_send_wr *bad_wr;
526         struct sk_buff *skb = tx_req->skb;
527
528         ipoib_build_sge(priv, tx_req);
529
530         priv->tx_wr.wr.wr_id    = wr_id;
531         priv->tx_wr.remote_qpn  = qpn;
532         priv->tx_wr.ah          = address;
533
534         if (head) {
535                 priv->tx_wr.mss         = skb_shinfo(skb)->gso_size;
536                 priv->tx_wr.header      = head;
537                 priv->tx_wr.hlen        = hlen;
538                 priv->tx_wr.wr.opcode   = IB_WR_LSO;
539         } else
540                 priv->tx_wr.wr.opcode   = IB_WR_SEND;
541
542         return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
543 }
544
545 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
546                 struct ipoib_ah *address, u32 qpn)
547 {
548         struct ipoib_dev_priv *priv = netdev_priv(dev);
549         struct ipoib_tx_buf *tx_req;
550         int hlen, rc;
551         void *phead;
552         unsigned usable_sge = priv->max_send_sge - !!skb_headlen(skb);
553
554         if (skb_is_gso(skb)) {
555                 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
556                 phead = skb->data;
557                 if (unlikely(!skb_pull(skb, hlen))) {
558                         ipoib_warn(priv, "linear data too small\n");
559                         ++dev->stats.tx_dropped;
560                         ++dev->stats.tx_errors;
561                         dev_kfree_skb_any(skb);
562                         return;
563                 }
564         } else {
565                 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
566                         ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
567                                    skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
568                         ++dev->stats.tx_dropped;
569                         ++dev->stats.tx_errors;
570                         ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
571                         return;
572                 }
573                 phead = NULL;
574                 hlen  = 0;
575         }
576         if (skb_shinfo(skb)->nr_frags > usable_sge) {
577                 if (skb_linearize(skb) < 0) {
578                         ipoib_warn(priv, "skb could not be linearized\n");
579                         ++dev->stats.tx_dropped;
580                         ++dev->stats.tx_errors;
581                         dev_kfree_skb_any(skb);
582                         return;
583                 }
584                 /* Does skb_linearize return ok without reducing nr_frags? */
585                 if (skb_shinfo(skb)->nr_frags > usable_sge) {
586                         ipoib_warn(priv, "too many frags after skb linearize\n");
587                         ++dev->stats.tx_dropped;
588                         ++dev->stats.tx_errors;
589                         dev_kfree_skb_any(skb);
590                         return;
591                 }
592         }
593
594         ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
595                        skb->len, address, qpn);
596
597         /*
598          * We put the skb into the tx_ring _before_ we call post_send()
599          * because it's entirely possible that the completion handler will
600          * run before we execute anything after the post_send().  That
601          * means we have to make sure everything is properly recorded and
602          * our state is consistent before we call post_send().
603          */
604         tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
605         tx_req->skb = skb;
606         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
607                 ++dev->stats.tx_errors;
608                 dev_kfree_skb_any(skb);
609                 return;
610         }
611
612         if (skb->ip_summed == CHECKSUM_PARTIAL)
613                 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
614         else
615                 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
616
617         if (++priv->tx_outstanding == ipoib_sendq_size) {
618                 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
619                 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
620                         ipoib_warn(priv, "request notify on send CQ failed\n");
621                 netif_stop_queue(dev);
622         }
623
624         skb_orphan(skb);
625         skb_dst_drop(skb);
626
627         rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
628                        address->ah, qpn, tx_req, phead, hlen);
629         if (unlikely(rc)) {
630                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
631                 ++dev->stats.tx_errors;
632                 --priv->tx_outstanding;
633                 ipoib_dma_unmap_tx(priv, tx_req);
634                 dev_kfree_skb_any(skb);
635                 if (netif_queue_stopped(dev))
636                         netif_wake_queue(dev);
637         } else {
638                 netif_trans_update(dev);
639
640                 address->last_send = priv->tx_head;
641                 ++priv->tx_head;
642         }
643
644         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
645                 while (poll_tx(priv))
646                         ; /* nothing */
647 }
648
649 static void __ipoib_reap_ah(struct net_device *dev)
650 {
651         struct ipoib_dev_priv *priv = netdev_priv(dev);
652         struct ipoib_ah *ah, *tah;
653         LIST_HEAD(remove_list);
654         unsigned long flags;
655
656         netif_tx_lock_bh(dev);
657         spin_lock_irqsave(&priv->lock, flags);
658
659         list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
660                 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
661                         list_del(&ah->list);
662                         ib_destroy_ah(ah->ah);
663                         kfree(ah);
664                 }
665
666         spin_unlock_irqrestore(&priv->lock, flags);
667         netif_tx_unlock_bh(dev);
668 }
669
670 void ipoib_reap_ah(struct work_struct *work)
671 {
672         struct ipoib_dev_priv *priv =
673                 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
674         struct net_device *dev = priv->dev;
675
676         __ipoib_reap_ah(dev);
677
678         if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
679                 queue_delayed_work(priv->wq, &priv->ah_reap_task,
680                                    round_jiffies_relative(HZ));
681 }
682
683 static void ipoib_flush_ah(struct net_device *dev)
684 {
685         struct ipoib_dev_priv *priv = netdev_priv(dev);
686
687         cancel_delayed_work(&priv->ah_reap_task);
688         flush_workqueue(priv->wq);
689         ipoib_reap_ah(&priv->ah_reap_task.work);
690 }
691
692 static void ipoib_stop_ah(struct net_device *dev)
693 {
694         struct ipoib_dev_priv *priv = netdev_priv(dev);
695
696         set_bit(IPOIB_STOP_REAPER, &priv->flags);
697         ipoib_flush_ah(dev);
698 }
699
700 static void ipoib_ib_tx_timer_func(unsigned long ctx)
701 {
702         drain_tx_cq((struct net_device *)ctx);
703 }
704
705 int ipoib_ib_dev_open(struct net_device *dev)
706 {
707         struct ipoib_dev_priv *priv = netdev_priv(dev);
708         int ret;
709
710         ipoib_pkey_dev_check_presence(dev);
711
712         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
713                 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
714                            (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
715                 return -1;
716         }
717
718         ret = ipoib_init_qp(dev);
719         if (ret) {
720                 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
721                 return -1;
722         }
723
724         ret = ipoib_ib_post_receives(dev);
725         if (ret) {
726                 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
727                 goto dev_stop;
728         }
729
730         ret = ipoib_cm_dev_open(dev);
731         if (ret) {
732                 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
733                 goto dev_stop;
734         }
735
736         clear_bit(IPOIB_STOP_REAPER, &priv->flags);
737         queue_delayed_work(priv->wq, &priv->ah_reap_task,
738                            round_jiffies_relative(HZ));
739
740         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
741                 napi_enable(&priv->napi);
742
743         return 0;
744 dev_stop:
745         if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
746                 napi_enable(&priv->napi);
747         ipoib_ib_dev_stop(dev);
748         return -1;
749 }
750
751 void ipoib_pkey_dev_check_presence(struct net_device *dev)
752 {
753         struct ipoib_dev_priv *priv = netdev_priv(dev);
754
755         if (!(priv->pkey & 0x7fff) ||
756             ib_find_pkey(priv->ca, priv->port, priv->pkey,
757                          &priv->pkey_index))
758                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
759         else
760                 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
761 }
762
763 int ipoib_ib_dev_up(struct net_device *dev)
764 {
765         struct ipoib_dev_priv *priv = netdev_priv(dev);
766
767         ipoib_pkey_dev_check_presence(dev);
768
769         if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
770                 ipoib_dbg(priv, "PKEY is not assigned.\n");
771                 return 0;
772         }
773
774         set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
775
776         return ipoib_mcast_start_thread(dev);
777 }
778
779 int ipoib_ib_dev_down(struct net_device *dev)
780 {
781         struct ipoib_dev_priv *priv = netdev_priv(dev);
782
783         ipoib_dbg(priv, "downing ib_dev\n");
784
785         clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
786         netif_carrier_off(dev);
787
788         ipoib_mcast_stop_thread(dev);
789         ipoib_mcast_dev_flush(dev);
790
791         ipoib_flush_paths(dev);
792
793         return 0;
794 }
795
796 static int recvs_pending(struct net_device *dev)
797 {
798         struct ipoib_dev_priv *priv = netdev_priv(dev);
799         int pending = 0;
800         int i;
801
802         for (i = 0; i < ipoib_recvq_size; ++i)
803                 if (priv->rx_ring[i].skb)
804                         ++pending;
805
806         return pending;
807 }
808
809 void ipoib_drain_cq(struct net_device *dev)
810 {
811         struct ipoib_dev_priv *priv = netdev_priv(dev);
812         int i, n;
813
814         /*
815          * We call completion handling routines that expect to be
816          * called from the BH-disabled NAPI poll context, so disable
817          * BHs here too.
818          */
819         local_bh_disable();
820
821         do {
822                 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
823                 for (i = 0; i < n; ++i) {
824                         /*
825                          * Convert any successful completions to flush
826                          * errors to avoid passing packets up the
827                          * stack after bringing the device down.
828                          */
829                         if (priv->ibwc[i].status == IB_WC_SUCCESS)
830                                 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
831
832                         if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
833                                 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
834                                         ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
835                                 else
836                                         ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
837                         } else
838                                 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
839                 }
840         } while (n == IPOIB_NUM_WC);
841
842         while (poll_tx(priv))
843                 ; /* nothing */
844
845         local_bh_enable();
846 }
847
848 int ipoib_ib_dev_stop(struct net_device *dev)
849 {
850         struct ipoib_dev_priv *priv = netdev_priv(dev);
851         struct ib_qp_attr qp_attr;
852         unsigned long begin;
853         struct ipoib_tx_buf *tx_req;
854         int i;
855
856         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
857                 napi_disable(&priv->napi);
858
859         ipoib_cm_dev_stop(dev);
860
861         /*
862          * Move our QP to the error state and then reinitialize in
863          * when all work requests have completed or have been flushed.
864          */
865         qp_attr.qp_state = IB_QPS_ERR;
866         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
867                 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
868
869         /* Wait for all sends and receives to complete */
870         begin = jiffies;
871
872         while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
873                 if (time_after(jiffies, begin + 5 * HZ)) {
874                         ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
875                                    priv->tx_head - priv->tx_tail, recvs_pending(dev));
876
877                         /*
878                          * assume the HW is wedged and just free up
879                          * all our pending work requests.
880                          */
881                         while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
882                                 tx_req = &priv->tx_ring[priv->tx_tail &
883                                                         (ipoib_sendq_size - 1)];
884                                 ipoib_dma_unmap_tx(priv, tx_req);
885                                 dev_kfree_skb_any(tx_req->skb);
886                                 ++priv->tx_tail;
887                                 --priv->tx_outstanding;
888                         }
889
890                         for (i = 0; i < ipoib_recvq_size; ++i) {
891                                 struct ipoib_rx_buf *rx_req;
892
893                                 rx_req = &priv->rx_ring[i];
894                                 if (!rx_req->skb)
895                                         continue;
896                                 ipoib_ud_dma_unmap_rx(priv,
897                                                       priv->rx_ring[i].mapping);
898                                 dev_kfree_skb_any(rx_req->skb);
899                                 rx_req->skb = NULL;
900                         }
901
902                         goto timeout;
903                 }
904
905                 ipoib_drain_cq(dev);
906
907                 msleep(1);
908         }
909
910         ipoib_dbg(priv, "All sends and receives done.\n");
911
912 timeout:
913         del_timer_sync(&priv->poll_timer);
914         qp_attr.qp_state = IB_QPS_RESET;
915         if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
916                 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
917
918         ipoib_flush_ah(dev);
919
920         ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
921
922         return 0;
923 }
924
925 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
926 {
927         struct ipoib_dev_priv *priv = netdev_priv(dev);
928
929         priv->ca = ca;
930         priv->port = port;
931         priv->qp = NULL;
932
933         if (ipoib_transport_dev_init(dev, ca)) {
934                 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
935                 return -ENODEV;
936         }
937
938         setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
939                     (unsigned long) dev);
940
941         if (dev->flags & IFF_UP) {
942                 if (ipoib_ib_dev_open(dev)) {
943                         ipoib_transport_dev_cleanup(dev);
944                         return -ENODEV;
945                 }
946         }
947
948         return 0;
949 }
950
951 /*
952  * Takes whatever value which is in pkey index 0 and updates priv->pkey
953  * returns 0 if the pkey value was changed.
954  */
955 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
956 {
957         int result;
958         u16 prev_pkey;
959
960         prev_pkey = priv->pkey;
961         result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
962         if (result) {
963                 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
964                            priv->port, result);
965                 return result;
966         }
967
968         priv->pkey |= 0x8000;
969
970         if (prev_pkey != priv->pkey) {
971                 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
972                           prev_pkey, priv->pkey);
973                 /*
974                  * Update the pkey in the broadcast address, while making sure to set
975                  * the full membership bit, so that we join the right broadcast group.
976                  */
977                 priv->dev->broadcast[8] = priv->pkey >> 8;
978                 priv->dev->broadcast[9] = priv->pkey & 0xff;
979                 return 0;
980         }
981
982         return 1;
983 }
984 /*
985  * returns 0 if pkey value was found in a different slot.
986  */
987 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
988 {
989         u16 old_index = priv->pkey_index;
990
991         priv->pkey_index = 0;
992         ipoib_pkey_dev_check_presence(priv->dev);
993
994         if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
995             (old_index == priv->pkey_index))
996                 return 1;
997         return 0;
998 }
999
1000 /*
1001  * returns true if the device address of the ipoib interface has changed and the
1002  * new address is a valid one (i.e in the gid table), return false otherwise.
1003  */
1004 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)
1005 {
1006         union ib_gid search_gid;
1007         union ib_gid gid0;
1008         union ib_gid *netdev_gid;
1009         int err;
1010         u16 index;
1011         u8 port;
1012         bool ret = false;
1013
1014         netdev_gid = (union ib_gid *)(priv->dev->dev_addr + 4);
1015         if (ib_query_gid(priv->ca, priv->port, 0, &gid0, NULL))
1016                 return false;
1017
1018         netif_addr_lock(priv->dev);
1019
1020         /* The subnet prefix may have changed, update it now so we won't have
1021          * to do it later
1022          */
1023         priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1024         netdev_gid->global.subnet_prefix = gid0.global.subnet_prefix;
1025         search_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1026
1027         search_gid.global.interface_id = priv->local_gid.global.interface_id;
1028
1029         netif_addr_unlock(priv->dev);
1030
1031         err = ib_find_gid(priv->ca, &search_gid, IB_GID_TYPE_IB,
1032                           priv->dev, &port, &index);
1033
1034         netif_addr_lock(priv->dev);
1035
1036         if (search_gid.global.interface_id !=
1037             priv->local_gid.global.interface_id)
1038                 /* There was a change while we were looking up the gid, bail
1039                  * here and let the next work sort this out
1040                  */
1041                 goto out;
1042
1043         /* The next section of code needs some background:
1044          * Per IB spec the port GUID can't change if the HCA is powered on.
1045          * port GUID is the basis for GID at index 0 which is the basis for
1046          * the default device address of a ipoib interface.
1047          *
1048          * so it seems the flow should be:
1049          * if user_changed_dev_addr && gid in gid tbl
1050          *      set bit dev_addr_set
1051          *      return true
1052          * else
1053          *      return false
1054          *
1055          * The issue is that there are devices that don't follow the spec,
1056          * they change the port GUID when the HCA is powered, so in order
1057          * not to break userspace applications, We need to check if the
1058          * user wanted to control the device address and we assume that
1059          * if he sets the device address back to be based on GID index 0,
1060          * he no longer wishs to control it.
1061          *
1062          * If the user doesn't control the the device address,
1063          * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1064          * the port GUID has changed and GID at index 0 has changed
1065          * so we need to change priv->local_gid and priv->dev->dev_addr
1066          * to reflect the new GID.
1067          */
1068         if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
1069                 if (!err && port == priv->port) {
1070                         set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1071                         if (index == 0)
1072                                 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,
1073                                           &priv->flags);
1074                         else
1075                                 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);
1076                         ret = true;
1077                 } else {
1078                         ret = false;
1079                 }
1080         } else {
1081                 if (!err && port == priv->port) {
1082                         ret = true;
1083                 } else {
1084                         if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {
1085                                 memcpy(&priv->local_gid, &gid0,
1086                                        sizeof(priv->local_gid));
1087                                 memcpy(priv->dev->dev_addr + 4, &gid0,
1088                                        sizeof(priv->local_gid));
1089                                 ret = true;
1090                         }
1091                 }
1092         }
1093
1094 out:
1095         netif_addr_unlock(priv->dev);
1096
1097         return ret;
1098 }
1099
1100 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1101                                 enum ipoib_flush_level level,
1102                                 int nesting)
1103 {
1104         struct ipoib_dev_priv *cpriv;
1105         struct net_device *dev = priv->dev;
1106         int result;
1107
1108         down_read_nested(&priv->vlan_rwsem, nesting);
1109
1110         /*
1111          * Flush any child interfaces too -- they might be up even if
1112          * the parent is down.
1113          */
1114         list_for_each_entry(cpriv, &priv->child_intfs, list)
1115                 __ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1116
1117         up_read(&priv->vlan_rwsem);
1118
1119         if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1120             level != IPOIB_FLUSH_HEAVY) {
1121                 /* Make sure the dev_addr is set even if not flushing */
1122                 if (level == IPOIB_FLUSH_LIGHT)
1123                         ipoib_dev_addr_changed_valid(priv);
1124                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1125                 return;
1126         }
1127
1128         if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1129                 /* interface is down. update pkey and leave. */
1130                 if (level == IPOIB_FLUSH_HEAVY) {
1131                         if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1132                                 update_parent_pkey(priv);
1133                         else
1134                                 update_child_pkey(priv);
1135                 } else if (level == IPOIB_FLUSH_LIGHT)
1136                         ipoib_dev_addr_changed_valid(priv);
1137                 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1138                 return;
1139         }
1140
1141         if (level == IPOIB_FLUSH_HEAVY) {
1142                 /* child devices chase their origin pkey value, while non-child
1143                  * (parent) devices should always takes what present in pkey index 0
1144                  */
1145                 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1146                         result = update_child_pkey(priv);
1147                         if (result) {
1148                                 /* restart QP only if P_Key index is changed */
1149                                 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1150                                 return;
1151                         }
1152
1153                 } else {
1154                         result = update_parent_pkey(priv);
1155                         /* restart QP only if P_Key value changed */
1156                         if (result) {
1157                                 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1158                                 return;
1159                         }
1160                 }
1161         }
1162
1163         if (level == IPOIB_FLUSH_LIGHT) {
1164                 ipoib_mark_paths_invalid(dev);
1165                 ipoib_mcast_dev_flush(dev);
1166                 ipoib_flush_ah(dev);
1167         }
1168
1169         if (level >= IPOIB_FLUSH_NORMAL)
1170                 ipoib_ib_dev_down(dev);
1171
1172         if (level == IPOIB_FLUSH_HEAVY) {
1173                 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1174                         ipoib_ib_dev_stop(dev);
1175                 if (ipoib_ib_dev_open(dev) != 0)
1176                         return;
1177                 if (netif_queue_stopped(dev))
1178                         netif_start_queue(dev);
1179         }
1180
1181         /*
1182          * The device could have been brought down between the start and when
1183          * we get here, don't bring it back up if it's not configured up
1184          */
1185         if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1186                 if (level >= IPOIB_FLUSH_NORMAL)
1187                         ipoib_ib_dev_up(dev);
1188                 if (ipoib_dev_addr_changed_valid(priv))
1189                         ipoib_mcast_restart_task(&priv->restart_task);
1190         }
1191 }
1192
1193 void ipoib_ib_dev_flush_light(struct work_struct *work)
1194 {
1195         struct ipoib_dev_priv *priv =
1196                 container_of(work, struct ipoib_dev_priv, flush_light);
1197
1198         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1199 }
1200
1201 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1202 {
1203         struct ipoib_dev_priv *priv =
1204                 container_of(work, struct ipoib_dev_priv, flush_normal);
1205
1206         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1207 }
1208
1209 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1210 {
1211         struct ipoib_dev_priv *priv =
1212                 container_of(work, struct ipoib_dev_priv, flush_heavy);
1213
1214         __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1215 }
1216
1217 void ipoib_ib_dev_cleanup(struct net_device *dev)
1218 {
1219         struct ipoib_dev_priv *priv = netdev_priv(dev);
1220
1221         ipoib_dbg(priv, "cleaning up ib_dev\n");
1222         /*
1223          * We must make sure there are no more (path) completions
1224          * that may wish to touch priv fields that are no longer valid
1225          */
1226         ipoib_flush_paths(dev);
1227
1228         ipoib_mcast_stop_thread(dev);
1229         ipoib_mcast_dev_flush(dev);
1230
1231         /*
1232          * All of our ah references aren't free until after
1233          * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1234          * the neighbor garbage collection is stopped and reaped.
1235          * That should all be done now, so make a final ah flush.
1236          */
1237         ipoib_stop_ah(dev);
1238
1239         ipoib_transport_dev_cleanup(dev);
1240 }
1241
1242