?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge remote-tracking branch 'moduleh/module.h-split'
[linux-2.6/next.git] / drivers / infiniband / ulp / ipoib / ipoib_ib.c
blob00435be4a44b00a395d2cf77a0154b99cd02b36c
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/dma-mapping.h>
38 #include <linux/slab.h>
39
40 #include <linux/ip.h>
41 #include <linux/tcp.h>
42
43 #include "ipoib.h"
44
45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
46 static int data_debug_level;
47
48 module_param(data_debug_level, int, 0644);
49 MODULE_PARM_DESC(data_debug_level,
50 "Enable data path debug tracing if > 0");
51 #endif
52
53 static DEFINE_MUTEX(pkey_mutex);
54
55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
56 struct ib_pd *pd, struct ib_ah_attr *attr)
57 {
58 struct ipoib_ah *ah;
59
60 ah = kmalloc(sizeof *ah, GFP_KERNEL);
61 if (!ah)
62 return NULL;
63
64 ah->dev = dev;
65 ah->last_send = 0;
66 kref_init(&ah->ref);
67
68 ah->ah = ib_create_ah(pd, attr);
69 if (IS_ERR(ah->ah)) {
70 kfree(ah);
71 ah = NULL;
72 } else
73 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
74
75 return ah;
76 }
77
78 void ipoib_free_ah(struct kref *kref)
79 {
80 struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
81 struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
82
83 unsigned long flags;
84
85 spin_lock_irqsave(&priv->lock, flags);
86 list_add_tail(&ah->list, &priv->dead_ahs);
87 spin_unlock_irqrestore(&priv->lock, flags);
88 }
89
90 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
91 u64 mapping[IPOIB_UD_RX_SG])
92 {
93 if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
94 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
95 DMA_FROM_DEVICE);
96 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
97 DMA_FROM_DEVICE);
98 } else
99 ib_dma_unmap_single(priv->ca, mapping[0],
100 IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
101 DMA_FROM_DEVICE);
102 }
103
104 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
105 struct sk_buff *skb,
106 unsigned int length)
107 {
108 if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
109 skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
110 unsigned int size;
111 /*
112 * There is only two buffers needed for max_payload = 4K,
113 * first buf size is IPOIB_UD_HEAD_SIZE
114 */
115 skb->tail += IPOIB_UD_HEAD_SIZE;
116 skb->len += length;
117
118 size = length - IPOIB_UD_HEAD_SIZE;
119
120 frag->size = size;
121 skb->data_len += size;
122 skb->truesize += size;
123 } else
124 skb_put(skb, length);
125
126 }
127
128 static int ipoib_ib_post_receive(struct net_device *dev, int id)
129 {
130 struct ipoib_dev_priv *priv = netdev_priv(dev);
131 struct ib_recv_wr *bad_wr;
132 int ret;
133
134 priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
135 priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
136 priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
137
138
139 ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
140 if (unlikely(ret)) {
141 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
142 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
143 dev_kfree_skb_any(priv->rx_ring[id].skb);
144 priv->rx_ring[id].skb = NULL;
145 }
146
147 return ret;
148 }
149
150 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
151 {
152 struct ipoib_dev_priv *priv = netdev_priv(dev);
153 struct sk_buff *skb;
154 int buf_size;
155 u64 *mapping;
156
157 if (ipoib_ud_need_sg(priv->max_ib_mtu))
158 buf_size = IPOIB_UD_HEAD_SIZE;
159 else
160 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
161
162 skb = dev_alloc_skb(buf_size + 4);
163 if (unlikely(!skb))
164 return NULL;
165
166 /*
167 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
168 * header. So we need 4 more bytes to get to 48 and align the
169 * IP header to a multiple of 16.
170 */
171 skb_reserve(skb, 4);
172
173 mapping = priv->rx_ring[id].mapping;
174 mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
175 DMA_FROM_DEVICE);
176 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
177 goto error;
178
179 if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
180 struct page *page = alloc_page(GFP_ATOMIC);
181 if (!page)
182 goto partial_error;
183 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
184 mapping[1] =
185 ib_dma_map_page(priv->ca, page,
186 0, PAGE_SIZE, DMA_FROM_DEVICE);
187 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
188 goto partial_error;
189 }
190
191 priv->rx_ring[id].skb = skb;
192 return skb;
193
194 partial_error:
195 ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
196 error:
197 dev_kfree_skb_any(skb);
198 return NULL;
199 }
200
201 static int ipoib_ib_post_receives(struct net_device *dev)
202 {
203 struct ipoib_dev_priv *priv = netdev_priv(dev);
204 int i;
205
206 for (i = 0; i < ipoib_recvq_size; ++i) {
207 if (!ipoib_alloc_rx_skb(dev, i)) {
208 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
209 return -ENOMEM;
210 }
211 if (ipoib_ib_post_receive(dev, i)) {
212 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
213 return -EIO;
214 }
215 }
216
217 return 0;
218 }
219
220 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
221 {
222 struct ipoib_dev_priv *priv = netdev_priv(dev);
223 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
224 struct sk_buff *skb;
225 u64 mapping[IPOIB_UD_RX_SG];
226 union ib_gid *dgid;
227
228 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
229 wr_id, wc->status);
230
231 if (unlikely(wr_id >= ipoib_recvq_size)) {
232 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
233 wr_id, ipoib_recvq_size);
234 return;
235 }
236
237 skb = priv->rx_ring[wr_id].skb;
238
239 if (unlikely(wc->status != IB_WC_SUCCESS)) {
240 if (wc->status != IB_WC_WR_FLUSH_ERR)
241 ipoib_warn(priv, "failed recv event "
242 "(status=%d, wrid=%d vend_err %x)\n",
243 wc->status, wr_id, wc->vendor_err);
244 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
245 dev_kfree_skb_any(skb);
246 priv->rx_ring[wr_id].skb = NULL;
247 return;
248 }
249
250 /*
251 * Drop packets that this interface sent, ie multicast packets
252 * that the HCA has replicated.
253 */
254 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
255 goto repost;
256
257 memcpy(mapping, priv->rx_ring[wr_id].mapping,
258 IPOIB_UD_RX_SG * sizeof *mapping);
259
260 /*
261 * If we can't allocate a new RX buffer, dump
262 * this packet and reuse the old buffer.
263 */
264 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
265 ++dev->stats.rx_dropped;
266 goto repost;
267 }
268
269 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
270 wc->byte_len, wc->slid);
271
272 ipoib_ud_dma_unmap_rx(priv, mapping);
273 ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
274
275 /* First byte of dgid signals multicast when 0xff */
276 dgid = &((struct ib_grh *)skb->data)->dgid;
277
278 if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
279 skb->pkt_type = PACKET_HOST;
280 else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
281 skb->pkt_type = PACKET_BROADCAST;
282 else
283 skb->pkt_type = PACKET_MULTICAST;
284
285 skb_pull(skb, IB_GRH_BYTES);
286
287 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
288 skb_reset_mac_header(skb);
289 skb_pull(skb, IPOIB_ENCAP_LEN);
290
291 ++dev->stats.rx_packets;
292 dev->stats.rx_bytes += skb->len;
293
294 skb->dev = dev;
295 if ((dev->features & NETIF_F_RXCSUM) && likely(wc->csum_ok))
296 skb->ip_summed = CHECKSUM_UNNECESSARY;
297
298 napi_gro_receive(&priv->napi, skb);
299
300 repost:
301 if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
302 ipoib_warn(priv, "ipoib_ib_post_receive failed "
303 "for buf %d\n", wr_id);
304 }
305
306 static int ipoib_dma_map_tx(struct ib_device *ca,
307 struct ipoib_tx_buf *tx_req)
308 {
309 struct sk_buff *skb = tx_req->skb;
310 u64 *mapping = tx_req->mapping;
311 int i;
312 int off;
313
314 if (skb_headlen(skb)) {
315 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
316 DMA_TO_DEVICE);
317 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
318 return -EIO;
319
320 off = 1;
321 } else
322 off = 0;
323
324 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
325 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
326 mapping[i + off] = ib_dma_map_page(ca,
327 skb_frag_page(frag),
328 frag->page_offset, frag->size,
329 DMA_TO_DEVICE);
330 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
331 goto partial_error;
332 }
333 return 0;
334
335 partial_error:
336 for (; i > 0; --i) {
337 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
338 ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
339 }
340
341 if (off)
342 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
343
344 return -EIO;
345 }
346
347 static void ipoib_dma_unmap_tx(struct ib_device *ca,
348 struct ipoib_tx_buf *tx_req)
349 {
350 struct sk_buff *skb = tx_req->skb;
351 u64 *mapping = tx_req->mapping;
352 int i;
353 int off;
354
355 if (skb_headlen(skb)) {
356 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
357 off = 1;
358 } else
359 off = 0;
360
361 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
362 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
363 ib_dma_unmap_page(ca, mapping[i + off], frag->size,
364 DMA_TO_DEVICE);
365 }
366 }
367
368 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
369 {
370 struct ipoib_dev_priv *priv = netdev_priv(dev);
371 unsigned int wr_id = wc->wr_id;
372 struct ipoib_tx_buf *tx_req;
373
374 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
375 wr_id, wc->status);
376
377 if (unlikely(wr_id >= ipoib_sendq_size)) {
378 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
379 wr_id, ipoib_sendq_size);
380 return;
381 }
382
383 tx_req = &priv->tx_ring[wr_id];
384
385 ipoib_dma_unmap_tx(priv->ca, tx_req);
386
387 ++dev->stats.tx_packets;
388 dev->stats.tx_bytes += tx_req->skb->len;
389
390 dev_kfree_skb_any(tx_req->skb);
391
392 ++priv->tx_tail;
393 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
394 netif_queue_stopped(dev) &&
395 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
396 netif_wake_queue(dev);
397
398 if (wc->status != IB_WC_SUCCESS &&
399 wc->status != IB_WC_WR_FLUSH_ERR)
400 ipoib_warn(priv, "failed send event "
401 "(status=%d, wrid=%d vend_err %x)\n",
402 wc->status, wr_id, wc->vendor_err);
403 }
404
405 static int poll_tx(struct ipoib_dev_priv *priv)
406 {
407 int n, i;
408
409 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
410 for (i = 0; i < n; ++i)
411 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
412
413 return n == MAX_SEND_CQE;
414 }
415
416 int ipoib_poll(struct napi_struct *napi, int budget)
417 {
418 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
419 struct net_device *dev = priv->dev;
420 int done;
421 int t;
422 int n, i;
423
424 done = 0;
425
426 poll_more:
427 while (done < budget) {
428 int max = (budget - done);
429
430 t = min(IPOIB_NUM_WC, max);
431 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
432
433 for (i = 0; i < n; i++) {
434 struct ib_wc *wc = priv->ibwc + i;
435
436 if (wc->wr_id & IPOIB_OP_RECV) {
437 ++done;
438 if (wc->wr_id & IPOIB_OP_CM)
439 ipoib_cm_handle_rx_wc(dev, wc);
440 else
441 ipoib_ib_handle_rx_wc(dev, wc);
442 } else
443 ipoib_cm_handle_tx_wc(priv->dev, wc);
444 }
445
446 if (n != t)
447 break;
448 }
449
450 if (done < budget) {
451 napi_complete(napi);
452 if (unlikely(ib_req_notify_cq(priv->recv_cq,
453 IB_CQ_NEXT_COMP |
454 IB_CQ_REPORT_MISSED_EVENTS)) &&
455 napi_reschedule(napi))
456 goto poll_more;
457 }
458
459 return done;
460 }
461
462 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
463 {
464 struct net_device *dev = dev_ptr;
465 struct ipoib_dev_priv *priv = netdev_priv(dev);
466
467 napi_schedule(&priv->napi);
468 }
469
470 static void drain_tx_cq(struct net_device *dev)
471 {
472 struct ipoib_dev_priv *priv = netdev_priv(dev);
473
474 netif_tx_lock(dev);
475 while (poll_tx(priv))
476 ; /* nothing */
477
478 if (netif_queue_stopped(dev))
479 mod_timer(&priv->poll_timer, jiffies + 1);
480
481 netif_tx_unlock(dev);
482 }
483
484 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
485 {
486 struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
487
488 mod_timer(&priv->poll_timer, jiffies);
489 }
490
491 static inline int post_send(struct ipoib_dev_priv *priv,
492 unsigned int wr_id,
493 struct ib_ah *address, u32 qpn,
494 struct ipoib_tx_buf *tx_req,
495 void *head, int hlen)
496 {
497 struct ib_send_wr *bad_wr;
498 int i, off;
499 struct sk_buff *skb = tx_req->skb;
500 skb_frag_t *frags = skb_shinfo(skb)->frags;
501 int nr_frags = skb_shinfo(skb)->nr_frags;
502 u64 *mapping = tx_req->mapping;
503
504 if (skb_headlen(skb)) {
505 priv->tx_sge[0].addr = mapping[0];
506 priv->tx_sge[0].length = skb_headlen(skb);
507 off = 1;
508 } else
509 off = 0;
510
511 for (i = 0; i < nr_frags; ++i) {
512 priv->tx_sge[i + off].addr = mapping[i + off];
513 priv->tx_sge[i + off].length = frags[i].size;
514 }
515 priv->tx_wr.num_sge = nr_frags + off;
516 priv->tx_wr.wr_id = wr_id;
517 priv->tx_wr.wr.ud.remote_qpn = qpn;
518 priv->tx_wr.wr.ud.ah = address;
519
520 if (head) {
521 priv->tx_wr.wr.ud.mss = skb_shinfo(skb)->gso_size;
522 priv->tx_wr.wr.ud.header = head;
523 priv->tx_wr.wr.ud.hlen = hlen;
524 priv->tx_wr.opcode = IB_WR_LSO;
525 } else
526 priv->tx_wr.opcode = IB_WR_SEND;
527
528 return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
529 }
530
531 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
532 struct ipoib_ah *address, u32 qpn)
533 {
534 struct ipoib_dev_priv *priv = netdev_priv(dev);
535 struct ipoib_tx_buf *tx_req;
536 int hlen, rc;
537 void *phead;
538
539 if (skb_is_gso(skb)) {
540 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
541 phead = skb->data;
542 if (unlikely(!skb_pull(skb, hlen))) {
543 ipoib_warn(priv, "linear data too small\n");
544 ++dev->stats.tx_dropped;
545 ++dev->stats.tx_errors;
546 dev_kfree_skb_any(skb);
547 return;
548 }
549 } else {
550 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
551 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
552 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
553 ++dev->stats.tx_dropped;
554 ++dev->stats.tx_errors;
555 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
556 return;
557 }
558 phead = NULL;
559 hlen = 0;
560 }
561
562 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
563 skb->len, address, qpn);
564
565 /*
566 * We put the skb into the tx_ring _before_ we call post_send()
567 * because it's entirely possible that the completion handler will
568 * run before we execute anything after the post_send(). That
569 * means we have to make sure everything is properly recorded and
570 * our state is consistent before we call post_send().
571 */
572 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
573 tx_req->skb = skb;
574 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
575 ++dev->stats.tx_errors;
576 dev_kfree_skb_any(skb);
577 return;
578 }
579
580 if (skb->ip_summed == CHECKSUM_PARTIAL)
581 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
582 else
583 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
584
585 if (++priv->tx_outstanding == ipoib_sendq_size) {
586 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
587 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
588 ipoib_warn(priv, "request notify on send CQ failed\n");
589 netif_stop_queue(dev);
590 }
591
592 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
593 address->ah, qpn, tx_req, phead, hlen);
594 if (unlikely(rc)) {
595 ipoib_warn(priv, "post_send failed, error %d\n", rc);
596 ++dev->stats.tx_errors;
597 --priv->tx_outstanding;
598 ipoib_dma_unmap_tx(priv->ca, tx_req);
599 dev_kfree_skb_any(skb);
600 if (netif_queue_stopped(dev))
601 netif_wake_queue(dev);
602 } else {
603 dev->trans_start = jiffies;
604
605 address->last_send = priv->tx_head;
606 ++priv->tx_head;
607 skb_orphan(skb);
608
609 }
610
611 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
612 while (poll_tx(priv))
613 ; /* nothing */
614 }
615
616 static void __ipoib_reap_ah(struct net_device *dev)
617 {
618 struct ipoib_dev_priv *priv = netdev_priv(dev);
619 struct ipoib_ah *ah, *tah;
620 LIST_HEAD(remove_list);
621 unsigned long flags;
622
623 netif_tx_lock_bh(dev);
624 spin_lock_irqsave(&priv->lock, flags);
625
626 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
627 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
628 list_del(&ah->list);
629 ib_destroy_ah(ah->ah);
630 kfree(ah);
631 }
632
633 spin_unlock_irqrestore(&priv->lock, flags);
634 netif_tx_unlock_bh(dev);
635 }
636
637 void ipoib_reap_ah(struct work_struct *work)
638 {
639 struct ipoib_dev_priv *priv =
640 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
641 struct net_device *dev = priv->dev;
642
643 __ipoib_reap_ah(dev);
644
645 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
646 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
647 round_jiffies_relative(HZ));
648 }
649
650 static void ipoib_ib_tx_timer_func(unsigned long ctx)
651 {
652 drain_tx_cq((struct net_device *)ctx);
653 }
654
655 int ipoib_ib_dev_open(struct net_device *dev)
656 {
657 struct ipoib_dev_priv *priv = netdev_priv(dev);
658 int ret;
659
660 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
661 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
662 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
663 return -1;
664 }
665 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
666
667 ret = ipoib_init_qp(dev);
668 if (ret) {
669 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
670 return -1;
671 }
672
673 ret = ipoib_ib_post_receives(dev);
674 if (ret) {
675 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
676 ipoib_ib_dev_stop(dev, 1);
677 return -1;
678 }
679
680 ret = ipoib_cm_dev_open(dev);
681 if (ret) {
682 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
683 ipoib_ib_dev_stop(dev, 1);
684 return -1;
685 }
686
687 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
688 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
689 round_jiffies_relative(HZ));
690
691 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
692 napi_enable(&priv->napi);
693
694 return 0;
695 }
696
697 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
698 {
699 struct ipoib_dev_priv *priv = netdev_priv(dev);
700 u16 pkey_index = 0;
701
702 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
703 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
704 else
705 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
706 }
707
708 int ipoib_ib_dev_up(struct net_device *dev)
709 {
710 struct ipoib_dev_priv *priv = netdev_priv(dev);
711
712 ipoib_pkey_dev_check_presence(dev);
713
714 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
715 ipoib_dbg(priv, "PKEY is not assigned.\n");
716 return 0;
717 }
718
719 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
720
721 return ipoib_mcast_start_thread(dev);
722 }
723
724 int ipoib_ib_dev_down(struct net_device *dev, int flush)
725 {
726 struct ipoib_dev_priv *priv = netdev_priv(dev);
727
728 ipoib_dbg(priv, "downing ib_dev\n");
729
730 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
731 netif_carrier_off(dev);
732
733 /* Shutdown the P_Key thread if still active */
734 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
735 mutex_lock(&pkey_mutex);
736 set_bit(IPOIB_PKEY_STOP, &priv->flags);
737 cancel_delayed_work(&priv->pkey_poll_task);
738 mutex_unlock(&pkey_mutex);
739 if (flush)
740 flush_workqueue(ipoib_workqueue);
741 }
742
743 ipoib_mcast_stop_thread(dev, flush);
744 ipoib_mcast_dev_flush(dev);
745
746 ipoib_flush_paths(dev);
747
748 return 0;
749 }
750
751 static int recvs_pending(struct net_device *dev)
752 {
753 struct ipoib_dev_priv *priv = netdev_priv(dev);
754 int pending = 0;
755 int i;
756
757 for (i = 0; i < ipoib_recvq_size; ++i)
758 if (priv->rx_ring[i].skb)
759 ++pending;
760
761 return pending;
762 }
763
764 void ipoib_drain_cq(struct net_device *dev)
765 {
766 struct ipoib_dev_priv *priv = netdev_priv(dev);
767 int i, n;
768
769 /*
770 * We call completion handling routines that expect to be
771 * called from the BH-disabled NAPI poll context, so disable
772 * BHs here too.
773 */
774 local_bh_disable();
775
776 do {
777 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
778 for (i = 0; i < n; ++i) {
779 /*
780 * Convert any successful completions to flush
781 * errors to avoid passing packets up the
782 * stack after bringing the device down.
783 */
784 if (priv->ibwc[i].status == IB_WC_SUCCESS)
785 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
786
787 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
788 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
789 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
790 else
791 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
792 } else
793 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
794 }
795 } while (n == IPOIB_NUM_WC);
796
797 while (poll_tx(priv))
798 ; /* nothing */
799
800 local_bh_enable();
801 }
802
803 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
804 {
805 struct ipoib_dev_priv *priv = netdev_priv(dev);
806 struct ib_qp_attr qp_attr;
807 unsigned long begin;
808 struct ipoib_tx_buf *tx_req;
809 int i;
810
811 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
812 napi_disable(&priv->napi);
813
814 ipoib_cm_dev_stop(dev);
815
816 /*
817 * Move our QP to the error state and then reinitialize in
818 * when all work requests have completed or have been flushed.
819 */
820 qp_attr.qp_state = IB_QPS_ERR;
821 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
822 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
823
824 /* Wait for all sends and receives to complete */
825 begin = jiffies;
826
827 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
828 if (time_after(jiffies, begin + 5 * HZ)) {
829 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
830 priv->tx_head - priv->tx_tail, recvs_pending(dev));
831
832 /*
833 * assume the HW is wedged and just free up
834 * all our pending work requests.
835 */
836 while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
837 tx_req = &priv->tx_ring[priv->tx_tail &
838 (ipoib_sendq_size - 1)];
839 ipoib_dma_unmap_tx(priv->ca, tx_req);
840 dev_kfree_skb_any(tx_req->skb);
841 ++priv->tx_tail;
842 --priv->tx_outstanding;
843 }
844
845 for (i = 0; i < ipoib_recvq_size; ++i) {
846 struct ipoib_rx_buf *rx_req;
847
848 rx_req = &priv->rx_ring[i];
849 if (!rx_req->skb)
850 continue;
851 ipoib_ud_dma_unmap_rx(priv,
852 priv->rx_ring[i].mapping);
853 dev_kfree_skb_any(rx_req->skb);
854 rx_req->skb = NULL;
855 }
856
857 goto timeout;
858 }
859
860 ipoib_drain_cq(dev);
861
862 msleep(1);
863 }
864
865 ipoib_dbg(priv, "All sends and receives done.\n");
866
867 timeout:
868 del_timer_sync(&priv->poll_timer);
869 qp_attr.qp_state = IB_QPS_RESET;
870 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
871 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
872
873 /* Wait for all AHs to be reaped */
874 set_bit(IPOIB_STOP_REAPER, &priv->flags);
875 cancel_delayed_work(&priv->ah_reap_task);
876 if (flush)
877 flush_workqueue(ipoib_workqueue);
878
879 begin = jiffies;
880
881 while (!list_empty(&priv->dead_ahs)) {
882 __ipoib_reap_ah(dev);
883
884 if (time_after(jiffies, begin + HZ)) {
885 ipoib_warn(priv, "timing out; will leak address handles\n");
886 break;
887 }
888
889 msleep(1);
890 }
891
892 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
893
894 return 0;
895 }
896
897 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
898 {
899 struct ipoib_dev_priv *priv = netdev_priv(dev);
900
901 priv->ca = ca;
902 priv->port = port;
903 priv->qp = NULL;
904
905 if (ipoib_transport_dev_init(dev, ca)) {
906 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
907 return -ENODEV;
908 }
909
910 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
911 (unsigned long) dev);
912
913 if (dev->flags & IFF_UP) {
914 if (ipoib_ib_dev_open(dev)) {
915 ipoib_transport_dev_cleanup(dev);
916 return -ENODEV;
917 }
918 }
919
920 return 0;
921 }
922
923 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
924 enum ipoib_flush_level level)
925 {
926 struct ipoib_dev_priv *cpriv;
927 struct net_device *dev = priv->dev;
928 u16 new_index;
929
930 mutex_lock(&priv->vlan_mutex);
931
932 /*
933 * Flush any child interfaces too -- they might be up even if
934 * the parent is down.
935 */
936 list_for_each_entry(cpriv, &priv->child_intfs, list)
937 __ipoib_ib_dev_flush(cpriv, level);
938
939 mutex_unlock(&priv->vlan_mutex);
940
941 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
942 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
943 return;
944 }
945
946 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
947 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
948 return;
949 }
950
951 if (level == IPOIB_FLUSH_HEAVY) {
952 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
953 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
954 ipoib_ib_dev_down(dev, 0);
955 ipoib_ib_dev_stop(dev, 0);
956 if (ipoib_pkey_dev_delay_open(dev))
957 return;
958 }
959
960 /* restart QP only if P_Key index is changed */
961 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
962 new_index == priv->pkey_index) {
963 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
964 return;
965 }
966 priv->pkey_index = new_index;
967 }
968
969 if (level == IPOIB_FLUSH_LIGHT) {
970 ipoib_mark_paths_invalid(dev);
971 ipoib_mcast_dev_flush(dev);
972 }
973
974 if (level >= IPOIB_FLUSH_NORMAL)
975 ipoib_ib_dev_down(dev, 0);
976
977 if (level == IPOIB_FLUSH_HEAVY) {
978 ipoib_ib_dev_stop(dev, 0);
979 ipoib_ib_dev_open(dev);
980 }
981
982 /*
983 * The device could have been brought down between the start and when
984 * we get here, don't bring it back up if it's not configured up
985 */
986 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
987 if (level >= IPOIB_FLUSH_NORMAL)
988 ipoib_ib_dev_up(dev);
989 ipoib_mcast_restart_task(&priv->restart_task);
990 }
991 }
992
993 void ipoib_ib_dev_flush_light(struct work_struct *work)
994 {
995 struct ipoib_dev_priv *priv =
996 container_of(work, struct ipoib_dev_priv, flush_light);
997
998 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
999 }
1000
1001 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1002 {
1003 struct ipoib_dev_priv *priv =
1004 container_of(work, struct ipoib_dev_priv, flush_normal);
1005
1006 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1007 }
1008
1009 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1010 {
1011 struct ipoib_dev_priv *priv =
1012 container_of(work, struct ipoib_dev_priv, flush_heavy);
1013
1014 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1015 }
1016
1017 void ipoib_ib_dev_cleanup(struct net_device *dev)
1018 {
1019 struct ipoib_dev_priv *priv = netdev_priv(dev);
1020
1021 ipoib_dbg(priv, "cleaning up ib_dev\n");
1022
1023 ipoib_mcast_stop_thread(dev, 1);
1024 ipoib_mcast_dev_flush(dev);
1025
1026 ipoib_transport_dev_cleanup(dev);
1027 }
1028
1029 /*
1030 * Delayed P_Key Assigment Interim Support
1031 *
1032 * The following is initial implementation of delayed P_Key assigment
1033 * mechanism. It is using the same approach implemented for the multicast
1034 * group join. The single goal of this implementation is to quickly address
1035 * Bug #2507. This implementation will probably be removed when the P_Key
1036 * change async notification is available.
1037 */
1038
1039 void ipoib_pkey_poll(struct work_struct *work)
1040 {
1041 struct ipoib_dev_priv *priv =
1042 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1043 struct net_device *dev = priv->dev;
1044
1045 ipoib_pkey_dev_check_presence(dev);
1046
1047 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1048 ipoib_open(dev);
1049 else {
1050 mutex_lock(&pkey_mutex);
1051 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1052 queue_delayed_work(ipoib_workqueue,
1053 &priv->pkey_poll_task,
1054 HZ);
1055 mutex_unlock(&pkey_mutex);
1056 }
1057 }
1058
1059 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1060 {
1061 struct ipoib_dev_priv *priv = netdev_priv(dev);
1062
1063 /* Look for the interface pkey value in the IB Port P_Key table and */
1064 /* set the interface pkey assigment flag */
1065 ipoib_pkey_dev_check_presence(dev);
1066
1067 /* P_Key value not assigned yet - start polling */
1068 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1069 mutex_lock(&pkey_mutex);
1070 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1071 queue_delayed_work(ipoib_workqueue,
1072 &priv->pkey_poll_task,
1073 HZ);
1074 mutex_unlock(&pkey_mutex);
1075 return 1;
1076 }
1077
1078 return 0;
1079 }
linux-next