?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'fix/pcm-hwptr' into for-linus
[linux/fpc-iii.git] / drivers / infiniband / ulp / ipoib / ipoib_ib.c
blobe7e5adf84e840e3f13aa54dc1c1036f6d4a81846
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
39 #include <rdma/ib_cache.h>
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, skb_shinfo(skb)->frags[0].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
227 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
228 wr_id, wc->status);
229
230 if (unlikely(wr_id >= ipoib_recvq_size)) {
231 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
232 wr_id, ipoib_recvq_size);
233 return;
234 }
235
236 skb = priv->rx_ring[wr_id].skb;
237
238 if (unlikely(wc->status != IB_WC_SUCCESS)) {
239 if (wc->status != IB_WC_WR_FLUSH_ERR)
240 ipoib_warn(priv, "failed recv event "
241 "(status=%d, wrid=%d vend_err %x)\n",
242 wc->status, wr_id, wc->vendor_err);
243 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
244 dev_kfree_skb_any(skb);
245 priv->rx_ring[wr_id].skb = NULL;
246 return;
247 }
248
249 /*
250 * Drop packets that this interface sent, ie multicast packets
251 * that the HCA has replicated.
252 */
253 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
254 goto repost;
255
256 memcpy(mapping, priv->rx_ring[wr_id].mapping,
257 IPOIB_UD_RX_SG * sizeof *mapping);
258
259 /*
260 * If we can't allocate a new RX buffer, dump
261 * this packet and reuse the old buffer.
262 */
263 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
264 ++dev->stats.rx_dropped;
265 goto repost;
266 }
267
268 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
269 wc->byte_len, wc->slid);
270
271 ipoib_ud_dma_unmap_rx(priv, mapping);
272 ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
273
274 skb_pull(skb, IB_GRH_BYTES);
275
276 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
277 skb_reset_mac_header(skb);
278 skb_pull(skb, IPOIB_ENCAP_LEN);
279
280 dev->last_rx = jiffies;
281 ++dev->stats.rx_packets;
282 dev->stats.rx_bytes += skb->len;
283
284 skb->dev = dev;
285 /* XXX get correct PACKET_ type here */
286 skb->pkt_type = PACKET_HOST;
287
288 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
289 skb->ip_summed = CHECKSUM_UNNECESSARY;
290
291 if (dev->features & NETIF_F_LRO)
292 lro_receive_skb(&priv->lro.lro_mgr, skb, NULL);
293 else
294 netif_receive_skb(skb);
295
296 repost:
297 if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
298 ipoib_warn(priv, "ipoib_ib_post_receive failed "
299 "for buf %d\n", wr_id);
300 }
301
302 static int ipoib_dma_map_tx(struct ib_device *ca,
303 struct ipoib_tx_buf *tx_req)
304 {
305 struct sk_buff *skb = tx_req->skb;
306 u64 *mapping = tx_req->mapping;
307 int i;
308 int off;
309
310 if (skb_headlen(skb)) {
311 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
312 DMA_TO_DEVICE);
313 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
314 return -EIO;
315
316 off = 1;
317 } else
318 off = 0;
319
320 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
321 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
322 mapping[i + off] = ib_dma_map_page(ca, frag->page,
323 frag->page_offset, frag->size,
324 DMA_TO_DEVICE);
325 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
326 goto partial_error;
327 }
328 return 0;
329
330 partial_error:
331 for (; i > 0; --i) {
332 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
333 ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE);
334 }
335
336 if (off)
337 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
338
339 return -EIO;
340 }
341
342 static void ipoib_dma_unmap_tx(struct ib_device *ca,
343 struct ipoib_tx_buf *tx_req)
344 {
345 struct sk_buff *skb = tx_req->skb;
346 u64 *mapping = tx_req->mapping;
347 int i;
348 int off;
349
350 if (skb_headlen(skb)) {
351 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
352 off = 1;
353 } else
354 off = 0;
355
356 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
357 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
358 ib_dma_unmap_page(ca, mapping[i + off], frag->size,
359 DMA_TO_DEVICE);
360 }
361 }
362
363 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
364 {
365 struct ipoib_dev_priv *priv = netdev_priv(dev);
366 unsigned int wr_id = wc->wr_id;
367 struct ipoib_tx_buf *tx_req;
368
369 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
370 wr_id, wc->status);
371
372 if (unlikely(wr_id >= ipoib_sendq_size)) {
373 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
374 wr_id, ipoib_sendq_size);
375 return;
376 }
377
378 tx_req = &priv->tx_ring[wr_id];
379
380 ipoib_dma_unmap_tx(priv->ca, tx_req);
381
382 ++dev->stats.tx_packets;
383 dev->stats.tx_bytes += tx_req->skb->len;
384
385 dev_kfree_skb_any(tx_req->skb);
386
387 ++priv->tx_tail;
388 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
389 netif_queue_stopped(dev) &&
390 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
391 netif_wake_queue(dev);
392
393 if (wc->status != IB_WC_SUCCESS &&
394 wc->status != IB_WC_WR_FLUSH_ERR)
395 ipoib_warn(priv, "failed send event "
396 "(status=%d, wrid=%d vend_err %x)\n",
397 wc->status, wr_id, wc->vendor_err);
398 }
399
400 static int poll_tx(struct ipoib_dev_priv *priv)
401 {
402 int n, i;
403
404 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
405 for (i = 0; i < n; ++i)
406 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
407
408 return n == MAX_SEND_CQE;
409 }
410
411 int ipoib_poll(struct napi_struct *napi, int budget)
412 {
413 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
414 struct net_device *dev = priv->dev;
415 int done;
416 int t;
417 int n, i;
418
419 done = 0;
420
421 poll_more:
422 while (done < budget) {
423 int max = (budget - done);
424
425 t = min(IPOIB_NUM_WC, max);
426 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
427
428 for (i = 0; i < n; i++) {
429 struct ib_wc *wc = priv->ibwc + i;
430
431 if (wc->wr_id & IPOIB_OP_RECV) {
432 ++done;
433 if (wc->wr_id & IPOIB_OP_CM)
434 ipoib_cm_handle_rx_wc(dev, wc);
435 else
436 ipoib_ib_handle_rx_wc(dev, wc);
437 } else
438 ipoib_cm_handle_tx_wc(priv->dev, wc);
439 }
440
441 if (n != t)
442 break;
443 }
444
445 if (done < budget) {
446 if (dev->features & NETIF_F_LRO)
447 lro_flush_all(&priv->lro.lro_mgr);
448
449 napi_complete(napi);
450 if (unlikely(ib_req_notify_cq(priv->recv_cq,
451 IB_CQ_NEXT_COMP |
452 IB_CQ_REPORT_MISSED_EVENTS)) &&
453 napi_reschedule(napi))
454 goto poll_more;
455 }
456
457 return done;
458 }
459
460 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
461 {
462 struct net_device *dev = dev_ptr;
463 struct ipoib_dev_priv *priv = netdev_priv(dev);
464
465 napi_schedule(&priv->napi);
466 }
467
468 static void drain_tx_cq(struct net_device *dev)
469 {
470 struct ipoib_dev_priv *priv = netdev_priv(dev);
471
472 netif_tx_lock(dev);
473 while (poll_tx(priv))
474 ; /* nothing */
475
476 if (netif_queue_stopped(dev))
477 mod_timer(&priv->poll_timer, jiffies + 1);
478
479 netif_tx_unlock(dev);
480 }
481
482 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
483 {
484 struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
485
486 mod_timer(&priv->poll_timer, jiffies);
487 }
488
489 static inline int post_send(struct ipoib_dev_priv *priv,
490 unsigned int wr_id,
491 struct ib_ah *address, u32 qpn,
492 struct ipoib_tx_buf *tx_req,
493 void *head, int hlen)
494 {
495 struct ib_send_wr *bad_wr;
496 int i, off;
497 struct sk_buff *skb = tx_req->skb;
498 skb_frag_t *frags = skb_shinfo(skb)->frags;
499 int nr_frags = skb_shinfo(skb)->nr_frags;
500 u64 *mapping = tx_req->mapping;
501
502 if (skb_headlen(skb)) {
503 priv->tx_sge[0].addr = mapping[0];
504 priv->tx_sge[0].length = skb_headlen(skb);
505 off = 1;
506 } else
507 off = 0;
508
509 for (i = 0; i < nr_frags; ++i) {
510 priv->tx_sge[i + off].addr = mapping[i + off];
511 priv->tx_sge[i + off].length = frags[i].size;
512 }
513 priv->tx_wr.num_sge = nr_frags + off;
514 priv->tx_wr.wr_id = wr_id;
515 priv->tx_wr.wr.ud.remote_qpn = qpn;
516 priv->tx_wr.wr.ud.ah = address;
517
518 if (head) {
519 priv->tx_wr.wr.ud.mss = skb_shinfo(skb)->gso_size;
520 priv->tx_wr.wr.ud.header = head;
521 priv->tx_wr.wr.ud.hlen = hlen;
522 priv->tx_wr.opcode = IB_WR_LSO;
523 } else
524 priv->tx_wr.opcode = IB_WR_SEND;
525
526 return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
527 }
528
529 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
530 struct ipoib_ah *address, u32 qpn)
531 {
532 struct ipoib_dev_priv *priv = netdev_priv(dev);
533 struct ipoib_tx_buf *tx_req;
534 int hlen;
535 void *phead;
536
537 if (skb_is_gso(skb)) {
538 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
539 phead = skb->data;
540 if (unlikely(!skb_pull(skb, hlen))) {
541 ipoib_warn(priv, "linear data too small\n");
542 ++dev->stats.tx_dropped;
543 ++dev->stats.tx_errors;
544 dev_kfree_skb_any(skb);
545 return;
546 }
547 } else {
548 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
549 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
550 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
551 ++dev->stats.tx_dropped;
552 ++dev->stats.tx_errors;
553 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
554 return;
555 }
556 phead = NULL;
557 hlen = 0;
558 }
559
560 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
561 skb->len, address, qpn);
562
563 /*
564 * We put the skb into the tx_ring _before_ we call post_send()
565 * because it's entirely possible that the completion handler will
566 * run before we execute anything after the post_send(). That
567 * means we have to make sure everything is properly recorded and
568 * our state is consistent before we call post_send().
569 */
570 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
571 tx_req->skb = skb;
572 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
573 ++dev->stats.tx_errors;
574 dev_kfree_skb_any(skb);
575 return;
576 }
577
578 if (skb->ip_summed == CHECKSUM_PARTIAL)
579 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
580 else
581 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
582
583 if (++priv->tx_outstanding == ipoib_sendq_size) {
584 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
585 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
586 ipoib_warn(priv, "request notify on send CQ failed\n");
587 netif_stop_queue(dev);
588 }
589
590 if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
591 address->ah, qpn, tx_req, phead, hlen))) {
592 ipoib_warn(priv, "post_send failed\n");
593 ++dev->stats.tx_errors;
594 --priv->tx_outstanding;
595 ipoib_dma_unmap_tx(priv->ca, tx_req);
596 dev_kfree_skb_any(skb);
597 if (netif_queue_stopped(dev))
598 netif_wake_queue(dev);
599 } else {
600 dev->trans_start = jiffies;
601
602 address->last_send = priv->tx_head;
603 ++priv->tx_head;
604 skb_orphan(skb);
605
606 }
607
608 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
609 while (poll_tx(priv))
610 ; /* nothing */
611 }
612
613 static void __ipoib_reap_ah(struct net_device *dev)
614 {
615 struct ipoib_dev_priv *priv = netdev_priv(dev);
616 struct ipoib_ah *ah, *tah;
617 LIST_HEAD(remove_list);
618 unsigned long flags;
619
620 netif_tx_lock_bh(dev);
621 spin_lock_irqsave(&priv->lock, flags);
622
623 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
624 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
625 list_del(&ah->list);
626 ib_destroy_ah(ah->ah);
627 kfree(ah);
628 }
629
630 spin_unlock_irqrestore(&priv->lock, flags);
631 netif_tx_unlock_bh(dev);
632 }
633
634 void ipoib_reap_ah(struct work_struct *work)
635 {
636 struct ipoib_dev_priv *priv =
637 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
638 struct net_device *dev = priv->dev;
639
640 __ipoib_reap_ah(dev);
641
642 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
643 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
644 round_jiffies_relative(HZ));
645 }
646
647 static void ipoib_ib_tx_timer_func(unsigned long ctx)
648 {
649 drain_tx_cq((struct net_device *)ctx);
650 }
651
652 int ipoib_ib_dev_open(struct net_device *dev)
653 {
654 struct ipoib_dev_priv *priv = netdev_priv(dev);
655 int ret;
656
657 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) {
658 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey);
659 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
660 return -1;
661 }
662 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
663
664 ret = ipoib_init_qp(dev);
665 if (ret) {
666 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
667 return -1;
668 }
669
670 ret = ipoib_ib_post_receives(dev);
671 if (ret) {
672 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
673 ipoib_ib_dev_stop(dev, 1);
674 return -1;
675 }
676
677 ret = ipoib_cm_dev_open(dev);
678 if (ret) {
679 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
680 ipoib_ib_dev_stop(dev, 1);
681 return -1;
682 }
683
684 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
685 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
686 round_jiffies_relative(HZ));
687
688 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
689 napi_enable(&priv->napi);
690
691 return 0;
692 }
693
694 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
695 {
696 struct ipoib_dev_priv *priv = netdev_priv(dev);
697 u16 pkey_index = 0;
698
699 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
700 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
701 else
702 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
703 }
704
705 int ipoib_ib_dev_up(struct net_device *dev)
706 {
707 struct ipoib_dev_priv *priv = netdev_priv(dev);
708
709 ipoib_pkey_dev_check_presence(dev);
710
711 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
712 ipoib_dbg(priv, "PKEY is not assigned.\n");
713 return 0;
714 }
715
716 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
717
718 return ipoib_mcast_start_thread(dev);
719 }
720
721 int ipoib_ib_dev_down(struct net_device *dev, int flush)
722 {
723 struct ipoib_dev_priv *priv = netdev_priv(dev);
724
725 ipoib_dbg(priv, "downing ib_dev\n");
726
727 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
728 netif_carrier_off(dev);
729
730 /* Shutdown the P_Key thread if still active */
731 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
732 mutex_lock(&pkey_mutex);
733 set_bit(IPOIB_PKEY_STOP, &priv->flags);
734 cancel_delayed_work(&priv->pkey_poll_task);
735 mutex_unlock(&pkey_mutex);
736 if (flush)
737 flush_workqueue(ipoib_workqueue);
738 }
739
740 ipoib_mcast_stop_thread(dev, flush);
741 ipoib_mcast_dev_flush(dev);
742
743 ipoib_flush_paths(dev);
744
745 return 0;
746 }
747
748 static int recvs_pending(struct net_device *dev)
749 {
750 struct ipoib_dev_priv *priv = netdev_priv(dev);
751 int pending = 0;
752 int i;
753
754 for (i = 0; i < ipoib_recvq_size; ++i)
755 if (priv->rx_ring[i].skb)
756 ++pending;
757
758 return pending;
759 }
760
761 void ipoib_drain_cq(struct net_device *dev)
762 {
763 struct ipoib_dev_priv *priv = netdev_priv(dev);
764 int i, n;
765
766 /*
767 * We call completion handling routines that expect to be
768 * called from the BH-disabled NAPI poll context, so disable
769 * BHs here too.
770 */
771 local_bh_disable();
772
773 do {
774 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
775 for (i = 0; i < n; ++i) {
776 /*
777 * Convert any successful completions to flush
778 * errors to avoid passing packets up the
779 * stack after bringing the device down.
780 */
781 if (priv->ibwc[i].status == IB_WC_SUCCESS)
782 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
783
784 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
785 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
786 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
787 else
788 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
789 } else
790 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
791 }
792 } while (n == IPOIB_NUM_WC);
793
794 while (poll_tx(priv))
795 ; /* nothing */
796
797 local_bh_enable();
798 }
799
800 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
801 {
802 struct ipoib_dev_priv *priv = netdev_priv(dev);
803 struct ib_qp_attr qp_attr;
804 unsigned long begin;
805 struct ipoib_tx_buf *tx_req;
806 int i;
807
808 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
809 napi_disable(&priv->napi);
810
811 ipoib_cm_dev_stop(dev);
812
813 /*
814 * Move our QP to the error state and then reinitialize in
815 * when all work requests have completed or have been flushed.
816 */
817 qp_attr.qp_state = IB_QPS_ERR;
818 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
819 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
820
821 /* Wait for all sends and receives to complete */
822 begin = jiffies;
823
824 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
825 if (time_after(jiffies, begin + 5 * HZ)) {
826 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
827 priv->tx_head - priv->tx_tail, recvs_pending(dev));
828
829 /*
830 * assume the HW is wedged and just free up
831 * all our pending work requests.
832 */
833 while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
834 tx_req = &priv->tx_ring[priv->tx_tail &
835 (ipoib_sendq_size - 1)];
836 ipoib_dma_unmap_tx(priv->ca, tx_req);
837 dev_kfree_skb_any(tx_req->skb);
838 ++priv->tx_tail;
839 --priv->tx_outstanding;
840 }
841
842 for (i = 0; i < ipoib_recvq_size; ++i) {
843 struct ipoib_rx_buf *rx_req;
844
845 rx_req = &priv->rx_ring[i];
846 if (!rx_req->skb)
847 continue;
848 ipoib_ud_dma_unmap_rx(priv,
849 priv->rx_ring[i].mapping);
850 dev_kfree_skb_any(rx_req->skb);
851 rx_req->skb = NULL;
852 }
853
854 goto timeout;
855 }
856
857 ipoib_drain_cq(dev);
858
859 msleep(1);
860 }
861
862 ipoib_dbg(priv, "All sends and receives done.\n");
863
864 timeout:
865 del_timer_sync(&priv->poll_timer);
866 qp_attr.qp_state = IB_QPS_RESET;
867 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
868 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
869
870 /* Wait for all AHs to be reaped */
871 set_bit(IPOIB_STOP_REAPER, &priv->flags);
872 cancel_delayed_work(&priv->ah_reap_task);
873 if (flush)
874 flush_workqueue(ipoib_workqueue);
875
876 begin = jiffies;
877
878 while (!list_empty(&priv->dead_ahs)) {
879 __ipoib_reap_ah(dev);
880
881 if (time_after(jiffies, begin + HZ)) {
882 ipoib_warn(priv, "timing out; will leak address handles\n");
883 break;
884 }
885
886 msleep(1);
887 }
888
889 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
890
891 return 0;
892 }
893
894 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
895 {
896 struct ipoib_dev_priv *priv = netdev_priv(dev);
897
898 priv->ca = ca;
899 priv->port = port;
900 priv->qp = NULL;
901
902 if (ipoib_transport_dev_init(dev, ca)) {
903 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
904 return -ENODEV;
905 }
906
907 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
908 (unsigned long) dev);
909
910 if (dev->flags & IFF_UP) {
911 if (ipoib_ib_dev_open(dev)) {
912 ipoib_transport_dev_cleanup(dev);
913 return -ENODEV;
914 }
915 }
916
917 return 0;
918 }
919
920 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
921 enum ipoib_flush_level level)
922 {
923 struct ipoib_dev_priv *cpriv;
924 struct net_device *dev = priv->dev;
925 u16 new_index;
926
927 mutex_lock(&priv->vlan_mutex);
928
929 /*
930 * Flush any child interfaces too -- they might be up even if
931 * the parent is down.
932 */
933 list_for_each_entry(cpriv, &priv->child_intfs, list)
934 __ipoib_ib_dev_flush(cpriv, level);
935
936 mutex_unlock(&priv->vlan_mutex);
937
938 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) {
939 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
940 return;
941 }
942
943 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
944 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
945 return;
946 }
947
948 if (level == IPOIB_FLUSH_HEAVY) {
949 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) {
950 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
951 ipoib_ib_dev_down(dev, 0);
952 ipoib_ib_dev_stop(dev, 0);
953 if (ipoib_pkey_dev_delay_open(dev))
954 return;
955 }
956
957 /* restart QP only if P_Key index is changed */
958 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
959 new_index == priv->pkey_index) {
960 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
961 return;
962 }
963 priv->pkey_index = new_index;
964 }
965
966 if (level == IPOIB_FLUSH_LIGHT) {
967 ipoib_mark_paths_invalid(dev);
968 ipoib_mcast_dev_flush(dev);
969 }
970
971 if (level >= IPOIB_FLUSH_NORMAL)
972 ipoib_ib_dev_down(dev, 0);
973
974 if (level == IPOIB_FLUSH_HEAVY) {
975 ipoib_ib_dev_stop(dev, 0);
976 ipoib_ib_dev_open(dev);
977 }
978
979 /*
980 * The device could have been brought down between the start and when
981 * we get here, don't bring it back up if it's not configured up
982 */
983 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
984 if (level >= IPOIB_FLUSH_NORMAL)
985 ipoib_ib_dev_up(dev);
986 ipoib_mcast_restart_task(&priv->restart_task);
987 }
988 }
989
990 void ipoib_ib_dev_flush_light(struct work_struct *work)
991 {
992 struct ipoib_dev_priv *priv =
993 container_of(work, struct ipoib_dev_priv, flush_light);
994
995 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
996 }
997
998 void ipoib_ib_dev_flush_normal(struct work_struct *work)
999 {
1000 struct ipoib_dev_priv *priv =
1001 container_of(work, struct ipoib_dev_priv, flush_normal);
1002
1003 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1004 }
1005
1006 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1007 {
1008 struct ipoib_dev_priv *priv =
1009 container_of(work, struct ipoib_dev_priv, flush_heavy);
1010
1011 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1012 }
1013
1014 void ipoib_ib_dev_cleanup(struct net_device *dev)
1015 {
1016 struct ipoib_dev_priv *priv = netdev_priv(dev);
1017
1018 ipoib_dbg(priv, "cleaning up ib_dev\n");
1019
1020 ipoib_mcast_stop_thread(dev, 1);
1021 ipoib_mcast_dev_flush(dev);
1022
1023 ipoib_transport_dev_cleanup(dev);
1024 }
1025
1026 /*
1027 * Delayed P_Key Assigment Interim Support
1028 *
1029 * The following is initial implementation of delayed P_Key assigment
1030 * mechanism. It is using the same approach implemented for the multicast
1031 * group join. The single goal of this implementation is to quickly address
1032 * Bug #2507. This implementation will probably be removed when the P_Key
1033 * change async notification is available.
1034 */
1035
1036 void ipoib_pkey_poll(struct work_struct *work)
1037 {
1038 struct ipoib_dev_priv *priv =
1039 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work);
1040 struct net_device *dev = priv->dev;
1041
1042 ipoib_pkey_dev_check_presence(dev);
1043
1044 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
1045 ipoib_open(dev);
1046 else {
1047 mutex_lock(&pkey_mutex);
1048 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
1049 queue_delayed_work(ipoib_workqueue,
1050 &priv->pkey_poll_task,
1051 HZ);
1052 mutex_unlock(&pkey_mutex);
1053 }
1054 }
1055
1056 int ipoib_pkey_dev_delay_open(struct net_device *dev)
1057 {
1058 struct ipoib_dev_priv *priv = netdev_priv(dev);
1059
1060 /* Look for the interface pkey value in the IB Port P_Key table and */
1061 /* set the interface pkey assigment flag */
1062 ipoib_pkey_dev_check_presence(dev);
1063
1064 /* P_Key value not assigned yet - start polling */
1065 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
1066 mutex_lock(&pkey_mutex);
1067 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
1068 queue_delayed_work(ipoib_workqueue,
1069 &priv->pkey_poll_task,
1070 HZ);
1071 mutex_unlock(&pkey_mutex);
1072 return 1;
1073 }
1074
1075 return 0;
1076 }
Linux with added FPC-III support