?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
powerpc/mm/4k: don't allocate larger pmd page table for 4k
[linux/fpc-iii.git] / drivers / infiniband / ulp / ipoib / ipoib_ib.c
blob5038f9d2d753f06d7a9c5f26d9f2e864d76ca570
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_HARD_LEN);
132 if (unlikely(!skb))
133 return NULL;
134
135 /*
136 * the IP header will be at IPOIP_HARD_LEN + IB_GRH_BYTES, that is
137 * 64 bytes aligned
138 */
139 skb_reserve(skb, sizeof(struct ipoib_pseudo_header));
140
141 mapping = priv->rx_ring[id].mapping;
142 mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
143 DMA_FROM_DEVICE);
144 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
145 goto error;
146
147 priv->rx_ring[id].skb = skb;
148 return skb;
149 error:
150 dev_kfree_skb_any(skb);
151 return NULL;
152 }
153
154 static int ipoib_ib_post_receives(struct net_device *dev)
155 {
156 struct ipoib_dev_priv *priv = netdev_priv(dev);
157 int i;
158
159 for (i = 0; i < ipoib_recvq_size; ++i) {
160 if (!ipoib_alloc_rx_skb(dev, i)) {
161 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
162 return -ENOMEM;
163 }
164 if (ipoib_ib_post_receive(dev, i)) {
165 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
166 return -EIO;
167 }
168 }
169
170 return 0;
171 }
172
173 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
174 {
175 struct ipoib_dev_priv *priv = netdev_priv(dev);
176 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
177 struct sk_buff *skb;
178 u64 mapping[IPOIB_UD_RX_SG];
179 union ib_gid *dgid;
180 union ib_gid *sgid;
181
182 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
183 wr_id, wc->status);
184
185 if (unlikely(wr_id >= ipoib_recvq_size)) {
186 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
187 wr_id, ipoib_recvq_size);
188 return;
189 }
190
191 skb = priv->rx_ring[wr_id].skb;
192
193 if (unlikely(wc->status != IB_WC_SUCCESS)) {
194 if (wc->status != IB_WC_WR_FLUSH_ERR)
195 ipoib_warn(priv, "failed recv event "
196 "(status=%d, wrid=%d vend_err %x)\n",
197 wc->status, wr_id, wc->vendor_err);
198 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
199 dev_kfree_skb_any(skb);
200 priv->rx_ring[wr_id].skb = NULL;
201 return;
202 }
203
204 memcpy(mapping, priv->rx_ring[wr_id].mapping,
205 IPOIB_UD_RX_SG * sizeof *mapping);
206
207 /*
208 * If we can't allocate a new RX buffer, dump
209 * this packet and reuse the old buffer.
210 */
211 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
212 ++dev->stats.rx_dropped;
213 goto repost;
214 }
215
216 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
217 wc->byte_len, wc->slid);
218
219 ipoib_ud_dma_unmap_rx(priv, mapping);
220
221 skb_put(skb, wc->byte_len);
222
223 /* First byte of dgid signals multicast when 0xff */
224 dgid = &((struct ib_grh *)skb->data)->dgid;
225
226 if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
227 skb->pkt_type = PACKET_HOST;
228 else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
229 skb->pkt_type = PACKET_BROADCAST;
230 else
231 skb->pkt_type = PACKET_MULTICAST;
232
233 sgid = &((struct ib_grh *)skb->data)->sgid;
234
235 /*
236 * Drop packets that this interface sent, ie multicast packets
237 * that the HCA has replicated.
238 */
239 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) {
240 int need_repost = 1;
241
242 if ((wc->wc_flags & IB_WC_GRH) &&
243 sgid->global.interface_id != priv->local_gid.global.interface_id)
244 need_repost = 0;
245
246 if (need_repost) {
247 dev_kfree_skb_any(skb);
248 goto repost;
249 }
250 }
251
252 skb_pull(skb, IB_GRH_BYTES);
253
254 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
255 skb_add_pseudo_hdr(skb);
256
257 ++dev->stats.rx_packets;
258 dev->stats.rx_bytes += skb->len;
259
260 skb->dev = dev;
261 if ((dev->features & NETIF_F_RXCSUM) &&
262 likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
263 skb->ip_summed = CHECKSUM_UNNECESSARY;
264
265 napi_gro_receive(&priv->napi, skb);
266
267 repost:
268 if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
269 ipoib_warn(priv, "ipoib_ib_post_receive failed "
270 "for buf %d\n", wr_id);
271 }
272
273 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
274 {
275 struct sk_buff *skb = tx_req->skb;
276 u64 *mapping = tx_req->mapping;
277 int i;
278 int off;
279
280 if (skb_headlen(skb)) {
281 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
282 DMA_TO_DEVICE);
283 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
284 return -EIO;
285
286 off = 1;
287 } else
288 off = 0;
289
290 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
291 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
292 mapping[i + off] = ib_dma_map_page(ca,
293 skb_frag_page(frag),
294 frag->page_offset, skb_frag_size(frag),
295 DMA_TO_DEVICE);
296 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
297 goto partial_error;
298 }
299 return 0;
300
301 partial_error:
302 for (; i > 0; --i) {
303 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
304
305 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
306 }
307
308 if (off)
309 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
310
311 return -EIO;
312 }
313
314 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
315 struct ipoib_tx_buf *tx_req)
316 {
317 struct sk_buff *skb = tx_req->skb;
318 u64 *mapping = tx_req->mapping;
319 int i;
320 int off;
321
322 if (skb_headlen(skb)) {
323 ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
324 DMA_TO_DEVICE);
325 off = 1;
326 } else
327 off = 0;
328
329 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
330 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
331
332 ib_dma_unmap_page(priv->ca, mapping[i + off],
333 skb_frag_size(frag), DMA_TO_DEVICE);
334 }
335 }
336
337 /*
338 * As the result of a completion error the QP Can be transferred to SQE states.
339 * The function checks if the (send)QP is in SQE state and
340 * moves it back to RTS state, that in order to have it functional again.
341 */
342 static void ipoib_qp_state_validate_work(struct work_struct *work)
343 {
344 struct ipoib_qp_state_validate *qp_work =
345 container_of(work, struct ipoib_qp_state_validate, work);
346
347 struct ipoib_dev_priv *priv = qp_work->priv;
348 struct ib_qp_attr qp_attr;
349 struct ib_qp_init_attr query_init_attr;
350 int ret;
351
352 ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
353 if (ret) {
354 ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
355 __func__, ret);
356 goto free_res;
357 }
358 pr_info("%s: QP: 0x%x is in state: %d\n",
359 __func__, priv->qp->qp_num, qp_attr.qp_state);
360
361 /* currently support only in SQE->RTS transition*/
362 if (qp_attr.qp_state == IB_QPS_SQE) {
363 qp_attr.qp_state = IB_QPS_RTS;
364
365 ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
366 if (ret) {
367 pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
368 ret, priv->qp->qp_num);
369 goto free_res;
370 }
371 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
372 __func__, priv->qp->qp_num);
373 } else {
374 pr_warn("QP (%d) will stay in state: %d\n",
375 priv->qp->qp_num, qp_attr.qp_state);
376 }
377
378 free_res:
379 kfree(qp_work);
380 }
381
382 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
383 {
384 struct ipoib_dev_priv *priv = netdev_priv(dev);
385 unsigned int wr_id = wc->wr_id;
386 struct ipoib_tx_buf *tx_req;
387
388 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
389 wr_id, wc->status);
390
391 if (unlikely(wr_id >= ipoib_sendq_size)) {
392 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
393 wr_id, ipoib_sendq_size);
394 return;
395 }
396
397 tx_req = &priv->tx_ring[wr_id];
398
399 ipoib_dma_unmap_tx(priv, tx_req);
400
401 ++dev->stats.tx_packets;
402 dev->stats.tx_bytes += tx_req->skb->len;
403
404 dev_kfree_skb_any(tx_req->skb);
405
406 ++priv->tx_tail;
407 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
408 netif_queue_stopped(dev) &&
409 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
410 netif_wake_queue(dev);
411
412 if (wc->status != IB_WC_SUCCESS &&
413 wc->status != IB_WC_WR_FLUSH_ERR) {
414 struct ipoib_qp_state_validate *qp_work;
415 ipoib_warn(priv, "failed send event "
416 "(status=%d, wrid=%d vend_err %x)\n",
417 wc->status, wr_id, wc->vendor_err);
418 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
419 if (!qp_work)
420 return;
421
422 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
423 qp_work->priv = priv;
424 queue_work(priv->wq, &qp_work->work);
425 }
426 }
427
428 static int poll_tx(struct ipoib_dev_priv *priv)
429 {
430 int n, i;
431
432 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
433 for (i = 0; i < n; ++i)
434 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
435
436 return n == MAX_SEND_CQE;
437 }
438
439 int ipoib_poll(struct napi_struct *napi, int budget)
440 {
441 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
442 struct net_device *dev = priv->dev;
443 int done;
444 int t;
445 int n, i;
446
447 done = 0;
448
449 poll_more:
450 while (done < budget) {
451 int max = (budget - done);
452
453 t = min(IPOIB_NUM_WC, max);
454 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
455
456 for (i = 0; i < n; i++) {
457 struct ib_wc *wc = priv->ibwc + i;
458
459 if (wc->wr_id & IPOIB_OP_RECV) {
460 ++done;
461 if (wc->wr_id & IPOIB_OP_CM)
462 ipoib_cm_handle_rx_wc(dev, wc);
463 else
464 ipoib_ib_handle_rx_wc(dev, wc);
465 } else
466 ipoib_cm_handle_tx_wc(priv->dev, wc);
467 }
468
469 if (n != t)
470 break;
471 }
472
473 if (done < budget) {
474 napi_complete(napi);
475 if (unlikely(ib_req_notify_cq(priv->recv_cq,
476 IB_CQ_NEXT_COMP |
477 IB_CQ_REPORT_MISSED_EVENTS)) &&
478 napi_reschedule(napi))
479 goto poll_more;
480 }
481
482 return done;
483 }
484
485 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
486 {
487 struct net_device *dev = dev_ptr;
488 struct ipoib_dev_priv *priv = netdev_priv(dev);
489
490 napi_schedule(&priv->napi);
491 }
492
493 static void drain_tx_cq(struct net_device *dev)
494 {
495 struct ipoib_dev_priv *priv = netdev_priv(dev);
496
497 netif_tx_lock(dev);
498 while (poll_tx(priv))
499 ; /* nothing */
500
501 if (netif_queue_stopped(dev))
502 mod_timer(&priv->poll_timer, jiffies + 1);
503
504 netif_tx_unlock(dev);
505 }
506
507 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
508 {
509 struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
510
511 mod_timer(&priv->poll_timer, jiffies);
512 }
513
514 static inline int post_send(struct ipoib_dev_priv *priv,
515 unsigned int wr_id,
516 struct ib_ah *address, u32 qpn,
517 struct ipoib_tx_buf *tx_req,
518 void *head, int hlen)
519 {
520 struct ib_send_wr *bad_wr;
521 struct sk_buff *skb = tx_req->skb;
522
523 ipoib_build_sge(priv, tx_req);
524
525 priv->tx_wr.wr.wr_id = wr_id;
526 priv->tx_wr.remote_qpn = qpn;
527 priv->tx_wr.ah = address;
528
529 if (head) {
530 priv->tx_wr.mss = skb_shinfo(skb)->gso_size;
531 priv->tx_wr.header = head;
532 priv->tx_wr.hlen = hlen;
533 priv->tx_wr.wr.opcode = IB_WR_LSO;
534 } else
535 priv->tx_wr.wr.opcode = IB_WR_SEND;
536
537 return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
538 }
539
540 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
541 struct ipoib_ah *address, u32 qpn)
542 {
543 struct ipoib_dev_priv *priv = netdev_priv(dev);
544 struct ipoib_tx_buf *tx_req;
545 int hlen, rc;
546 void *phead;
547 unsigned usable_sge = priv->max_send_sge - !!skb_headlen(skb);
548
549 if (skb_is_gso(skb)) {
550 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
551 phead = skb->data;
552 if (unlikely(!skb_pull(skb, hlen))) {
553 ipoib_warn(priv, "linear data too small\n");
554 ++dev->stats.tx_dropped;
555 ++dev->stats.tx_errors;
556 dev_kfree_skb_any(skb);
557 return;
558 }
559 } else {
560 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
561 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
562 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
563 ++dev->stats.tx_dropped;
564 ++dev->stats.tx_errors;
565 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
566 return;
567 }
568 phead = NULL;
569 hlen = 0;
570 }
571 if (skb_shinfo(skb)->nr_frags > usable_sge) {
572 if (skb_linearize(skb) < 0) {
573 ipoib_warn(priv, "skb could not be linearized\n");
574 ++dev->stats.tx_dropped;
575 ++dev->stats.tx_errors;
576 dev_kfree_skb_any(skb);
577 return;
578 }
579 /* Does skb_linearize return ok without reducing nr_frags? */
580 if (skb_shinfo(skb)->nr_frags > usable_sge) {
581 ipoib_warn(priv, "too many frags after skb linearize\n");
582 ++dev->stats.tx_dropped;
583 ++dev->stats.tx_errors;
584 dev_kfree_skb_any(skb);
585 return;
586 }
587 }
588
589 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
590 skb->len, address, qpn);
591
592 /*
593 * We put the skb into the tx_ring _before_ we call post_send()
594 * because it's entirely possible that the completion handler will
595 * run before we execute anything after the post_send(). That
596 * means we have to make sure everything is properly recorded and
597 * our state is consistent before we call post_send().
598 */
599 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
600 tx_req->skb = skb;
601 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
602 ++dev->stats.tx_errors;
603 dev_kfree_skb_any(skb);
604 return;
605 }
606
607 if (skb->ip_summed == CHECKSUM_PARTIAL)
608 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
609 else
610 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
611
612 if (++priv->tx_outstanding == ipoib_sendq_size) {
613 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
614 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
615 ipoib_warn(priv, "request notify on send CQ failed\n");
616 netif_stop_queue(dev);
617 }
618
619 skb_orphan(skb);
620 skb_dst_drop(skb);
621
622 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
623 address->ah, qpn, tx_req, phead, hlen);
624 if (unlikely(rc)) {
625 ipoib_warn(priv, "post_send failed, error %d\n", rc);
626 ++dev->stats.tx_errors;
627 --priv->tx_outstanding;
628 ipoib_dma_unmap_tx(priv, tx_req);
629 dev_kfree_skb_any(skb);
630 if (netif_queue_stopped(dev))
631 netif_wake_queue(dev);
632 } else {
633 netif_trans_update(dev);
634
635 address->last_send = priv->tx_head;
636 ++priv->tx_head;
637 }
638
639 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
640 while (poll_tx(priv))
641 ; /* nothing */
642 }
643
644 static void __ipoib_reap_ah(struct net_device *dev)
645 {
646 struct ipoib_dev_priv *priv = netdev_priv(dev);
647 struct ipoib_ah *ah, *tah;
648 LIST_HEAD(remove_list);
649 unsigned long flags;
650
651 netif_tx_lock_bh(dev);
652 spin_lock_irqsave(&priv->lock, flags);
653
654 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
655 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
656 list_del(&ah->list);
657 ib_destroy_ah(ah->ah);
658 kfree(ah);
659 }
660
661 spin_unlock_irqrestore(&priv->lock, flags);
662 netif_tx_unlock_bh(dev);
663 }
664
665 void ipoib_reap_ah(struct work_struct *work)
666 {
667 struct ipoib_dev_priv *priv =
668 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
669 struct net_device *dev = priv->dev;
670
671 __ipoib_reap_ah(dev);
672
673 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
674 queue_delayed_work(priv->wq, &priv->ah_reap_task,
675 round_jiffies_relative(HZ));
676 }
677
678 static void ipoib_flush_ah(struct net_device *dev)
679 {
680 struct ipoib_dev_priv *priv = netdev_priv(dev);
681
682 cancel_delayed_work(&priv->ah_reap_task);
683 flush_workqueue(priv->wq);
684 ipoib_reap_ah(&priv->ah_reap_task.work);
685 }
686
687 static void ipoib_stop_ah(struct net_device *dev)
688 {
689 struct ipoib_dev_priv *priv = netdev_priv(dev);
690
691 set_bit(IPOIB_STOP_REAPER, &priv->flags);
692 ipoib_flush_ah(dev);
693 }
694
695 static void ipoib_ib_tx_timer_func(unsigned long ctx)
696 {
697 drain_tx_cq((struct net_device *)ctx);
698 }
699
700 int ipoib_ib_dev_open(struct net_device *dev)
701 {
702 struct ipoib_dev_priv *priv = netdev_priv(dev);
703 int ret;
704
705 ipoib_pkey_dev_check_presence(dev);
706
707 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
708 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
709 (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
710 return -1;
711 }
712
713 ret = ipoib_init_qp(dev);
714 if (ret) {
715 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
716 return -1;
717 }
718
719 ret = ipoib_ib_post_receives(dev);
720 if (ret) {
721 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
722 goto dev_stop;
723 }
724
725 ret = ipoib_cm_dev_open(dev);
726 if (ret) {
727 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
728 goto dev_stop;
729 }
730
731 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
732 queue_delayed_work(priv->wq, &priv->ah_reap_task,
733 round_jiffies_relative(HZ));
734
735 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
736 napi_enable(&priv->napi);
737
738 return 0;
739 dev_stop:
740 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
741 napi_enable(&priv->napi);
742 ipoib_ib_dev_stop(dev);
743 return -1;
744 }
745
746 void ipoib_pkey_dev_check_presence(struct net_device *dev)
747 {
748 struct ipoib_dev_priv *priv = netdev_priv(dev);
749
750 if (!(priv->pkey & 0x7fff) ||
751 ib_find_pkey(priv->ca, priv->port, priv->pkey,
752 &priv->pkey_index))
753 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
754 else
755 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
756 }
757
758 int ipoib_ib_dev_up(struct net_device *dev)
759 {
760 struct ipoib_dev_priv *priv = netdev_priv(dev);
761
762 ipoib_pkey_dev_check_presence(dev);
763
764 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
765 ipoib_dbg(priv, "PKEY is not assigned.\n");
766 return 0;
767 }
768
769 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
770
771 return ipoib_mcast_start_thread(dev);
772 }
773
774 int ipoib_ib_dev_down(struct net_device *dev)
775 {
776 struct ipoib_dev_priv *priv = netdev_priv(dev);
777
778 ipoib_dbg(priv, "downing ib_dev\n");
779
780 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
781 netif_carrier_off(dev);
782
783 ipoib_mcast_stop_thread(dev);
784 ipoib_mcast_dev_flush(dev);
785
786 ipoib_flush_paths(dev);
787
788 return 0;
789 }
790
791 static int recvs_pending(struct net_device *dev)
792 {
793 struct ipoib_dev_priv *priv = netdev_priv(dev);
794 int pending = 0;
795 int i;
796
797 for (i = 0; i < ipoib_recvq_size; ++i)
798 if (priv->rx_ring[i].skb)
799 ++pending;
800
801 return pending;
802 }
803
804 void ipoib_drain_cq(struct net_device *dev)
805 {
806 struct ipoib_dev_priv *priv = netdev_priv(dev);
807 int i, n;
808
809 /*
810 * We call completion handling routines that expect to be
811 * called from the BH-disabled NAPI poll context, so disable
812 * BHs here too.
813 */
814 local_bh_disable();
815
816 do {
817 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
818 for (i = 0; i < n; ++i) {
819 /*
820 * Convert any successful completions to flush
821 * errors to avoid passing packets up the
822 * stack after bringing the device down.
823 */
824 if (priv->ibwc[i].status == IB_WC_SUCCESS)
825 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
826
827 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
828 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
829 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
830 else
831 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
832 } else
833 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
834 }
835 } while (n == IPOIB_NUM_WC);
836
837 while (poll_tx(priv))
838 ; /* nothing */
839
840 local_bh_enable();
841 }
842
843 int ipoib_ib_dev_stop(struct net_device *dev)
844 {
845 struct ipoib_dev_priv *priv = netdev_priv(dev);
846 struct ib_qp_attr qp_attr;
847 unsigned long begin;
848 struct ipoib_tx_buf *tx_req;
849 int i;
850
851 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
852 napi_disable(&priv->napi);
853
854 ipoib_cm_dev_stop(dev);
855
856 /*
857 * Move our QP to the error state and then reinitialize in
858 * when all work requests have completed or have been flushed.
859 */
860 qp_attr.qp_state = IB_QPS_ERR;
861 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
862 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
863
864 /* Wait for all sends and receives to complete */
865 begin = jiffies;
866
867 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
868 if (time_after(jiffies, begin + 5 * HZ)) {
869 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
870 priv->tx_head - priv->tx_tail, recvs_pending(dev));
871
872 /*
873 * assume the HW is wedged and just free up
874 * all our pending work requests.
875 */
876 while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
877 tx_req = &priv->tx_ring[priv->tx_tail &
878 (ipoib_sendq_size - 1)];
879 ipoib_dma_unmap_tx(priv, tx_req);
880 dev_kfree_skb_any(tx_req->skb);
881 ++priv->tx_tail;
882 --priv->tx_outstanding;
883 }
884
885 for (i = 0; i < ipoib_recvq_size; ++i) {
886 struct ipoib_rx_buf *rx_req;
887
888 rx_req = &priv->rx_ring[i];
889 if (!rx_req->skb)
890 continue;
891 ipoib_ud_dma_unmap_rx(priv,
892 priv->rx_ring[i].mapping);
893 dev_kfree_skb_any(rx_req->skb);
894 rx_req->skb = NULL;
895 }
896
897 goto timeout;
898 }
899
900 ipoib_drain_cq(dev);
901
902 msleep(1);
903 }
904
905 ipoib_dbg(priv, "All sends and receives done.\n");
906
907 timeout:
908 del_timer_sync(&priv->poll_timer);
909 qp_attr.qp_state = IB_QPS_RESET;
910 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
911 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
912
913 ipoib_flush_ah(dev);
914
915 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
916
917 return 0;
918 }
919
920 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
921 {
922 struct ipoib_dev_priv *priv = netdev_priv(dev);
923
924 priv->ca = ca;
925 priv->port = port;
926 priv->qp = NULL;
927
928 if (ipoib_transport_dev_init(dev, ca)) {
929 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
930 return -ENODEV;
931 }
932
933 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
934 (unsigned long) dev);
935
936 if (dev->flags & IFF_UP) {
937 if (ipoib_ib_dev_open(dev)) {
938 ipoib_transport_dev_cleanup(dev);
939 return -ENODEV;
940 }
941 }
942
943 return 0;
944 }
945
946 /*
947 * Takes whatever value which is in pkey index 0 and updates priv->pkey
948 * returns 0 if the pkey value was changed.
949 */
950 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
951 {
952 int result;
953 u16 prev_pkey;
954
955 prev_pkey = priv->pkey;
956 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
957 if (result) {
958 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
959 priv->port, result);
960 return result;
961 }
962
963 priv->pkey |= 0x8000;
964
965 if (prev_pkey != priv->pkey) {
966 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
967 prev_pkey, priv->pkey);
968 /*
969 * Update the pkey in the broadcast address, while making sure to set
970 * the full membership bit, so that we join the right broadcast group.
971 */
972 priv->dev->broadcast[8] = priv->pkey >> 8;
973 priv->dev->broadcast[9] = priv->pkey & 0xff;
974 return 0;
975 }
976
977 return 1;
978 }
979 /*
980 * returns 0 if pkey value was found in a different slot.
981 */
982 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
983 {
984 u16 old_index = priv->pkey_index;
985
986 priv->pkey_index = 0;
987 ipoib_pkey_dev_check_presence(priv->dev);
988
989 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
990 (old_index == priv->pkey_index))
991 return 1;
992 return 0;
993 }
994
995 /*
996 * returns true if the device address of the ipoib interface has changed and the
997 * new address is a valid one (i.e in the gid table), return false otherwise.
998 */
999 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)
1000 {
1001 union ib_gid search_gid;
1002 union ib_gid gid0;
1003 union ib_gid *netdev_gid;
1004 int err;
1005 u16 index;
1006 u8 port;
1007 bool ret = false;
1008
1009 netdev_gid = (union ib_gid *)(priv->dev->dev_addr + 4);
1010 if (ib_query_gid(priv->ca, priv->port, 0, &gid0, NULL))
1011 return false;
1012
1013 netif_addr_lock_bh(priv->dev);
1014
1015 /* The subnet prefix may have changed, update it now so we won't have
1016 * to do it later
1017 */
1018 priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1019 netdev_gid->global.subnet_prefix = gid0.global.subnet_prefix;
1020 search_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1021
1022 search_gid.global.interface_id = priv->local_gid.global.interface_id;
1023
1024 netif_addr_unlock_bh(priv->dev);
1025
1026 err = ib_find_gid(priv->ca, &search_gid, IB_GID_TYPE_IB,
1027 priv->dev, &port, &index);
1028
1029 netif_addr_lock_bh(priv->dev);
1030
1031 if (search_gid.global.interface_id !=
1032 priv->local_gid.global.interface_id)
1033 /* There was a change while we were looking up the gid, bail
1034 * here and let the next work sort this out
1035 */
1036 goto out;
1037
1038 /* The next section of code needs some background:
1039 * Per IB spec the port GUID can't change if the HCA is powered on.
1040 * port GUID is the basis for GID at index 0 which is the basis for
1041 * the default device address of a ipoib interface.
1042 *
1043 * so it seems the flow should be:
1044 * if user_changed_dev_addr && gid in gid tbl
1045 * set bit dev_addr_set
1046 * return true
1047 * else
1048 * return false
1049 *
1050 * The issue is that there are devices that don't follow the spec,
1051 * they change the port GUID when the HCA is powered, so in order
1052 * not to break userspace applications, We need to check if the
1053 * user wanted to control the device address and we assume that
1054 * if he sets the device address back to be based on GID index 0,
1055 * he no longer wishs to control it.
1056 *
1057 * If the user doesn't control the the device address,
1058 * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1059 * the port GUID has changed and GID at index 0 has changed
1060 * so we need to change priv->local_gid and priv->dev->dev_addr
1061 * to reflect the new GID.
1062 */
1063 if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
1064 if (!err && port == priv->port) {
1065 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1066 if (index == 0)
1067 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,
1068 &priv->flags);
1069 else
1070 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);
1071 ret = true;
1072 } else {
1073 ret = false;
1074 }
1075 } else {
1076 if (!err && port == priv->port) {
1077 ret = true;
1078 } else {
1079 if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {
1080 memcpy(&priv->local_gid, &gid0,
1081 sizeof(priv->local_gid));
1082 memcpy(priv->dev->dev_addr + 4, &gid0,
1083 sizeof(priv->local_gid));
1084 ret = true;
1085 }
1086 }
1087 }
1088
1089 out:
1090 netif_addr_unlock_bh(priv->dev);
1091
1092 return ret;
1093 }
1094
1095 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1096 enum ipoib_flush_level level,
1097 int nesting)
1098 {
1099 struct ipoib_dev_priv *cpriv;
1100 struct net_device *dev = priv->dev;
1101 int result;
1102
1103 down_read_nested(&priv->vlan_rwsem, nesting);
1104
1105 /*
1106 * Flush any child interfaces too -- they might be up even if
1107 * the parent is down.
1108 */
1109 list_for_each_entry(cpriv, &priv->child_intfs, list)
1110 __ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1111
1112 up_read(&priv->vlan_rwsem);
1113
1114 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1115 level != IPOIB_FLUSH_HEAVY) {
1116 /* Make sure the dev_addr is set even if not flushing */
1117 if (level == IPOIB_FLUSH_LIGHT)
1118 ipoib_dev_addr_changed_valid(priv);
1119 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1120 return;
1121 }
1122
1123 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1124 /* interface is down. update pkey and leave. */
1125 if (level == IPOIB_FLUSH_HEAVY) {
1126 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1127 update_parent_pkey(priv);
1128 else
1129 update_child_pkey(priv);
1130 } else if (level == IPOIB_FLUSH_LIGHT)
1131 ipoib_dev_addr_changed_valid(priv);
1132 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1133 return;
1134 }
1135
1136 if (level == IPOIB_FLUSH_HEAVY) {
1137 /* child devices chase their origin pkey value, while non-child
1138 * (parent) devices should always takes what present in pkey index 0
1139 */
1140 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1141 result = update_child_pkey(priv);
1142 if (result) {
1143 /* restart QP only if P_Key index is changed */
1144 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1145 return;
1146 }
1147
1148 } else {
1149 result = update_parent_pkey(priv);
1150 /* restart QP only if P_Key value changed */
1151 if (result) {
1152 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1153 return;
1154 }
1155 }
1156 }
1157
1158 if (level == IPOIB_FLUSH_LIGHT) {
1159 int oper_up;
1160 ipoib_mark_paths_invalid(dev);
1161 /* Set IPoIB operation as down to prevent races between:
1162 * the flush flow which leaves MCG and on the fly joins
1163 * which can happen during that time. mcast restart task
1164 * should deal with join requests we missed.
1165 */
1166 oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1167 ipoib_mcast_dev_flush(dev);
1168 if (oper_up)
1169 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1170 ipoib_flush_ah(dev);
1171 }
1172
1173 if (level >= IPOIB_FLUSH_NORMAL)
1174 ipoib_ib_dev_down(dev);
1175
1176 if (level == IPOIB_FLUSH_HEAVY) {
1177 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1178 ipoib_ib_dev_stop(dev);
1179 if (ipoib_ib_dev_open(dev) != 0)
1180 return;
1181 if (netif_queue_stopped(dev))
1182 netif_start_queue(dev);
1183 }
1184
1185 /*
1186 * The device could have been brought down between the start and when
1187 * we get here, don't bring it back up if it's not configured up
1188 */
1189 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1190 if (level >= IPOIB_FLUSH_NORMAL)
1191 ipoib_ib_dev_up(dev);
1192 if (ipoib_dev_addr_changed_valid(priv))
1193 ipoib_mcast_restart_task(&priv->restart_task);
1194 }
1195 }
1196
1197 void ipoib_ib_dev_flush_light(struct work_struct *work)
1198 {
1199 struct ipoib_dev_priv *priv =
1200 container_of(work, struct ipoib_dev_priv, flush_light);
1201
1202 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1203 }
1204
1205 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1206 {
1207 struct ipoib_dev_priv *priv =
1208 container_of(work, struct ipoib_dev_priv, flush_normal);
1209
1210 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1211 }
1212
1213 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1214 {
1215 struct ipoib_dev_priv *priv =
1216 container_of(work, struct ipoib_dev_priv, flush_heavy);
1217
1218 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1219 }
1220
1221 void ipoib_ib_dev_cleanup(struct net_device *dev)
1222 {
1223 struct ipoib_dev_priv *priv = netdev_priv(dev);
1224
1225 ipoib_dbg(priv, "cleaning up ib_dev\n");
1226 /*
1227 * We must make sure there are no more (path) completions
1228 * that may wish to touch priv fields that are no longer valid
1229 */
1230 ipoib_flush_paths(dev);
1231
1232 ipoib_mcast_stop_thread(dev);
1233 ipoib_mcast_dev_flush(dev);
1234
1235 /*
1236 * All of our ah references aren't free until after
1237 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1238 * the neighbor garbage collection is stopped and reaped.
1239 * That should all be done now, so make a final ah flush.
1240 */
1241 ipoib_stop_ah(dev);
1242
1243 ipoib_transport_dev_cleanup(dev);
1244 }
1245
1246
Linux with added FPC-III support