?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
IPoIB/cm: Use common CQ for CM send completions
[linux-2.6/verdex.git] / drivers / infiniband / ulp / ipoib / ipoib_cm.c
blob87610772a97944448ee2965e0ae5b9a5f787abb7
1 /*
2 * Copyright (c) 2006 Mellanox Technologies. All rights reserved
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 * $Id$
33 */
34
35 #include <rdma/ib_cm.h>
36 #include <rdma/ib_cache.h>
37 #include <net/dst.h>
38 #include <net/icmp.h>
39 #include <linux/icmpv6.h>
40 #include <linux/delay.h>
41
42 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
43 static int data_debug_level;
44
45 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
46 MODULE_PARM_DESC(cm_data_debug_level,
47 "Enable data path debug tracing for connected mode if > 0");
48 #endif
49
50 #include "ipoib.h"
51
52 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
53
54 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
55 #define IPOIB_CM_RX_TIMEOUT (2 * 256 * HZ)
56 #define IPOIB_CM_RX_DELAY (3 * 256 * HZ)
57 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
58
59 static struct ib_qp_attr ipoib_cm_err_attr = {
60 .qp_state = IB_QPS_ERR
61 };
62
63 #define IPOIB_CM_RX_DRAIN_WRID 0x7fffffff
64
65 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
66 .wr_id = IPOIB_CM_RX_DRAIN_WRID,
67 .opcode = IB_WR_SEND,
68 };
69
70 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
71 struct ib_cm_event *event);
72
73 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
74 u64 mapping[IPOIB_CM_RX_SG])
75 {
76 int i;
77
78 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
79
80 for (i = 0; i < frags; ++i)
81 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
82 }
83
84 static int ipoib_cm_post_receive(struct net_device *dev, int id)
85 {
86 struct ipoib_dev_priv *priv = netdev_priv(dev);
87 struct ib_recv_wr *bad_wr;
88 int i, ret;
89
90 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
91
92 for (i = 0; i < IPOIB_CM_RX_SG; ++i)
93 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
94
95 ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
96 if (unlikely(ret)) {
97 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
98 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
99 priv->cm.srq_ring[id].mapping);
100 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
101 priv->cm.srq_ring[id].skb = NULL;
102 }
103
104 return ret;
105 }
106
107 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, int id, int frags,
108 u64 mapping[IPOIB_CM_RX_SG])
109 {
110 struct ipoib_dev_priv *priv = netdev_priv(dev);
111 struct sk_buff *skb;
112 int i;
113
114 skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
115 if (unlikely(!skb))
116 return NULL;
117
118 /*
119 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
120 * IP header to a multiple of 16.
121 */
122 skb_reserve(skb, 12);
123
124 mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
125 DMA_FROM_DEVICE);
126 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
127 dev_kfree_skb_any(skb);
128 return NULL;
129 }
130
131 for (i = 0; i < frags; i++) {
132 struct page *page = alloc_page(GFP_ATOMIC);
133
134 if (!page)
135 goto partial_error;
136 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
137
138 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
139 0, PAGE_SIZE, DMA_FROM_DEVICE);
140 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
141 goto partial_error;
142 }
143
144 priv->cm.srq_ring[id].skb = skb;
145 return skb;
146
147 partial_error:
148
149 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
150
151 for (; i > 0; --i)
152 ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
153
154 dev_kfree_skb_any(skb);
155 return NULL;
156 }
157
158 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv* priv)
159 {
160 struct ib_send_wr *bad_wr;
161 struct ipoib_cm_rx *p;
162
163 /* We only reserved 1 extra slot in CQ for drain WRs, so
164 * make sure we have at most 1 outstanding WR. */
165 if (list_empty(&priv->cm.rx_flush_list) ||
166 !list_empty(&priv->cm.rx_drain_list))
167 return;
168
169 /*
170 * QPs on flush list are error state. This way, a "flush
171 * error" WC will be immediately generated for each WR we post.
172 */
173 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
174 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
175 ipoib_warn(priv, "failed to post drain wr\n");
176
177 list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
178 }
179
180 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
181 {
182 struct ipoib_cm_rx *p = ctx;
183 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
184 unsigned long flags;
185
186 if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
187 return;
188
189 spin_lock_irqsave(&priv->lock, flags);
190 list_move(&p->list, &priv->cm.rx_flush_list);
191 p->state = IPOIB_CM_RX_FLUSH;
192 ipoib_cm_start_rx_drain(priv);
193 spin_unlock_irqrestore(&priv->lock, flags);
194 }
195
196 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
197 struct ipoib_cm_rx *p)
198 {
199 struct ipoib_dev_priv *priv = netdev_priv(dev);
200 struct ib_qp_init_attr attr = {
201 .event_handler = ipoib_cm_rx_event_handler,
202 .send_cq = priv->cq, /* For drain WR */
203 .recv_cq = priv->cq,
204 .srq = priv->cm.srq,
205 .cap.max_send_wr = 1, /* For drain WR */
206 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
207 .sq_sig_type = IB_SIGNAL_ALL_WR,
208 .qp_type = IB_QPT_RC,
209 .qp_context = p,
210 };
211 return ib_create_qp(priv->pd, &attr);
212 }
213
214 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
215 struct ib_cm_id *cm_id, struct ib_qp *qp,
216 unsigned psn)
217 {
218 struct ipoib_dev_priv *priv = netdev_priv(dev);
219 struct ib_qp_attr qp_attr;
220 int qp_attr_mask, ret;
221
222 qp_attr.qp_state = IB_QPS_INIT;
223 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
224 if (ret) {
225 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
226 return ret;
227 }
228 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
229 if (ret) {
230 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
231 return ret;
232 }
233 qp_attr.qp_state = IB_QPS_RTR;
234 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
235 if (ret) {
236 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
237 return ret;
238 }
239 qp_attr.rq_psn = psn;
240 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
241 if (ret) {
242 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
243 return ret;
244 }
245
246 /*
247 * Current Mellanox HCA firmware won't generate completions
248 * with error for drain WRs unless the QP has been moved to
249 * RTS first. This work-around leaves a window where a QP has
250 * moved to error asynchronously, but this will eventually get
251 * fixed in firmware, so let's not error out if modify QP
252 * fails.
253 */
254 qp_attr.qp_state = IB_QPS_RTS;
255 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
256 if (ret) {
257 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
258 return 0;
259 }
260 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
261 if (ret) {
262 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
263 return 0;
264 }
265
266 return 0;
267 }
268
269 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
270 struct ib_qp *qp, struct ib_cm_req_event_param *req,
271 unsigned psn)
272 {
273 struct ipoib_dev_priv *priv = netdev_priv(dev);
274 struct ipoib_cm_data data = {};
275 struct ib_cm_rep_param rep = {};
276
277 data.qpn = cpu_to_be32(priv->qp->qp_num);
278 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
279
280 rep.private_data = &data;
281 rep.private_data_len = sizeof data;
282 rep.flow_control = 0;
283 rep.rnr_retry_count = req->rnr_retry_count;
284 rep.srq = 1;
285 rep.qp_num = qp->qp_num;
286 rep.starting_psn = psn;
287 return ib_send_cm_rep(cm_id, &rep);
288 }
289
290 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
291 {
292 struct net_device *dev = cm_id->context;
293 struct ipoib_dev_priv *priv = netdev_priv(dev);
294 struct ipoib_cm_rx *p;
295 unsigned psn;
296 int ret;
297
298 ipoib_dbg(priv, "REQ arrived\n");
299 p = kzalloc(sizeof *p, GFP_KERNEL);
300 if (!p)
301 return -ENOMEM;
302 p->dev = dev;
303 p->id = cm_id;
304 cm_id->context = p;
305 p->state = IPOIB_CM_RX_LIVE;
306 p->jiffies = jiffies;
307 INIT_LIST_HEAD(&p->list);
308
309 p->qp = ipoib_cm_create_rx_qp(dev, p);
310 if (IS_ERR(p->qp)) {
311 ret = PTR_ERR(p->qp);
312 goto err_qp;
313 }
314
315 psn = random32() & 0xffffff;
316 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
317 if (ret)
318 goto err_modify;
319
320 spin_lock_irq(&priv->lock);
321 queue_delayed_work(ipoib_workqueue,
322 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
323 /* Add this entry to passive ids list head, but do not re-add it
324 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
325 p->jiffies = jiffies;
326 if (p->state == IPOIB_CM_RX_LIVE)
327 list_move(&p->list, &priv->cm.passive_ids);
328 spin_unlock_irq(&priv->lock);
329
330 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
331 if (ret) {
332 ipoib_warn(priv, "failed to send REP: %d\n", ret);
333 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
334 ipoib_warn(priv, "unable to move qp to error state\n");
335 }
336 return 0;
337
338 err_modify:
339 ib_destroy_qp(p->qp);
340 err_qp:
341 kfree(p);
342 return ret;
343 }
344
345 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
346 struct ib_cm_event *event)
347 {
348 struct ipoib_cm_rx *p;
349 struct ipoib_dev_priv *priv;
350
351 switch (event->event) {
352 case IB_CM_REQ_RECEIVED:
353 return ipoib_cm_req_handler(cm_id, event);
354 case IB_CM_DREQ_RECEIVED:
355 p = cm_id->context;
356 ib_send_cm_drep(cm_id, NULL, 0);
357 /* Fall through */
358 case IB_CM_REJ_RECEIVED:
359 p = cm_id->context;
360 priv = netdev_priv(p->dev);
361 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
362 ipoib_warn(priv, "unable to move qp to error state\n");
363 /* Fall through */
364 default:
365 return 0;
366 }
367 }
368 /* Adjust length of skb with fragments to match received data */
369 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
370 unsigned int length, struct sk_buff *toskb)
371 {
372 int i, num_frags;
373 unsigned int size;
374
375 /* put header into skb */
376 size = min(length, hdr_space);
377 skb->tail += size;
378 skb->len += size;
379 length -= size;
380
381 num_frags = skb_shinfo(skb)->nr_frags;
382 for (i = 0; i < num_frags; i++) {
383 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
384
385 if (length == 0) {
386 /* don't need this page */
387 skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
388 --skb_shinfo(skb)->nr_frags;
389 } else {
390 size = min(length, (unsigned) PAGE_SIZE);
391
392 frag->size = size;
393 skb->data_len += size;
394 skb->truesize += size;
395 skb->len += size;
396 length -= size;
397 }
398 }
399 }
400
401 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
402 {
403 struct ipoib_dev_priv *priv = netdev_priv(dev);
404 unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
405 struct sk_buff *skb, *newskb;
406 struct ipoib_cm_rx *p;
407 unsigned long flags;
408 u64 mapping[IPOIB_CM_RX_SG];
409 int frags;
410
411 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
412 wr_id, wc->status);
413
414 if (unlikely(wr_id >= ipoib_recvq_size)) {
415 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
416 spin_lock_irqsave(&priv->lock, flags);
417 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
418 ipoib_cm_start_rx_drain(priv);
419 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
420 spin_unlock_irqrestore(&priv->lock, flags);
421 } else
422 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
423 wr_id, ipoib_recvq_size);
424 return;
425 }
426
427 skb = priv->cm.srq_ring[wr_id].skb;
428
429 if (unlikely(wc->status != IB_WC_SUCCESS)) {
430 ipoib_dbg(priv, "cm recv error "
431 "(status=%d, wrid=%d vend_err %x)\n",
432 wc->status, wr_id, wc->vendor_err);
433 ++dev->stats.rx_dropped;
434 goto repost;
435 }
436
437 if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
438 p = wc->qp->qp_context;
439 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
440 spin_lock_irqsave(&priv->lock, flags);
441 p->jiffies = jiffies;
442 /* Move this entry to list head, but do not re-add it
443 * if it has been moved out of list. */
444 if (p->state == IPOIB_CM_RX_LIVE)
445 list_move(&p->list, &priv->cm.passive_ids);
446 spin_unlock_irqrestore(&priv->lock, flags);
447 }
448 }
449
450 frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
451 (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
452
453 newskb = ipoib_cm_alloc_rx_skb(dev, wr_id, frags, mapping);
454 if (unlikely(!newskb)) {
455 /*
456 * If we can't allocate a new RX buffer, dump
457 * this packet and reuse the old buffer.
458 */
459 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
460 ++dev->stats.rx_dropped;
461 goto repost;
462 }
463
464 ipoib_cm_dma_unmap_rx(priv, frags, priv->cm.srq_ring[wr_id].mapping);
465 memcpy(priv->cm.srq_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
466
467 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
468 wc->byte_len, wc->slid);
469
470 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
471
472 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
473 skb_reset_mac_header(skb);
474 skb_pull(skb, IPOIB_ENCAP_LEN);
475
476 dev->last_rx = jiffies;
477 ++dev->stats.rx_packets;
478 dev->stats.rx_bytes += skb->len;
479
480 skb->dev = dev;
481 /* XXX get correct PACKET_ type here */
482 skb->pkt_type = PACKET_HOST;
483 netif_receive_skb(skb);
484
485 repost:
486 if (unlikely(ipoib_cm_post_receive(dev, wr_id)))
487 ipoib_warn(priv, "ipoib_cm_post_receive failed "
488 "for buf %d\n", wr_id);
489 }
490
491 static inline int post_send(struct ipoib_dev_priv *priv,
492 struct ipoib_cm_tx *tx,
493 unsigned int wr_id,
494 u64 addr, int len)
495 {
496 struct ib_send_wr *bad_wr;
497
498 priv->tx_sge.addr = addr;
499 priv->tx_sge.length = len;
500
501 priv->tx_wr.wr_id = wr_id | IPOIB_OP_CM;
502
503 return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
504 }
505
506 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
507 {
508 struct ipoib_dev_priv *priv = netdev_priv(dev);
509 struct ipoib_tx_buf *tx_req;
510 u64 addr;
511
512 if (unlikely(skb->len > tx->mtu)) {
513 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
514 skb->len, tx->mtu);
515 ++dev->stats.tx_dropped;
516 ++dev->stats.tx_errors;
517 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
518 return;
519 }
520
521 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
522 tx->tx_head, skb->len, tx->qp->qp_num);
523
524 /*
525 * We put the skb into the tx_ring _before_ we call post_send()
526 * because it's entirely possible that the completion handler will
527 * run before we execute anything after the post_send(). That
528 * means we have to make sure everything is properly recorded and
529 * our state is consistent before we call post_send().
530 */
531 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
532 tx_req->skb = skb;
533 addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
534 if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
535 ++dev->stats.tx_errors;
536 dev_kfree_skb_any(skb);
537 return;
538 }
539
540 tx_req->mapping = addr;
541
542 if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
543 addr, skb->len))) {
544 ipoib_warn(priv, "post_send failed\n");
545 ++dev->stats.tx_errors;
546 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
547 dev_kfree_skb_any(skb);
548 } else {
549 dev->trans_start = jiffies;
550 ++tx->tx_head;
551
552 if (++priv->tx_outstanding == ipoib_sendq_size) {
553 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
554 tx->qp->qp_num);
555 netif_stop_queue(dev);
556 }
557 }
558 }
559
560 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
561 {
562 struct ipoib_dev_priv *priv = netdev_priv(dev);
563 struct ipoib_cm_tx *tx = wc->qp->qp_context;
564 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
565 struct ipoib_tx_buf *tx_req;
566 unsigned long flags;
567
568 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
569 wr_id, wc->status);
570
571 if (unlikely(wr_id >= ipoib_sendq_size)) {
572 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
573 wr_id, ipoib_sendq_size);
574 return;
575 }
576
577 tx_req = &tx->tx_ring[wr_id];
578
579 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
580
581 /* FIXME: is this right? Shouldn't we only increment on success? */
582 ++dev->stats.tx_packets;
583 dev->stats.tx_bytes += tx_req->skb->len;
584
585 dev_kfree_skb_any(tx_req->skb);
586
587 spin_lock_irqsave(&priv->tx_lock, flags);
588 ++tx->tx_tail;
589 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
590 netif_queue_stopped(dev) &&
591 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
592 netif_wake_queue(dev);
593
594 if (wc->status != IB_WC_SUCCESS &&
595 wc->status != IB_WC_WR_FLUSH_ERR) {
596 struct ipoib_neigh *neigh;
597
598 ipoib_dbg(priv, "failed cm send event "
599 "(status=%d, wrid=%d vend_err %x)\n",
600 wc->status, wr_id, wc->vendor_err);
601
602 spin_lock(&priv->lock);
603 neigh = tx->neigh;
604
605 if (neigh) {
606 neigh->cm = NULL;
607 list_del(&neigh->list);
608 if (neigh->ah)
609 ipoib_put_ah(neigh->ah);
610 ipoib_neigh_free(dev, neigh);
611
612 tx->neigh = NULL;
613 }
614
615 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
616 list_move(&tx->list, &priv->cm.reap_list);
617 queue_work(ipoib_workqueue, &priv->cm.reap_task);
618 }
619
620 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
621
622 spin_unlock(&priv->lock);
623 }
624
625 spin_unlock_irqrestore(&priv->tx_lock, flags);
626 }
627
628 int ipoib_cm_dev_open(struct net_device *dev)
629 {
630 struct ipoib_dev_priv *priv = netdev_priv(dev);
631 int ret;
632
633 if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
634 return 0;
635
636 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
637 if (IS_ERR(priv->cm.id)) {
638 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
639 ret = PTR_ERR(priv->cm.id);
640 goto err_cm;
641 }
642
643 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
644 0, NULL);
645 if (ret) {
646 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
647 IPOIB_CM_IETF_ID | priv->qp->qp_num);
648 goto err_listen;
649 }
650
651 return 0;
652
653 err_listen:
654 ib_destroy_cm_id(priv->cm.id);
655 err_cm:
656 priv->cm.id = NULL;
657 return ret;
658 }
659
660 void ipoib_cm_dev_stop(struct net_device *dev)
661 {
662 struct ipoib_dev_priv *priv = netdev_priv(dev);
663 struct ipoib_cm_rx *p, *n;
664 unsigned long begin;
665 LIST_HEAD(list);
666 int ret;
667
668 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
669 return;
670
671 ib_destroy_cm_id(priv->cm.id);
672 priv->cm.id = NULL;
673
674 spin_lock_irq(&priv->lock);
675 while (!list_empty(&priv->cm.passive_ids)) {
676 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
677 list_move(&p->list, &priv->cm.rx_error_list);
678 p->state = IPOIB_CM_RX_ERROR;
679 spin_unlock_irq(&priv->lock);
680 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
681 if (ret)
682 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
683 spin_lock_irq(&priv->lock);
684 }
685
686 /* Wait for all RX to be drained */
687 begin = jiffies;
688
689 while (!list_empty(&priv->cm.rx_error_list) ||
690 !list_empty(&priv->cm.rx_flush_list) ||
691 !list_empty(&priv->cm.rx_drain_list)) {
692 if (time_after(jiffies, begin + 5 * HZ)) {
693 ipoib_warn(priv, "RX drain timing out\n");
694
695 /*
696 * assume the HW is wedged and just free up everything.
697 */
698 list_splice_init(&priv->cm.rx_flush_list, &list);
699 list_splice_init(&priv->cm.rx_error_list, &list);
700 list_splice_init(&priv->cm.rx_drain_list, &list);
701 break;
702 }
703 spin_unlock_irq(&priv->lock);
704 msleep(1);
705 ipoib_drain_cq(dev);
706 spin_lock_irq(&priv->lock);
707 }
708
709 list_splice_init(&priv->cm.rx_reap_list, &list);
710
711 spin_unlock_irq(&priv->lock);
712
713 list_for_each_entry_safe(p, n, &list, list) {
714 ib_destroy_cm_id(p->id);
715 ib_destroy_qp(p->qp);
716 kfree(p);
717 }
718
719 cancel_delayed_work(&priv->cm.stale_task);
720 }
721
722 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
723 {
724 struct ipoib_cm_tx *p = cm_id->context;
725 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
726 struct ipoib_cm_data *data = event->private_data;
727 struct sk_buff_head skqueue;
728 struct ib_qp_attr qp_attr;
729 int qp_attr_mask, ret;
730 struct sk_buff *skb;
731
732 p->mtu = be32_to_cpu(data->mtu);
733
734 if (p->mtu <= IPOIB_ENCAP_LEN) {
735 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
736 p->mtu, IPOIB_ENCAP_LEN);
737 return -EINVAL;
738 }
739
740 qp_attr.qp_state = IB_QPS_RTR;
741 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
742 if (ret) {
743 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
744 return ret;
745 }
746
747 qp_attr.rq_psn = 0 /* FIXME */;
748 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
749 if (ret) {
750 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
751 return ret;
752 }
753
754 qp_attr.qp_state = IB_QPS_RTS;
755 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
756 if (ret) {
757 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
758 return ret;
759 }
760 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
761 if (ret) {
762 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
763 return ret;
764 }
765
766 skb_queue_head_init(&skqueue);
767
768 spin_lock_irq(&priv->lock);
769 set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
770 if (p->neigh)
771 while ((skb = __skb_dequeue(&p->neigh->queue)))
772 __skb_queue_tail(&skqueue, skb);
773 spin_unlock_irq(&priv->lock);
774
775 while ((skb = __skb_dequeue(&skqueue))) {
776 skb->dev = p->dev;
777 if (dev_queue_xmit(skb))
778 ipoib_warn(priv, "dev_queue_xmit failed "
779 "to requeue packet\n");
780 }
781
782 ret = ib_send_cm_rtu(cm_id, NULL, 0);
783 if (ret) {
784 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
785 return ret;
786 }
787 return 0;
788 }
789
790 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
791 {
792 struct ipoib_dev_priv *priv = netdev_priv(dev);
793 struct ib_qp_init_attr attr = {
794 .send_cq = priv->cq,
795 .recv_cq = priv->cq,
796 .srq = priv->cm.srq,
797 .cap.max_send_wr = ipoib_sendq_size,
798 .cap.max_send_sge = 1,
799 .sq_sig_type = IB_SIGNAL_ALL_WR,
800 .qp_type = IB_QPT_RC,
801 .qp_context = tx
802 };
803
804 return ib_create_qp(priv->pd, &attr);
805 }
806
807 static int ipoib_cm_send_req(struct net_device *dev,
808 struct ib_cm_id *id, struct ib_qp *qp,
809 u32 qpn,
810 struct ib_sa_path_rec *pathrec)
811 {
812 struct ipoib_dev_priv *priv = netdev_priv(dev);
813 struct ipoib_cm_data data = {};
814 struct ib_cm_req_param req = {};
815
816 data.qpn = cpu_to_be32(priv->qp->qp_num);
817 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
818
819 req.primary_path = pathrec;
820 req.alternate_path = NULL;
821 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
822 req.qp_num = qp->qp_num;
823 req.qp_type = qp->qp_type;
824 req.private_data = &data;
825 req.private_data_len = sizeof data;
826 req.flow_control = 0;
827
828 req.starting_psn = 0; /* FIXME */
829
830 /*
831 * Pick some arbitrary defaults here; we could make these
832 * module parameters if anyone cared about setting them.
833 */
834 req.responder_resources = 4;
835 req.remote_cm_response_timeout = 20;
836 req.local_cm_response_timeout = 20;
837 req.retry_count = 0; /* RFC draft warns against retries */
838 req.rnr_retry_count = 0; /* RFC draft warns against retries */
839 req.max_cm_retries = 15;
840 req.srq = 1;
841 return ib_send_cm_req(id, &req);
842 }
843
844 static int ipoib_cm_modify_tx_init(struct net_device *dev,
845 struct ib_cm_id *cm_id, struct ib_qp *qp)
846 {
847 struct ipoib_dev_priv *priv = netdev_priv(dev);
848 struct ib_qp_attr qp_attr;
849 int qp_attr_mask, ret;
850 ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
851 if (ret) {
852 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
853 return ret;
854 }
855
856 qp_attr.qp_state = IB_QPS_INIT;
857 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
858 qp_attr.port_num = priv->port;
859 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
860
861 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
862 if (ret) {
863 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
864 return ret;
865 }
866 return 0;
867 }
868
869 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
870 struct ib_sa_path_rec *pathrec)
871 {
872 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
873 int ret;
874
875 p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
876 GFP_KERNEL);
877 if (!p->tx_ring) {
878 ipoib_warn(priv, "failed to allocate tx ring\n");
879 ret = -ENOMEM;
880 goto err_tx;
881 }
882
883 p->qp = ipoib_cm_create_tx_qp(p->dev, p);
884 if (IS_ERR(p->qp)) {
885 ret = PTR_ERR(p->qp);
886 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
887 goto err_qp;
888 }
889
890 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
891 if (IS_ERR(p->id)) {
892 ret = PTR_ERR(p->id);
893 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
894 goto err_id;
895 }
896
897 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp);
898 if (ret) {
899 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
900 goto err_modify;
901 }
902
903 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
904 if (ret) {
905 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
906 goto err_send_cm;
907 }
908
909 ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
910 p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
911
912 return 0;
913
914 err_send_cm:
915 err_modify:
916 ib_destroy_cm_id(p->id);
917 err_id:
918 p->id = NULL;
919 ib_destroy_qp(p->qp);
920 err_qp:
921 p->qp = NULL;
922 err_tx:
923 return ret;
924 }
925
926 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
927 {
928 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
929 struct ipoib_tx_buf *tx_req;
930 unsigned long flags;
931 unsigned long begin;
932
933 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
934 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
935
936 if (p->id)
937 ib_destroy_cm_id(p->id);
938
939 if (p->tx_ring) {
940 /* Wait for all sends to complete */
941 begin = jiffies;
942 while ((int) p->tx_tail - (int) p->tx_head < 0) {
943 if (time_after(jiffies, begin + 5 * HZ)) {
944 ipoib_warn(priv, "timing out; %d sends not completed\n",
945 p->tx_head - p->tx_tail);
946 goto timeout;
947 }
948
949 msleep(1);
950 }
951 }
952
953 timeout:
954
955 while ((int) p->tx_tail - (int) p->tx_head < 0) {
956 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
957 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
958 DMA_TO_DEVICE);
959 dev_kfree_skb_any(tx_req->skb);
960 ++p->tx_tail;
961 spin_lock_irqsave(&priv->tx_lock, flags);
962 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
963 netif_queue_stopped(p->dev) &&
964 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
965 netif_wake_queue(p->dev);
966 spin_unlock_irqrestore(&priv->tx_lock, flags);
967 }
968
969 if (p->qp)
970 ib_destroy_qp(p->qp);
971
972 kfree(p->tx_ring);
973 kfree(p);
974 }
975
976 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
977 struct ib_cm_event *event)
978 {
979 struct ipoib_cm_tx *tx = cm_id->context;
980 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
981 struct net_device *dev = priv->dev;
982 struct ipoib_neigh *neigh;
983 int ret;
984
985 switch (event->event) {
986 case IB_CM_DREQ_RECEIVED:
987 ipoib_dbg(priv, "DREQ received.\n");
988 ib_send_cm_drep(cm_id, NULL, 0);
989 break;
990 case IB_CM_REP_RECEIVED:
991 ipoib_dbg(priv, "REP received.\n");
992 ret = ipoib_cm_rep_handler(cm_id, event);
993 if (ret)
994 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
995 NULL, 0, NULL, 0);
996 break;
997 case IB_CM_REQ_ERROR:
998 case IB_CM_REJ_RECEIVED:
999 case IB_CM_TIMEWAIT_EXIT:
1000 ipoib_dbg(priv, "CM error %d.\n", event->event);
1001 spin_lock_irq(&priv->tx_lock);
1002 spin_lock(&priv->lock);
1003 neigh = tx->neigh;
1004
1005 if (neigh) {
1006 neigh->cm = NULL;
1007 list_del(&neigh->list);
1008 if (neigh->ah)
1009 ipoib_put_ah(neigh->ah);
1010 ipoib_neigh_free(dev, neigh);
1011
1012 tx->neigh = NULL;
1013 }
1014
1015 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1016 list_move(&tx->list, &priv->cm.reap_list);
1017 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1018 }
1019
1020 spin_unlock(&priv->lock);
1021 spin_unlock_irq(&priv->tx_lock);
1022 break;
1023 default:
1024 break;
1025 }
1026
1027 return 0;
1028 }
1029
1030 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1031 struct ipoib_neigh *neigh)
1032 {
1033 struct ipoib_dev_priv *priv = netdev_priv(dev);
1034 struct ipoib_cm_tx *tx;
1035
1036 tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1037 if (!tx)
1038 return NULL;
1039
1040 neigh->cm = tx;
1041 tx->neigh = neigh;
1042 tx->path = path;
1043 tx->dev = dev;
1044 list_add(&tx->list, &priv->cm.start_list);
1045 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1046 queue_work(ipoib_workqueue, &priv->cm.start_task);
1047 return tx;
1048 }
1049
1050 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1051 {
1052 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1053 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1054 list_move(&tx->list, &priv->cm.reap_list);
1055 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1056 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1057 IPOIB_GID_ARG(tx->neigh->dgid));
1058 tx->neigh = NULL;
1059 }
1060 }
1061
1062 static void ipoib_cm_tx_start(struct work_struct *work)
1063 {
1064 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1065 cm.start_task);
1066 struct net_device *dev = priv->dev;
1067 struct ipoib_neigh *neigh;
1068 struct ipoib_cm_tx *p;
1069 unsigned long flags;
1070 int ret;
1071
1072 struct ib_sa_path_rec pathrec;
1073 u32 qpn;
1074
1075 spin_lock_irqsave(&priv->tx_lock, flags);
1076 spin_lock(&priv->lock);
1077 while (!list_empty(&priv->cm.start_list)) {
1078 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1079 list_del_init(&p->list);
1080 neigh = p->neigh;
1081 qpn = IPOIB_QPN(neigh->neighbour->ha);
1082 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1083 spin_unlock(&priv->lock);
1084 spin_unlock_irqrestore(&priv->tx_lock, flags);
1085 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1086 spin_lock_irqsave(&priv->tx_lock, flags);
1087 spin_lock(&priv->lock);
1088 if (ret) {
1089 neigh = p->neigh;
1090 if (neigh) {
1091 neigh->cm = NULL;
1092 list_del(&neigh->list);
1093 if (neigh->ah)
1094 ipoib_put_ah(neigh->ah);
1095 ipoib_neigh_free(dev, neigh);
1096 }
1097 list_del(&p->list);
1098 kfree(p);
1099 }
1100 }
1101 spin_unlock(&priv->lock);
1102 spin_unlock_irqrestore(&priv->tx_lock, flags);
1103 }
1104
1105 static void ipoib_cm_tx_reap(struct work_struct *work)
1106 {
1107 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1108 cm.reap_task);
1109 struct ipoib_cm_tx *p;
1110
1111 spin_lock_irq(&priv->tx_lock);
1112 spin_lock(&priv->lock);
1113 while (!list_empty(&priv->cm.reap_list)) {
1114 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1115 list_del(&p->list);
1116 spin_unlock(&priv->lock);
1117 spin_unlock_irq(&priv->tx_lock);
1118 ipoib_cm_tx_destroy(p);
1119 spin_lock_irq(&priv->tx_lock);
1120 spin_lock(&priv->lock);
1121 }
1122 spin_unlock(&priv->lock);
1123 spin_unlock_irq(&priv->tx_lock);
1124 }
1125
1126 static void ipoib_cm_skb_reap(struct work_struct *work)
1127 {
1128 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1129 cm.skb_task);
1130 struct sk_buff *skb;
1131
1132 unsigned mtu = priv->mcast_mtu;
1133
1134 spin_lock_irq(&priv->tx_lock);
1135 spin_lock(&priv->lock);
1136 while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1137 spin_unlock(&priv->lock);
1138 spin_unlock_irq(&priv->tx_lock);
1139 if (skb->protocol == htons(ETH_P_IP))
1140 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1141 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1142 else if (skb->protocol == htons(ETH_P_IPV6))
1143 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1144 #endif
1145 dev_kfree_skb_any(skb);
1146 spin_lock_irq(&priv->tx_lock);
1147 spin_lock(&priv->lock);
1148 }
1149 spin_unlock(&priv->lock);
1150 spin_unlock_irq(&priv->tx_lock);
1151 }
1152
1153 void ipoib_cm_skb_too_long(struct net_device* dev, struct sk_buff *skb,
1154 unsigned int mtu)
1155 {
1156 struct ipoib_dev_priv *priv = netdev_priv(dev);
1157 int e = skb_queue_empty(&priv->cm.skb_queue);
1158
1159 if (skb->dst)
1160 skb->dst->ops->update_pmtu(skb->dst, mtu);
1161
1162 skb_queue_tail(&priv->cm.skb_queue, skb);
1163 if (e)
1164 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1165 }
1166
1167 static void ipoib_cm_rx_reap(struct work_struct *work)
1168 {
1169 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1170 cm.rx_reap_task);
1171 struct ipoib_cm_rx *p, *n;
1172 LIST_HEAD(list);
1173
1174 spin_lock_irq(&priv->lock);
1175 list_splice_init(&priv->cm.rx_reap_list, &list);
1176 spin_unlock_irq(&priv->lock);
1177
1178 list_for_each_entry_safe(p, n, &list, list) {
1179 ib_destroy_cm_id(p->id);
1180 ib_destroy_qp(p->qp);
1181 kfree(p);
1182 }
1183 }
1184
1185 static void ipoib_cm_stale_task(struct work_struct *work)
1186 {
1187 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1188 cm.stale_task.work);
1189 struct ipoib_cm_rx *p;
1190 int ret;
1191
1192 spin_lock_irq(&priv->lock);
1193 while (!list_empty(&priv->cm.passive_ids)) {
1194 /* List is sorted by LRU, start from tail,
1195 * stop when we see a recently used entry */
1196 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1197 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1198 break;
1199 list_move(&p->list, &priv->cm.rx_error_list);
1200 p->state = IPOIB_CM_RX_ERROR;
1201 spin_unlock_irq(&priv->lock);
1202 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1203 if (ret)
1204 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1205 spin_lock_irq(&priv->lock);
1206 }
1207
1208 if (!list_empty(&priv->cm.passive_ids))
1209 queue_delayed_work(ipoib_workqueue,
1210 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1211 spin_unlock_irq(&priv->lock);
1212 }
1213
1214
1215 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1216 char *buf)
1217 {
1218 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1219
1220 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1221 return sprintf(buf, "connected\n");
1222 else
1223 return sprintf(buf, "datagram\n");
1224 }
1225
1226 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1227 const char *buf, size_t count)
1228 {
1229 struct net_device *dev = to_net_dev(d);
1230 struct ipoib_dev_priv *priv = netdev_priv(dev);
1231
1232 /* flush paths if we switch modes so that connections are restarted */
1233 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1234 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1235 ipoib_warn(priv, "enabling connected mode "
1236 "will cause multicast packet drops\n");
1237 ipoib_flush_paths(dev);
1238 return count;
1239 }
1240
1241 if (!strcmp(buf, "datagram\n")) {
1242 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1243 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1244 ipoib_flush_paths(dev);
1245 return count;
1246 }
1247
1248 return -EINVAL;
1249 }
1250
1251 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1252
1253 int ipoib_cm_add_mode_attr(struct net_device *dev)
1254 {
1255 return device_create_file(&dev->dev, &dev_attr_mode);
1256 }
1257
1258 int ipoib_cm_dev_init(struct net_device *dev)
1259 {
1260 struct ipoib_dev_priv *priv = netdev_priv(dev);
1261 struct ib_srq_init_attr srq_init_attr = {
1262 .attr = {
1263 .max_wr = ipoib_recvq_size,
1264 .max_sge = IPOIB_CM_RX_SG
1265 }
1266 };
1267 int ret, i;
1268
1269 INIT_LIST_HEAD(&priv->cm.passive_ids);
1270 INIT_LIST_HEAD(&priv->cm.reap_list);
1271 INIT_LIST_HEAD(&priv->cm.start_list);
1272 INIT_LIST_HEAD(&priv->cm.rx_error_list);
1273 INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1274 INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1275 INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1276 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1277 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1278 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1279 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1280 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1281
1282 skb_queue_head_init(&priv->cm.skb_queue);
1283
1284 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1285 if (IS_ERR(priv->cm.srq)) {
1286 ret = PTR_ERR(priv->cm.srq);
1287 priv->cm.srq = NULL;
1288 return ret;
1289 }
1290
1291 priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1292 GFP_KERNEL);
1293 if (!priv->cm.srq_ring) {
1294 printk(KERN_WARNING "%s: failed to allocate CM ring (%d entries)\n",
1295 priv->ca->name, ipoib_recvq_size);
1296 ipoib_cm_dev_cleanup(dev);
1297 return -ENOMEM;
1298 }
1299
1300 for (i = 0; i < IPOIB_CM_RX_SG; ++i)
1301 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1302
1303 priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1304 for (i = 1; i < IPOIB_CM_RX_SG; ++i)
1305 priv->cm.rx_sge[i].length = PAGE_SIZE;
1306 priv->cm.rx_wr.next = NULL;
1307 priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1308 priv->cm.rx_wr.num_sge = IPOIB_CM_RX_SG;
1309
1310 for (i = 0; i < ipoib_recvq_size; ++i) {
1311 if (!ipoib_cm_alloc_rx_skb(dev, i, IPOIB_CM_RX_SG - 1,
1312 priv->cm.srq_ring[i].mapping)) {
1313 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
1314 ipoib_cm_dev_cleanup(dev);
1315 return -ENOMEM;
1316 }
1317 if (ipoib_cm_post_receive(dev, i)) {
1318 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
1319 ipoib_cm_dev_cleanup(dev);
1320 return -EIO;
1321 }
1322 }
1323
1324 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1325 return 0;
1326 }
1327
1328 void ipoib_cm_dev_cleanup(struct net_device *dev)
1329 {
1330 struct ipoib_dev_priv *priv = netdev_priv(dev);
1331 int i, ret;
1332
1333 if (!priv->cm.srq)
1334 return;
1335
1336 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1337
1338 ret = ib_destroy_srq(priv->cm.srq);
1339 if (ret)
1340 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1341
1342 priv->cm.srq = NULL;
1343 if (!priv->cm.srq_ring)
1344 return;
1345 for (i = 0; i < ipoib_recvq_size; ++i)
1346 if (priv->cm.srq_ring[i].skb) {
1347 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
1348 priv->cm.srq_ring[i].mapping);
1349 dev_kfree_skb_any(priv->cm.srq_ring[i].skb);
1350 priv->cm.srq_ring[i].skb = NULL;
1351 }
1352 kfree(priv->cm.srq_ring);
1353 priv->cm.srq_ring = NULL;
1354 }
Verdex Pro support