2 * Copyright (c) 2006 Mellanox Technologies. All rights reserved
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <rdma/ib_cm.h>
36 #include <linux/icmpv6.h>
37 #include <linux/delay.h>
38 #include <linux/slab.h>
39 #include <linux/vmalloc.h>
40 #include <linux/moduleparam.h>
44 int ipoib_max_conn_qp = 128;
46 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
47 MODULE_PARM_DESC(max_nonsrq_conn_qp,
48 "Max number of connected-mode QPs per interface "
49 "(applied only if shared receive queue is not available)");
51 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
52 static int data_debug_level;
54 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
55 MODULE_PARM_DESC(cm_data_debug_level,
56 "Enable data path debug tracing for connected mode if > 0");
59 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
61 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
62 #define IPOIB_CM_RX_TIMEOUT (2 * 256 * HZ)
63 #define IPOIB_CM_RX_DELAY (3 * 256 * HZ)
64 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
66 static struct ib_qp_attr ipoib_cm_err_attr = {
67 .qp_state = IB_QPS_ERR
70 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
72 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
73 .wr_id = IPOIB_CM_RX_DRAIN_WRID,
77 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
78 struct ib_cm_event *event);
80 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
81 u64 mapping[IPOIB_CM_RX_SG])
85 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
87 for (i = 0; i < frags; ++i)
88 ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
91 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
93 struct ipoib_dev_priv *priv = netdev_priv(dev);
94 struct ib_recv_wr *bad_wr;
97 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
99 for (i = 0; i < priv->cm.num_frags; ++i)
100 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
102 ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
104 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
105 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
106 priv->cm.srq_ring[id].mapping);
107 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
108 priv->cm.srq_ring[id].skb = NULL;
114 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
115 struct ipoib_cm_rx *rx,
116 struct ib_recv_wr *wr,
117 struct ib_sge *sge, int id)
119 struct ipoib_dev_priv *priv = netdev_priv(dev);
120 struct ib_recv_wr *bad_wr;
123 wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
125 for (i = 0; i < IPOIB_CM_RX_SG; ++i)
126 sge[i].addr = rx->rx_ring[id].mapping[i];
128 ret = ib_post_recv(rx->qp, wr, &bad_wr);
130 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
131 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
132 rx->rx_ring[id].mapping);
133 dev_kfree_skb_any(rx->rx_ring[id].skb);
134 rx->rx_ring[id].skb = NULL;
140 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
141 struct ipoib_cm_rx_buf *rx_ring,
143 u64 mapping[IPOIB_CM_RX_SG])
145 struct ipoib_dev_priv *priv = netdev_priv(dev);
149 skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
154 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
155 * IP header to a multiple of 16.
157 skb_reserve(skb, 12);
159 mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
161 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
162 dev_kfree_skb_any(skb);
166 for (i = 0; i < frags; i++) {
167 struct page *page = alloc_page(GFP_ATOMIC);
171 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
173 mapping[i + 1] = ib_dma_map_page(priv->ca, page,
174 0, PAGE_SIZE, DMA_FROM_DEVICE);
175 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
179 rx_ring[id].skb = skb;
184 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
187 ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
189 dev_kfree_skb_any(skb);
193 static void ipoib_cm_free_rx_ring(struct net_device *dev,
194 struct ipoib_cm_rx_buf *rx_ring)
196 struct ipoib_dev_priv *priv = netdev_priv(dev);
199 for (i = 0; i < ipoib_recvq_size; ++i)
200 if (rx_ring[i].skb) {
201 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
203 dev_kfree_skb_any(rx_ring[i].skb);
209 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
211 struct ib_send_wr *bad_wr;
212 struct ipoib_cm_rx *p;
214 /* We only reserved 1 extra slot in CQ for drain WRs, so
215 * make sure we have at most 1 outstanding WR. */
216 if (list_empty(&priv->cm.rx_flush_list) ||
217 !list_empty(&priv->cm.rx_drain_list))
221 * QPs on flush list are error state. This way, a "flush
222 * error" WC will be immediately generated for each WR we post.
224 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
225 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
226 ipoib_warn(priv, "failed to post drain wr\n");
228 list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
231 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
233 struct ipoib_cm_rx *p = ctx;
234 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
237 if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
240 spin_lock_irqsave(&priv->lock, flags);
241 list_move(&p->list, &priv->cm.rx_flush_list);
242 p->state = IPOIB_CM_RX_FLUSH;
243 ipoib_cm_start_rx_drain(priv);
244 spin_unlock_irqrestore(&priv->lock, flags);
247 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
248 struct ipoib_cm_rx *p)
250 struct ipoib_dev_priv *priv = netdev_priv(dev);
251 struct ib_qp_init_attr attr = {
252 .event_handler = ipoib_cm_rx_event_handler,
253 .send_cq = priv->recv_cq, /* For drain WR */
254 .recv_cq = priv->recv_cq,
256 .cap.max_send_wr = 1, /* For drain WR */
257 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
258 .sq_sig_type = IB_SIGNAL_ALL_WR,
259 .qp_type = IB_QPT_RC,
263 if (!ipoib_cm_has_srq(dev)) {
264 attr.cap.max_recv_wr = ipoib_recvq_size;
265 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
268 return ib_create_qp(priv->pd, &attr);
271 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
272 struct ib_cm_id *cm_id, struct ib_qp *qp,
275 struct ipoib_dev_priv *priv = netdev_priv(dev);
276 struct ib_qp_attr qp_attr;
277 int qp_attr_mask, ret;
279 qp_attr.qp_state = IB_QPS_INIT;
280 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
282 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
285 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
287 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
290 qp_attr.qp_state = IB_QPS_RTR;
291 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
293 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
296 qp_attr.rq_psn = psn;
297 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
299 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
304 * Current Mellanox HCA firmware won't generate completions
305 * with error for drain WRs unless the QP has been moved to
306 * RTS first. This work-around leaves a window where a QP has
307 * moved to error asynchronously, but this will eventually get
308 * fixed in firmware, so let's not error out if modify QP
311 qp_attr.qp_state = IB_QPS_RTS;
312 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
314 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
317 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
319 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
326 static void ipoib_cm_init_rx_wr(struct net_device *dev,
327 struct ib_recv_wr *wr,
330 struct ipoib_dev_priv *priv = netdev_priv(dev);
333 for (i = 0; i < priv->cm.num_frags; ++i)
334 sge[i].lkey = priv->mr->lkey;
336 sge[0].length = IPOIB_CM_HEAD_SIZE;
337 for (i = 1; i < priv->cm.num_frags; ++i)
338 sge[i].length = PAGE_SIZE;
342 wr->num_sge = priv->cm.num_frags;
345 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
346 struct ipoib_cm_rx *rx)
348 struct ipoib_dev_priv *priv = netdev_priv(dev);
350 struct ib_recv_wr wr;
351 struct ib_sge sge[IPOIB_CM_RX_SG];
356 rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
358 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
359 priv->ca->name, ipoib_recvq_size);
363 t = kmalloc(sizeof *t, GFP_KERNEL);
369 ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
371 spin_lock_irq(&priv->lock);
373 if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
374 spin_unlock_irq(&priv->lock);
375 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
379 ++priv->cm.nonsrq_conn_qp;
381 spin_unlock_irq(&priv->lock);
383 for (i = 0; i < ipoib_recvq_size; ++i) {
384 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
385 rx->rx_ring[i].mapping)) {
386 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
390 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
392 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
393 "failed for buf %d\n", i);
399 rx->recv_count = ipoib_recvq_size;
406 spin_lock_irq(&priv->lock);
407 --priv->cm.nonsrq_conn_qp;
408 spin_unlock_irq(&priv->lock);
412 ipoib_cm_free_rx_ring(dev, rx->rx_ring);
417 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
418 struct ib_qp *qp, struct ib_cm_req_event_param *req,
421 struct ipoib_dev_priv *priv = netdev_priv(dev);
422 struct ipoib_cm_data data = {};
423 struct ib_cm_rep_param rep = {};
425 data.qpn = cpu_to_be32(priv->qp->qp_num);
426 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
428 rep.private_data = &data;
429 rep.private_data_len = sizeof data;
430 rep.flow_control = 0;
431 rep.rnr_retry_count = req->rnr_retry_count;
432 rep.srq = ipoib_cm_has_srq(dev);
433 rep.qp_num = qp->qp_num;
434 rep.starting_psn = psn;
435 return ib_send_cm_rep(cm_id, &rep);
438 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
440 struct net_device *dev = cm_id->context;
441 struct ipoib_dev_priv *priv = netdev_priv(dev);
442 struct ipoib_cm_rx *p;
446 ipoib_dbg(priv, "REQ arrived\n");
447 p = kzalloc(sizeof *p, GFP_KERNEL);
453 p->state = IPOIB_CM_RX_LIVE;
454 p->jiffies = jiffies;
455 INIT_LIST_HEAD(&p->list);
457 p->qp = ipoib_cm_create_rx_qp(dev, p);
459 ret = PTR_ERR(p->qp);
463 psn = random32() & 0xffffff;
464 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
468 if (!ipoib_cm_has_srq(dev)) {
469 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
474 spin_lock_irq(&priv->lock);
475 queue_delayed_work(ipoib_workqueue,
476 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
477 /* Add this entry to passive ids list head, but do not re-add it
478 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
479 p->jiffies = jiffies;
480 if (p->state == IPOIB_CM_RX_LIVE)
481 list_move(&p->list, &priv->cm.passive_ids);
482 spin_unlock_irq(&priv->lock);
484 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
486 ipoib_warn(priv, "failed to send REP: %d\n", ret);
487 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
488 ipoib_warn(priv, "unable to move qp to error state\n");
493 ib_destroy_qp(p->qp);
499 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
500 struct ib_cm_event *event)
502 struct ipoib_cm_rx *p;
503 struct ipoib_dev_priv *priv;
505 switch (event->event) {
506 case IB_CM_REQ_RECEIVED:
507 return ipoib_cm_req_handler(cm_id, event);
508 case IB_CM_DREQ_RECEIVED:
509 ib_send_cm_drep(cm_id, NULL, 0);
511 case IB_CM_REJ_RECEIVED:
513 priv = netdev_priv(p->dev);
514 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
515 ipoib_warn(priv, "unable to move qp to error state\n");
521 /* Adjust length of skb with fragments to match received data */
522 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
523 unsigned int length, struct sk_buff *toskb)
528 /* put header into skb */
529 size = min(length, hdr_space);
534 num_frags = skb_shinfo(skb)->nr_frags;
535 for (i = 0; i < num_frags; i++) {
536 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
539 /* don't need this page */
540 skb_fill_page_desc(toskb, i, skb_frag_page(frag),
542 --skb_shinfo(skb)->nr_frags;
544 size = min(length, (unsigned) PAGE_SIZE);
546 skb_frag_size_set(frag, size);
547 skb->data_len += size;
548 skb->truesize += size;
555 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
557 struct ipoib_dev_priv *priv = netdev_priv(dev);
558 struct ipoib_cm_rx_buf *rx_ring;
559 unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
560 struct sk_buff *skb, *newskb;
561 struct ipoib_cm_rx *p;
563 u64 mapping[IPOIB_CM_RX_SG];
566 struct sk_buff *small_skb;
568 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
571 if (unlikely(wr_id >= ipoib_recvq_size)) {
572 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
573 spin_lock_irqsave(&priv->lock, flags);
574 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
575 ipoib_cm_start_rx_drain(priv);
576 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
577 spin_unlock_irqrestore(&priv->lock, flags);
579 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
580 wr_id, ipoib_recvq_size);
584 p = wc->qp->qp_context;
586 has_srq = ipoib_cm_has_srq(dev);
587 rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
589 skb = rx_ring[wr_id].skb;
591 if (unlikely(wc->status != IB_WC_SUCCESS)) {
592 ipoib_dbg(priv, "cm recv error "
593 "(status=%d, wrid=%d vend_err %x)\n",
594 wc->status, wr_id, wc->vendor_err);
595 ++dev->stats.rx_dropped;
599 if (!--p->recv_count) {
600 spin_lock_irqsave(&priv->lock, flags);
601 list_move(&p->list, &priv->cm.rx_reap_list);
602 spin_unlock_irqrestore(&priv->lock, flags);
603 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
609 if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
610 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
611 spin_lock_irqsave(&priv->lock, flags);
612 p->jiffies = jiffies;
613 /* Move this entry to list head, but do not re-add it
614 * if it has been moved out of list. */
615 if (p->state == IPOIB_CM_RX_LIVE)
616 list_move(&p->list, &priv->cm.passive_ids);
617 spin_unlock_irqrestore(&priv->lock, flags);
621 if (wc->byte_len < IPOIB_CM_COPYBREAK) {
622 int dlen = wc->byte_len;
624 small_skb = dev_alloc_skb(dlen + 12);
626 skb_reserve(small_skb, 12);
627 ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
628 dlen, DMA_FROM_DEVICE);
629 skb_copy_from_linear_data(skb, small_skb->data, dlen);
630 ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
631 dlen, DMA_FROM_DEVICE);
632 skb_put(small_skb, dlen);
638 frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
639 (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
641 newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
642 if (unlikely(!newskb)) {
644 * If we can't allocate a new RX buffer, dump
645 * this packet and reuse the old buffer.
647 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
648 ++dev->stats.rx_dropped;
652 ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
653 memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
655 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
656 wc->byte_len, wc->slid);
658 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
661 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
662 skb_reset_mac_header(skb);
663 skb_pull(skb, IPOIB_ENCAP_LEN);
665 ++dev->stats.rx_packets;
666 dev->stats.rx_bytes += skb->len;
669 /* XXX get correct PACKET_ type here */
670 skb->pkt_type = PACKET_HOST;
671 netif_receive_skb(skb);
675 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
676 ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
677 "for buf %d\n", wr_id);
679 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
684 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
685 "for buf %d\n", wr_id);
690 static inline int post_send(struct ipoib_dev_priv *priv,
691 struct ipoib_cm_tx *tx,
695 struct ib_send_wr *bad_wr;
697 priv->tx_sge[0].addr = addr;
698 priv->tx_sge[0].length = len;
700 priv->tx_wr.num_sge = 1;
701 priv->tx_wr.wr_id = wr_id | IPOIB_OP_CM;
703 return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
706 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
708 struct ipoib_dev_priv *priv = netdev_priv(dev);
709 struct ipoib_cm_tx_buf *tx_req;
713 if (unlikely(skb->len > tx->mtu)) {
714 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
716 ++dev->stats.tx_dropped;
717 ++dev->stats.tx_errors;
718 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
722 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
723 tx->tx_head, skb->len, tx->qp->qp_num);
726 * We put the skb into the tx_ring _before_ we call post_send()
727 * because it's entirely possible that the completion handler will
728 * run before we execute anything after the post_send(). That
729 * means we have to make sure everything is properly recorded and
730 * our state is consistent before we call post_send().
732 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
734 addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
735 if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
736 ++dev->stats.tx_errors;
737 dev_kfree_skb_any(skb);
741 tx_req->mapping = addr;
743 rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
746 ipoib_warn(priv, "post_send failed, error %d\n", rc);
747 ++dev->stats.tx_errors;
748 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
749 dev_kfree_skb_any(skb);
751 dev->trans_start = jiffies;
754 if (++priv->tx_outstanding == ipoib_sendq_size) {
755 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
757 netif_stop_queue(dev);
758 rc = ib_req_notify_cq(priv->send_cq,
759 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
761 ipoib_warn(priv, "request notify on send CQ failed\n");
763 ipoib_send_comp_handler(priv->send_cq, dev);
768 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
770 struct ipoib_dev_priv *priv = netdev_priv(dev);
771 struct ipoib_cm_tx *tx = wc->qp->qp_context;
772 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
773 struct ipoib_cm_tx_buf *tx_req;
776 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
779 if (unlikely(wr_id >= ipoib_sendq_size)) {
780 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
781 wr_id, ipoib_sendq_size);
785 tx_req = &tx->tx_ring[wr_id];
787 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
789 /* FIXME: is this right? Shouldn't we only increment on success? */
790 ++dev->stats.tx_packets;
791 dev->stats.tx_bytes += tx_req->skb->len;
793 dev_kfree_skb_any(tx_req->skb);
798 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
799 netif_queue_stopped(dev) &&
800 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
801 netif_wake_queue(dev);
803 if (wc->status != IB_WC_SUCCESS &&
804 wc->status != IB_WC_WR_FLUSH_ERR) {
805 struct ipoib_neigh *neigh;
807 ipoib_dbg(priv, "failed cm send event "
808 "(status=%d, wrid=%d vend_err %x)\n",
809 wc->status, wr_id, wc->vendor_err);
811 spin_lock_irqsave(&priv->lock, flags);
816 list_del(&neigh->list);
818 ipoib_put_ah(neigh->ah);
819 ipoib_neigh_free(dev, neigh);
824 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
825 list_move(&tx->list, &priv->cm.reap_list);
826 queue_work(ipoib_workqueue, &priv->cm.reap_task);
829 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
831 spin_unlock_irqrestore(&priv->lock, flags);
834 netif_tx_unlock(dev);
837 int ipoib_cm_dev_open(struct net_device *dev)
839 struct ipoib_dev_priv *priv = netdev_priv(dev);
842 if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
845 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
846 if (IS_ERR(priv->cm.id)) {
847 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
848 ret = PTR_ERR(priv->cm.id);
852 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
855 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
856 IPOIB_CM_IETF_ID | priv->qp->qp_num);
863 ib_destroy_cm_id(priv->cm.id);
869 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
871 struct ipoib_dev_priv *priv = netdev_priv(dev);
872 struct ipoib_cm_rx *rx, *n;
875 spin_lock_irq(&priv->lock);
876 list_splice_init(&priv->cm.rx_reap_list, &list);
877 spin_unlock_irq(&priv->lock);
879 list_for_each_entry_safe(rx, n, &list, list) {
880 ib_destroy_cm_id(rx->id);
881 ib_destroy_qp(rx->qp);
882 if (!ipoib_cm_has_srq(dev)) {
883 ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
884 spin_lock_irq(&priv->lock);
885 --priv->cm.nonsrq_conn_qp;
886 spin_unlock_irq(&priv->lock);
892 void ipoib_cm_dev_stop(struct net_device *dev)
894 struct ipoib_dev_priv *priv = netdev_priv(dev);
895 struct ipoib_cm_rx *p;
899 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
902 ib_destroy_cm_id(priv->cm.id);
905 spin_lock_irq(&priv->lock);
906 while (!list_empty(&priv->cm.passive_ids)) {
907 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
908 list_move(&p->list, &priv->cm.rx_error_list);
909 p->state = IPOIB_CM_RX_ERROR;
910 spin_unlock_irq(&priv->lock);
911 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
913 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
914 spin_lock_irq(&priv->lock);
917 /* Wait for all RX to be drained */
920 while (!list_empty(&priv->cm.rx_error_list) ||
921 !list_empty(&priv->cm.rx_flush_list) ||
922 !list_empty(&priv->cm.rx_drain_list)) {
923 if (time_after(jiffies, begin + 5 * HZ)) {
924 ipoib_warn(priv, "RX drain timing out\n");
927 * assume the HW is wedged and just free up everything.
929 list_splice_init(&priv->cm.rx_flush_list,
930 &priv->cm.rx_reap_list);
931 list_splice_init(&priv->cm.rx_error_list,
932 &priv->cm.rx_reap_list);
933 list_splice_init(&priv->cm.rx_drain_list,
934 &priv->cm.rx_reap_list);
937 spin_unlock_irq(&priv->lock);
940 spin_lock_irq(&priv->lock);
943 spin_unlock_irq(&priv->lock);
945 ipoib_cm_free_rx_reap_list(dev);
947 cancel_delayed_work(&priv->cm.stale_task);
950 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
952 struct ipoib_cm_tx *p = cm_id->context;
953 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
954 struct ipoib_cm_data *data = event->private_data;
955 struct sk_buff_head skqueue;
956 struct ib_qp_attr qp_attr;
957 int qp_attr_mask, ret;
960 p->mtu = be32_to_cpu(data->mtu);
962 if (p->mtu <= IPOIB_ENCAP_LEN) {
963 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
964 p->mtu, IPOIB_ENCAP_LEN);
968 qp_attr.qp_state = IB_QPS_RTR;
969 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
971 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
975 qp_attr.rq_psn = 0 /* FIXME */;
976 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
978 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
982 qp_attr.qp_state = IB_QPS_RTS;
983 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
985 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
988 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
990 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
994 skb_queue_head_init(&skqueue);
996 spin_lock_irq(&priv->lock);
997 set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
999 while ((skb = __skb_dequeue(&p->neigh->queue)))
1000 __skb_queue_tail(&skqueue, skb);
1001 spin_unlock_irq(&priv->lock);
1003 while ((skb = __skb_dequeue(&skqueue))) {
1005 if (dev_queue_xmit(skb))
1006 ipoib_warn(priv, "dev_queue_xmit failed "
1007 "to requeue packet\n");
1010 ret = ib_send_cm_rtu(cm_id, NULL, 0);
1012 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1018 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1020 struct ipoib_dev_priv *priv = netdev_priv(dev);
1021 struct ib_qp_init_attr attr = {
1022 .send_cq = priv->recv_cq,
1023 .recv_cq = priv->recv_cq,
1024 .srq = priv->cm.srq,
1025 .cap.max_send_wr = ipoib_sendq_size,
1026 .cap.max_send_sge = 1,
1027 .sq_sig_type = IB_SIGNAL_ALL_WR,
1028 .qp_type = IB_QPT_RC,
1032 return ib_create_qp(priv->pd, &attr);
1035 static int ipoib_cm_send_req(struct net_device *dev,
1036 struct ib_cm_id *id, struct ib_qp *qp,
1038 struct ib_sa_path_rec *pathrec)
1040 struct ipoib_dev_priv *priv = netdev_priv(dev);
1041 struct ipoib_cm_data data = {};
1042 struct ib_cm_req_param req = {};
1044 data.qpn = cpu_to_be32(priv->qp->qp_num);
1045 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1047 req.primary_path = pathrec;
1048 req.alternate_path = NULL;
1049 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1050 req.qp_num = qp->qp_num;
1051 req.qp_type = qp->qp_type;
1052 req.private_data = &data;
1053 req.private_data_len = sizeof data;
1054 req.flow_control = 0;
1056 req.starting_psn = 0; /* FIXME */
1059 * Pick some arbitrary defaults here; we could make these
1060 * module parameters if anyone cared about setting them.
1062 req.responder_resources = 4;
1063 req.remote_cm_response_timeout = 20;
1064 req.local_cm_response_timeout = 20;
1065 req.retry_count = 0; /* RFC draft warns against retries */
1066 req.rnr_retry_count = 0; /* RFC draft warns against retries */
1067 req.max_cm_retries = 15;
1068 req.srq = ipoib_cm_has_srq(dev);
1069 return ib_send_cm_req(id, &req);
1072 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1073 struct ib_cm_id *cm_id, struct ib_qp *qp)
1075 struct ipoib_dev_priv *priv = netdev_priv(dev);
1076 struct ib_qp_attr qp_attr;
1077 int qp_attr_mask, ret;
1078 ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1080 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1084 qp_attr.qp_state = IB_QPS_INIT;
1085 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1086 qp_attr.port_num = priv->port;
1087 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1089 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1091 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1097 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1098 struct ib_sa_path_rec *pathrec)
1100 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1103 p->tx_ring = vzalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1105 ipoib_warn(priv, "failed to allocate tx ring\n");
1110 p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1111 if (IS_ERR(p->qp)) {
1112 ret = PTR_ERR(p->qp);
1113 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1117 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1118 if (IS_ERR(p->id)) {
1119 ret = PTR_ERR(p->id);
1120 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1124 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp);
1126 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1130 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1132 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1136 ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1137 p->qp->qp_num, pathrec->dgid.raw, qpn);
1143 ib_destroy_cm_id(p->id);
1146 ib_destroy_qp(p->qp);
1154 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1156 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1157 struct ipoib_cm_tx_buf *tx_req;
1158 unsigned long begin;
1160 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1161 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1164 ib_destroy_cm_id(p->id);
1167 /* Wait for all sends to complete */
1169 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1170 if (time_after(jiffies, begin + 5 * HZ)) {
1171 ipoib_warn(priv, "timing out; %d sends not completed\n",
1172 p->tx_head - p->tx_tail);
1182 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1183 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1184 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1186 dev_kfree_skb_any(tx_req->skb);
1188 netif_tx_lock_bh(p->dev);
1189 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1190 netif_queue_stopped(p->dev) &&
1191 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1192 netif_wake_queue(p->dev);
1193 netif_tx_unlock_bh(p->dev);
1197 ib_destroy_qp(p->qp);
1203 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1204 struct ib_cm_event *event)
1206 struct ipoib_cm_tx *tx = cm_id->context;
1207 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1208 struct net_device *dev = priv->dev;
1209 struct ipoib_neigh *neigh;
1210 unsigned long flags;
1213 switch (event->event) {
1214 case IB_CM_DREQ_RECEIVED:
1215 ipoib_dbg(priv, "DREQ received.\n");
1216 ib_send_cm_drep(cm_id, NULL, 0);
1218 case IB_CM_REP_RECEIVED:
1219 ipoib_dbg(priv, "REP received.\n");
1220 ret = ipoib_cm_rep_handler(cm_id, event);
1222 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1225 case IB_CM_REQ_ERROR:
1226 case IB_CM_REJ_RECEIVED:
1227 case IB_CM_TIMEWAIT_EXIT:
1228 ipoib_dbg(priv, "CM error %d.\n", event->event);
1229 netif_tx_lock_bh(dev);
1230 spin_lock_irqsave(&priv->lock, flags);
1235 list_del(&neigh->list);
1237 ipoib_put_ah(neigh->ah);
1238 ipoib_neigh_free(dev, neigh);
1243 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1244 list_move(&tx->list, &priv->cm.reap_list);
1245 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1248 spin_unlock_irqrestore(&priv->lock, flags);
1249 netif_tx_unlock_bh(dev);
1258 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1259 struct ipoib_neigh *neigh)
1261 struct ipoib_dev_priv *priv = netdev_priv(dev);
1262 struct ipoib_cm_tx *tx;
1264 tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1272 list_add(&tx->list, &priv->cm.start_list);
1273 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1274 queue_work(ipoib_workqueue, &priv->cm.start_task);
1278 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1280 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1281 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1282 list_move(&tx->list, &priv->cm.reap_list);
1283 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1284 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1285 tx->neigh->dgid.raw);
1290 #define QPN_AND_OPTIONS_OFFSET 4
1292 static void ipoib_cm_tx_start(struct work_struct *work)
1294 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1296 struct net_device *dev = priv->dev;
1297 struct ipoib_neigh *neigh;
1298 struct ipoib_cm_tx *p;
1299 unsigned long flags;
1300 struct ipoib_path *path;
1303 struct ib_sa_path_rec pathrec;
1306 netif_tx_lock_bh(dev);
1307 spin_lock_irqsave(&priv->lock, flags);
1309 while (!list_empty(&priv->cm.start_list)) {
1310 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1311 list_del_init(&p->list);
1314 qpn = IPOIB_QPN(neigh->neighbour->ha);
1316 * As long as the search is with these 2 locks,
1317 * path existence indicates its validity.
1319 path = __path_find(dev, neigh->neighbour->ha + QPN_AND_OPTIONS_OFFSET);
1321 pr_info("%s ignore not valid path %pI6\n",
1323 neigh->neighbour->ha + QPN_AND_OPTIONS_OFFSET);
1326 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1328 spin_unlock_irqrestore(&priv->lock, flags);
1329 netif_tx_unlock_bh(dev);
1331 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1333 netif_tx_lock_bh(dev);
1334 spin_lock_irqsave(&priv->lock, flags);
1341 list_del(&neigh->list);
1343 ipoib_put_ah(neigh->ah);
1344 ipoib_neigh_free(dev, neigh);
1351 spin_unlock_irqrestore(&priv->lock, flags);
1352 netif_tx_unlock_bh(dev);
1355 static void ipoib_cm_tx_reap(struct work_struct *work)
1357 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1359 struct net_device *dev = priv->dev;
1360 struct ipoib_cm_tx *p;
1361 unsigned long flags;
1363 netif_tx_lock_bh(dev);
1364 spin_lock_irqsave(&priv->lock, flags);
1366 while (!list_empty(&priv->cm.reap_list)) {
1367 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1368 list_del_init(&p->list);
1369 spin_unlock_irqrestore(&priv->lock, flags);
1370 netif_tx_unlock_bh(dev);
1371 ipoib_cm_tx_destroy(p);
1372 netif_tx_lock_bh(dev);
1373 spin_lock_irqsave(&priv->lock, flags);
1376 spin_unlock_irqrestore(&priv->lock, flags);
1377 netif_tx_unlock_bh(dev);
1380 static void ipoib_cm_skb_reap(struct work_struct *work)
1382 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1384 struct net_device *dev = priv->dev;
1385 struct sk_buff *skb;
1386 unsigned long flags;
1387 unsigned mtu = priv->mcast_mtu;
1389 netif_tx_lock_bh(dev);
1390 spin_lock_irqsave(&priv->lock, flags);
1392 while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1393 spin_unlock_irqrestore(&priv->lock, flags);
1394 netif_tx_unlock_bh(dev);
1396 if (skb->protocol == htons(ETH_P_IP))
1397 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1398 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1399 else if (skb->protocol == htons(ETH_P_IPV6))
1400 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1402 dev_kfree_skb_any(skb);
1404 netif_tx_lock_bh(dev);
1405 spin_lock_irqsave(&priv->lock, flags);
1408 spin_unlock_irqrestore(&priv->lock, flags);
1409 netif_tx_unlock_bh(dev);
1412 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1415 struct ipoib_dev_priv *priv = netdev_priv(dev);
1416 int e = skb_queue_empty(&priv->cm.skb_queue);
1419 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu);
1421 skb_queue_tail(&priv->cm.skb_queue, skb);
1423 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1426 static void ipoib_cm_rx_reap(struct work_struct *work)
1428 ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1429 cm.rx_reap_task)->dev);
1432 static void ipoib_cm_stale_task(struct work_struct *work)
1434 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1435 cm.stale_task.work);
1436 struct ipoib_cm_rx *p;
1439 spin_lock_irq(&priv->lock);
1440 while (!list_empty(&priv->cm.passive_ids)) {
1441 /* List is sorted by LRU, start from tail,
1442 * stop when we see a recently used entry */
1443 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1444 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1446 list_move(&p->list, &priv->cm.rx_error_list);
1447 p->state = IPOIB_CM_RX_ERROR;
1448 spin_unlock_irq(&priv->lock);
1449 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1451 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1452 spin_lock_irq(&priv->lock);
1455 if (!list_empty(&priv->cm.passive_ids))
1456 queue_delayed_work(ipoib_workqueue,
1457 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1458 spin_unlock_irq(&priv->lock);
1462 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1465 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1467 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1468 return sprintf(buf, "connected\n");
1470 return sprintf(buf, "datagram\n");
1473 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1474 const char *buf, size_t count)
1476 struct net_device *dev = to_net_dev(d);
1477 struct ipoib_dev_priv *priv = netdev_priv(dev);
1479 if (!rtnl_trylock())
1480 return restart_syscall();
1482 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
1483 !strcmp(buf, "connected\n")) ||
1484 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
1485 !strcmp(buf, "datagram\n"))) {
1490 /* flush paths if we switch modes so that connections are restarted */
1491 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1492 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1493 ipoib_warn(priv, "enabling connected mode "
1494 "will cause multicast packet drops\n");
1495 netdev_update_features(dev);
1497 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1499 ipoib_flush_paths(dev);
1503 if (!strcmp(buf, "datagram\n")) {
1504 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1505 netdev_update_features(dev);
1506 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1508 ipoib_flush_paths(dev);
1517 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1519 int ipoib_cm_add_mode_attr(struct net_device *dev)
1521 return device_create_file(&dev->dev, &dev_attr_mode);
1524 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1526 struct ipoib_dev_priv *priv = netdev_priv(dev);
1527 struct ib_srq_init_attr srq_init_attr = {
1528 .srq_type = IB_SRQT_BASIC,
1530 .max_wr = ipoib_recvq_size,
1535 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1536 if (IS_ERR(priv->cm.srq)) {
1537 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1538 printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1539 priv->ca->name, PTR_ERR(priv->cm.srq));
1540 priv->cm.srq = NULL;
1544 priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1545 if (!priv->cm.srq_ring) {
1546 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1547 priv->ca->name, ipoib_recvq_size);
1548 ib_destroy_srq(priv->cm.srq);
1549 priv->cm.srq = NULL;
1555 int ipoib_cm_dev_init(struct net_device *dev)
1557 struct ipoib_dev_priv *priv = netdev_priv(dev);
1559 struct ib_device_attr attr;
1561 INIT_LIST_HEAD(&priv->cm.passive_ids);
1562 INIT_LIST_HEAD(&priv->cm.reap_list);
1563 INIT_LIST_HEAD(&priv->cm.start_list);
1564 INIT_LIST_HEAD(&priv->cm.rx_error_list);
1565 INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1566 INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1567 INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1568 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1569 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1570 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1571 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1572 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1574 skb_queue_head_init(&priv->cm.skb_queue);
1576 ret = ib_query_device(priv->ca, &attr);
1578 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1582 ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1584 attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1585 ipoib_cm_create_srq(dev, attr.max_srq_sge);
1586 if (ipoib_cm_has_srq(dev)) {
1587 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1588 priv->cm.num_frags = attr.max_srq_sge;
1589 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1590 priv->cm.max_cm_mtu, priv->cm.num_frags);
1592 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1593 priv->cm.num_frags = IPOIB_CM_RX_SG;
1596 ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1598 if (ipoib_cm_has_srq(dev)) {
1599 for (i = 0; i < ipoib_recvq_size; ++i) {
1600 if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1601 priv->cm.num_frags - 1,
1602 priv->cm.srq_ring[i].mapping)) {
1603 ipoib_warn(priv, "failed to allocate "
1604 "receive buffer %d\n", i);
1605 ipoib_cm_dev_cleanup(dev);
1609 if (ipoib_cm_post_receive_srq(dev, i)) {
1610 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1611 "failed for buf %d\n", i);
1612 ipoib_cm_dev_cleanup(dev);
1618 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1622 void ipoib_cm_dev_cleanup(struct net_device *dev)
1624 struct ipoib_dev_priv *priv = netdev_priv(dev);
1630 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1632 ret = ib_destroy_srq(priv->cm.srq);
1634 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1636 priv->cm.srq = NULL;
1637 if (!priv->cm.srq_ring)
1640 ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1641 priv->cm.srq_ring = NULL;