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
35 #include <rdma/ib_cm.h>
36 #include <rdma/ib_cache.h>
39 #include <linux/icmpv6.h>
40 #include <linux/delay.h>
42 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
43 static int data_debug_level;
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");
52 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
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)
66 static struct ib_qp_attr ipoib_cm_err_attr = {
67 .qp_state = IB_QPS_ERR
70 #define IPOIB_CM_RX_DRAIN_WRID 0x7fffffff
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_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
91 static int ipoib_cm_post_receive(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_CM_OP_SRQ;
99 for (i = 0; i < IPOIB_CM_RX_SG; ++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, IPOIB_CM_RX_SG - 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 struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, int id, int frags,
115 u64 mapping[IPOIB_CM_RX_SG])
117 struct ipoib_dev_priv *priv = netdev_priv(dev);
121 skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
126 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
127 * IP header to a multiple of 16.
129 skb_reserve(skb, 12);
131 mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
133 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
134 dev_kfree_skb_any(skb);
138 for (i = 0; i < frags; i++) {
139 struct page *page = alloc_page(GFP_ATOMIC);
143 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
145 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
146 0, PAGE_SIZE, DMA_FROM_DEVICE);
147 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
151 priv->cm.srq_ring[id].skb = skb;
156 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
159 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
161 dev_kfree_skb_any(skb);
165 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv* priv)
167 struct ib_send_wr *bad_wr;
168 struct ipoib_cm_rx *p;
170 /* We only reserved 1 extra slot in CQ for drain WRs, so
171 * make sure we have at most 1 outstanding WR. */
172 if (list_empty(&priv->cm.rx_flush_list) ||
173 !list_empty(&priv->cm.rx_drain_list))
177 * QPs on flush list are error state. This way, a "flush
178 * error" WC will be immediately generated for each WR we post.
180 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
181 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
182 ipoib_warn(priv, "failed to post drain wr\n");
184 list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
187 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
189 struct ipoib_cm_rx *p = ctx;
190 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
193 if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
196 spin_lock_irqsave(&priv->lock, flags);
197 list_move(&p->list, &priv->cm.rx_flush_list);
198 p->state = IPOIB_CM_RX_FLUSH;
199 ipoib_cm_start_rx_drain(priv);
200 spin_unlock_irqrestore(&priv->lock, flags);
203 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
204 struct ipoib_cm_rx *p)
206 struct ipoib_dev_priv *priv = netdev_priv(dev);
207 struct ib_qp_init_attr attr = {
208 .event_handler = ipoib_cm_rx_event_handler,
209 .send_cq = priv->cq, /* For drain WR */
212 .cap.max_send_wr = 1, /* For drain WR */
213 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
214 .sq_sig_type = IB_SIGNAL_ALL_WR,
215 .qp_type = IB_QPT_RC,
218 return ib_create_qp(priv->pd, &attr);
221 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
222 struct ib_cm_id *cm_id, struct ib_qp *qp,
225 struct ipoib_dev_priv *priv = netdev_priv(dev);
226 struct ib_qp_attr qp_attr;
227 int qp_attr_mask, ret;
229 qp_attr.qp_state = IB_QPS_INIT;
230 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
232 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
235 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
237 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
240 qp_attr.qp_state = IB_QPS_RTR;
241 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
243 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
246 qp_attr.rq_psn = psn;
247 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
249 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
254 * Current Mellanox HCA firmware won't generate completions
255 * with error for drain WRs unless the QP has been moved to
256 * RTS first. This work-around leaves a window where a QP has
257 * moved to error asynchronously, but this will eventually get
258 * fixed in firmware, so let's not error out if modify QP
261 qp_attr.qp_state = IB_QPS_RTS;
262 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
264 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
267 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
269 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
276 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
277 struct ib_qp *qp, struct ib_cm_req_event_param *req,
280 struct ipoib_dev_priv *priv = netdev_priv(dev);
281 struct ipoib_cm_data data = {};
282 struct ib_cm_rep_param rep = {};
284 data.qpn = cpu_to_be32(priv->qp->qp_num);
285 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
287 rep.private_data = &data;
288 rep.private_data_len = sizeof data;
289 rep.flow_control = 0;
290 rep.rnr_retry_count = req->rnr_retry_count;
291 rep.target_ack_delay = 20; /* FIXME */
293 rep.qp_num = qp->qp_num;
294 rep.starting_psn = psn;
295 return ib_send_cm_rep(cm_id, &rep);
298 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
300 struct net_device *dev = cm_id->context;
301 struct ipoib_dev_priv *priv = netdev_priv(dev);
302 struct ipoib_cm_rx *p;
306 ipoib_dbg(priv, "REQ arrived\n");
307 p = kzalloc(sizeof *p, GFP_KERNEL);
313 p->state = IPOIB_CM_RX_LIVE;
314 p->jiffies = jiffies;
315 INIT_LIST_HEAD(&p->list);
317 p->qp = ipoib_cm_create_rx_qp(dev, p);
319 ret = PTR_ERR(p->qp);
323 psn = random32() & 0xffffff;
324 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
328 spin_lock_irq(&priv->lock);
329 queue_delayed_work(ipoib_workqueue,
330 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
331 /* Add this entry to passive ids list head, but do not re-add it
332 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
333 p->jiffies = jiffies;
334 if (p->state == IPOIB_CM_RX_LIVE)
335 list_move(&p->list, &priv->cm.passive_ids);
336 spin_unlock_irq(&priv->lock);
338 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
340 ipoib_warn(priv, "failed to send REP: %d\n", ret);
341 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
342 ipoib_warn(priv, "unable to move qp to error state\n");
347 ib_destroy_qp(p->qp);
353 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
354 struct ib_cm_event *event)
356 struct ipoib_cm_rx *p;
357 struct ipoib_dev_priv *priv;
359 switch (event->event) {
360 case IB_CM_REQ_RECEIVED:
361 return ipoib_cm_req_handler(cm_id, event);
362 case IB_CM_DREQ_RECEIVED:
364 ib_send_cm_drep(cm_id, NULL, 0);
366 case IB_CM_REJ_RECEIVED:
368 priv = netdev_priv(p->dev);
369 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
370 ipoib_warn(priv, "unable to move qp to error state\n");
376 /* Adjust length of skb with fragments to match received data */
377 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
378 unsigned int length, struct sk_buff *toskb)
383 /* put header into skb */
384 size = min(length, hdr_space);
389 num_frags = skb_shinfo(skb)->nr_frags;
390 for (i = 0; i < num_frags; i++) {
391 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
394 /* don't need this page */
395 skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
396 --skb_shinfo(skb)->nr_frags;
398 size = min(length, (unsigned) PAGE_SIZE);
401 skb->data_len += size;
402 skb->truesize += size;
409 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
411 struct ipoib_dev_priv *priv = netdev_priv(dev);
412 unsigned int wr_id = wc->wr_id & ~IPOIB_CM_OP_SRQ;
413 struct sk_buff *skb, *newskb;
414 struct ipoib_cm_rx *p;
416 u64 mapping[IPOIB_CM_RX_SG];
419 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
422 if (unlikely(wr_id >= ipoib_recvq_size)) {
423 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~IPOIB_CM_OP_SRQ)) {
424 spin_lock_irqsave(&priv->lock, flags);
425 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
426 ipoib_cm_start_rx_drain(priv);
427 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
428 spin_unlock_irqrestore(&priv->lock, flags);
430 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
431 wr_id, ipoib_recvq_size);
435 skb = priv->cm.srq_ring[wr_id].skb;
437 if (unlikely(wc->status != IB_WC_SUCCESS)) {
438 ipoib_dbg(priv, "cm recv error "
439 "(status=%d, wrid=%d vend_err %x)\n",
440 wc->status, wr_id, wc->vendor_err);
441 ++priv->stats.rx_dropped;
445 if (!likely(wr_id & IPOIB_CM_RX_UPDATE_MASK)) {
446 p = wc->qp->qp_context;
447 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
448 spin_lock_irqsave(&priv->lock, flags);
449 p->jiffies = jiffies;
450 /* Move this entry to list head, but do not re-add it
451 * if it has been moved out of list. */
452 if (p->state == IPOIB_CM_RX_LIVE)
453 list_move(&p->list, &priv->cm.passive_ids);
454 spin_unlock_irqrestore(&priv->lock, flags);
458 frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
459 (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
461 newskb = ipoib_cm_alloc_rx_skb(dev, wr_id, frags, mapping);
462 if (unlikely(!newskb)) {
464 * If we can't allocate a new RX buffer, dump
465 * this packet and reuse the old buffer.
467 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
468 ++priv->stats.rx_dropped;
472 ipoib_cm_dma_unmap_rx(priv, frags, priv->cm.srq_ring[wr_id].mapping);
473 memcpy(priv->cm.srq_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
475 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
476 wc->byte_len, wc->slid);
478 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
480 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
481 skb_reset_mac_header(skb);
482 skb_pull(skb, IPOIB_ENCAP_LEN);
484 dev->last_rx = jiffies;
485 ++priv->stats.rx_packets;
486 priv->stats.rx_bytes += skb->len;
489 /* XXX get correct PACKET_ type here */
490 skb->pkt_type = PACKET_HOST;
491 netif_receive_skb(skb);
494 if (unlikely(ipoib_cm_post_receive(dev, wr_id)))
495 ipoib_warn(priv, "ipoib_cm_post_receive failed "
496 "for buf %d\n", wr_id);
499 static inline int post_send(struct ipoib_dev_priv *priv,
500 struct ipoib_cm_tx *tx,
504 struct ib_send_wr *bad_wr;
506 priv->tx_sge.addr = addr;
507 priv->tx_sge.length = len;
509 priv->tx_wr.wr_id = wr_id;
511 return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
514 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
516 struct ipoib_dev_priv *priv = netdev_priv(dev);
517 struct ipoib_tx_buf *tx_req;
520 if (unlikely(skb->len > tx->mtu)) {
521 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
523 ++priv->stats.tx_dropped;
524 ++priv->stats.tx_errors;
525 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
529 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
530 tx->tx_head, skb->len, tx->qp->qp_num);
533 * We put the skb into the tx_ring _before_ we call post_send()
534 * because it's entirely possible that the completion handler will
535 * run before we execute anything after the post_send(). That
536 * means we have to make sure everything is properly recorded and
537 * our state is consistent before we call post_send().
539 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
541 addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
542 if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
543 ++priv->stats.tx_errors;
544 dev_kfree_skb_any(skb);
548 tx_req->mapping = addr;
550 if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
552 ipoib_warn(priv, "post_send failed\n");
553 ++priv->stats.tx_errors;
554 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
555 dev_kfree_skb_any(skb);
557 dev->trans_start = jiffies;
560 if (tx->tx_head - tx->tx_tail == ipoib_sendq_size) {
561 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
563 netif_stop_queue(dev);
564 set_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
569 static void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ipoib_cm_tx *tx,
572 struct ipoib_dev_priv *priv = netdev_priv(dev);
573 unsigned int wr_id = wc->wr_id;
574 struct ipoib_tx_buf *tx_req;
577 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
580 if (unlikely(wr_id >= ipoib_sendq_size)) {
581 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
582 wr_id, ipoib_sendq_size);
586 tx_req = &tx->tx_ring[wr_id];
588 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
590 /* FIXME: is this right? Shouldn't we only increment on success? */
591 ++priv->stats.tx_packets;
592 priv->stats.tx_bytes += tx_req->skb->len;
594 dev_kfree_skb_any(tx_req->skb);
596 spin_lock_irqsave(&priv->tx_lock, flags);
598 if (unlikely(test_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags)) &&
599 tx->tx_head - tx->tx_tail <= ipoib_sendq_size >> 1) {
600 clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags);
601 netif_wake_queue(dev);
604 if (wc->status != IB_WC_SUCCESS &&
605 wc->status != IB_WC_WR_FLUSH_ERR) {
606 struct ipoib_neigh *neigh;
608 ipoib_dbg(priv, "failed cm send event "
609 "(status=%d, wrid=%d vend_err %x)\n",
610 wc->status, wr_id, wc->vendor_err);
612 spin_lock(&priv->lock);
617 list_del(&neigh->list);
619 ipoib_put_ah(neigh->ah);
620 ipoib_neigh_free(dev, neigh);
625 /* queue would be re-started anyway when TX is destroyed,
626 * but it makes sense to do it ASAP here. */
627 if (test_and_clear_bit(IPOIB_FLAG_NETIF_STOPPED, &tx->flags))
628 netif_wake_queue(dev);
630 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
631 list_move(&tx->list, &priv->cm.reap_list);
632 queue_work(ipoib_workqueue, &priv->cm.reap_task);
635 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
637 spin_unlock(&priv->lock);
640 spin_unlock_irqrestore(&priv->tx_lock, flags);
643 static void ipoib_cm_tx_completion(struct ib_cq *cq, void *tx_ptr)
645 struct ipoib_cm_tx *tx = tx_ptr;
648 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
650 n = ib_poll_cq(cq, IPOIB_NUM_WC, tx->ibwc);
651 for (i = 0; i < n; ++i)
652 ipoib_cm_handle_tx_wc(tx->dev, tx, tx->ibwc + i);
653 } while (n == IPOIB_NUM_WC);
656 int ipoib_cm_dev_open(struct net_device *dev)
658 struct ipoib_dev_priv *priv = netdev_priv(dev);
661 if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
664 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
665 if (IS_ERR(priv->cm.id)) {
666 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
667 ret = PTR_ERR(priv->cm.id);
671 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
674 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
675 IPOIB_CM_IETF_ID | priv->qp->qp_num);
682 ib_destroy_cm_id(priv->cm.id);
688 void ipoib_cm_dev_stop(struct net_device *dev)
690 struct ipoib_dev_priv *priv = netdev_priv(dev);
691 struct ipoib_cm_rx *p, *n;
696 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
699 ib_destroy_cm_id(priv->cm.id);
702 spin_lock_irq(&priv->lock);
703 while (!list_empty(&priv->cm.passive_ids)) {
704 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
705 list_move(&p->list, &priv->cm.rx_error_list);
706 p->state = IPOIB_CM_RX_ERROR;
707 spin_unlock_irq(&priv->lock);
708 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
710 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
711 spin_lock_irq(&priv->lock);
714 /* Wait for all RX to be drained */
717 while (!list_empty(&priv->cm.rx_error_list) ||
718 !list_empty(&priv->cm.rx_flush_list) ||
719 !list_empty(&priv->cm.rx_drain_list)) {
720 if (time_after(jiffies, begin + 5 * HZ)) {
721 ipoib_warn(priv, "RX drain timing out\n");
724 * assume the HW is wedged and just free up everything.
726 list_splice_init(&priv->cm.rx_flush_list, &list);
727 list_splice_init(&priv->cm.rx_error_list, &list);
728 list_splice_init(&priv->cm.rx_drain_list, &list);
731 spin_unlock_irq(&priv->lock);
734 spin_lock_irq(&priv->lock);
737 list_splice_init(&priv->cm.rx_reap_list, &list);
739 spin_unlock_irq(&priv->lock);
741 list_for_each_entry_safe(p, n, &list, list) {
742 ib_destroy_cm_id(p->id);
743 ib_destroy_qp(p->qp);
747 cancel_delayed_work(&priv->cm.stale_task);
750 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
752 struct ipoib_cm_tx *p = cm_id->context;
753 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
754 struct ipoib_cm_data *data = event->private_data;
755 struct sk_buff_head skqueue;
756 struct ib_qp_attr qp_attr;
757 int qp_attr_mask, ret;
760 p->mtu = be32_to_cpu(data->mtu);
762 if (p->mtu <= IPOIB_ENCAP_LEN) {
763 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
764 p->mtu, IPOIB_ENCAP_LEN);
768 qp_attr.qp_state = IB_QPS_RTR;
769 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
771 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
775 qp_attr.rq_psn = 0 /* FIXME */;
776 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
778 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
782 qp_attr.qp_state = IB_QPS_RTS;
783 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
785 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
788 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
790 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
794 skb_queue_head_init(&skqueue);
796 spin_lock_irq(&priv->lock);
797 set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
799 while ((skb = __skb_dequeue(&p->neigh->queue)))
800 __skb_queue_tail(&skqueue, skb);
801 spin_unlock_irq(&priv->lock);
803 while ((skb = __skb_dequeue(&skqueue))) {
805 if (dev_queue_xmit(skb))
806 ipoib_warn(priv, "dev_queue_xmit failed "
807 "to requeue packet\n");
810 ret = ib_send_cm_rtu(cm_id, NULL, 0);
812 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
818 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ib_cq *cq)
820 struct ipoib_dev_priv *priv = netdev_priv(dev);
821 struct ib_qp_init_attr attr = {};
822 attr.recv_cq = priv->cq;
823 attr.srq = priv->cm.srq;
824 attr.cap.max_send_wr = ipoib_sendq_size;
825 attr.cap.max_send_sge = 1;
826 attr.sq_sig_type = IB_SIGNAL_ALL_WR;
827 attr.qp_type = IB_QPT_RC;
829 return ib_create_qp(priv->pd, &attr);
832 static int ipoib_cm_send_req(struct net_device *dev,
833 struct ib_cm_id *id, struct ib_qp *qp,
835 struct ib_sa_path_rec *pathrec)
837 struct ipoib_dev_priv *priv = netdev_priv(dev);
838 struct ipoib_cm_data data = {};
839 struct ib_cm_req_param req = {};
841 data.qpn = cpu_to_be32(priv->qp->qp_num);
842 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
844 req.primary_path = pathrec;
845 req.alternate_path = NULL;
846 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
847 req.qp_num = qp->qp_num;
848 req.qp_type = qp->qp_type;
849 req.private_data = &data;
850 req.private_data_len = sizeof data;
851 req.flow_control = 0;
853 req.starting_psn = 0; /* FIXME */
856 * Pick some arbitrary defaults here; we could make these
857 * module parameters if anyone cared about setting them.
859 req.responder_resources = 4;
860 req.remote_cm_response_timeout = 20;
861 req.local_cm_response_timeout = 20;
862 req.retry_count = 0; /* RFC draft warns against retries */
863 req.rnr_retry_count = 0; /* RFC draft warns against retries */
864 req.max_cm_retries = 15;
866 return ib_send_cm_req(id, &req);
869 static int ipoib_cm_modify_tx_init(struct net_device *dev,
870 struct ib_cm_id *cm_id, struct ib_qp *qp)
872 struct ipoib_dev_priv *priv = netdev_priv(dev);
873 struct ib_qp_attr qp_attr;
874 int qp_attr_mask, ret;
875 ret = ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
877 ipoib_warn(priv, "pkey 0x%x not in cache: %d\n", priv->pkey, ret);
881 qp_attr.qp_state = IB_QPS_INIT;
882 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
883 qp_attr.port_num = priv->port;
884 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
886 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
888 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
894 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
895 struct ib_sa_path_rec *pathrec)
897 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
900 p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring,
903 ipoib_warn(priv, "failed to allocate tx ring\n");
908 p->cq = ib_create_cq(priv->ca, ipoib_cm_tx_completion, NULL, p,
909 ipoib_sendq_size + 1, 0);
911 ret = PTR_ERR(p->cq);
912 ipoib_warn(priv, "failed to allocate tx cq: %d\n", ret);
916 ret = ib_req_notify_cq(p->cq, IB_CQ_NEXT_COMP);
918 ipoib_warn(priv, "failed to request completion notification: %d\n", ret);
922 p->qp = ipoib_cm_create_tx_qp(p->dev, p->cq);
924 ret = PTR_ERR(p->qp);
925 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
929 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
931 ret = PTR_ERR(p->id);
932 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
936 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp);
938 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
942 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
944 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
948 ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
949 p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
955 ib_destroy_cm_id(p->id);
958 ib_destroy_qp(p->qp);
962 ib_destroy_cq(p->cq);
969 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
971 struct ipoib_dev_priv *priv = netdev_priv(p->dev);
972 struct ipoib_tx_buf *tx_req;
974 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
975 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
978 ib_destroy_cm_id(p->id);
981 ib_destroy_qp(p->qp);
984 ib_destroy_cq(p->cq);
986 if (test_bit(IPOIB_FLAG_NETIF_STOPPED, &p->flags))
987 netif_wake_queue(p->dev);
990 while ((int) p->tx_tail - (int) p->tx_head < 0) {
991 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
992 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
994 dev_kfree_skb_any(tx_req->skb);
1004 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1005 struct ib_cm_event *event)
1007 struct ipoib_cm_tx *tx = cm_id->context;
1008 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1009 struct net_device *dev = priv->dev;
1010 struct ipoib_neigh *neigh;
1013 switch (event->event) {
1014 case IB_CM_DREQ_RECEIVED:
1015 ipoib_dbg(priv, "DREQ received.\n");
1016 ib_send_cm_drep(cm_id, NULL, 0);
1018 case IB_CM_REP_RECEIVED:
1019 ipoib_dbg(priv, "REP received.\n");
1020 ret = ipoib_cm_rep_handler(cm_id, event);
1022 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1025 case IB_CM_REQ_ERROR:
1026 case IB_CM_REJ_RECEIVED:
1027 case IB_CM_TIMEWAIT_EXIT:
1028 ipoib_dbg(priv, "CM error %d.\n", event->event);
1029 spin_lock_irq(&priv->tx_lock);
1030 spin_lock(&priv->lock);
1035 list_del(&neigh->list);
1037 ipoib_put_ah(neigh->ah);
1038 ipoib_neigh_free(dev, neigh);
1043 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1044 list_move(&tx->list, &priv->cm.reap_list);
1045 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1048 spin_unlock(&priv->lock);
1049 spin_unlock_irq(&priv->tx_lock);
1058 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1059 struct ipoib_neigh *neigh)
1061 struct ipoib_dev_priv *priv = netdev_priv(dev);
1062 struct ipoib_cm_tx *tx;
1064 tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1072 list_add(&tx->list, &priv->cm.start_list);
1073 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1074 queue_work(ipoib_workqueue, &priv->cm.start_task);
1078 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1080 struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1081 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1082 list_move(&tx->list, &priv->cm.reap_list);
1083 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1084 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1085 IPOIB_GID_ARG(tx->neigh->dgid));
1090 static void ipoib_cm_tx_start(struct work_struct *work)
1092 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1094 struct net_device *dev = priv->dev;
1095 struct ipoib_neigh *neigh;
1096 struct ipoib_cm_tx *p;
1097 unsigned long flags;
1100 struct ib_sa_path_rec pathrec;
1103 spin_lock_irqsave(&priv->tx_lock, flags);
1104 spin_lock(&priv->lock);
1105 while (!list_empty(&priv->cm.start_list)) {
1106 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1107 list_del_init(&p->list);
1109 qpn = IPOIB_QPN(neigh->neighbour->ha);
1110 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1111 spin_unlock(&priv->lock);
1112 spin_unlock_irqrestore(&priv->tx_lock, flags);
1113 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1114 spin_lock_irqsave(&priv->tx_lock, flags);
1115 spin_lock(&priv->lock);
1120 list_del(&neigh->list);
1122 ipoib_put_ah(neigh->ah);
1123 ipoib_neigh_free(dev, neigh);
1129 spin_unlock(&priv->lock);
1130 spin_unlock_irqrestore(&priv->tx_lock, flags);
1133 static void ipoib_cm_tx_reap(struct work_struct *work)
1135 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1137 struct ipoib_cm_tx *p;
1139 spin_lock_irq(&priv->tx_lock);
1140 spin_lock(&priv->lock);
1141 while (!list_empty(&priv->cm.reap_list)) {
1142 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1144 spin_unlock(&priv->lock);
1145 spin_unlock_irq(&priv->tx_lock);
1146 ipoib_cm_tx_destroy(p);
1147 spin_lock_irq(&priv->tx_lock);
1148 spin_lock(&priv->lock);
1150 spin_unlock(&priv->lock);
1151 spin_unlock_irq(&priv->tx_lock);
1154 static void ipoib_cm_skb_reap(struct work_struct *work)
1156 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1158 struct net_device *dev = priv->dev;
1159 struct sk_buff *skb;
1161 unsigned mtu = priv->mcast_mtu;
1163 spin_lock_irq(&priv->tx_lock);
1164 spin_lock(&priv->lock);
1165 while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1166 spin_unlock(&priv->lock);
1167 spin_unlock_irq(&priv->tx_lock);
1168 if (skb->protocol == htons(ETH_P_IP))
1169 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1170 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1171 else if (skb->protocol == htons(ETH_P_IPV6))
1172 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, dev);
1174 dev_kfree_skb_any(skb);
1175 spin_lock_irq(&priv->tx_lock);
1176 spin_lock(&priv->lock);
1178 spin_unlock(&priv->lock);
1179 spin_unlock_irq(&priv->tx_lock);
1182 void ipoib_cm_skb_too_long(struct net_device* dev, struct sk_buff *skb,
1185 struct ipoib_dev_priv *priv = netdev_priv(dev);
1186 int e = skb_queue_empty(&priv->cm.skb_queue);
1189 skb->dst->ops->update_pmtu(skb->dst, mtu);
1191 skb_queue_tail(&priv->cm.skb_queue, skb);
1193 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1196 static void ipoib_cm_rx_reap(struct work_struct *work)
1198 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1200 struct ipoib_cm_rx *p, *n;
1203 spin_lock_irq(&priv->lock);
1204 list_splice_init(&priv->cm.rx_reap_list, &list);
1205 spin_unlock_irq(&priv->lock);
1207 list_for_each_entry_safe(p, n, &list, list) {
1208 ib_destroy_cm_id(p->id);
1209 ib_destroy_qp(p->qp);
1214 static void ipoib_cm_stale_task(struct work_struct *work)
1216 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1217 cm.stale_task.work);
1218 struct ipoib_cm_rx *p;
1221 spin_lock_irq(&priv->lock);
1222 while (!list_empty(&priv->cm.passive_ids)) {
1223 /* List is sorted by LRU, start from tail,
1224 * stop when we see a recently used entry */
1225 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1226 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1228 list_move(&p->list, &priv->cm.rx_error_list);
1229 p->state = IPOIB_CM_RX_ERROR;
1230 spin_unlock_irq(&priv->lock);
1231 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1233 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1234 spin_lock_irq(&priv->lock);
1237 if (!list_empty(&priv->cm.passive_ids))
1238 queue_delayed_work(ipoib_workqueue,
1239 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1240 spin_unlock_irq(&priv->lock);
1244 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1247 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1249 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1250 return sprintf(buf, "connected\n");
1252 return sprintf(buf, "datagram\n");
1255 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1256 const char *buf, size_t count)
1258 struct net_device *dev = to_net_dev(d);
1259 struct ipoib_dev_priv *priv = netdev_priv(dev);
1261 /* flush paths if we switch modes so that connections are restarted */
1262 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1263 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1264 ipoib_warn(priv, "enabling connected mode "
1265 "will cause multicast packet drops\n");
1266 ipoib_flush_paths(dev);
1270 if (!strcmp(buf, "datagram\n")) {
1271 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1272 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1273 ipoib_flush_paths(dev);
1280 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1282 int ipoib_cm_add_mode_attr(struct net_device *dev)
1284 return device_create_file(&dev->dev, &dev_attr_mode);
1287 int ipoib_cm_dev_init(struct net_device *dev)
1289 struct ipoib_dev_priv *priv = netdev_priv(dev);
1290 struct ib_srq_init_attr srq_init_attr = {
1292 .max_wr = ipoib_recvq_size,
1293 .max_sge = IPOIB_CM_RX_SG
1298 INIT_LIST_HEAD(&priv->cm.passive_ids);
1299 INIT_LIST_HEAD(&priv->cm.reap_list);
1300 INIT_LIST_HEAD(&priv->cm.start_list);
1301 INIT_LIST_HEAD(&priv->cm.rx_error_list);
1302 INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1303 INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1304 INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1305 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1306 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1307 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1308 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1309 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1311 skb_queue_head_init(&priv->cm.skb_queue);
1313 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1314 if (IS_ERR(priv->cm.srq)) {
1315 ret = PTR_ERR(priv->cm.srq);
1316 priv->cm.srq = NULL;
1320 priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1322 if (!priv->cm.srq_ring) {
1323 printk(KERN_WARNING "%s: failed to allocate CM ring (%d entries)\n",
1324 priv->ca->name, ipoib_recvq_size);
1325 ipoib_cm_dev_cleanup(dev);
1329 for (i = 0; i < IPOIB_CM_RX_SG; ++i)
1330 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1332 priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1333 for (i = 1; i < IPOIB_CM_RX_SG; ++i)
1334 priv->cm.rx_sge[i].length = PAGE_SIZE;
1335 priv->cm.rx_wr.next = NULL;
1336 priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1337 priv->cm.rx_wr.num_sge = IPOIB_CM_RX_SG;
1339 for (i = 0; i < ipoib_recvq_size; ++i) {
1340 if (!ipoib_cm_alloc_rx_skb(dev, i, IPOIB_CM_RX_SG - 1,
1341 priv->cm.srq_ring[i].mapping)) {
1342 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
1343 ipoib_cm_dev_cleanup(dev);
1346 if (ipoib_cm_post_receive(dev, i)) {
1347 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
1348 ipoib_cm_dev_cleanup(dev);
1353 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1357 void ipoib_cm_dev_cleanup(struct net_device *dev)
1359 struct ipoib_dev_priv *priv = netdev_priv(dev);
1365 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1367 ret = ib_destroy_srq(priv->cm.srq);
1369 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1371 priv->cm.srq = NULL;
1372 if (!priv->cm.srq_ring)
1374 for (i = 0; i < ipoib_recvq_size; ++i)
1375 if (priv->cm.srq_ring[i].skb) {
1376 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
1377 priv->cm.srq_ring[i].mapping);
1378 dev_kfree_skb_any(priv->cm.srq_ring[i].skb);
1379 priv->cm.srq_ring[i].skb = NULL;
1381 kfree(priv->cm.srq_ring);
1382 priv->cm.srq_ring = NULL;