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>
41 #include <linux/sched/signal.h>
42 #include <linux/sched/mm.h>
46 int ipoib_max_conn_qp = 128;
48 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
49 MODULE_PARM_DESC(max_nonsrq_conn_qp,
50 "Max number of connected-mode QPs per interface "
51 "(applied only if shared receive queue is not available)");
53 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
54 static int data_debug_level;
56 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
57 MODULE_PARM_DESC(cm_data_debug_level,
58 "Enable data path debug tracing for connected mode if > 0");
61 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
63 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
64 #define IPOIB_CM_RX_TIMEOUT (2 * 256 * HZ)
65 #define IPOIB_CM_RX_DELAY (3 * 256 * HZ)
66 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
68 #define IPOIB_CM_RX_RESERVE (ALIGN(IPOIB_HARD_LEN, 16) - IPOIB_ENCAP_LEN)
70 static struct ib_qp_attr ipoib_cm_err_attr = {
71 .qp_state = IB_QPS_ERR
74 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
76 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
80 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
81 const struct ib_cm_event *event);
83 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
84 u64 mapping[IPOIB_CM_RX_SG])
88 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
90 for (i = 0; i < frags; ++i)
91 ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
94 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
96 struct ipoib_dev_priv *priv = ipoib_priv(dev);
99 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
101 for (i = 0; i < priv->cm.num_frags; ++i)
102 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
104 ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, NULL);
106 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
107 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
108 priv->cm.srq_ring[id].mapping);
109 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
110 priv->cm.srq_ring[id].skb = NULL;
116 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
117 struct ipoib_cm_rx *rx,
118 struct ib_recv_wr *wr,
119 struct ib_sge *sge, int id)
121 struct ipoib_dev_priv *priv = ipoib_priv(dev);
124 wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
126 for (i = 0; i < IPOIB_CM_RX_SG; ++i)
127 sge[i].addr = rx->rx_ring[id].mapping[i];
129 ret = ib_post_recv(rx->qp, wr, NULL);
131 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
132 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
133 rx->rx_ring[id].mapping);
134 dev_kfree_skb_any(rx->rx_ring[id].skb);
135 rx->rx_ring[id].skb = NULL;
141 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
142 struct ipoib_cm_rx_buf *rx_ring,
144 u64 mapping[IPOIB_CM_RX_SG],
147 struct ipoib_dev_priv *priv = ipoib_priv(dev);
151 skb = dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE + IPOIB_PSEUDO_LEN, 16));
156 * IPoIB adds a IPOIB_ENCAP_LEN byte header, this will align the
157 * IP header to a multiple of 16.
159 skb_reserve(skb, IPOIB_CM_RX_RESERVE);
161 mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
163 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
164 dev_kfree_skb_any(skb);
168 for (i = 0; i < frags; i++) {
169 struct page *page = alloc_page(gfp);
173 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
175 mapping[i + 1] = ib_dma_map_page(priv->ca, page,
176 0, PAGE_SIZE, DMA_FROM_DEVICE);
177 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
181 rx_ring[id].skb = skb;
186 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
189 ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
191 dev_kfree_skb_any(skb);
195 static void ipoib_cm_free_rx_ring(struct net_device *dev,
196 struct ipoib_cm_rx_buf *rx_ring)
198 struct ipoib_dev_priv *priv = ipoib_priv(dev);
201 for (i = 0; i < ipoib_recvq_size; ++i)
202 if (rx_ring[i].skb) {
203 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
205 dev_kfree_skb_any(rx_ring[i].skb);
211 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
213 struct ipoib_cm_rx *p;
215 /* We only reserved 1 extra slot in CQ for drain WRs, so
216 * make sure we have at most 1 outstanding WR. */
217 if (list_empty(&priv->cm.rx_flush_list) ||
218 !list_empty(&priv->cm.rx_drain_list))
222 * QPs on flush list are error state. This way, a "flush
223 * error" WC will be immediately generated for each WR we post.
225 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
226 ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
227 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, NULL))
228 ipoib_warn(priv, "failed to post drain wr\n");
230 list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
233 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
235 struct ipoib_cm_rx *p = ctx;
236 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
239 if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
242 spin_lock_irqsave(&priv->lock, flags);
243 list_move(&p->list, &priv->cm.rx_flush_list);
244 p->state = IPOIB_CM_RX_FLUSH;
245 ipoib_cm_start_rx_drain(priv);
246 spin_unlock_irqrestore(&priv->lock, flags);
249 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
250 struct ipoib_cm_rx *p)
252 struct ipoib_dev_priv *priv = ipoib_priv(dev);
253 struct ib_qp_init_attr attr = {
254 .event_handler = ipoib_cm_rx_event_handler,
255 .send_cq = priv->recv_cq, /* For drain WR */
256 .recv_cq = priv->recv_cq,
258 .cap.max_send_wr = 1, /* For drain WR */
259 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
260 .sq_sig_type = IB_SIGNAL_ALL_WR,
261 .qp_type = IB_QPT_RC,
265 if (!ipoib_cm_has_srq(dev)) {
266 attr.cap.max_recv_wr = ipoib_recvq_size;
267 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
270 return ib_create_qp(priv->pd, &attr);
273 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
274 struct ib_cm_id *cm_id, struct ib_qp *qp,
277 struct ipoib_dev_priv *priv = ipoib_priv(dev);
278 struct ib_qp_attr qp_attr;
279 int qp_attr_mask, ret;
281 qp_attr.qp_state = IB_QPS_INIT;
282 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
284 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
287 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
289 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
292 qp_attr.qp_state = IB_QPS_RTR;
293 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
295 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
298 qp_attr.rq_psn = psn;
299 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
301 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
306 * Current Mellanox HCA firmware won't generate completions
307 * with error for drain WRs unless the QP has been moved to
308 * RTS first. This work-around leaves a window where a QP has
309 * moved to error asynchronously, but this will eventually get
310 * fixed in firmware, so let's not error out if modify QP
313 qp_attr.qp_state = IB_QPS_RTS;
314 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
316 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
319 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
321 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
328 static void ipoib_cm_init_rx_wr(struct net_device *dev,
329 struct ib_recv_wr *wr,
332 struct ipoib_dev_priv *priv = ipoib_priv(dev);
335 for (i = 0; i < priv->cm.num_frags; ++i)
336 sge[i].lkey = priv->pd->local_dma_lkey;
338 sge[0].length = IPOIB_CM_HEAD_SIZE;
339 for (i = 1; i < priv->cm.num_frags; ++i)
340 sge[i].length = PAGE_SIZE;
344 wr->num_sge = priv->cm.num_frags;
347 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
348 struct ipoib_cm_rx *rx)
350 struct ipoib_dev_priv *priv = ipoib_priv(dev);
352 struct ib_recv_wr wr;
353 struct ib_sge sge[IPOIB_CM_RX_SG];
358 rx->rx_ring = vzalloc(array_size(ipoib_recvq_size,
359 sizeof(*rx->rx_ring)));
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,
387 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
391 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
393 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
394 "failed for buf %d\n", i);
400 rx->recv_count = ipoib_recvq_size;
407 spin_lock_irq(&priv->lock);
408 --priv->cm.nonsrq_conn_qp;
409 spin_unlock_irq(&priv->lock);
415 ipoib_cm_free_rx_ring(dev, rx->rx_ring);
420 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
422 const struct ib_cm_req_event_param *req,
425 struct ipoib_dev_priv *priv = ipoib_priv(dev);
426 struct ipoib_cm_data data = {};
427 struct ib_cm_rep_param rep = {};
429 data.qpn = cpu_to_be32(priv->qp->qp_num);
430 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
432 rep.private_data = &data;
433 rep.private_data_len = sizeof(data);
434 rep.flow_control = 0;
435 rep.rnr_retry_count = req->rnr_retry_count;
436 rep.srq = ipoib_cm_has_srq(dev);
437 rep.qp_num = qp->qp_num;
438 rep.starting_psn = psn;
439 return ib_send_cm_rep(cm_id, &rep);
442 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id,
443 const struct ib_cm_event *event)
445 struct net_device *dev = cm_id->context;
446 struct ipoib_dev_priv *priv = ipoib_priv(dev);
447 struct ipoib_cm_rx *p;
451 ipoib_dbg(priv, "REQ arrived\n");
452 p = kzalloc(sizeof(*p), GFP_KERNEL);
458 p->state = IPOIB_CM_RX_LIVE;
459 p->jiffies = jiffies;
460 INIT_LIST_HEAD(&p->list);
462 p->qp = ipoib_cm_create_rx_qp(dev, p);
464 ret = PTR_ERR(p->qp);
468 psn = prandom_u32() & 0xffffff;
469 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
473 if (!ipoib_cm_has_srq(dev)) {
474 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
479 spin_lock_irq(&priv->lock);
480 queue_delayed_work(priv->wq,
481 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
482 /* Add this entry to passive ids list head, but do not re-add it
483 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
484 p->jiffies = jiffies;
485 if (p->state == IPOIB_CM_RX_LIVE)
486 list_move(&p->list, &priv->cm.passive_ids);
487 spin_unlock_irq(&priv->lock);
489 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
491 ipoib_warn(priv, "failed to send REP: %d\n", ret);
492 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
493 ipoib_warn(priv, "unable to move qp to error state\n");
498 ib_destroy_qp(p->qp);
504 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
505 const struct ib_cm_event *event)
507 struct ipoib_cm_rx *p;
508 struct ipoib_dev_priv *priv;
510 switch (event->event) {
511 case IB_CM_REQ_RECEIVED:
512 return ipoib_cm_req_handler(cm_id, event);
513 case IB_CM_DREQ_RECEIVED:
514 ib_send_cm_drep(cm_id, NULL, 0);
516 case IB_CM_REJ_RECEIVED:
518 priv = ipoib_priv(p->dev);
519 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
520 ipoib_warn(priv, "unable to move qp to error state\n");
526 /* Adjust length of skb with fragments to match received data */
527 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
528 unsigned int length, struct sk_buff *toskb)
533 /* put header into skb */
534 size = min(length, hdr_space);
539 num_frags = skb_shinfo(skb)->nr_frags;
540 for (i = 0; i < num_frags; i++) {
541 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
544 /* don't need this page */
545 skb_fill_page_desc(toskb, i, skb_frag_page(frag),
547 --skb_shinfo(skb)->nr_frags;
549 size = min_t(unsigned int, length, PAGE_SIZE);
551 skb_frag_size_set(frag, size);
552 skb->data_len += size;
553 skb->truesize += size;
560 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
562 struct ipoib_dev_priv *priv = ipoib_priv(dev);
563 struct ipoib_cm_rx_buf *rx_ring;
564 unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
565 struct sk_buff *skb, *newskb;
566 struct ipoib_cm_rx *p;
568 u64 mapping[IPOIB_CM_RX_SG];
571 struct sk_buff *small_skb;
573 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
576 if (unlikely(wr_id >= ipoib_recvq_size)) {
577 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
578 spin_lock_irqsave(&priv->lock, flags);
579 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
580 ipoib_cm_start_rx_drain(priv);
581 queue_work(priv->wq, &priv->cm.rx_reap_task);
582 spin_unlock_irqrestore(&priv->lock, flags);
584 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
585 wr_id, ipoib_recvq_size);
589 p = wc->qp->qp_context;
591 has_srq = ipoib_cm_has_srq(dev);
592 rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
594 skb = rx_ring[wr_id].skb;
596 if (unlikely(wc->status != IB_WC_SUCCESS)) {
598 "cm recv error (status=%d, wrid=%d vend_err %#x)\n",
599 wc->status, wr_id, wc->vendor_err);
600 ++dev->stats.rx_dropped;
604 if (!--p->recv_count) {
605 spin_lock_irqsave(&priv->lock, flags);
606 list_move(&p->list, &priv->cm.rx_reap_list);
607 spin_unlock_irqrestore(&priv->lock, flags);
608 queue_work(priv->wq, &priv->cm.rx_reap_task);
614 if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
615 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
616 spin_lock_irqsave(&priv->lock, flags);
617 p->jiffies = jiffies;
618 /* Move this entry to list head, but do not re-add it
619 * if it has been moved out of list. */
620 if (p->state == IPOIB_CM_RX_LIVE)
621 list_move(&p->list, &priv->cm.passive_ids);
622 spin_unlock_irqrestore(&priv->lock, flags);
626 if (wc->byte_len < IPOIB_CM_COPYBREAK) {
627 int dlen = wc->byte_len;
629 small_skb = dev_alloc_skb(dlen + IPOIB_CM_RX_RESERVE);
631 skb_reserve(small_skb, IPOIB_CM_RX_RESERVE);
632 ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
633 dlen, DMA_FROM_DEVICE);
634 skb_copy_from_linear_data(skb, small_skb->data, dlen);
635 ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
636 dlen, DMA_FROM_DEVICE);
637 skb_put(small_skb, dlen);
643 frags = PAGE_ALIGN(wc->byte_len -
644 min_t(u32, wc->byte_len, IPOIB_CM_HEAD_SIZE)) /
647 newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
648 mapping, GFP_ATOMIC);
649 if (unlikely(!newskb)) {
651 * If we can't allocate a new RX buffer, dump
652 * this packet and reuse the old buffer.
654 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
655 ++dev->stats.rx_dropped;
659 ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
660 memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof(*mapping));
662 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
663 wc->byte_len, wc->slid);
665 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
668 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
669 skb_add_pseudo_hdr(skb);
671 ++dev->stats.rx_packets;
672 dev->stats.rx_bytes += skb->len;
675 /* XXX get correct PACKET_ type here */
676 skb->pkt_type = PACKET_HOST;
677 netif_receive_skb(skb);
681 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
682 ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
683 "for buf %d\n", wr_id);
685 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
690 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
691 "for buf %d\n", wr_id);
696 static inline int post_send(struct ipoib_dev_priv *priv,
697 struct ipoib_cm_tx *tx,
699 struct ipoib_tx_buf *tx_req)
701 ipoib_build_sge(priv, tx_req);
703 priv->tx_wr.wr.wr_id = wr_id | IPOIB_OP_CM;
705 return ib_post_send(tx->qp, &priv->tx_wr.wr, NULL);
708 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
710 struct ipoib_dev_priv *priv = ipoib_priv(dev);
711 struct ipoib_tx_buf *tx_req;
713 unsigned int usable_sge = tx->max_send_sge - !!skb_headlen(skb);
715 if (unlikely(skb->len > tx->mtu)) {
716 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
718 ++dev->stats.tx_dropped;
719 ++dev->stats.tx_errors;
720 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
723 if (skb_shinfo(skb)->nr_frags > usable_sge) {
724 if (skb_linearize(skb) < 0) {
725 ipoib_warn(priv, "skb could not be linearized\n");
726 ++dev->stats.tx_dropped;
727 ++dev->stats.tx_errors;
728 dev_kfree_skb_any(skb);
731 /* Does skb_linearize return ok without reducing nr_frags? */
732 if (skb_shinfo(skb)->nr_frags > usable_sge) {
733 ipoib_warn(priv, "too many frags after skb linearize\n");
734 ++dev->stats.tx_dropped;
735 ++dev->stats.tx_errors;
736 dev_kfree_skb_any(skb);
740 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
741 tx->tx_head, skb->len, tx->qp->qp_num);
744 * We put the skb into the tx_ring _before_ we call post_send()
745 * because it's entirely possible that the completion handler will
746 * run before we execute anything after the post_send(). That
747 * means we have to make sure everything is properly recorded and
748 * our state is consistent before we call post_send().
750 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
753 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
754 ++dev->stats.tx_errors;
755 dev_kfree_skb_any(skb);
759 if ((priv->tx_head - priv->tx_tail) == ipoib_sendq_size - 1) {
760 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
762 netif_stop_queue(dev);
768 if (netif_queue_stopped(dev)) {
769 rc = ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
770 IB_CQ_REPORT_MISSED_EVENTS);
771 if (unlikely(rc < 0))
772 ipoib_warn(priv, "IPoIB/CM:request notify on send CQ failed\n");
774 napi_schedule(&priv->send_napi);
777 rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
779 ipoib_warn(priv, "IPoIB/CM:post_send failed, error %d\n", rc);
780 ++dev->stats.tx_errors;
781 ipoib_dma_unmap_tx(priv, tx_req);
782 dev_kfree_skb_any(skb);
784 if (netif_queue_stopped(dev))
785 netif_wake_queue(dev);
787 netif_trans_update(dev);
793 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
795 struct ipoib_dev_priv *priv = ipoib_priv(dev);
796 struct ipoib_cm_tx *tx = wc->qp->qp_context;
797 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
798 struct ipoib_tx_buf *tx_req;
801 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
804 if (unlikely(wr_id >= ipoib_sendq_size)) {
805 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
806 wr_id, ipoib_sendq_size);
810 tx_req = &tx->tx_ring[wr_id];
812 ipoib_dma_unmap_tx(priv, tx_req);
814 /* FIXME: is this right? Shouldn't we only increment on success? */
815 ++dev->stats.tx_packets;
816 dev->stats.tx_bytes += tx_req->skb->len;
818 dev_kfree_skb_any(tx_req->skb);
825 if (unlikely(netif_queue_stopped(dev) &&
826 (priv->tx_head - priv->tx_tail) <= ipoib_sendq_size >> 1 &&
827 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
828 netif_wake_queue(dev);
830 if (wc->status != IB_WC_SUCCESS &&
831 wc->status != IB_WC_WR_FLUSH_ERR) {
832 struct ipoib_neigh *neigh;
834 /* IB_WC[_RNR]_RETRY_EXC_ERR error is part of the life cycle,
835 * so don't make waves.
837 if (wc->status == IB_WC_RNR_RETRY_EXC_ERR ||
838 wc->status == IB_WC_RETRY_EXC_ERR)
840 "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
841 __func__, wc->status, wr_id, wc->vendor_err);
844 "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
845 __func__, wc->status, wr_id, wc->vendor_err);
847 spin_lock_irqsave(&priv->lock, flags);
852 ipoib_neigh_free(neigh);
857 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
858 list_move(&tx->list, &priv->cm.reap_list);
859 queue_work(priv->wq, &priv->cm.reap_task);
862 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
864 spin_unlock_irqrestore(&priv->lock, flags);
867 netif_tx_unlock(dev);
870 int ipoib_cm_dev_open(struct net_device *dev)
872 struct ipoib_dev_priv *priv = ipoib_priv(dev);
875 if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
878 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
879 if (IS_ERR(priv->cm.id)) {
880 pr_warn("%s: failed to create CM ID\n", priv->ca->name);
881 ret = PTR_ERR(priv->cm.id);
885 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
888 pr_warn("%s: failed to listen on ID 0x%llx\n", priv->ca->name,
889 IPOIB_CM_IETF_ID | priv->qp->qp_num);
896 ib_destroy_cm_id(priv->cm.id);
902 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
904 struct ipoib_dev_priv *priv = ipoib_priv(dev);
905 struct ipoib_cm_rx *rx, *n;
908 spin_lock_irq(&priv->lock);
909 list_splice_init(&priv->cm.rx_reap_list, &list);
910 spin_unlock_irq(&priv->lock);
912 list_for_each_entry_safe(rx, n, &list, list) {
913 ib_destroy_cm_id(rx->id);
914 ib_destroy_qp(rx->qp);
915 if (!ipoib_cm_has_srq(dev)) {
916 ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
917 spin_lock_irq(&priv->lock);
918 --priv->cm.nonsrq_conn_qp;
919 spin_unlock_irq(&priv->lock);
925 void ipoib_cm_dev_stop(struct net_device *dev)
927 struct ipoib_dev_priv *priv = ipoib_priv(dev);
928 struct ipoib_cm_rx *p;
932 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
935 ib_destroy_cm_id(priv->cm.id);
938 spin_lock_irq(&priv->lock);
939 while (!list_empty(&priv->cm.passive_ids)) {
940 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
941 list_move(&p->list, &priv->cm.rx_error_list);
942 p->state = IPOIB_CM_RX_ERROR;
943 spin_unlock_irq(&priv->lock);
944 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
946 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
947 spin_lock_irq(&priv->lock);
950 /* Wait for all RX to be drained */
953 while (!list_empty(&priv->cm.rx_error_list) ||
954 !list_empty(&priv->cm.rx_flush_list) ||
955 !list_empty(&priv->cm.rx_drain_list)) {
956 if (time_after(jiffies, begin + 5 * HZ)) {
957 ipoib_warn(priv, "RX drain timing out\n");
960 * assume the HW is wedged and just free up everything.
962 list_splice_init(&priv->cm.rx_flush_list,
963 &priv->cm.rx_reap_list);
964 list_splice_init(&priv->cm.rx_error_list,
965 &priv->cm.rx_reap_list);
966 list_splice_init(&priv->cm.rx_drain_list,
967 &priv->cm.rx_reap_list);
970 spin_unlock_irq(&priv->lock);
971 usleep_range(1000, 2000);
973 spin_lock_irq(&priv->lock);
976 spin_unlock_irq(&priv->lock);
978 ipoib_cm_free_rx_reap_list(dev);
980 cancel_delayed_work(&priv->cm.stale_task);
983 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id,
984 const struct ib_cm_event *event)
986 struct ipoib_cm_tx *p = cm_id->context;
987 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
988 struct ipoib_cm_data *data = event->private_data;
989 struct sk_buff_head skqueue;
990 struct ib_qp_attr qp_attr;
991 int qp_attr_mask, ret;
994 p->mtu = be32_to_cpu(data->mtu);
996 if (p->mtu <= IPOIB_ENCAP_LEN) {
997 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
998 p->mtu, IPOIB_ENCAP_LEN);
1002 qp_attr.qp_state = IB_QPS_RTR;
1003 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1005 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
1009 qp_attr.rq_psn = 0 /* FIXME */;
1010 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1012 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
1016 qp_attr.qp_state = IB_QPS_RTS;
1017 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1019 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
1022 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1024 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1028 skb_queue_head_init(&skqueue);
1030 netif_tx_lock_bh(p->dev);
1031 spin_lock_irq(&priv->lock);
1032 set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1034 while ((skb = __skb_dequeue(&p->neigh->queue)))
1035 __skb_queue_tail(&skqueue, skb);
1036 spin_unlock_irq(&priv->lock);
1037 netif_tx_unlock_bh(p->dev);
1039 while ((skb = __skb_dequeue(&skqueue))) {
1041 ret = dev_queue_xmit(skb);
1043 ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1047 ret = ib_send_cm_rtu(cm_id, NULL, 0);
1049 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1055 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1057 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1058 struct ib_qp_init_attr attr = {
1059 .send_cq = priv->send_cq,
1060 .recv_cq = priv->recv_cq,
1061 .srq = priv->cm.srq,
1062 .cap.max_send_wr = ipoib_sendq_size,
1063 .cap.max_send_sge = 1,
1064 .sq_sig_type = IB_SIGNAL_ALL_WR,
1065 .qp_type = IB_QPT_RC,
1069 struct ib_qp *tx_qp;
1071 if (dev->features & NETIF_F_SG)
1072 attr.cap.max_send_sge = min_t(u32, priv->ca->attrs.max_send_sge,
1075 tx_qp = ib_create_qp(priv->pd, &attr);
1076 tx->max_send_sge = attr.cap.max_send_sge;
1080 static int ipoib_cm_send_req(struct net_device *dev,
1081 struct ib_cm_id *id, struct ib_qp *qp,
1083 struct sa_path_rec *pathrec)
1085 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1086 struct ipoib_cm_data data = {};
1087 struct ib_cm_req_param req = {};
1089 data.qpn = cpu_to_be32(priv->qp->qp_num);
1090 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1092 req.primary_path = pathrec;
1093 req.alternate_path = NULL;
1094 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1095 req.qp_num = qp->qp_num;
1096 req.qp_type = qp->qp_type;
1097 req.private_data = &data;
1098 req.private_data_len = sizeof(data);
1099 req.flow_control = 0;
1101 req.starting_psn = 0; /* FIXME */
1104 * Pick some arbitrary defaults here; we could make these
1105 * module parameters if anyone cared about setting them.
1107 req.responder_resources = 4;
1108 req.remote_cm_response_timeout = 20;
1109 req.local_cm_response_timeout = 20;
1110 req.retry_count = 0; /* RFC draft warns against retries */
1111 req.rnr_retry_count = 0; /* RFC draft warns against retries */
1112 req.max_cm_retries = 15;
1113 req.srq = ipoib_cm_has_srq(dev);
1114 return ib_send_cm_req(id, &req);
1117 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1118 struct ib_cm_id *cm_id, struct ib_qp *qp)
1120 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1121 struct ib_qp_attr qp_attr;
1122 int qp_attr_mask, ret;
1123 ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1125 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1129 qp_attr.qp_state = IB_QPS_INIT;
1130 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1131 qp_attr.port_num = priv->port;
1132 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1134 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1136 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1142 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1143 struct sa_path_rec *pathrec)
1145 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1146 unsigned int noio_flag;
1149 noio_flag = memalloc_noio_save();
1150 p->tx_ring = vzalloc(array_size(ipoib_sendq_size, sizeof(*p->tx_ring)));
1152 memalloc_noio_restore(noio_flag);
1156 memset(p->tx_ring, 0, ipoib_sendq_size * sizeof(*p->tx_ring));
1158 p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1159 memalloc_noio_restore(noio_flag);
1160 if (IS_ERR(p->qp)) {
1161 ret = PTR_ERR(p->qp);
1162 ipoib_warn(priv, "failed to create tx qp: %d\n", ret);
1166 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1167 if (IS_ERR(p->id)) {
1168 ret = PTR_ERR(p->id);
1169 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1173 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp);
1175 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1176 goto err_modify_send;
1179 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1181 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1182 goto err_modify_send;
1185 ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1186 p->qp->qp_num, pathrec->dgid.raw, qpn);
1191 ib_destroy_cm_id(p->id);
1194 ib_destroy_qp(p->qp);
1202 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1204 struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1205 struct ipoib_tx_buf *tx_req;
1206 unsigned long begin;
1208 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1209 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1212 ib_destroy_cm_id(p->id);
1215 /* Wait for all sends to complete */
1217 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1218 if (time_after(jiffies, begin + 5 * HZ)) {
1219 ipoib_warn(priv, "timing out; %d sends not completed\n",
1220 p->tx_head - p->tx_tail);
1224 usleep_range(1000, 2000);
1230 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1231 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1232 ipoib_dma_unmap_tx(priv, tx_req);
1233 dev_kfree_skb_any(tx_req->skb);
1234 netif_tx_lock_bh(p->dev);
1237 if (unlikely(priv->tx_head - priv->tx_tail == ipoib_sendq_size >> 1) &&
1238 netif_queue_stopped(p->dev) &&
1239 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1240 netif_wake_queue(p->dev);
1241 netif_tx_unlock_bh(p->dev);
1245 ib_destroy_qp(p->qp);
1251 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1252 const struct ib_cm_event *event)
1254 struct ipoib_cm_tx *tx = cm_id->context;
1255 struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1256 struct net_device *dev = priv->dev;
1257 struct ipoib_neigh *neigh;
1258 unsigned long flags;
1261 switch (event->event) {
1262 case IB_CM_DREQ_RECEIVED:
1263 ipoib_dbg(priv, "DREQ received.\n");
1264 ib_send_cm_drep(cm_id, NULL, 0);
1266 case IB_CM_REP_RECEIVED:
1267 ipoib_dbg(priv, "REP received.\n");
1268 ret = ipoib_cm_rep_handler(cm_id, event);
1270 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1273 case IB_CM_REQ_ERROR:
1274 case IB_CM_REJ_RECEIVED:
1275 case IB_CM_TIMEWAIT_EXIT:
1276 ipoib_dbg(priv, "CM error %d.\n", event->event);
1277 netif_tx_lock_bh(dev);
1278 spin_lock_irqsave(&priv->lock, flags);
1283 ipoib_neigh_free(neigh);
1288 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1289 list_move(&tx->list, &priv->cm.reap_list);
1290 queue_work(priv->wq, &priv->cm.reap_task);
1293 spin_unlock_irqrestore(&priv->lock, flags);
1294 netif_tx_unlock_bh(dev);
1303 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1304 struct ipoib_neigh *neigh)
1306 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1307 struct ipoib_cm_tx *tx;
1309 tx = kzalloc(sizeof(*tx), GFP_ATOMIC);
1317 list_add(&tx->list, &priv->cm.start_list);
1318 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1319 queue_work(priv->wq, &priv->cm.start_task);
1323 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1325 struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1326 unsigned long flags;
1327 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1328 spin_lock_irqsave(&priv->lock, flags);
1329 list_move(&tx->list, &priv->cm.reap_list);
1330 queue_work(priv->wq, &priv->cm.reap_task);
1331 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1332 tx->neigh->daddr + 4);
1334 spin_unlock_irqrestore(&priv->lock, flags);
1338 #define QPN_AND_OPTIONS_OFFSET 4
1340 static void ipoib_cm_tx_start(struct work_struct *work)
1342 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1344 struct net_device *dev = priv->dev;
1345 struct ipoib_neigh *neigh;
1346 struct ipoib_cm_tx *p;
1347 unsigned long flags;
1348 struct ipoib_path *path;
1351 struct sa_path_rec pathrec;
1354 netif_tx_lock_bh(dev);
1355 spin_lock_irqsave(&priv->lock, flags);
1357 while (!list_empty(&priv->cm.start_list)) {
1358 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1359 list_del_init(&p->list);
1362 qpn = IPOIB_QPN(neigh->daddr);
1364 * As long as the search is with these 2 locks,
1365 * path existence indicates its validity.
1367 path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1369 pr_info("%s ignore not valid path %pI6\n",
1371 neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1374 memcpy(&pathrec, &p->path->pathrec, sizeof(pathrec));
1376 spin_unlock_irqrestore(&priv->lock, flags);
1377 netif_tx_unlock_bh(dev);
1379 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1381 netif_tx_lock_bh(dev);
1382 spin_lock_irqsave(&priv->lock, flags);
1389 ipoib_neigh_free(neigh);
1396 spin_unlock_irqrestore(&priv->lock, flags);
1397 netif_tx_unlock_bh(dev);
1400 static void ipoib_cm_tx_reap(struct work_struct *work)
1402 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1404 struct net_device *dev = priv->dev;
1405 struct ipoib_cm_tx *p;
1406 unsigned long flags;
1408 netif_tx_lock_bh(dev);
1409 spin_lock_irqsave(&priv->lock, flags);
1411 while (!list_empty(&priv->cm.reap_list)) {
1412 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1413 list_del_init(&p->list);
1414 spin_unlock_irqrestore(&priv->lock, flags);
1415 netif_tx_unlock_bh(dev);
1416 ipoib_cm_tx_destroy(p);
1417 netif_tx_lock_bh(dev);
1418 spin_lock_irqsave(&priv->lock, flags);
1421 spin_unlock_irqrestore(&priv->lock, flags);
1422 netif_tx_unlock_bh(dev);
1425 static void ipoib_cm_skb_reap(struct work_struct *work)
1427 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1429 struct net_device *dev = priv->dev;
1430 struct sk_buff *skb;
1431 unsigned long flags;
1432 unsigned int mtu = priv->mcast_mtu;
1434 netif_tx_lock_bh(dev);
1435 spin_lock_irqsave(&priv->lock, flags);
1437 while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1438 spin_unlock_irqrestore(&priv->lock, flags);
1439 netif_tx_unlock_bh(dev);
1441 if (skb->protocol == htons(ETH_P_IP))
1442 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1443 #if IS_ENABLED(CONFIG_IPV6)
1444 else if (skb->protocol == htons(ETH_P_IPV6))
1445 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1447 dev_kfree_skb_any(skb);
1449 netif_tx_lock_bh(dev);
1450 spin_lock_irqsave(&priv->lock, flags);
1453 spin_unlock_irqrestore(&priv->lock, flags);
1454 netif_tx_unlock_bh(dev);
1457 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1460 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1461 int e = skb_queue_empty(&priv->cm.skb_queue);
1463 skb_dst_update_pmtu(skb, mtu);
1465 skb_queue_tail(&priv->cm.skb_queue, skb);
1467 queue_work(priv->wq, &priv->cm.skb_task);
1470 static void ipoib_cm_rx_reap(struct work_struct *work)
1472 ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1473 cm.rx_reap_task)->dev);
1476 static void ipoib_cm_stale_task(struct work_struct *work)
1478 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1479 cm.stale_task.work);
1480 struct ipoib_cm_rx *p;
1483 spin_lock_irq(&priv->lock);
1484 while (!list_empty(&priv->cm.passive_ids)) {
1485 /* List is sorted by LRU, start from tail,
1486 * stop when we see a recently used entry */
1487 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1488 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1490 list_move(&p->list, &priv->cm.rx_error_list);
1491 p->state = IPOIB_CM_RX_ERROR;
1492 spin_unlock_irq(&priv->lock);
1493 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1495 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1496 spin_lock_irq(&priv->lock);
1499 if (!list_empty(&priv->cm.passive_ids))
1500 queue_delayed_work(priv->wq,
1501 &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1502 spin_unlock_irq(&priv->lock);
1505 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1508 struct net_device *dev = to_net_dev(d);
1509 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1511 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1512 return sprintf(buf, "connected\n");
1514 return sprintf(buf, "datagram\n");
1517 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1518 const char *buf, size_t count)
1520 struct net_device *dev = to_net_dev(d);
1523 if (!rtnl_trylock()) {
1524 return restart_syscall();
1527 if (dev->reg_state != NETREG_REGISTERED) {
1532 ret = ipoib_set_mode(dev, buf);
1534 /* The assumption is that the function ipoib_set_mode returned
1535 * with the rtnl held by it, if not the value -EBUSY returned,
1536 * then no need to rtnl_unlock
1541 return (!ret || ret == -EBUSY) ? count : ret;
1544 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1546 int ipoib_cm_add_mode_attr(struct net_device *dev)
1548 return device_create_file(&dev->dev, &dev_attr_mode);
1551 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1553 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1554 struct ib_srq_init_attr srq_init_attr = {
1555 .srq_type = IB_SRQT_BASIC,
1557 .max_wr = ipoib_recvq_size,
1562 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1563 if (IS_ERR(priv->cm.srq)) {
1564 if (PTR_ERR(priv->cm.srq) != -EOPNOTSUPP)
1565 pr_warn("%s: failed to allocate SRQ, error %ld\n",
1566 priv->ca->name, PTR_ERR(priv->cm.srq));
1567 priv->cm.srq = NULL;
1571 priv->cm.srq_ring = vzalloc(array_size(ipoib_recvq_size,
1572 sizeof(*priv->cm.srq_ring)));
1573 if (!priv->cm.srq_ring) {
1574 ib_destroy_srq(priv->cm.srq);
1575 priv->cm.srq = NULL;
1581 int ipoib_cm_dev_init(struct net_device *dev)
1583 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1586 INIT_LIST_HEAD(&priv->cm.passive_ids);
1587 INIT_LIST_HEAD(&priv->cm.reap_list);
1588 INIT_LIST_HEAD(&priv->cm.start_list);
1589 INIT_LIST_HEAD(&priv->cm.rx_error_list);
1590 INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1591 INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1592 INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1593 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1594 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1595 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1596 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1597 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1599 skb_queue_head_init(&priv->cm.skb_queue);
1601 ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1603 max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1604 ipoib_cm_create_srq(dev, max_srq_sge);
1605 if (ipoib_cm_has_srq(dev)) {
1606 priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1607 priv->cm.num_frags = max_srq_sge;
1608 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1609 priv->cm.max_cm_mtu, priv->cm.num_frags);
1611 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1612 priv->cm.num_frags = IPOIB_CM_RX_SG;
1615 ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1617 if (ipoib_cm_has_srq(dev)) {
1618 for (i = 0; i < ipoib_recvq_size; ++i) {
1619 if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1620 priv->cm.num_frags - 1,
1621 priv->cm.srq_ring[i].mapping,
1623 ipoib_warn(priv, "failed to allocate "
1624 "receive buffer %d\n", i);
1625 ipoib_cm_dev_cleanup(dev);
1629 if (ipoib_cm_post_receive_srq(dev, i)) {
1630 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1631 "failed for buf %d\n", i);
1632 ipoib_cm_dev_cleanup(dev);
1638 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1642 void ipoib_cm_dev_cleanup(struct net_device *dev)
1644 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1650 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1652 ret = ib_destroy_srq(priv->cm.srq);
1654 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1656 priv->cm.srq = NULL;
1657 if (!priv->cm.srq_ring)
1660 ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1661 priv->cm.srq_ring = NULL;