2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/delay.h>
37 #include <linux/moduleparam.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/slab.h>
42 #include <linux/tcp.h>
46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
47 static int data_debug_level;
49 module_param(data_debug_level, int, 0644);
50 MODULE_PARM_DESC(data_debug_level,
51 "Enable data path debug tracing if > 0");
54 static DEFINE_MUTEX(pkey_mutex);
56 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
57 struct ib_pd *pd, struct ib_ah_attr *attr)
62 ah = kmalloc(sizeof *ah, GFP_KERNEL);
64 return ERR_PTR(-ENOMEM);
70 vah = ib_create_ah(pd, attr);
73 ah = (struct ipoib_ah *)vah;
76 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah);
82 void ipoib_free_ah(struct kref *kref)
84 struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
85 struct ipoib_dev_priv *priv = netdev_priv(ah->dev);
89 spin_lock_irqsave(&priv->lock, flags);
90 list_add_tail(&ah->list, &priv->dead_ahs);
91 spin_unlock_irqrestore(&priv->lock, flags);
94 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
95 u64 mapping[IPOIB_UD_RX_SG])
97 ib_dma_unmap_single(priv->ca, mapping[0],
98 IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
102 static int ipoib_ib_post_receive(struct net_device *dev, int id)
104 struct ipoib_dev_priv *priv = netdev_priv(dev);
105 struct ib_recv_wr *bad_wr;
108 priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
109 priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
110 priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
113 ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
115 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
116 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
117 dev_kfree_skb_any(priv->rx_ring[id].skb);
118 priv->rx_ring[id].skb = NULL;
124 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
126 struct ipoib_dev_priv *priv = netdev_priv(dev);
131 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
133 skb = dev_alloc_skb(buf_size + IPOIB_ENCAP_LEN);
138 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
139 * header. So we need 4 more bytes to get to 48 and align the
140 * IP header to a multiple of 16.
144 mapping = priv->rx_ring[id].mapping;
145 mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
147 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
150 priv->rx_ring[id].skb = skb;
153 dev_kfree_skb_any(skb);
157 static int ipoib_ib_post_receives(struct net_device *dev)
159 struct ipoib_dev_priv *priv = netdev_priv(dev);
162 for (i = 0; i < ipoib_recvq_size; ++i) {
163 if (!ipoib_alloc_rx_skb(dev, i)) {
164 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
167 if (ipoib_ib_post_receive(dev, i)) {
168 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
176 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
178 struct ipoib_dev_priv *priv = netdev_priv(dev);
179 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
181 u64 mapping[IPOIB_UD_RX_SG];
184 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
187 if (unlikely(wr_id >= ipoib_recvq_size)) {
188 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
189 wr_id, ipoib_recvq_size);
193 skb = priv->rx_ring[wr_id].skb;
195 if (unlikely(wc->status != IB_WC_SUCCESS)) {
196 if (wc->status != IB_WC_WR_FLUSH_ERR)
197 ipoib_warn(priv, "failed recv event "
198 "(status=%d, wrid=%d vend_err %x)\n",
199 wc->status, wr_id, wc->vendor_err);
200 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
201 dev_kfree_skb_any(skb);
202 priv->rx_ring[wr_id].skb = NULL;
207 * Drop packets that this interface sent, ie multicast packets
208 * that the HCA has replicated.
210 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
213 memcpy(mapping, priv->rx_ring[wr_id].mapping,
214 IPOIB_UD_RX_SG * sizeof *mapping);
217 * If we can't allocate a new RX buffer, dump
218 * this packet and reuse the old buffer.
220 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
221 ++dev->stats.rx_dropped;
225 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
226 wc->byte_len, wc->slid);
228 ipoib_ud_dma_unmap_rx(priv, mapping);
230 skb_put(skb, wc->byte_len);
232 /* First byte of dgid signals multicast when 0xff */
233 dgid = &((struct ib_grh *)skb->data)->dgid;
235 if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
236 skb->pkt_type = PACKET_HOST;
237 else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
238 skb->pkt_type = PACKET_BROADCAST;
240 skb->pkt_type = PACKET_MULTICAST;
242 skb_pull(skb, IB_GRH_BYTES);
244 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
245 skb_reset_mac_header(skb);
246 skb_pull(skb, IPOIB_ENCAP_LEN);
248 skb->truesize = SKB_TRUESIZE(skb->len);
250 ++dev->stats.rx_packets;
251 dev->stats.rx_bytes += skb->len;
254 if ((dev->features & NETIF_F_RXCSUM) &&
255 likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
256 skb->ip_summed = CHECKSUM_UNNECESSARY;
258 napi_gro_receive(&priv->napi, skb);
261 if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
262 ipoib_warn(priv, "ipoib_ib_post_receive failed "
263 "for buf %d\n", wr_id);
266 static int ipoib_dma_map_tx(struct ib_device *ca,
267 struct ipoib_tx_buf *tx_req)
269 struct sk_buff *skb = tx_req->skb;
270 u64 *mapping = tx_req->mapping;
274 if (skb_headlen(skb)) {
275 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
277 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
284 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
285 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
286 mapping[i + off] = ib_dma_map_page(ca,
288 frag->page_offset, skb_frag_size(frag),
290 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
297 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
299 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
303 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
308 static void ipoib_dma_unmap_tx(struct ib_device *ca,
309 struct ipoib_tx_buf *tx_req)
311 struct sk_buff *skb = tx_req->skb;
312 u64 *mapping = tx_req->mapping;
316 if (skb_headlen(skb)) {
317 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
322 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
323 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
325 ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
330 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
332 struct ipoib_dev_priv *priv = netdev_priv(dev);
333 unsigned int wr_id = wc->wr_id;
334 struct ipoib_tx_buf *tx_req;
336 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
339 if (unlikely(wr_id >= ipoib_sendq_size)) {
340 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
341 wr_id, ipoib_sendq_size);
345 tx_req = &priv->tx_ring[wr_id];
347 ipoib_dma_unmap_tx(priv->ca, tx_req);
349 ++dev->stats.tx_packets;
350 dev->stats.tx_bytes += tx_req->skb->len;
352 dev_kfree_skb_any(tx_req->skb);
355 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
356 netif_queue_stopped(dev) &&
357 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
358 netif_wake_queue(dev);
360 if (wc->status != IB_WC_SUCCESS &&
361 wc->status != IB_WC_WR_FLUSH_ERR)
362 ipoib_warn(priv, "failed send event "
363 "(status=%d, wrid=%d vend_err %x)\n",
364 wc->status, wr_id, wc->vendor_err);
367 static int poll_tx(struct ipoib_dev_priv *priv)
371 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
372 for (i = 0; i < n; ++i)
373 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
375 return n == MAX_SEND_CQE;
378 int ipoib_poll(struct napi_struct *napi, int budget)
380 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
381 struct net_device *dev = priv->dev;
389 while (done < budget) {
390 int max = (budget - done);
392 t = min(IPOIB_NUM_WC, max);
393 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
395 for (i = 0; i < n; i++) {
396 struct ib_wc *wc = priv->ibwc + i;
398 if (wc->wr_id & IPOIB_OP_RECV) {
400 if (wc->wr_id & IPOIB_OP_CM)
401 ipoib_cm_handle_rx_wc(dev, wc);
403 ipoib_ib_handle_rx_wc(dev, wc);
405 ipoib_cm_handle_tx_wc(priv->dev, wc);
414 if (unlikely(ib_req_notify_cq(priv->recv_cq,
416 IB_CQ_REPORT_MISSED_EVENTS)) &&
417 napi_reschedule(napi))
424 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
426 struct net_device *dev = dev_ptr;
427 struct ipoib_dev_priv *priv = netdev_priv(dev);
429 napi_schedule(&priv->napi);
432 static void drain_tx_cq(struct net_device *dev)
434 struct ipoib_dev_priv *priv = netdev_priv(dev);
437 while (poll_tx(priv))
440 if (netif_queue_stopped(dev))
441 mod_timer(&priv->poll_timer, jiffies + 1);
443 netif_tx_unlock(dev);
446 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
448 struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
450 mod_timer(&priv->poll_timer, jiffies);
453 static inline int post_send(struct ipoib_dev_priv *priv,
455 struct ib_ah *address, u32 qpn,
456 struct ipoib_tx_buf *tx_req,
457 void *head, int hlen)
459 struct ib_send_wr *bad_wr;
461 struct sk_buff *skb = tx_req->skb;
462 skb_frag_t *frags = skb_shinfo(skb)->frags;
463 int nr_frags = skb_shinfo(skb)->nr_frags;
464 u64 *mapping = tx_req->mapping;
466 if (skb_headlen(skb)) {
467 priv->tx_sge[0].addr = mapping[0];
468 priv->tx_sge[0].length = skb_headlen(skb);
473 for (i = 0; i < nr_frags; ++i) {
474 priv->tx_sge[i + off].addr = mapping[i + off];
475 priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
477 priv->tx_wr.num_sge = nr_frags + off;
478 priv->tx_wr.wr_id = wr_id;
479 priv->tx_wr.wr.ud.remote_qpn = qpn;
480 priv->tx_wr.wr.ud.ah = address;
483 priv->tx_wr.wr.ud.mss = skb_shinfo(skb)->gso_size;
484 priv->tx_wr.wr.ud.header = head;
485 priv->tx_wr.wr.ud.hlen = hlen;
486 priv->tx_wr.opcode = IB_WR_LSO;
488 priv->tx_wr.opcode = IB_WR_SEND;
490 return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
493 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
494 struct ipoib_ah *address, u32 qpn)
496 struct ipoib_dev_priv *priv = netdev_priv(dev);
497 struct ipoib_tx_buf *tx_req;
501 if (skb_is_gso(skb)) {
502 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
504 if (unlikely(!skb_pull(skb, hlen))) {
505 ipoib_warn(priv, "linear data too small\n");
506 ++dev->stats.tx_dropped;
507 ++dev->stats.tx_errors;
508 dev_kfree_skb_any(skb);
512 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
513 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
514 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
515 ++dev->stats.tx_dropped;
516 ++dev->stats.tx_errors;
517 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
524 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
525 skb->len, address, qpn);
528 * We put the skb into the tx_ring _before_ we call post_send()
529 * because it's entirely possible that the completion handler will
530 * run before we execute anything after the post_send(). That
531 * means we have to make sure everything is properly recorded and
532 * our state is consistent before we call post_send().
534 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
536 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
537 ++dev->stats.tx_errors;
538 dev_kfree_skb_any(skb);
542 if (skb->ip_summed == CHECKSUM_PARTIAL)
543 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
545 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
547 if (++priv->tx_outstanding == ipoib_sendq_size) {
548 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
549 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
550 ipoib_warn(priv, "request notify on send CQ failed\n");
551 netif_stop_queue(dev);
557 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
558 address->ah, qpn, tx_req, phead, hlen);
560 ipoib_warn(priv, "post_send failed, error %d\n", rc);
561 ++dev->stats.tx_errors;
562 --priv->tx_outstanding;
563 ipoib_dma_unmap_tx(priv->ca, tx_req);
564 dev_kfree_skb_any(skb);
565 if (netif_queue_stopped(dev))
566 netif_wake_queue(dev);
568 dev->trans_start = jiffies;
570 address->last_send = priv->tx_head;
574 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
575 while (poll_tx(priv))
579 static void __ipoib_reap_ah(struct net_device *dev)
581 struct ipoib_dev_priv *priv = netdev_priv(dev);
582 struct ipoib_ah *ah, *tah;
583 LIST_HEAD(remove_list);
586 netif_tx_lock_bh(dev);
587 spin_lock_irqsave(&priv->lock, flags);
589 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
590 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
592 ib_destroy_ah(ah->ah);
596 spin_unlock_irqrestore(&priv->lock, flags);
597 netif_tx_unlock_bh(dev);
600 void ipoib_reap_ah(struct work_struct *work)
602 struct ipoib_dev_priv *priv =
603 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
604 struct net_device *dev = priv->dev;
606 __ipoib_reap_ah(dev);
608 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
609 queue_delayed_work(priv->wq, &priv->ah_reap_task,
610 round_jiffies_relative(HZ));
613 static void ipoib_flush_ah(struct net_device *dev)
615 struct ipoib_dev_priv *priv = netdev_priv(dev);
617 cancel_delayed_work(&priv->ah_reap_task);
618 flush_workqueue(priv->wq);
619 ipoib_reap_ah(&priv->ah_reap_task.work);
622 static void ipoib_stop_ah(struct net_device *dev)
624 struct ipoib_dev_priv *priv = netdev_priv(dev);
626 set_bit(IPOIB_STOP_REAPER, &priv->flags);
630 static void ipoib_ib_tx_timer_func(unsigned long ctx)
632 drain_tx_cq((struct net_device *)ctx);
635 int ipoib_ib_dev_open(struct net_device *dev)
637 struct ipoib_dev_priv *priv = netdev_priv(dev);
640 ipoib_pkey_dev_check_presence(dev);
642 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
643 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
644 (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
648 ret = ipoib_init_qp(dev);
650 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
654 ret = ipoib_ib_post_receives(dev);
656 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
660 ret = ipoib_cm_dev_open(dev);
662 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
666 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
667 queue_delayed_work(priv->wq, &priv->ah_reap_task,
668 round_jiffies_relative(HZ));
670 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
671 napi_enable(&priv->napi);
675 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
676 napi_enable(&priv->napi);
677 ipoib_ib_dev_stop(dev);
681 void ipoib_pkey_dev_check_presence(struct net_device *dev)
683 struct ipoib_dev_priv *priv = netdev_priv(dev);
685 if (!(priv->pkey & 0x7fff) ||
686 ib_find_pkey(priv->ca, priv->port, priv->pkey,
688 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
690 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
693 int ipoib_ib_dev_up(struct net_device *dev)
695 struct ipoib_dev_priv *priv = netdev_priv(dev);
697 ipoib_pkey_dev_check_presence(dev);
699 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
700 ipoib_dbg(priv, "PKEY is not assigned.\n");
704 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
706 return ipoib_mcast_start_thread(dev);
709 int ipoib_ib_dev_down(struct net_device *dev)
711 struct ipoib_dev_priv *priv = netdev_priv(dev);
713 ipoib_dbg(priv, "downing ib_dev\n");
715 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
716 netif_carrier_off(dev);
718 ipoib_mcast_stop_thread(dev);
719 ipoib_mcast_dev_flush(dev);
721 ipoib_flush_paths(dev);
726 static int recvs_pending(struct net_device *dev)
728 struct ipoib_dev_priv *priv = netdev_priv(dev);
732 for (i = 0; i < ipoib_recvq_size; ++i)
733 if (priv->rx_ring[i].skb)
739 void ipoib_drain_cq(struct net_device *dev)
741 struct ipoib_dev_priv *priv = netdev_priv(dev);
745 * We call completion handling routines that expect to be
746 * called from the BH-disabled NAPI poll context, so disable
752 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
753 for (i = 0; i < n; ++i) {
755 * Convert any successful completions to flush
756 * errors to avoid passing packets up the
757 * stack after bringing the device down.
759 if (priv->ibwc[i].status == IB_WC_SUCCESS)
760 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
762 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
763 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
764 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
766 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
768 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
770 } while (n == IPOIB_NUM_WC);
772 while (poll_tx(priv))
778 int ipoib_ib_dev_stop(struct net_device *dev)
780 struct ipoib_dev_priv *priv = netdev_priv(dev);
781 struct ib_qp_attr qp_attr;
783 struct ipoib_tx_buf *tx_req;
786 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
787 napi_disable(&priv->napi);
789 ipoib_cm_dev_stop(dev);
792 * Move our QP to the error state and then reinitialize in
793 * when all work requests have completed or have been flushed.
795 qp_attr.qp_state = IB_QPS_ERR;
796 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
797 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
799 /* Wait for all sends and receives to complete */
802 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
803 if (time_after(jiffies, begin + 5 * HZ)) {
804 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
805 priv->tx_head - priv->tx_tail, recvs_pending(dev));
808 * assume the HW is wedged and just free up
809 * all our pending work requests.
811 while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
812 tx_req = &priv->tx_ring[priv->tx_tail &
813 (ipoib_sendq_size - 1)];
814 ipoib_dma_unmap_tx(priv->ca, tx_req);
815 dev_kfree_skb_any(tx_req->skb);
817 --priv->tx_outstanding;
820 for (i = 0; i < ipoib_recvq_size; ++i) {
821 struct ipoib_rx_buf *rx_req;
823 rx_req = &priv->rx_ring[i];
826 ipoib_ud_dma_unmap_rx(priv,
827 priv->rx_ring[i].mapping);
828 dev_kfree_skb_any(rx_req->skb);
840 ipoib_dbg(priv, "All sends and receives done.\n");
843 del_timer_sync(&priv->poll_timer);
844 qp_attr.qp_state = IB_QPS_RESET;
845 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
846 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
850 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
855 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
857 struct ipoib_dev_priv *priv = netdev_priv(dev);
863 if (ipoib_transport_dev_init(dev, ca)) {
864 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
868 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
869 (unsigned long) dev);
871 if (dev->flags & IFF_UP) {
872 if (ipoib_ib_dev_open(dev)) {
873 ipoib_transport_dev_cleanup(dev);
882 * Takes whatever value which is in pkey index 0 and updates priv->pkey
883 * returns 0 if the pkey value was changed.
885 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
890 prev_pkey = priv->pkey;
891 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
893 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
898 priv->pkey |= 0x8000;
900 if (prev_pkey != priv->pkey) {
901 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
902 prev_pkey, priv->pkey);
904 * Update the pkey in the broadcast address, while making sure to set
905 * the full membership bit, so that we join the right broadcast group.
907 priv->dev->broadcast[8] = priv->pkey >> 8;
908 priv->dev->broadcast[9] = priv->pkey & 0xff;
915 * returns 0 if pkey value was found in a different slot.
917 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
919 u16 old_index = priv->pkey_index;
921 priv->pkey_index = 0;
922 ipoib_pkey_dev_check_presence(priv->dev);
924 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
925 (old_index == priv->pkey_index))
930 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
931 enum ipoib_flush_level level)
933 struct ipoib_dev_priv *cpriv;
934 struct net_device *dev = priv->dev;
937 down_read(&priv->vlan_rwsem);
940 * Flush any child interfaces too -- they might be up even if
941 * the parent is down.
943 list_for_each_entry(cpriv, &priv->child_intfs, list)
944 __ipoib_ib_dev_flush(cpriv, level);
946 up_read(&priv->vlan_rwsem);
948 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
949 level != IPOIB_FLUSH_HEAVY) {
950 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
954 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
955 /* interface is down. update pkey and leave. */
956 if (level == IPOIB_FLUSH_HEAVY) {
957 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
958 update_parent_pkey(priv);
960 update_child_pkey(priv);
962 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
966 if (level == IPOIB_FLUSH_HEAVY) {
967 /* child devices chase their origin pkey value, while non-child
968 * (parent) devices should always takes what present in pkey index 0
970 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
971 result = update_child_pkey(priv);
973 /* restart QP only if P_Key index is changed */
974 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
979 result = update_parent_pkey(priv);
980 /* restart QP only if P_Key value changed */
982 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
988 if (level == IPOIB_FLUSH_LIGHT) {
989 ipoib_mark_paths_invalid(dev);
990 ipoib_mcast_dev_flush(dev);
994 if (level >= IPOIB_FLUSH_NORMAL)
995 ipoib_ib_dev_down(dev);
997 if (level == IPOIB_FLUSH_HEAVY) {
998 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
999 ipoib_ib_dev_stop(dev);
1000 if (ipoib_ib_dev_open(dev) != 0)
1002 if (netif_queue_stopped(dev))
1003 netif_start_queue(dev);
1007 * The device could have been brought down between the start and when
1008 * we get here, don't bring it back up if it's not configured up
1010 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1011 if (level >= IPOIB_FLUSH_NORMAL)
1012 ipoib_ib_dev_up(dev);
1013 ipoib_mcast_restart_task(&priv->restart_task);
1017 void ipoib_ib_dev_flush_light(struct work_struct *work)
1019 struct ipoib_dev_priv *priv =
1020 container_of(work, struct ipoib_dev_priv, flush_light);
1022 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1025 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1027 struct ipoib_dev_priv *priv =
1028 container_of(work, struct ipoib_dev_priv, flush_normal);
1030 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1033 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1035 struct ipoib_dev_priv *priv =
1036 container_of(work, struct ipoib_dev_priv, flush_heavy);
1038 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1041 void ipoib_ib_dev_cleanup(struct net_device *dev)
1043 struct ipoib_dev_priv *priv = netdev_priv(dev);
1045 ipoib_dbg(priv, "cleaning up ib_dev\n");
1047 * We must make sure there are no more (path) completions
1048 * that may wish to touch priv fields that are no longer valid
1050 ipoib_flush_paths(dev);
1052 ipoib_mcast_stop_thread(dev);
1053 ipoib_mcast_dev_flush(dev);
1056 * All of our ah references aren't free until after
1057 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1058 * the neighbor garbage collection is stopped and reaped.
1059 * That should all be done now, so make a final ah flush.
1063 ipoib_transport_dev_cleanup(dev);