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 if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
98 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE,
100 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE,
103 ib_dma_unmap_single(priv->ca, mapping[0],
104 IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
108 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv,
112 if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
113 skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
116 * There is only two buffers needed for max_payload = 4K,
117 * first buf size is IPOIB_UD_HEAD_SIZE
119 skb->tail += IPOIB_UD_HEAD_SIZE;
122 size = length - IPOIB_UD_HEAD_SIZE;
124 skb_frag_size_set(frag, size);
125 skb->data_len += size;
126 skb->truesize += PAGE_SIZE;
128 skb_put(skb, length);
132 static int ipoib_ib_post_receive(struct net_device *dev, int id)
134 struct ipoib_dev_priv *priv = netdev_priv(dev);
135 struct ib_recv_wr *bad_wr;
138 priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
139 priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
140 priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
143 ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
145 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
146 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
147 dev_kfree_skb_any(priv->rx_ring[id].skb);
148 priv->rx_ring[id].skb = NULL;
154 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
156 struct ipoib_dev_priv *priv = netdev_priv(dev);
162 if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
163 buf_size = IPOIB_UD_HEAD_SIZE;
164 tailroom = 128; /* reserve some tailroom for IP/TCP headers */
166 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
170 skb = dev_alloc_skb(buf_size + tailroom + 4);
175 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
176 * header. So we need 4 more bytes to get to 48 and align the
177 * IP header to a multiple of 16.
181 mapping = priv->rx_ring[id].mapping;
182 mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
184 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
187 if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
188 struct page *page = alloc_page(GFP_ATOMIC);
191 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE);
193 ib_dma_map_page(priv->ca, page,
194 0, PAGE_SIZE, DMA_FROM_DEVICE);
195 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1])))
199 priv->rx_ring[id].skb = skb;
203 ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE);
205 dev_kfree_skb_any(skb);
209 static int ipoib_ib_post_receives(struct net_device *dev)
211 struct ipoib_dev_priv *priv = netdev_priv(dev);
214 for (i = 0; i < ipoib_recvq_size; ++i) {
215 if (!ipoib_alloc_rx_skb(dev, i)) {
216 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
219 if (ipoib_ib_post_receive(dev, i)) {
220 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
228 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
230 struct ipoib_dev_priv *priv = netdev_priv(dev);
231 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
233 u64 mapping[IPOIB_UD_RX_SG];
236 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
239 if (unlikely(wr_id >= ipoib_recvq_size)) {
240 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
241 wr_id, ipoib_recvq_size);
245 skb = priv->rx_ring[wr_id].skb;
247 if (unlikely(wc->status != IB_WC_SUCCESS)) {
248 if (wc->status != IB_WC_WR_FLUSH_ERR)
249 ipoib_warn(priv, "failed recv event "
250 "(status=%d, wrid=%d vend_err %x)\n",
251 wc->status, wr_id, wc->vendor_err);
252 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
253 dev_kfree_skb_any(skb);
254 priv->rx_ring[wr_id].skb = NULL;
259 * Drop packets that this interface sent, ie multicast packets
260 * that the HCA has replicated.
262 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num)
265 memcpy(mapping, priv->rx_ring[wr_id].mapping,
266 IPOIB_UD_RX_SG * sizeof *mapping);
269 * If we can't allocate a new RX buffer, dump
270 * this packet and reuse the old buffer.
272 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
273 ++dev->stats.rx_dropped;
277 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
278 wc->byte_len, wc->slid);
280 ipoib_ud_dma_unmap_rx(priv, mapping);
281 ipoib_ud_skb_put_frags(priv, skb, wc->byte_len);
283 /* First byte of dgid signals multicast when 0xff */
284 dgid = &((struct ib_grh *)skb->data)->dgid;
286 if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
287 skb->pkt_type = PACKET_HOST;
288 else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
289 skb->pkt_type = PACKET_BROADCAST;
291 skb->pkt_type = PACKET_MULTICAST;
293 skb_pull(skb, IB_GRH_BYTES);
295 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
296 skb_reset_mac_header(skb);
297 skb_pull(skb, IPOIB_ENCAP_LEN);
299 ++dev->stats.rx_packets;
300 dev->stats.rx_bytes += skb->len;
303 if ((dev->features & NETIF_F_RXCSUM) &&
304 likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
305 skb->ip_summed = CHECKSUM_UNNECESSARY;
307 napi_gro_receive(&priv->napi, skb);
310 if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
311 ipoib_warn(priv, "ipoib_ib_post_receive failed "
312 "for buf %d\n", wr_id);
315 static int ipoib_dma_map_tx(struct ib_device *ca,
316 struct ipoib_tx_buf *tx_req)
318 struct sk_buff *skb = tx_req->skb;
319 u64 *mapping = tx_req->mapping;
323 if (skb_headlen(skb)) {
324 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
326 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
333 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
334 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
335 mapping[i + off] = ib_dma_map_page(ca,
337 frag->page_offset, skb_frag_size(frag),
339 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
346 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
348 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
352 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
357 static void ipoib_dma_unmap_tx(struct ib_device *ca,
358 struct ipoib_tx_buf *tx_req)
360 struct sk_buff *skb = tx_req->skb;
361 u64 *mapping = tx_req->mapping;
365 if (skb_headlen(skb)) {
366 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
371 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
372 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
374 ib_dma_unmap_page(ca, mapping[i + off], skb_frag_size(frag),
379 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
381 struct ipoib_dev_priv *priv = netdev_priv(dev);
382 unsigned int wr_id = wc->wr_id;
383 struct ipoib_tx_buf *tx_req;
385 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
388 if (unlikely(wr_id >= ipoib_sendq_size)) {
389 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
390 wr_id, ipoib_sendq_size);
394 tx_req = &priv->tx_ring[wr_id];
396 ipoib_dma_unmap_tx(priv->ca, tx_req);
398 ++dev->stats.tx_packets;
399 dev->stats.tx_bytes += tx_req->skb->len;
401 dev_kfree_skb_any(tx_req->skb);
404 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
405 netif_queue_stopped(dev) &&
406 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
407 netif_wake_queue(dev);
409 if (wc->status != IB_WC_SUCCESS &&
410 wc->status != IB_WC_WR_FLUSH_ERR)
411 ipoib_warn(priv, "failed send event "
412 "(status=%d, wrid=%d vend_err %x)\n",
413 wc->status, wr_id, wc->vendor_err);
416 static int poll_tx(struct ipoib_dev_priv *priv)
420 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
421 for (i = 0; i < n; ++i)
422 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
424 return n == MAX_SEND_CQE;
427 int ipoib_poll(struct napi_struct *napi, int budget)
429 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi);
430 struct net_device *dev = priv->dev;
438 while (done < budget) {
439 int max = (budget - done);
441 t = min(IPOIB_NUM_WC, max);
442 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
444 for (i = 0; i < n; i++) {
445 struct ib_wc *wc = priv->ibwc + i;
447 if (wc->wr_id & IPOIB_OP_RECV) {
449 if (wc->wr_id & IPOIB_OP_CM)
450 ipoib_cm_handle_rx_wc(dev, wc);
452 ipoib_ib_handle_rx_wc(dev, wc);
454 ipoib_cm_handle_tx_wc(priv->dev, wc);
463 if (unlikely(ib_req_notify_cq(priv->recv_cq,
465 IB_CQ_REPORT_MISSED_EVENTS)) &&
466 napi_reschedule(napi))
473 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
475 struct net_device *dev = dev_ptr;
476 struct ipoib_dev_priv *priv = netdev_priv(dev);
478 napi_schedule(&priv->napi);
481 static void drain_tx_cq(struct net_device *dev)
483 struct ipoib_dev_priv *priv = netdev_priv(dev);
486 while (poll_tx(priv))
489 if (netif_queue_stopped(dev))
490 mod_timer(&priv->poll_timer, jiffies + 1);
492 netif_tx_unlock(dev);
495 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr)
497 struct ipoib_dev_priv *priv = netdev_priv(dev_ptr);
499 mod_timer(&priv->poll_timer, jiffies);
502 static inline int post_send(struct ipoib_dev_priv *priv,
504 struct ib_ah *address, u32 qpn,
505 struct ipoib_tx_buf *tx_req,
506 void *head, int hlen)
508 struct ib_send_wr *bad_wr;
510 struct sk_buff *skb = tx_req->skb;
511 skb_frag_t *frags = skb_shinfo(skb)->frags;
512 int nr_frags = skb_shinfo(skb)->nr_frags;
513 u64 *mapping = tx_req->mapping;
515 if (skb_headlen(skb)) {
516 priv->tx_sge[0].addr = mapping[0];
517 priv->tx_sge[0].length = skb_headlen(skb);
522 for (i = 0; i < nr_frags; ++i) {
523 priv->tx_sge[i + off].addr = mapping[i + off];
524 priv->tx_sge[i + off].length = skb_frag_size(&frags[i]);
526 priv->tx_wr.num_sge = nr_frags + off;
527 priv->tx_wr.wr_id = wr_id;
528 priv->tx_wr.wr.ud.remote_qpn = qpn;
529 priv->tx_wr.wr.ud.ah = address;
532 priv->tx_wr.wr.ud.mss = skb_shinfo(skb)->gso_size;
533 priv->tx_wr.wr.ud.header = head;
534 priv->tx_wr.wr.ud.hlen = hlen;
535 priv->tx_wr.opcode = IB_WR_LSO;
537 priv->tx_wr.opcode = IB_WR_SEND;
539 return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
542 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
543 struct ipoib_ah *address, u32 qpn)
545 struct ipoib_dev_priv *priv = netdev_priv(dev);
546 struct ipoib_tx_buf *tx_req;
550 if (skb_is_gso(skb)) {
551 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
553 if (unlikely(!skb_pull(skb, hlen))) {
554 ipoib_warn(priv, "linear data too small\n");
555 ++dev->stats.tx_dropped;
556 ++dev->stats.tx_errors;
557 dev_kfree_skb_any(skb);
561 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
562 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
563 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
564 ++dev->stats.tx_dropped;
565 ++dev->stats.tx_errors;
566 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
573 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
574 skb->len, address, qpn);
577 * We put the skb into the tx_ring _before_ we call post_send()
578 * because it's entirely possible that the completion handler will
579 * run before we execute anything after the post_send(). That
580 * means we have to make sure everything is properly recorded and
581 * our state is consistent before we call post_send().
583 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
585 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
586 ++dev->stats.tx_errors;
587 dev_kfree_skb_any(skb);
591 if (skb->ip_summed == CHECKSUM_PARTIAL)
592 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM;
594 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
596 if (++priv->tx_outstanding == ipoib_sendq_size) {
597 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
598 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
599 ipoib_warn(priv, "request notify on send CQ failed\n");
600 netif_stop_queue(dev);
606 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
607 address->ah, qpn, tx_req, phead, hlen);
609 ipoib_warn(priv, "post_send failed, error %d\n", rc);
610 ++dev->stats.tx_errors;
611 --priv->tx_outstanding;
612 ipoib_dma_unmap_tx(priv->ca, tx_req);
613 dev_kfree_skb_any(skb);
614 if (netif_queue_stopped(dev))
615 netif_wake_queue(dev);
617 dev->trans_start = jiffies;
619 address->last_send = priv->tx_head;
623 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE))
624 while (poll_tx(priv))
628 static void __ipoib_reap_ah(struct net_device *dev)
630 struct ipoib_dev_priv *priv = netdev_priv(dev);
631 struct ipoib_ah *ah, *tah;
632 LIST_HEAD(remove_list);
635 netif_tx_lock_bh(dev);
636 spin_lock_irqsave(&priv->lock, flags);
638 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
639 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
641 ib_destroy_ah(ah->ah);
645 spin_unlock_irqrestore(&priv->lock, flags);
646 netif_tx_unlock_bh(dev);
649 void ipoib_reap_ah(struct work_struct *work)
651 struct ipoib_dev_priv *priv =
652 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
653 struct net_device *dev = priv->dev;
655 __ipoib_reap_ah(dev);
657 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
658 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
659 round_jiffies_relative(HZ));
662 static void ipoib_flush_ah(struct net_device *dev, int flush)
664 struct ipoib_dev_priv *priv = netdev_priv(dev);
666 cancel_delayed_work(&priv->ah_reap_task);
668 flush_workqueue(ipoib_workqueue);
669 ipoib_reap_ah(&priv->ah_reap_task.work);
672 static void ipoib_stop_ah(struct net_device *dev, int flush)
674 struct ipoib_dev_priv *priv = netdev_priv(dev);
676 set_bit(IPOIB_STOP_REAPER, &priv->flags);
677 ipoib_flush_ah(dev, flush);
680 static void ipoib_ib_tx_timer_func(unsigned long ctx)
682 drain_tx_cq((struct net_device *)ctx);
685 int ipoib_ib_dev_open(struct net_device *dev, int flush)
687 struct ipoib_dev_priv *priv = netdev_priv(dev);
690 ipoib_pkey_dev_check_presence(dev);
692 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
693 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
694 (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
698 ret = ipoib_init_qp(dev);
700 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
704 ret = ipoib_ib_post_receives(dev);
706 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
710 ret = ipoib_cm_dev_open(dev);
712 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
716 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
717 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task,
718 round_jiffies_relative(HZ));
720 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
721 napi_enable(&priv->napi);
725 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
726 napi_enable(&priv->napi);
727 ipoib_ib_dev_stop(dev, flush);
731 void ipoib_pkey_dev_check_presence(struct net_device *dev)
733 struct ipoib_dev_priv *priv = netdev_priv(dev);
735 if (!(priv->pkey & 0x7fff) ||
736 ib_find_pkey(priv->ca, priv->port, priv->pkey,
738 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
740 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
743 int ipoib_ib_dev_up(struct net_device *dev)
745 struct ipoib_dev_priv *priv = netdev_priv(dev);
747 ipoib_pkey_dev_check_presence(dev);
749 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
750 ipoib_dbg(priv, "PKEY is not assigned.\n");
754 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
756 return ipoib_mcast_start_thread(dev);
759 int ipoib_ib_dev_down(struct net_device *dev, int flush)
761 struct ipoib_dev_priv *priv = netdev_priv(dev);
763 ipoib_dbg(priv, "downing ib_dev\n");
765 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
766 netif_carrier_off(dev);
768 ipoib_mcast_stop_thread(dev, flush);
769 ipoib_mcast_dev_flush(dev);
771 ipoib_flush_paths(dev);
776 static int recvs_pending(struct net_device *dev)
778 struct ipoib_dev_priv *priv = netdev_priv(dev);
782 for (i = 0; i < ipoib_recvq_size; ++i)
783 if (priv->rx_ring[i].skb)
789 void ipoib_drain_cq(struct net_device *dev)
791 struct ipoib_dev_priv *priv = netdev_priv(dev);
795 * We call completion handling routines that expect to be
796 * called from the BH-disabled NAPI poll context, so disable
802 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
803 for (i = 0; i < n; ++i) {
805 * Convert any successful completions to flush
806 * errors to avoid passing packets up the
807 * stack after bringing the device down.
809 if (priv->ibwc[i].status == IB_WC_SUCCESS)
810 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
812 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
813 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
814 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
816 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
818 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i);
820 } while (n == IPOIB_NUM_WC);
822 while (poll_tx(priv))
828 int ipoib_ib_dev_stop(struct net_device *dev, int flush)
830 struct ipoib_dev_priv *priv = netdev_priv(dev);
831 struct ib_qp_attr qp_attr;
833 struct ipoib_tx_buf *tx_req;
836 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
837 napi_disable(&priv->napi);
839 ipoib_cm_dev_stop(dev);
842 * Move our QP to the error state and then reinitialize in
843 * when all work requests have completed or have been flushed.
845 qp_attr.qp_state = IB_QPS_ERR;
846 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
847 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
849 /* Wait for all sends and receives to complete */
852 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
853 if (time_after(jiffies, begin + 5 * HZ)) {
854 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
855 priv->tx_head - priv->tx_tail, recvs_pending(dev));
858 * assume the HW is wedged and just free up
859 * all our pending work requests.
861 while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
862 tx_req = &priv->tx_ring[priv->tx_tail &
863 (ipoib_sendq_size - 1)];
864 ipoib_dma_unmap_tx(priv->ca, tx_req);
865 dev_kfree_skb_any(tx_req->skb);
867 --priv->tx_outstanding;
870 for (i = 0; i < ipoib_recvq_size; ++i) {
871 struct ipoib_rx_buf *rx_req;
873 rx_req = &priv->rx_ring[i];
876 ipoib_ud_dma_unmap_rx(priv,
877 priv->rx_ring[i].mapping);
878 dev_kfree_skb_any(rx_req->skb);
890 ipoib_dbg(priv, "All sends and receives done.\n");
893 del_timer_sync(&priv->poll_timer);
894 qp_attr.qp_state = IB_QPS_RESET;
895 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
896 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
898 ipoib_flush_ah(dev, flush);
900 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
905 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
907 struct ipoib_dev_priv *priv = netdev_priv(dev);
913 if (ipoib_transport_dev_init(dev, ca)) {
914 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
918 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func,
919 (unsigned long) dev);
921 if (dev->flags & IFF_UP) {
922 if (ipoib_ib_dev_open(dev, 1)) {
923 ipoib_transport_dev_cleanup(dev);
932 * Takes whatever value which is in pkey index 0 and updates priv->pkey
933 * returns 0 if the pkey value was changed.
935 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
940 prev_pkey = priv->pkey;
941 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
943 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
948 priv->pkey |= 0x8000;
950 if (prev_pkey != priv->pkey) {
951 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
952 prev_pkey, priv->pkey);
954 * Update the pkey in the broadcast address, while making sure to set
955 * the full membership bit, so that we join the right broadcast group.
957 priv->dev->broadcast[8] = priv->pkey >> 8;
958 priv->dev->broadcast[9] = priv->pkey & 0xff;
965 * returns 0 if pkey value was found in a different slot.
967 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
969 u16 old_index = priv->pkey_index;
971 priv->pkey_index = 0;
972 ipoib_pkey_dev_check_presence(priv->dev);
974 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
975 (old_index == priv->pkey_index))
980 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
981 enum ipoib_flush_level level)
983 struct ipoib_dev_priv *cpriv;
984 struct net_device *dev = priv->dev;
987 down_read(&priv->vlan_rwsem);
990 * Flush any child interfaces too -- they might be up even if
991 * the parent is down.
993 list_for_each_entry(cpriv, &priv->child_intfs, list)
994 __ipoib_ib_dev_flush(cpriv, level);
996 up_read(&priv->vlan_rwsem);
998 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
999 level != IPOIB_FLUSH_HEAVY) {
1000 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1004 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1005 /* interface is down. update pkey and leave. */
1006 if (level == IPOIB_FLUSH_HEAVY) {
1007 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1008 update_parent_pkey(priv);
1010 update_child_pkey(priv);
1012 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1016 if (level == IPOIB_FLUSH_HEAVY) {
1017 /* child devices chase their origin pkey value, while non-child
1018 * (parent) devices should always takes what present in pkey index 0
1020 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1021 result = update_child_pkey(priv);
1023 /* restart QP only if P_Key index is changed */
1024 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1029 result = update_parent_pkey(priv);
1030 /* restart QP only if P_Key value changed */
1032 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1038 if (level == IPOIB_FLUSH_LIGHT) {
1039 ipoib_mark_paths_invalid(dev);
1040 ipoib_mcast_dev_flush(dev);
1041 ipoib_flush_ah(dev, 0);
1044 if (level >= IPOIB_FLUSH_NORMAL)
1045 ipoib_ib_dev_down(dev, 0);
1047 if (level == IPOIB_FLUSH_HEAVY) {
1048 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1049 ipoib_ib_dev_stop(dev, 0);
1050 if (ipoib_ib_dev_open(dev, 0) != 0)
1052 if (netif_queue_stopped(dev))
1053 netif_start_queue(dev);
1057 * The device could have been brought down between the start and when
1058 * we get here, don't bring it back up if it's not configured up
1060 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1061 if (level >= IPOIB_FLUSH_NORMAL)
1062 ipoib_ib_dev_up(dev);
1063 ipoib_mcast_restart_task(&priv->restart_task);
1067 void ipoib_ib_dev_flush_light(struct work_struct *work)
1069 struct ipoib_dev_priv *priv =
1070 container_of(work, struct ipoib_dev_priv, flush_light);
1072 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT);
1075 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1077 struct ipoib_dev_priv *priv =
1078 container_of(work, struct ipoib_dev_priv, flush_normal);
1080 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL);
1083 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1085 struct ipoib_dev_priv *priv =
1086 container_of(work, struct ipoib_dev_priv, flush_heavy);
1088 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY);
1091 void ipoib_ib_dev_cleanup(struct net_device *dev)
1093 struct ipoib_dev_priv *priv = netdev_priv(dev);
1095 ipoib_dbg(priv, "cleaning up ib_dev\n");
1097 * We must make sure there are no more (path) completions
1098 * that may wish to touch priv fields that are no longer valid
1100 ipoib_flush_paths(dev);
1102 ipoib_mcast_stop_thread(dev, 1);
1103 ipoib_mcast_dev_flush(dev);
1106 * All of our ah references aren't free until after
1107 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1108 * the neighbor garbage collection is stopped and reaped.
1109 * That should all be done now, so make a final ah flush.
1111 ipoib_stop_ah(dev, 1);
1113 ipoib_transport_dev_cleanup(dev);