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 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
55 struct ib_pd *pd, struct rdma_ah_attr *attr)
60 ah = kmalloc(sizeof *ah, GFP_KERNEL);
62 return ERR_PTR(-ENOMEM);
68 vah = rdma_create_ah(pd, attr);
71 ah = (struct ipoib_ah *)vah;
74 ipoib_dbg(ipoib_priv(dev), "Created ah %p\n", ah->ah);
80 void ipoib_free_ah(struct kref *kref)
82 struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
83 struct ipoib_dev_priv *priv = ipoib_priv(ah->dev);
87 spin_lock_irqsave(&priv->lock, flags);
88 list_add_tail(&ah->list, &priv->dead_ahs);
89 spin_unlock_irqrestore(&priv->lock, flags);
92 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
93 u64 mapping[IPOIB_UD_RX_SG])
95 ib_dma_unmap_single(priv->ca, mapping[0],
96 IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
100 static int ipoib_ib_post_receive(struct net_device *dev, int id)
102 struct ipoib_dev_priv *priv = ipoib_priv(dev);
103 struct ib_recv_wr *bad_wr;
106 priv->rx_wr.wr_id = id | IPOIB_OP_RECV;
107 priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];
108 priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];
111 ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr);
113 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
114 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);
115 dev_kfree_skb_any(priv->rx_ring[id].skb);
116 priv->rx_ring[id].skb = NULL;
122 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)
124 struct ipoib_dev_priv *priv = ipoib_priv(dev);
129 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
131 skb = dev_alloc_skb(buf_size + IPOIB_HARD_LEN);
136 * the IP header will be at IPOIP_HARD_LEN + IB_GRH_BYTES, that is
139 skb_reserve(skb, sizeof(struct ipoib_pseudo_header));
141 mapping = priv->rx_ring[id].mapping;
142 mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,
144 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))
147 priv->rx_ring[id].skb = skb;
150 dev_kfree_skb_any(skb);
154 static int ipoib_ib_post_receives(struct net_device *dev)
156 struct ipoib_dev_priv *priv = ipoib_priv(dev);
159 for (i = 0; i < ipoib_recvq_size; ++i) {
160 if (!ipoib_alloc_rx_skb(dev, i)) {
161 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
164 if (ipoib_ib_post_receive(dev, i)) {
165 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
173 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
175 struct ipoib_dev_priv *priv = ipoib_priv(dev);
176 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;
178 u64 mapping[IPOIB_UD_RX_SG];
182 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",
185 if (unlikely(wr_id >= ipoib_recvq_size)) {
186 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",
187 wr_id, ipoib_recvq_size);
191 skb = priv->rx_ring[wr_id].skb;
193 if (unlikely(wc->status != IB_WC_SUCCESS)) {
194 if (wc->status != IB_WC_WR_FLUSH_ERR)
196 "failed recv event (status=%d, wrid=%d vend_err %#x)\n",
197 wc->status, wr_id, wc->vendor_err);
198 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);
199 dev_kfree_skb_any(skb);
200 priv->rx_ring[wr_id].skb = NULL;
204 memcpy(mapping, priv->rx_ring[wr_id].mapping,
205 IPOIB_UD_RX_SG * sizeof *mapping);
208 * If we can't allocate a new RX buffer, dump
209 * this packet and reuse the old buffer.
211 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {
212 ++dev->stats.rx_dropped;
216 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
217 wc->byte_len, wc->slid);
219 ipoib_ud_dma_unmap_rx(priv, mapping);
221 skb_put(skb, wc->byte_len);
223 /* First byte of dgid signals multicast when 0xff */
224 dgid = &((struct ib_grh *)skb->data)->dgid;
226 if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)
227 skb->pkt_type = PACKET_HOST;
228 else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)
229 skb->pkt_type = PACKET_BROADCAST;
231 skb->pkt_type = PACKET_MULTICAST;
233 sgid = &((struct ib_grh *)skb->data)->sgid;
236 * Drop packets that this interface sent, ie multicast packets
237 * that the HCA has replicated.
239 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) {
242 if ((wc->wc_flags & IB_WC_GRH) &&
243 sgid->global.interface_id != priv->local_gid.global.interface_id)
247 dev_kfree_skb_any(skb);
252 skb_pull(skb, IB_GRH_BYTES);
254 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
255 skb_add_pseudo_hdr(skb);
257 ++dev->stats.rx_packets;
258 dev->stats.rx_bytes += skb->len;
259 if (skb->pkt_type == PACKET_MULTICAST)
260 dev->stats.multicast++;
263 if ((dev->features & NETIF_F_RXCSUM) &&
264 likely(wc->wc_flags & IB_WC_IP_CSUM_OK))
265 skb->ip_summed = CHECKSUM_UNNECESSARY;
267 napi_gro_receive(&priv->recv_napi, skb);
270 if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
271 ipoib_warn(priv, "ipoib_ib_post_receive failed "
272 "for buf %d\n", wr_id);
275 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)
277 struct sk_buff *skb = tx_req->skb;
278 u64 *mapping = tx_req->mapping;
282 if (skb_headlen(skb)) {
283 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),
285 if (unlikely(ib_dma_mapping_error(ca, mapping[0])))
292 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
293 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
294 mapping[i + off] = ib_dma_map_page(ca,
296 frag->page_offset, skb_frag_size(frag),
298 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
305 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
307 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
311 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
316 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
317 struct ipoib_tx_buf *tx_req)
319 struct sk_buff *skb = tx_req->skb;
320 u64 *mapping = tx_req->mapping;
324 if (skb_headlen(skb)) {
325 ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
331 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
332 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
334 ib_dma_unmap_page(priv->ca, mapping[i + off],
335 skb_frag_size(frag), DMA_TO_DEVICE);
340 * As the result of a completion error the QP Can be transferred to SQE states.
341 * The function checks if the (send)QP is in SQE state and
342 * moves it back to RTS state, that in order to have it functional again.
344 static void ipoib_qp_state_validate_work(struct work_struct *work)
346 struct ipoib_qp_state_validate *qp_work =
347 container_of(work, struct ipoib_qp_state_validate, work);
349 struct ipoib_dev_priv *priv = qp_work->priv;
350 struct ib_qp_attr qp_attr;
351 struct ib_qp_init_attr query_init_attr;
354 ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
356 ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
360 pr_info("%s: QP: 0x%x is in state: %d\n",
361 __func__, priv->qp->qp_num, qp_attr.qp_state);
363 /* currently support only in SQE->RTS transition*/
364 if (qp_attr.qp_state == IB_QPS_SQE) {
365 qp_attr.qp_state = IB_QPS_RTS;
367 ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
369 pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
370 ret, priv->qp->qp_num);
373 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
374 __func__, priv->qp->qp_num);
376 pr_warn("QP (%d) will stay in state: %d\n",
377 priv->qp->qp_num, qp_attr.qp_state);
384 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
386 struct ipoib_dev_priv *priv = ipoib_priv(dev);
387 unsigned int wr_id = wc->wr_id;
388 struct ipoib_tx_buf *tx_req;
390 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
393 if (unlikely(wr_id >= ipoib_sendq_size)) {
394 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
395 wr_id, ipoib_sendq_size);
399 tx_req = &priv->tx_ring[wr_id];
401 ipoib_dma_unmap_tx(priv, tx_req);
403 ++dev->stats.tx_packets;
404 dev->stats.tx_bytes += tx_req->skb->len;
406 dev_kfree_skb_any(tx_req->skb);
410 if (unlikely(netif_queue_stopped(dev) &&
411 ((priv->tx_head - priv->tx_tail) <= ipoib_sendq_size >> 1) &&
412 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
413 netif_wake_queue(dev);
415 if (wc->status != IB_WC_SUCCESS &&
416 wc->status != IB_WC_WR_FLUSH_ERR) {
417 struct ipoib_qp_state_validate *qp_work;
419 "failed send event (status=%d, wrid=%d vend_err %#x)\n",
420 wc->status, wr_id, wc->vendor_err);
421 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
425 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
426 qp_work->priv = priv;
427 queue_work(priv->wq, &qp_work->work);
431 static int poll_tx(struct ipoib_dev_priv *priv)
436 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
437 for (i = 0; i < n; ++i) {
438 wc = priv->send_wc + i;
439 if (wc->wr_id & IPOIB_OP_CM)
440 ipoib_cm_handle_tx_wc(priv->dev, priv->send_wc + i);
442 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
444 return n == MAX_SEND_CQE;
447 int ipoib_rx_poll(struct napi_struct *napi, int budget)
449 struct ipoib_dev_priv *priv =
450 container_of(napi, struct ipoib_dev_priv, recv_napi);
451 struct net_device *dev = priv->dev;
459 while (done < budget) {
460 int max = (budget - done);
462 t = min(IPOIB_NUM_WC, max);
463 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
465 for (i = 0; i < n; i++) {
466 struct ib_wc *wc = priv->ibwc + i;
468 if (wc->wr_id & IPOIB_OP_RECV) {
470 if (wc->wr_id & IPOIB_OP_CM)
471 ipoib_cm_handle_rx_wc(dev, wc);
473 ipoib_ib_handle_rx_wc(dev, wc);
475 pr_warn("%s: Got unexpected wqe id\n", __func__);
485 if (unlikely(ib_req_notify_cq(priv->recv_cq,
487 IB_CQ_REPORT_MISSED_EVENTS)) &&
488 napi_reschedule(napi))
495 int ipoib_tx_poll(struct napi_struct *napi, int budget)
497 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv,
499 struct net_device *dev = priv->dev;
504 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
506 for (i = 0; i < n; i++) {
507 wc = priv->send_wc + i;
508 if (wc->wr_id & IPOIB_OP_CM)
509 ipoib_cm_handle_tx_wc(dev, wc);
511 ipoib_ib_handle_tx_wc(dev, wc);
516 if (unlikely(ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
517 IB_CQ_REPORT_MISSED_EVENTS)) &&
518 napi_reschedule(napi))
521 return n < 0 ? 0 : n;
524 void ipoib_ib_rx_completion(struct ib_cq *cq, void *ctx_ptr)
526 struct ipoib_dev_priv *priv = ctx_ptr;
528 napi_schedule(&priv->recv_napi);
531 void ipoib_ib_tx_completion(struct ib_cq *cq, void *ctx_ptr)
533 struct ipoib_dev_priv *priv = ctx_ptr;
535 napi_schedule(&priv->send_napi);
538 static inline int post_send(struct ipoib_dev_priv *priv,
540 struct ib_ah *address, u32 dqpn,
541 struct ipoib_tx_buf *tx_req,
542 void *head, int hlen)
544 struct ib_send_wr *bad_wr;
545 struct sk_buff *skb = tx_req->skb;
547 ipoib_build_sge(priv, tx_req);
549 priv->tx_wr.wr.wr_id = wr_id;
550 priv->tx_wr.remote_qpn = dqpn;
551 priv->tx_wr.ah = address;
554 priv->tx_wr.mss = skb_shinfo(skb)->gso_size;
555 priv->tx_wr.header = head;
556 priv->tx_wr.hlen = hlen;
557 priv->tx_wr.wr.opcode = IB_WR_LSO;
559 priv->tx_wr.wr.opcode = IB_WR_SEND;
561 return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr);
564 int ipoib_send(struct net_device *dev, struct sk_buff *skb,
565 struct ib_ah *address, u32 dqpn)
567 struct ipoib_dev_priv *priv = ipoib_priv(dev);
568 struct ipoib_tx_buf *tx_req;
571 unsigned usable_sge = priv->max_send_sge - !!skb_headlen(skb);
573 if (skb_is_gso(skb)) {
574 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
576 if (unlikely(!skb_pull(skb, hlen))) {
577 ipoib_warn(priv, "linear data too small\n");
578 ++dev->stats.tx_dropped;
579 ++dev->stats.tx_errors;
580 dev_kfree_skb_any(skb);
584 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
585 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
586 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
587 ++dev->stats.tx_dropped;
588 ++dev->stats.tx_errors;
589 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
595 if (skb_shinfo(skb)->nr_frags > usable_sge) {
596 if (skb_linearize(skb) < 0) {
597 ipoib_warn(priv, "skb could not be linearized\n");
598 ++dev->stats.tx_dropped;
599 ++dev->stats.tx_errors;
600 dev_kfree_skb_any(skb);
603 /* Does skb_linearize return ok without reducing nr_frags? */
604 if (skb_shinfo(skb)->nr_frags > usable_sge) {
605 ipoib_warn(priv, "too many frags after skb linearize\n");
606 ++dev->stats.tx_dropped;
607 ++dev->stats.tx_errors;
608 dev_kfree_skb_any(skb);
614 "sending packet, length=%d address=%p dqpn=0x%06x\n",
615 skb->len, address, dqpn);
618 * We put the skb into the tx_ring _before_ we call post_send()
619 * because it's entirely possible that the completion handler will
620 * run before we execute anything after the post_send(). That
621 * means we have to make sure everything is properly recorded and
622 * our state is consistent before we call post_send().
624 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
626 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
627 ++dev->stats.tx_errors;
628 dev_kfree_skb_any(skb);
632 if (skb->ip_summed == CHECKSUM_PARTIAL)
633 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
635 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
636 /* increase the tx_head after send success, but use it for queue state */
637 if (priv->tx_head - priv->tx_tail == ipoib_sendq_size - 1) {
638 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
639 netif_stop_queue(dev);
645 if (netif_queue_stopped(dev))
646 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
647 IB_CQ_REPORT_MISSED_EVENTS))
648 ipoib_warn(priv, "request notify on send CQ failed\n");
650 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
651 address, dqpn, tx_req, phead, hlen);
653 ipoib_warn(priv, "post_send failed, error %d\n", rc);
654 ++dev->stats.tx_errors;
655 ipoib_dma_unmap_tx(priv, tx_req);
656 dev_kfree_skb_any(skb);
657 if (netif_queue_stopped(dev))
658 netif_wake_queue(dev);
661 netif_trans_update(dev);
669 static void __ipoib_reap_ah(struct net_device *dev)
671 struct ipoib_dev_priv *priv = ipoib_priv(dev);
672 struct ipoib_ah *ah, *tah;
673 LIST_HEAD(remove_list);
676 netif_tx_lock_bh(dev);
677 spin_lock_irqsave(&priv->lock, flags);
679 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
680 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
682 rdma_destroy_ah(ah->ah);
686 spin_unlock_irqrestore(&priv->lock, flags);
687 netif_tx_unlock_bh(dev);
690 void ipoib_reap_ah(struct work_struct *work)
692 struct ipoib_dev_priv *priv =
693 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
694 struct net_device *dev = priv->dev;
696 __ipoib_reap_ah(dev);
698 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
699 queue_delayed_work(priv->wq, &priv->ah_reap_task,
700 round_jiffies_relative(HZ));
703 static void ipoib_flush_ah(struct net_device *dev)
705 struct ipoib_dev_priv *priv = ipoib_priv(dev);
707 cancel_delayed_work(&priv->ah_reap_task);
708 flush_workqueue(priv->wq);
709 ipoib_reap_ah(&priv->ah_reap_task.work);
712 static void ipoib_stop_ah(struct net_device *dev)
714 struct ipoib_dev_priv *priv = ipoib_priv(dev);
716 set_bit(IPOIB_STOP_REAPER, &priv->flags);
720 static int recvs_pending(struct net_device *dev)
722 struct ipoib_dev_priv *priv = ipoib_priv(dev);
726 for (i = 0; i < ipoib_recvq_size; ++i)
727 if (priv->rx_ring[i].skb)
733 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,
735 enum ib_qp_state new_state)
737 struct ib_qp_attr qp_attr;
738 struct ib_qp_init_attr query_init_attr;
741 ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);
743 ipoib_warn(priv, "%s: Failed to query QP\n", __func__);
746 /* print according to the new-state and the previous state.*/
747 if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET)
748 ipoib_dbg(priv, "Failed modify QP, IB_QPS_RESET to IB_QPS_ERR, acceptable\n");
750 ipoib_warn(priv, "Failed to modify QP to state: %d from state: %d\n",
751 new_state, qp_attr.qp_state);
754 static void ipoib_napi_enable(struct net_device *dev)
756 struct ipoib_dev_priv *priv = ipoib_priv(dev);
758 napi_enable(&priv->recv_napi);
759 napi_enable(&priv->send_napi);
762 static void ipoib_napi_disable(struct net_device *dev)
764 struct ipoib_dev_priv *priv = ipoib_priv(dev);
766 napi_disable(&priv->recv_napi);
767 napi_disable(&priv->send_napi);
770 int ipoib_ib_dev_stop_default(struct net_device *dev)
772 struct ipoib_dev_priv *priv = ipoib_priv(dev);
773 struct ib_qp_attr qp_attr;
775 struct ipoib_tx_buf *tx_req;
778 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
779 ipoib_napi_disable(dev);
781 ipoib_cm_dev_stop(dev);
784 * Move our QP to the error state and then reinitialize in
785 * when all work requests have completed or have been flushed.
787 qp_attr.qp_state = IB_QPS_ERR;
788 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
789 check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);
791 /* Wait for all sends and receives to complete */
794 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
795 if (time_after(jiffies, begin + 5 * HZ)) {
797 "timing out; %d sends %d receives not completed\n",
798 priv->tx_head - priv->tx_tail,
802 * assume the HW is wedged and just free up
803 * all our pending work requests.
805 while ((int)priv->tx_tail - (int)priv->tx_head < 0) {
806 tx_req = &priv->tx_ring[priv->tx_tail &
807 (ipoib_sendq_size - 1)];
808 ipoib_dma_unmap_tx(priv, tx_req);
809 dev_kfree_skb_any(tx_req->skb);
813 for (i = 0; i < ipoib_recvq_size; ++i) {
814 struct ipoib_rx_buf *rx_req;
816 rx_req = &priv->rx_ring[i];
819 ipoib_ud_dma_unmap_rx(priv,
820 priv->rx_ring[i].mapping);
821 dev_kfree_skb_any(rx_req->skb);
830 usleep_range(1000, 2000);
833 ipoib_dbg(priv, "All sends and receives done.\n");
836 qp_attr.qp_state = IB_QPS_RESET;
837 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
838 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
840 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
845 int ipoib_ib_dev_stop(struct net_device *dev)
847 struct ipoib_dev_priv *priv = ipoib_priv(dev);
849 priv->rn_ops->ndo_stop(dev);
851 clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
857 int ipoib_ib_dev_open_default(struct net_device *dev)
859 struct ipoib_dev_priv *priv = ipoib_priv(dev);
862 ret = ipoib_init_qp(dev);
864 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
868 ret = ipoib_ib_post_receives(dev);
870 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
874 ret = ipoib_cm_dev_open(dev);
876 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
880 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
881 ipoib_napi_enable(dev);
888 int ipoib_ib_dev_open(struct net_device *dev)
890 struct ipoib_dev_priv *priv = ipoib_priv(dev);
892 ipoib_pkey_dev_check_presence(dev);
894 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
895 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
896 (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
900 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
901 queue_delayed_work(priv->wq, &priv->ah_reap_task,
902 round_jiffies_relative(HZ));
904 if (priv->rn_ops->ndo_open(dev)) {
905 pr_warn("%s: Failed to open dev\n", dev->name);
909 set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
914 set_bit(IPOIB_STOP_REAPER, &priv->flags);
915 cancel_delayed_work(&priv->ah_reap_task);
916 set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
917 ipoib_ib_dev_stop(dev);
921 void ipoib_pkey_dev_check_presence(struct net_device *dev)
923 struct ipoib_dev_priv *priv = ipoib_priv(dev);
924 struct rdma_netdev *rn = netdev_priv(dev);
926 if (!(priv->pkey & 0x7fff) ||
927 ib_find_pkey(priv->ca, priv->port, priv->pkey,
928 &priv->pkey_index)) {
929 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
932 rn->set_id(dev, priv->pkey_index);
933 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
937 void ipoib_ib_dev_up(struct net_device *dev)
939 struct ipoib_dev_priv *priv = ipoib_priv(dev);
941 ipoib_pkey_dev_check_presence(dev);
943 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
944 ipoib_dbg(priv, "PKEY is not assigned.\n");
948 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
950 ipoib_mcast_start_thread(dev);
953 void ipoib_ib_dev_down(struct net_device *dev)
955 struct ipoib_dev_priv *priv = ipoib_priv(dev);
957 ipoib_dbg(priv, "downing ib_dev\n");
959 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
960 netif_carrier_off(dev);
962 ipoib_mcast_stop_thread(dev);
963 ipoib_mcast_dev_flush(dev);
965 ipoib_flush_paths(dev);
968 void ipoib_drain_cq(struct net_device *dev)
970 struct ipoib_dev_priv *priv = ipoib_priv(dev);
974 * We call completion handling routines that expect to be
975 * called from the BH-disabled NAPI poll context, so disable
981 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
982 for (i = 0; i < n; ++i) {
984 * Convert any successful completions to flush
985 * errors to avoid passing packets up the
986 * stack after bringing the device down.
988 if (priv->ibwc[i].status == IB_WC_SUCCESS)
989 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
991 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
992 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
993 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
995 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
997 pr_warn("%s: Got unexpected wqe id\n", __func__);
1000 } while (n == IPOIB_NUM_WC);
1002 while (poll_tx(priv))
1009 * Takes whatever value which is in pkey index 0 and updates priv->pkey
1010 * returns 0 if the pkey value was changed.
1012 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
1017 prev_pkey = priv->pkey;
1018 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1020 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
1021 priv->port, result);
1025 priv->pkey |= 0x8000;
1027 if (prev_pkey != priv->pkey) {
1028 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
1029 prev_pkey, priv->pkey);
1031 * Update the pkey in the broadcast address, while making sure to set
1032 * the full membership bit, so that we join the right broadcast group.
1034 priv->dev->broadcast[8] = priv->pkey >> 8;
1035 priv->dev->broadcast[9] = priv->pkey & 0xff;
1042 * returns 0 if pkey value was found in a different slot.
1044 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
1046 u16 old_index = priv->pkey_index;
1048 priv->pkey_index = 0;
1049 ipoib_pkey_dev_check_presence(priv->dev);
1051 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
1052 (old_index == priv->pkey_index))
1058 * returns true if the device address of the ipoib interface has changed and the
1059 * new address is a valid one (i.e in the gid table), return false otherwise.
1061 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)
1063 union ib_gid search_gid;
1065 union ib_gid *netdev_gid;
1071 netdev_gid = (union ib_gid *)(priv->dev->dev_addr + 4);
1072 if (ib_query_gid(priv->ca, priv->port, 0, &gid0, NULL))
1075 netif_addr_lock_bh(priv->dev);
1077 /* The subnet prefix may have changed, update it now so we won't have
1080 priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1081 netdev_gid->global.subnet_prefix = gid0.global.subnet_prefix;
1082 search_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1084 search_gid.global.interface_id = priv->local_gid.global.interface_id;
1086 netif_addr_unlock_bh(priv->dev);
1088 err = ib_find_gid(priv->ca, &search_gid, IB_GID_TYPE_IB,
1089 priv->dev, &port, &index);
1091 netif_addr_lock_bh(priv->dev);
1093 if (search_gid.global.interface_id !=
1094 priv->local_gid.global.interface_id)
1095 /* There was a change while we were looking up the gid, bail
1096 * here and let the next work sort this out
1100 /* The next section of code needs some background:
1101 * Per IB spec the port GUID can't change if the HCA is powered on.
1102 * port GUID is the basis for GID at index 0 which is the basis for
1103 * the default device address of a ipoib interface.
1105 * so it seems the flow should be:
1106 * if user_changed_dev_addr && gid in gid tbl
1107 * set bit dev_addr_set
1112 * The issue is that there are devices that don't follow the spec,
1113 * they change the port GUID when the HCA is powered, so in order
1114 * not to break userspace applications, We need to check if the
1115 * user wanted to control the device address and we assume that
1116 * if he sets the device address back to be based on GID index 0,
1117 * he no longer wishs to control it.
1119 * If the user doesn't control the the device address,
1120 * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1121 * the port GUID has changed and GID at index 0 has changed
1122 * so we need to change priv->local_gid and priv->dev->dev_addr
1123 * to reflect the new GID.
1125 if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
1126 if (!err && port == priv->port) {
1127 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1129 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,
1132 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);
1138 if (!err && port == priv->port) {
1141 if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {
1142 memcpy(&priv->local_gid, &gid0,
1143 sizeof(priv->local_gid));
1144 memcpy(priv->dev->dev_addr + 4, &gid0,
1145 sizeof(priv->local_gid));
1152 netif_addr_unlock_bh(priv->dev);
1157 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1158 enum ipoib_flush_level level,
1161 struct ipoib_dev_priv *cpriv;
1162 struct net_device *dev = priv->dev;
1165 down_read_nested(&priv->vlan_rwsem, nesting);
1168 * Flush any child interfaces too -- they might be up even if
1169 * the parent is down.
1171 list_for_each_entry(cpriv, &priv->child_intfs, list)
1172 __ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1174 up_read(&priv->vlan_rwsem);
1176 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1177 level != IPOIB_FLUSH_HEAVY) {
1178 /* Make sure the dev_addr is set even if not flushing */
1179 if (level == IPOIB_FLUSH_LIGHT)
1180 ipoib_dev_addr_changed_valid(priv);
1181 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1185 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1186 /* interface is down. update pkey and leave. */
1187 if (level == IPOIB_FLUSH_HEAVY) {
1188 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1189 update_parent_pkey(priv);
1191 update_child_pkey(priv);
1192 } else if (level == IPOIB_FLUSH_LIGHT)
1193 ipoib_dev_addr_changed_valid(priv);
1194 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1198 if (level == IPOIB_FLUSH_HEAVY) {
1199 /* child devices chase their origin pkey value, while non-child
1200 * (parent) devices should always takes what present in pkey index 0
1202 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1203 result = update_child_pkey(priv);
1205 /* restart QP only if P_Key index is changed */
1206 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1211 result = update_parent_pkey(priv);
1212 /* restart QP only if P_Key value changed */
1214 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1220 if (level == IPOIB_FLUSH_LIGHT) {
1222 ipoib_mark_paths_invalid(dev);
1223 /* Set IPoIB operation as down to prevent races between:
1224 * the flush flow which leaves MCG and on the fly joins
1225 * which can happen during that time. mcast restart task
1226 * should deal with join requests we missed.
1228 oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1229 ipoib_mcast_dev_flush(dev);
1231 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1232 ipoib_flush_ah(dev);
1235 if (level >= IPOIB_FLUSH_NORMAL)
1236 ipoib_ib_dev_down(dev);
1238 if (level == IPOIB_FLUSH_HEAVY) {
1239 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1240 ipoib_ib_dev_stop(dev);
1242 if (ipoib_ib_dev_open(dev))
1245 if (netif_queue_stopped(dev))
1246 netif_start_queue(dev);
1250 * The device could have been brought down between the start and when
1251 * we get here, don't bring it back up if it's not configured up
1253 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1254 if (level >= IPOIB_FLUSH_NORMAL)
1255 ipoib_ib_dev_up(dev);
1256 if (ipoib_dev_addr_changed_valid(priv))
1257 ipoib_mcast_restart_task(&priv->restart_task);
1261 void ipoib_ib_dev_flush_light(struct work_struct *work)
1263 struct ipoib_dev_priv *priv =
1264 container_of(work, struct ipoib_dev_priv, flush_light);
1266 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1269 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1271 struct ipoib_dev_priv *priv =
1272 container_of(work, struct ipoib_dev_priv, flush_normal);
1274 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1277 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1279 struct ipoib_dev_priv *priv =
1280 container_of(work, struct ipoib_dev_priv, flush_heavy);
1283 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1287 void ipoib_ib_dev_cleanup(struct net_device *dev)
1289 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1291 ipoib_dbg(priv, "cleaning up ib_dev\n");
1293 * We must make sure there are no more (path) completions
1294 * that may wish to touch priv fields that are no longer valid
1296 ipoib_flush_paths(dev);
1298 ipoib_mcast_stop_thread(dev);
1299 ipoib_mcast_dev_flush(dev);
1302 * All of our ah references aren't free until after
1303 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1304 * the neighbor garbage collection is stopped and reaped.
1305 * That should all be done now, so make a final ah flush.
1309 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
1311 priv->rn_ops->ndo_uninit(dev);
1314 ib_dealloc_pd(priv->pd);