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>
43 #include <rdma/ib_cache.h>
47 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
48 static int data_debug_level;
50 module_param(data_debug_level, int, 0644);
51 MODULE_PARM_DESC(data_debug_level,
52 "Enable data path debug tracing if > 0");
55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
56 struct ib_pd *pd, struct rdma_ah_attr *attr)
61 ah = kmalloc(sizeof(*ah), GFP_KERNEL);
63 return ERR_PTR(-ENOMEM);
69 vah = rdma_create_ah(pd, attr, RDMA_CREATE_AH_SLEEPABLE);
72 ah = (struct ipoib_ah *)vah;
75 ipoib_dbg(ipoib_priv(dev), "Created ah %p\n", ah->ah);
81 void ipoib_free_ah(struct kref *kref)
83 struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);
84 struct ipoib_dev_priv *priv = ipoib_priv(ah->dev);
88 spin_lock_irqsave(&priv->lock, flags);
89 list_add_tail(&ah->list, &priv->dead_ahs);
90 spin_unlock_irqrestore(&priv->lock, flags);
93 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,
94 u64 mapping[IPOIB_UD_RX_SG])
96 ib_dma_unmap_single(priv->ca, mapping[0],
97 IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),
101 static int ipoib_ib_post_receive(struct net_device *dev, int id)
103 struct ipoib_dev_priv *priv = ipoib_priv(dev);
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, NULL);
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,
299 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))
306 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
308 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);
312 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);
317 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,
318 struct ipoib_tx_buf *tx_req)
320 struct sk_buff *skb = tx_req->skb;
321 u64 *mapping = tx_req->mapping;
325 if (skb_headlen(skb)) {
326 ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),
332 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
333 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
335 ib_dma_unmap_page(priv->ca, mapping[i + off],
336 skb_frag_size(frag), DMA_TO_DEVICE);
341 * As the result of a completion error the QP Can be transferred to SQE states.
342 * The function checks if the (send)QP is in SQE state and
343 * moves it back to RTS state, that in order to have it functional again.
345 static void ipoib_qp_state_validate_work(struct work_struct *work)
347 struct ipoib_qp_state_validate *qp_work =
348 container_of(work, struct ipoib_qp_state_validate, work);
350 struct ipoib_dev_priv *priv = qp_work->priv;
351 struct ib_qp_attr qp_attr;
352 struct ib_qp_init_attr query_init_attr;
355 ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);
357 ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",
361 pr_info("%s: QP: 0x%x is in state: %d\n",
362 __func__, priv->qp->qp_num, qp_attr.qp_state);
364 /* currently support only in SQE->RTS transition*/
365 if (qp_attr.qp_state == IB_QPS_SQE) {
366 qp_attr.qp_state = IB_QPS_RTS;
368 ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);
370 pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",
371 ret, priv->qp->qp_num);
374 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",
375 __func__, priv->qp->qp_num);
377 pr_warn("QP (%d) will stay in state: %d\n",
378 priv->qp->qp_num, qp_attr.qp_state);
385 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
387 struct ipoib_dev_priv *priv = ipoib_priv(dev);
388 unsigned int wr_id = wc->wr_id;
389 struct ipoib_tx_buf *tx_req;
391 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",
394 if (unlikely(wr_id >= ipoib_sendq_size)) {
395 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",
396 wr_id, ipoib_sendq_size);
400 tx_req = &priv->tx_ring[wr_id];
402 ipoib_dma_unmap_tx(priv, tx_req);
404 ++dev->stats.tx_packets;
405 dev->stats.tx_bytes += tx_req->skb->len;
407 dev_kfree_skb_any(tx_req->skb);
410 ++priv->global_tx_tail;
412 if (unlikely(netif_queue_stopped(dev) &&
413 ((priv->global_tx_head - priv->global_tx_tail) <=
414 ipoib_sendq_size >> 1) &&
415 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
416 netif_wake_queue(dev);
418 if (wc->status != IB_WC_SUCCESS &&
419 wc->status != IB_WC_WR_FLUSH_ERR) {
420 struct ipoib_qp_state_validate *qp_work;
422 "failed send event (status=%d, wrid=%d vend_err %#x)\n",
423 wc->status, wr_id, wc->vendor_err);
424 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);
428 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);
429 qp_work->priv = priv;
430 queue_work(priv->wq, &qp_work->work);
434 static int poll_tx(struct ipoib_dev_priv *priv)
439 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
440 for (i = 0; i < n; ++i) {
441 wc = priv->send_wc + i;
442 if (wc->wr_id & IPOIB_OP_CM)
443 ipoib_cm_handle_tx_wc(priv->dev, priv->send_wc + i);
445 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);
447 return n == MAX_SEND_CQE;
450 int ipoib_rx_poll(struct napi_struct *napi, int budget)
452 struct ipoib_dev_priv *priv =
453 container_of(napi, struct ipoib_dev_priv, recv_napi);
454 struct net_device *dev = priv->dev;
462 while (done < budget) {
463 int max = (budget - done);
465 t = min(IPOIB_NUM_WC, max);
466 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);
468 for (i = 0; i < n; i++) {
469 struct ib_wc *wc = priv->ibwc + i;
471 if (wc->wr_id & IPOIB_OP_RECV) {
473 if (wc->wr_id & IPOIB_OP_CM)
474 ipoib_cm_handle_rx_wc(dev, wc);
476 ipoib_ib_handle_rx_wc(dev, wc);
478 pr_warn("%s: Got unexpected wqe id\n", __func__);
488 if (unlikely(ib_req_notify_cq(priv->recv_cq,
490 IB_CQ_REPORT_MISSED_EVENTS)) &&
498 int ipoib_tx_poll(struct napi_struct *napi, int budget)
500 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv,
502 struct net_device *dev = priv->dev;
507 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);
509 for (i = 0; i < n; i++) {
510 wc = priv->send_wc + i;
511 if (wc->wr_id & IPOIB_OP_CM)
512 ipoib_cm_handle_tx_wc(dev, wc);
514 ipoib_ib_handle_tx_wc(dev, wc);
519 if (unlikely(ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
520 IB_CQ_REPORT_MISSED_EVENTS)) &&
524 return n < 0 ? 0 : n;
527 void ipoib_ib_rx_completion(struct ib_cq *cq, void *ctx_ptr)
529 struct ipoib_dev_priv *priv = ctx_ptr;
531 napi_schedule(&priv->recv_napi);
534 void ipoib_ib_tx_completion(struct ib_cq *cq, void *ctx_ptr)
536 struct ipoib_dev_priv *priv = ctx_ptr;
538 napi_schedule(&priv->send_napi);
541 static inline int post_send(struct ipoib_dev_priv *priv,
543 struct ib_ah *address, u32 dqpn,
544 struct ipoib_tx_buf *tx_req,
545 void *head, int hlen)
547 struct sk_buff *skb = tx_req->skb;
549 ipoib_build_sge(priv, tx_req);
551 priv->tx_wr.wr.wr_id = wr_id;
552 priv->tx_wr.remote_qpn = dqpn;
553 priv->tx_wr.ah = address;
556 priv->tx_wr.mss = skb_shinfo(skb)->gso_size;
557 priv->tx_wr.header = head;
558 priv->tx_wr.hlen = hlen;
559 priv->tx_wr.wr.opcode = IB_WR_LSO;
561 priv->tx_wr.wr.opcode = IB_WR_SEND;
563 return ib_post_send(priv->qp, &priv->tx_wr.wr, NULL);
566 int ipoib_send(struct net_device *dev, struct sk_buff *skb,
567 struct ib_ah *address, u32 dqpn)
569 struct ipoib_dev_priv *priv = ipoib_priv(dev);
570 struct ipoib_tx_buf *tx_req;
573 unsigned int usable_sge = priv->max_send_sge - !!skb_headlen(skb);
575 if (skb_is_gso(skb)) {
576 hlen = skb_tcp_all_headers(skb);
578 if (unlikely(!skb_pull(skb, hlen))) {
579 ipoib_warn(priv, "linear data too small\n");
580 ++dev->stats.tx_dropped;
581 ++dev->stats.tx_errors;
582 dev_kfree_skb_any(skb);
586 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {
587 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
588 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);
589 ++dev->stats.tx_dropped;
590 ++dev->stats.tx_errors;
591 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);
597 if (skb_shinfo(skb)->nr_frags > usable_sge) {
598 if (skb_linearize(skb) < 0) {
599 ipoib_warn(priv, "skb could not be linearized\n");
600 ++dev->stats.tx_dropped;
601 ++dev->stats.tx_errors;
602 dev_kfree_skb_any(skb);
605 /* Does skb_linearize return ok without reducing nr_frags? */
606 if (skb_shinfo(skb)->nr_frags > usable_sge) {
607 ipoib_warn(priv, "too many frags after skb linearize\n");
608 ++dev->stats.tx_dropped;
609 ++dev->stats.tx_errors;
610 dev_kfree_skb_any(skb);
616 "sending packet, length=%d address=%p dqpn=0x%06x\n",
617 skb->len, address, dqpn);
620 * We put the skb into the tx_ring _before_ we call post_send()
621 * because it's entirely possible that the completion handler will
622 * run before we execute anything after the post_send(). That
623 * means we have to make sure everything is properly recorded and
624 * our state is consistent before we call post_send().
626 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
628 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
629 ++dev->stats.tx_errors;
630 dev_kfree_skb_any(skb);
634 if (skb->ip_summed == CHECKSUM_PARTIAL)
635 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;
637 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
638 /* increase the tx_head after send success, but use it for queue state */
639 if ((priv->global_tx_head - priv->global_tx_tail) ==
640 ipoib_sendq_size - 1) {
641 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
642 netif_stop_queue(dev);
648 if (netif_queue_stopped(dev))
649 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
650 IB_CQ_REPORT_MISSED_EVENTS) < 0)
651 ipoib_warn(priv, "request notify on send CQ failed\n");
653 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
654 address, dqpn, tx_req, phead, hlen);
656 ipoib_warn(priv, "post_send failed, error %d\n", rc);
657 ++dev->stats.tx_errors;
658 ipoib_dma_unmap_tx(priv, tx_req);
659 dev_kfree_skb_any(skb);
660 if (netif_queue_stopped(dev))
661 netif_wake_queue(dev);
664 netif_trans_update(dev);
668 ++priv->global_tx_head;
673 static void ipoib_reap_dead_ahs(struct ipoib_dev_priv *priv)
675 struct ipoib_ah *ah, *tah;
678 netif_tx_lock_bh(priv->dev);
679 spin_lock_irqsave(&priv->lock, flags);
681 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
682 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
684 rdma_destroy_ah(ah->ah, 0);
688 spin_unlock_irqrestore(&priv->lock, flags);
689 netif_tx_unlock_bh(priv->dev);
692 void ipoib_reap_ah(struct work_struct *work)
694 struct ipoib_dev_priv *priv =
695 container_of(work, struct ipoib_dev_priv, ah_reap_task.work);
697 ipoib_reap_dead_ahs(priv);
699 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
700 queue_delayed_work(priv->wq, &priv->ah_reap_task,
701 round_jiffies_relative(HZ));
704 static void ipoib_start_ah_reaper(struct ipoib_dev_priv *priv)
706 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
707 queue_delayed_work(priv->wq, &priv->ah_reap_task,
708 round_jiffies_relative(HZ));
711 static void ipoib_stop_ah_reaper(struct ipoib_dev_priv *priv)
713 set_bit(IPOIB_STOP_REAPER, &priv->flags);
714 cancel_delayed_work(&priv->ah_reap_task);
716 * After ipoib_stop_ah_reaper() we always go through
717 * ipoib_reap_dead_ahs() which ensures the work is really stopped and
718 * does a final flush out of the dead_ah's list
722 static int recvs_pending(struct net_device *dev)
724 struct ipoib_dev_priv *priv = ipoib_priv(dev);
728 for (i = 0; i < ipoib_recvq_size; ++i)
729 if (priv->rx_ring[i].skb)
735 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,
737 enum ib_qp_state new_state)
739 struct ib_qp_attr qp_attr;
740 struct ib_qp_init_attr query_init_attr;
743 ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);
745 ipoib_warn(priv, "%s: Failed to query QP\n", __func__);
748 /* print according to the new-state and the previous state.*/
749 if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET)
750 ipoib_dbg(priv, "Failed modify QP, IB_QPS_RESET to IB_QPS_ERR, acceptable\n");
752 ipoib_warn(priv, "Failed to modify QP to state: %d from state: %d\n",
753 new_state, qp_attr.qp_state);
756 static void ipoib_napi_enable(struct net_device *dev)
758 struct ipoib_dev_priv *priv = ipoib_priv(dev);
760 napi_enable(&priv->recv_napi);
761 napi_enable(&priv->send_napi);
764 static void ipoib_napi_disable(struct net_device *dev)
766 struct ipoib_dev_priv *priv = ipoib_priv(dev);
768 napi_disable(&priv->recv_napi);
769 napi_disable(&priv->send_napi);
772 int ipoib_ib_dev_stop_default(struct net_device *dev)
774 struct ipoib_dev_priv *priv = ipoib_priv(dev);
775 struct ib_qp_attr qp_attr;
777 struct ipoib_tx_buf *tx_req;
780 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
781 ipoib_napi_disable(dev);
783 ipoib_cm_dev_stop(dev);
786 * Move our QP to the error state and then reinitialize in
787 * when all work requests have completed or have been flushed.
789 qp_attr.qp_state = IB_QPS_ERR;
790 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
791 check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);
793 /* Wait for all sends and receives to complete */
796 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
797 if (time_after(jiffies, begin + 5 * HZ)) {
799 "timing out; %d sends %d receives not completed\n",
800 priv->tx_head - priv->tx_tail,
804 * assume the HW is wedged and just free up
805 * all our pending work requests.
807 while ((int)priv->tx_tail - (int)priv->tx_head < 0) {
808 tx_req = &priv->tx_ring[priv->tx_tail &
809 (ipoib_sendq_size - 1)];
810 ipoib_dma_unmap_tx(priv, tx_req);
811 dev_kfree_skb_any(tx_req->skb);
813 ++priv->global_tx_tail;
816 for (i = 0; i < ipoib_recvq_size; ++i) {
817 struct ipoib_rx_buf *rx_req;
819 rx_req = &priv->rx_ring[i];
822 ipoib_ud_dma_unmap_rx(priv,
823 priv->rx_ring[i].mapping);
824 dev_kfree_skb_any(rx_req->skb);
833 usleep_range(1000, 2000);
836 ipoib_dbg(priv, "All sends and receives done.\n");
839 qp_attr.qp_state = IB_QPS_RESET;
840 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
841 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
843 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);
848 int ipoib_ib_dev_open_default(struct net_device *dev)
850 struct ipoib_dev_priv *priv = ipoib_priv(dev);
853 ret = ipoib_init_qp(dev);
855 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
859 ret = ipoib_ib_post_receives(dev);
861 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
865 ret = ipoib_cm_dev_open(dev);
867 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);
871 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
872 ipoib_napi_enable(dev);
879 int ipoib_ib_dev_open(struct net_device *dev)
881 struct ipoib_dev_priv *priv = ipoib_priv(dev);
883 ipoib_pkey_dev_check_presence(dev);
885 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
886 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,
887 (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));
891 ipoib_start_ah_reaper(priv);
892 if (priv->rn_ops->ndo_open(dev)) {
893 pr_warn("%s: Failed to open dev\n", dev->name);
897 set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
902 ipoib_stop_ah_reaper(priv);
906 void ipoib_ib_dev_stop(struct net_device *dev)
908 struct ipoib_dev_priv *priv = ipoib_priv(dev);
910 priv->rn_ops->ndo_stop(dev);
912 clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
913 ipoib_stop_ah_reaper(priv);
916 void ipoib_pkey_dev_check_presence(struct net_device *dev)
918 struct ipoib_dev_priv *priv = ipoib_priv(dev);
919 struct rdma_netdev *rn = netdev_priv(dev);
921 if (!(priv->pkey & 0x7fff) ||
922 ib_find_pkey(priv->ca, priv->port, priv->pkey,
923 &priv->pkey_index)) {
924 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
927 rn->set_id(dev, priv->pkey_index);
928 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
932 void ipoib_ib_dev_up(struct net_device *dev)
934 struct ipoib_dev_priv *priv = ipoib_priv(dev);
936 ipoib_pkey_dev_check_presence(dev);
938 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
939 ipoib_dbg(priv, "PKEY is not assigned.\n");
943 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
945 ipoib_mcast_start_thread(dev);
948 void ipoib_ib_dev_down(struct net_device *dev)
950 struct ipoib_dev_priv *priv = ipoib_priv(dev);
952 ipoib_dbg(priv, "downing ib_dev\n");
954 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
955 netif_carrier_off(dev);
957 ipoib_mcast_stop_thread(dev);
958 ipoib_mcast_dev_flush(dev);
960 ipoib_flush_paths(dev);
963 void ipoib_drain_cq(struct net_device *dev)
965 struct ipoib_dev_priv *priv = ipoib_priv(dev);
969 * We call completion handling routines that expect to be
970 * called from the BH-disabled NAPI poll context, so disable
976 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);
977 for (i = 0; i < n; ++i) {
979 * Convert any successful completions to flush
980 * errors to avoid passing packets up the
981 * stack after bringing the device down.
983 if (priv->ibwc[i].status == IB_WC_SUCCESS)
984 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;
986 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {
987 if (priv->ibwc[i].wr_id & IPOIB_OP_CM)
988 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);
990 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);
992 pr_warn("%s: Got unexpected wqe id\n", __func__);
995 } while (n == IPOIB_NUM_WC);
997 while (poll_tx(priv))
1004 * Takes whatever value which is in pkey index 0 and updates priv->pkey
1005 * returns 0 if the pkey value was changed.
1007 static inline int update_parent_pkey(struct ipoib_dev_priv *priv)
1012 prev_pkey = priv->pkey;
1013 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1015 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",
1016 priv->port, result);
1020 priv->pkey |= 0x8000;
1022 if (prev_pkey != priv->pkey) {
1023 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",
1024 prev_pkey, priv->pkey);
1026 * Update the pkey in the broadcast address, while making sure to set
1027 * the full membership bit, so that we join the right broadcast group.
1029 priv->dev->broadcast[8] = priv->pkey >> 8;
1030 priv->dev->broadcast[9] = priv->pkey & 0xff;
1037 * returns 0 if pkey value was found in a different slot.
1039 static inline int update_child_pkey(struct ipoib_dev_priv *priv)
1041 u16 old_index = priv->pkey_index;
1043 priv->pkey_index = 0;
1044 ipoib_pkey_dev_check_presence(priv->dev);
1046 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&
1047 (old_index == priv->pkey_index))
1053 * returns true if the device address of the ipoib interface has changed and the
1054 * new address is a valid one (i.e in the gid table), return false otherwise.
1056 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)
1058 union ib_gid search_gid;
1065 if (rdma_query_gid(priv->ca, priv->port, 0, &gid0))
1068 netif_addr_lock_bh(priv->dev);
1070 /* The subnet prefix may have changed, update it now so we won't have
1073 priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1074 dev_addr_mod(priv->dev, 4, (u8 *)&gid0.global.subnet_prefix,
1075 sizeof(gid0.global.subnet_prefix));
1076 search_gid.global.subnet_prefix = gid0.global.subnet_prefix;
1078 search_gid.global.interface_id = priv->local_gid.global.interface_id;
1080 netif_addr_unlock_bh(priv->dev);
1082 err = ib_find_gid(priv->ca, &search_gid, &port, &index);
1084 netif_addr_lock_bh(priv->dev);
1086 if (search_gid.global.interface_id !=
1087 priv->local_gid.global.interface_id)
1088 /* There was a change while we were looking up the gid, bail
1089 * here and let the next work sort this out
1093 /* The next section of code needs some background:
1094 * Per IB spec the port GUID can't change if the HCA is powered on.
1095 * port GUID is the basis for GID at index 0 which is the basis for
1096 * the default device address of a ipoib interface.
1098 * so it seems the flow should be:
1099 * if user_changed_dev_addr && gid in gid tbl
1100 * set bit dev_addr_set
1105 * The issue is that there are devices that don't follow the spec,
1106 * they change the port GUID when the HCA is powered, so in order
1107 * not to break userspace applications, We need to check if the
1108 * user wanted to control the device address and we assume that
1109 * if he sets the device address back to be based on GID index 0,
1110 * he no longer wishs to control it.
1112 * If the user doesn't control the device address,
1113 * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means
1114 * the port GUID has changed and GID at index 0 has changed
1115 * so we need to change priv->local_gid and priv->dev->dev_addr
1116 * to reflect the new GID.
1118 if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
1119 if (!err && port == priv->port) {
1120 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1122 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,
1125 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);
1131 if (!err && port == priv->port) {
1134 if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {
1135 memcpy(&priv->local_gid, &gid0,
1136 sizeof(priv->local_gid));
1137 dev_addr_mod(priv->dev, 4, (u8 *)&gid0,
1138 sizeof(priv->local_gid));
1145 netif_addr_unlock_bh(priv->dev);
1150 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
1151 enum ipoib_flush_level level,
1154 struct ipoib_dev_priv *cpriv;
1155 struct net_device *dev = priv->dev;
1158 down_read_nested(&priv->vlan_rwsem, nesting);
1161 * Flush any child interfaces too -- they might be up even if
1162 * the parent is down.
1164 list_for_each_entry(cpriv, &priv->child_intfs, list)
1165 __ipoib_ib_dev_flush(cpriv, level, nesting + 1);
1167 up_read(&priv->vlan_rwsem);
1169 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&
1170 level != IPOIB_FLUSH_HEAVY) {
1171 /* Make sure the dev_addr is set even if not flushing */
1172 if (level == IPOIB_FLUSH_LIGHT)
1173 ipoib_dev_addr_changed_valid(priv);
1174 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
1178 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1179 /* interface is down. update pkey and leave. */
1180 if (level == IPOIB_FLUSH_HEAVY) {
1181 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1182 update_parent_pkey(priv);
1184 update_child_pkey(priv);
1185 } else if (level == IPOIB_FLUSH_LIGHT)
1186 ipoib_dev_addr_changed_valid(priv);
1187 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
1191 if (level == IPOIB_FLUSH_HEAVY) {
1192 /* child devices chase their origin pkey value, while non-child
1193 * (parent) devices should always takes what present in pkey index 0
1195 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
1196 result = update_child_pkey(priv);
1198 /* restart QP only if P_Key index is changed */
1199 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");
1204 result = update_parent_pkey(priv);
1205 /* restart QP only if P_Key value changed */
1207 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");
1213 if (level == IPOIB_FLUSH_LIGHT) {
1215 ipoib_mark_paths_invalid(dev);
1216 /* Set IPoIB operation as down to prevent races between:
1217 * the flush flow which leaves MCG and on the fly joins
1218 * which can happen during that time. mcast restart task
1219 * should deal with join requests we missed.
1221 oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1222 ipoib_mcast_dev_flush(dev);
1224 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
1225 ipoib_reap_dead_ahs(priv);
1228 if (level >= IPOIB_FLUSH_NORMAL)
1229 ipoib_ib_dev_down(dev);
1231 if (level == IPOIB_FLUSH_HEAVY) {
1232 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))
1233 ipoib_ib_dev_stop(dev);
1235 if (ipoib_ib_dev_open(dev))
1238 if (netif_queue_stopped(dev))
1239 netif_start_queue(dev);
1243 * The device could have been brought down between the start and when
1244 * we get here, don't bring it back up if it's not configured up
1246 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
1247 if (level >= IPOIB_FLUSH_NORMAL)
1248 ipoib_ib_dev_up(dev);
1249 if (ipoib_dev_addr_changed_valid(priv))
1250 ipoib_mcast_restart_task(&priv->restart_task);
1254 void ipoib_ib_dev_flush_light(struct work_struct *work)
1256 struct ipoib_dev_priv *priv =
1257 container_of(work, struct ipoib_dev_priv, flush_light);
1259 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);
1262 void ipoib_ib_dev_flush_normal(struct work_struct *work)
1264 struct ipoib_dev_priv *priv =
1265 container_of(work, struct ipoib_dev_priv, flush_normal);
1267 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);
1270 void ipoib_ib_dev_flush_heavy(struct work_struct *work)
1272 struct ipoib_dev_priv *priv =
1273 container_of(work, struct ipoib_dev_priv, flush_heavy);
1276 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);
1280 void ipoib_ib_dev_cleanup(struct net_device *dev)
1282 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1284 ipoib_dbg(priv, "cleaning up ib_dev\n");
1286 * We must make sure there are no more (path) completions
1287 * that may wish to touch priv fields that are no longer valid
1289 ipoib_flush_paths(dev);
1291 ipoib_mcast_stop_thread(dev);
1292 ipoib_mcast_dev_flush(dev);
1295 * All of our ah references aren't free until after
1296 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and
1297 * the neighbor garbage collection is stopped and reaped.
1298 * That should all be done now, so make a final ah flush.
1300 ipoib_reap_dead_ahs(priv);
1302 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
1304 priv->rn_ops->ndo_uninit(dev);
1307 ib_dealloc_pd(priv->pd);