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
35 * $Id: ipoib_ib.c 1386 2004-12-27 16:23:17Z roland $
38 #include <linux/delay.h>
39 #include <linux/dma-mapping.h>
41 #include <rdma/ib_cache.h>
45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
46 static int data_debug_level;
48 module_param(data_debug_level, int, 0644);
49 MODULE_PARM_DESC(data_debug_level,
50 "Enable data path debug tracing if > 0");
53 #define IPOIB_OP_RECV (1ul << 31)
55 static DEFINE_MUTEX(pkey_mutex);
57 struct ipoib_ah *ipoib_create_ah(struct net_device *dev,
58 struct ib_pd *pd, struct ib_ah_attr *attr)
62 ah = kmalloc(sizeof *ah, GFP_KERNEL);
70 ah->ah = ib_create_ah(pd, attr);
75 ipoib_dbg(netdev_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 = netdev_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 int ipoib_ib_post_receive(struct net_device *dev, int id)
94 struct ipoib_dev_priv *priv = netdev_priv(dev);
96 struct ib_recv_wr param;
97 struct ib_recv_wr *bad_wr;
100 list.addr = priv->rx_ring[id].mapping;
101 list.length = IPOIB_BUF_SIZE;
102 list.lkey = priv->mr->lkey;
105 param.wr_id = id | IPOIB_OP_RECV;
106 param.sg_list = &list;
109 ret = ib_post_recv(priv->qp, ¶m, &bad_wr);
111 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);
112 dma_unmap_single(priv->ca->dma_device,
113 priv->rx_ring[id].mapping,
114 IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
115 dev_kfree_skb_any(priv->rx_ring[id].skb);
116 priv->rx_ring[id].skb = NULL;
122 static int ipoib_alloc_rx_skb(struct net_device *dev, int id)
124 struct ipoib_dev_priv *priv = netdev_priv(dev);
128 skb = dev_alloc_skb(IPOIB_BUF_SIZE + 4);
133 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte
134 * header. So we need 4 more bytes to get to 48 and align the
135 * IP header to a multiple of 16.
139 addr = dma_map_single(priv->ca->dma_device,
140 skb->data, IPOIB_BUF_SIZE,
142 if (unlikely(dma_mapping_error(addr))) {
143 dev_kfree_skb_any(skb);
147 priv->rx_ring[id].skb = skb;
148 priv->rx_ring[id].mapping = addr;
153 static int ipoib_ib_post_receives(struct net_device *dev)
155 struct ipoib_dev_priv *priv = netdev_priv(dev);
158 for (i = 0; i < ipoib_recvq_size; ++i) {
159 if (ipoib_alloc_rx_skb(dev, i)) {
160 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
163 if (ipoib_ib_post_receive(dev, i)) {
164 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);
172 static void ipoib_ib_handle_wc(struct net_device *dev,
175 struct ipoib_dev_priv *priv = netdev_priv(dev);
176 unsigned int wr_id = wc->wr_id;
178 ipoib_dbg_data(priv, "called: id %d, op %d, status: %d\n",
179 wr_id, wc->opcode, wc->status);
181 if (wr_id & IPOIB_OP_RECV) {
182 wr_id &= ~IPOIB_OP_RECV;
184 if (wr_id < ipoib_recvq_size) {
185 struct sk_buff *skb = priv->rx_ring[wr_id].skb;
186 dma_addr_t addr = priv->rx_ring[wr_id].mapping;
188 if (unlikely(wc->status != IB_WC_SUCCESS)) {
189 if (wc->status != IB_WC_WR_FLUSH_ERR)
190 ipoib_warn(priv, "failed recv event "
191 "(status=%d, wrid=%d vend_err %x)\n",
192 wc->status, wr_id, wc->vendor_err);
193 dma_unmap_single(priv->ca->dma_device, addr,
194 IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
195 dev_kfree_skb_any(skb);
196 priv->rx_ring[wr_id].skb = NULL;
201 * If we can't allocate a new RX buffer, dump
202 * this packet and reuse the old buffer.
204 if (unlikely(ipoib_alloc_rx_skb(dev, wr_id))) {
205 ++priv->stats.rx_dropped;
209 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
210 wc->byte_len, wc->slid);
212 dma_unmap_single(priv->ca->dma_device, addr,
213 IPOIB_BUF_SIZE, DMA_FROM_DEVICE);
215 skb_put(skb, wc->byte_len);
216 skb_pull(skb, IB_GRH_BYTES);
218 if (wc->slid != priv->local_lid ||
219 wc->src_qp != priv->qp->qp_num) {
220 skb->protocol = ((struct ipoib_header *) skb->data)->proto;
221 skb->mac.raw = skb->data;
222 skb_pull(skb, IPOIB_ENCAP_LEN);
224 dev->last_rx = jiffies;
225 ++priv->stats.rx_packets;
226 priv->stats.rx_bytes += skb->len;
229 /* XXX get correct PACKET_ type here */
230 skb->pkt_type = PACKET_HOST;
233 ipoib_dbg_data(priv, "dropping loopback packet\n");
234 dev_kfree_skb_any(skb);
238 if (unlikely(ipoib_ib_post_receive(dev, wr_id)))
239 ipoib_warn(priv, "ipoib_ib_post_receive failed "
240 "for buf %d\n", wr_id);
242 ipoib_warn(priv, "completion event with wrid %d\n",
246 struct ipoib_tx_buf *tx_req;
249 if (wr_id >= ipoib_sendq_size) {
250 ipoib_warn(priv, "completion event with wrid %d (> %d)\n",
251 wr_id, ipoib_sendq_size);
255 ipoib_dbg_data(priv, "send complete, wrid %d\n", wr_id);
257 tx_req = &priv->tx_ring[wr_id];
259 dma_unmap_single(priv->ca->dma_device,
260 pci_unmap_addr(tx_req, mapping),
264 ++priv->stats.tx_packets;
265 priv->stats.tx_bytes += tx_req->skb->len;
267 dev_kfree_skb_any(tx_req->skb);
269 spin_lock_irqsave(&priv->tx_lock, flags);
271 if (netif_queue_stopped(dev) &&
272 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags) &&
273 priv->tx_head - priv->tx_tail <= ipoib_sendq_size >> 1)
274 netif_wake_queue(dev);
275 spin_unlock_irqrestore(&priv->tx_lock, flags);
277 if (wc->status != IB_WC_SUCCESS &&
278 wc->status != IB_WC_WR_FLUSH_ERR)
279 ipoib_warn(priv, "failed send event "
280 "(status=%d, wrid=%d vend_err %x)\n",
281 wc->status, wr_id, wc->vendor_err);
285 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr)
287 struct net_device *dev = (struct net_device *) dev_ptr;
288 struct ipoib_dev_priv *priv = netdev_priv(dev);
291 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
293 n = ib_poll_cq(cq, IPOIB_NUM_WC, priv->ibwc);
294 for (i = 0; i < n; ++i)
295 ipoib_ib_handle_wc(dev, priv->ibwc + i);
296 } while (n == IPOIB_NUM_WC);
299 static inline int post_send(struct ipoib_dev_priv *priv,
301 struct ib_ah *address, u32 qpn,
302 dma_addr_t addr, int len)
304 struct ib_send_wr *bad_wr;
306 priv->tx_sge.addr = addr;
307 priv->tx_sge.length = len;
309 priv->tx_wr.wr_id = wr_id;
310 priv->tx_wr.wr.ud.remote_qpn = qpn;
311 priv->tx_wr.wr.ud.ah = address;
313 return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr);
316 void ipoib_send(struct net_device *dev, struct sk_buff *skb,
317 struct ipoib_ah *address, u32 qpn)
319 struct ipoib_dev_priv *priv = netdev_priv(dev);
320 struct ipoib_tx_buf *tx_req;
323 if (skb->len > dev->mtu + INFINIBAND_ALEN) {
324 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
325 skb->len, dev->mtu + INFINIBAND_ALEN);
326 ++priv->stats.tx_dropped;
327 ++priv->stats.tx_errors;
328 dev_kfree_skb_any(skb);
332 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n",
333 skb->len, address, qpn);
336 * We put the skb into the tx_ring _before_ we call post_send()
337 * because it's entirely possible that the completion handler will
338 * run before we execute anything after the post_send(). That
339 * means we have to make sure everything is properly recorded and
340 * our state is consistent before we call post_send().
342 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];
344 addr = dma_map_single(priv->ca->dma_device, skb->data, skb->len,
346 pci_unmap_addr_set(tx_req, mapping, addr);
348 if (unlikely(post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),
349 address->ah, qpn, addr, skb->len))) {
350 ipoib_warn(priv, "post_send failed\n");
351 ++priv->stats.tx_errors;
352 dma_unmap_single(priv->ca->dma_device, addr, skb->len,
354 dev_kfree_skb_any(skb);
356 dev->trans_start = jiffies;
358 address->last_send = priv->tx_head;
361 if (priv->tx_head - priv->tx_tail == ipoib_sendq_size) {
362 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
363 netif_stop_queue(dev);
368 static void __ipoib_reap_ah(struct net_device *dev)
370 struct ipoib_dev_priv *priv = netdev_priv(dev);
371 struct ipoib_ah *ah, *tah;
372 LIST_HEAD(remove_list);
374 spin_lock_irq(&priv->tx_lock);
375 spin_lock(&priv->lock);
376 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)
377 if ((int) priv->tx_tail - (int) ah->last_send >= 0) {
379 ib_destroy_ah(ah->ah);
382 spin_unlock(&priv->lock);
383 spin_unlock_irq(&priv->tx_lock);
386 void ipoib_reap_ah(void *dev_ptr)
388 struct net_device *dev = dev_ptr;
389 struct ipoib_dev_priv *priv = netdev_priv(dev);
391 __ipoib_reap_ah(dev);
393 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))
394 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
397 int ipoib_ib_dev_open(struct net_device *dev)
399 struct ipoib_dev_priv *priv = netdev_priv(dev);
402 ret = ipoib_init_qp(dev);
404 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);
408 ret = ipoib_ib_post_receives(dev);
410 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);
411 ipoib_ib_dev_stop(dev);
415 clear_bit(IPOIB_STOP_REAPER, &priv->flags);
416 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ);
418 set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
423 static void ipoib_pkey_dev_check_presence(struct net_device *dev)
425 struct ipoib_dev_priv *priv = netdev_priv(dev);
428 if (ib_find_cached_pkey(priv->ca, priv->port, priv->pkey, &pkey_index))
429 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
431 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
434 int ipoib_ib_dev_up(struct net_device *dev)
436 struct ipoib_dev_priv *priv = netdev_priv(dev);
438 ipoib_pkey_dev_check_presence(dev);
440 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
441 ipoib_dbg(priv, "PKEY is not assigned.\n");
445 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
447 return ipoib_mcast_start_thread(dev);
450 int ipoib_ib_dev_down(struct net_device *dev, int flush)
452 struct ipoib_dev_priv *priv = netdev_priv(dev);
454 ipoib_dbg(priv, "downing ib_dev\n");
456 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
457 netif_carrier_off(dev);
459 /* Shutdown the P_Key thread if still active */
460 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
461 mutex_lock(&pkey_mutex);
462 set_bit(IPOIB_PKEY_STOP, &priv->flags);
463 cancel_delayed_work(&priv->pkey_task);
464 mutex_unlock(&pkey_mutex);
466 flush_workqueue(ipoib_workqueue);
469 ipoib_mcast_stop_thread(dev, flush);
470 ipoib_mcast_dev_flush(dev);
472 ipoib_flush_paths(dev);
477 static int recvs_pending(struct net_device *dev)
479 struct ipoib_dev_priv *priv = netdev_priv(dev);
483 for (i = 0; i < ipoib_recvq_size; ++i)
484 if (priv->rx_ring[i].skb)
490 int ipoib_ib_dev_stop(struct net_device *dev)
492 struct ipoib_dev_priv *priv = netdev_priv(dev);
493 struct ib_qp_attr qp_attr;
495 struct ipoib_tx_buf *tx_req;
498 clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);
501 * Move our QP to the error state and then reinitialize in
502 * when all work requests have completed or have been flushed.
504 qp_attr.qp_state = IB_QPS_ERR;
505 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
506 ipoib_warn(priv, "Failed to modify QP to ERROR state\n");
508 /* Wait for all sends and receives to complete */
511 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {
512 if (time_after(jiffies, begin + 5 * HZ)) {
513 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n",
514 priv->tx_head - priv->tx_tail, recvs_pending(dev));
517 * assume the HW is wedged and just free up
518 * all our pending work requests.
520 while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
521 tx_req = &priv->tx_ring[priv->tx_tail &
522 (ipoib_sendq_size - 1)];
523 dma_unmap_single(priv->ca->dma_device,
524 pci_unmap_addr(tx_req, mapping),
527 dev_kfree_skb_any(tx_req->skb);
531 for (i = 0; i < ipoib_recvq_size; ++i)
532 if (priv->rx_ring[i].skb) {
533 dma_unmap_single(priv->ca->dma_device,
534 pci_unmap_addr(&priv->rx_ring[i],
538 dev_kfree_skb_any(priv->rx_ring[i].skb);
539 priv->rx_ring[i].skb = NULL;
548 ipoib_dbg(priv, "All sends and receives done.\n");
551 qp_attr.qp_state = IB_QPS_RESET;
552 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
553 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
555 /* Wait for all AHs to be reaped */
556 set_bit(IPOIB_STOP_REAPER, &priv->flags);
557 cancel_delayed_work(&priv->ah_reap_task);
558 flush_workqueue(ipoib_workqueue);
562 while (!list_empty(&priv->dead_ahs)) {
563 __ipoib_reap_ah(dev);
565 if (time_after(jiffies, begin + HZ)) {
566 ipoib_warn(priv, "timing out; will leak address handles\n");
576 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
578 struct ipoib_dev_priv *priv = netdev_priv(dev);
584 if (ipoib_transport_dev_init(dev, ca)) {
585 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name);
589 if (dev->flags & IFF_UP) {
590 if (ipoib_ib_dev_open(dev)) {
591 ipoib_transport_dev_cleanup(dev);
599 void ipoib_ib_dev_flush(void *_dev)
601 struct net_device *dev = (struct net_device *)_dev;
602 struct ipoib_dev_priv *priv = netdev_priv(dev), *cpriv;
604 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) ) {
605 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");
609 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {
610 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");
614 ipoib_dbg(priv, "flushing\n");
616 ipoib_ib_dev_down(dev, 0);
619 * The device could have been brought down between the start and when
620 * we get here, don't bring it back up if it's not configured up
622 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
623 ipoib_ib_dev_up(dev);
625 mutex_lock(&priv->vlan_mutex);
627 /* Flush any child interfaces too */
628 list_for_each_entry(cpriv, &priv->child_intfs, list)
629 ipoib_ib_dev_flush(cpriv->dev);
631 mutex_unlock(&priv->vlan_mutex);
634 void ipoib_ib_dev_cleanup(struct net_device *dev)
636 struct ipoib_dev_priv *priv = netdev_priv(dev);
638 ipoib_dbg(priv, "cleaning up ib_dev\n");
640 ipoib_mcast_stop_thread(dev, 1);
641 ipoib_mcast_dev_flush(dev);
643 ipoib_transport_dev_cleanup(dev);
647 * Delayed P_Key Assigment Interim Support
649 * The following is initial implementation of delayed P_Key assigment
650 * mechanism. It is using the same approach implemented for the multicast
651 * group join. The single goal of this implementation is to quickly address
652 * Bug #2507. This implementation will probably be removed when the P_Key
653 * change async notification is available.
656 void ipoib_pkey_poll(void *dev_ptr)
658 struct net_device *dev = dev_ptr;
659 struct ipoib_dev_priv *priv = netdev_priv(dev);
661 ipoib_pkey_dev_check_presence(dev);
663 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
666 mutex_lock(&pkey_mutex);
667 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags))
668 queue_delayed_work(ipoib_workqueue,
671 mutex_unlock(&pkey_mutex);
675 int ipoib_pkey_dev_delay_open(struct net_device *dev)
677 struct ipoib_dev_priv *priv = netdev_priv(dev);
679 /* Look for the interface pkey value in the IB Port P_Key table and */
680 /* set the interface pkey assigment flag */
681 ipoib_pkey_dev_check_presence(dev);
683 /* P_Key value not assigned yet - start polling */
684 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {
685 mutex_lock(&pkey_mutex);
686 clear_bit(IPOIB_PKEY_STOP, &priv->flags);
687 queue_delayed_work(ipoib_workqueue,
690 mutex_unlock(&pkey_mutex);