ipoib_multicast.c

   1/*
   2 * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   5 *
   6 * This software is available to you under a choice of one of two
   7 * licenses.  You may choose to be licensed under the terms of the GNU
   8 * General Public License (GPL) Version 2, available from the file
   9 * COPYING in the main directory of this source tree, or the
  10 * OpenIB.org BSD license below:
  11 *
  12 *     Redistribution and use in source and binary forms, with or
  13 *     without modification, are permitted provided that the following
  14 *     conditions are met:
  15 *
  16 *      - Redistributions of source code must retain the above
  17 *        copyright notice, this list of conditions and the following
  18 *        disclaimer.
  19 *
  20 *      - Redistributions in binary form must reproduce the above
  21 *        copyright notice, this list of conditions and the following
  22 *        disclaimer in the documentation and/or other materials
  23 *        provided with the distribution.
  24 *
  25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32 * SOFTWARE.
  33 */
  34
  35#include <linux/skbuff.h>
  36#include <linux/rtnetlink.h>
  37#include <linux/moduleparam.h>
  38#include <linux/ip.h>
  39#include <linux/in.h>
  40#include <linux/igmp.h>
  41#include <linux/inetdevice.h>
  42#include <linux/delay.h>
  43#include <linux/completion.h>
  44#include <linux/slab.h>
  45
  46#include <net/dst.h>
  47
  48#include "ipoib.h"
  49
  50#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
  51static int mcast_debug_level;
  52
  53module_param(mcast_debug_level, int, 0644);
  54MODULE_PARM_DESC(mcast_debug_level,
  55		 "Enable multicast debug tracing if > 0");
  56#endif
  57
  58struct ipoib_mcast_iter {
  59	struct net_device *dev;
  60	union ib_gid       mgid;
  61	unsigned long      created;
  62	unsigned int       queuelen;
  63	unsigned int       complete;
  64	unsigned int       send_only;
  65};
  66
  67/* join state that allows creating mcg with sendonly member request */
  68#define SENDONLY_FULLMEMBER_JOIN	8
  69
  70/*
  71 * This should be called with the priv->lock held
  72 */
  73static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
  74					       struct ipoib_mcast *mcast,
  75					       bool delay)
  76{
  77	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
  78		return;
  79
  80	/*
  81	 * We will be scheduling *something*, so cancel whatever is
  82	 * currently scheduled first
  83	 */
  84	cancel_delayed_work(&priv->mcast_task);
  85	if (mcast && delay) {
  86		/*
  87		 * We had a failure and want to schedule a retry later
  88		 */
  89		mcast->backoff *= 2;
  90		if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
  91			mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
  92		mcast->delay_until = jiffies + (mcast->backoff * HZ);
  93		/*
  94		 * Mark this mcast for its delay, but restart the
  95		 * task immediately.  The join task will make sure to
  96		 * clear out all entries without delays, and then
  97		 * schedule itself to run again when the earliest
  98		 * delay expires
  99		 */
 100		queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 101	} else if (delay) {
 102		/*
 103		 * Special case of retrying after a failure to
 104		 * allocate the broadcast multicast group, wait
 105		 * 1 second and try again
 106		 */
 107		queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
 108	} else
 109		queue_delayed_work(priv->wq, &priv->mcast_task, 0);
 110}
 111
 112static void ipoib_mcast_free(struct ipoib_mcast *mcast)
 113{
 114	struct net_device *dev = mcast->dev;
 115	int tx_dropped = 0;
 116
 117	ipoib_dbg_mcast(ipoib_priv(dev), "deleting multicast group %pI6\n",
 118			mcast->mcmember.mgid.raw);
 119
 120	/* remove all neigh connected to this mcast */
 121	ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
 122
 123	if (mcast->ah)
 124		ipoib_put_ah(mcast->ah);
 125
 126	while (!skb_queue_empty(&mcast->pkt_queue)) {
 127		++tx_dropped;
 128		dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 129	}
 130
 131	netif_tx_lock_bh(dev);
 132	dev->stats.tx_dropped += tx_dropped;
 133	netif_tx_unlock_bh(dev);
 134
 135	kfree(mcast);
 136}
 137
 138static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev)
 139{
 140	struct ipoib_mcast *mcast;
 141
 142	mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 143	if (!mcast)
 144		return NULL;
 145
 146	mcast->dev = dev;
 147	mcast->created = jiffies;
 148	mcast->delay_until = jiffies;
 149	mcast->backoff = 1;
 150
 151	INIT_LIST_HEAD(&mcast->list);
 152	INIT_LIST_HEAD(&mcast->neigh_list);
 153	skb_queue_head_init(&mcast->pkt_queue);
 154
 155	return mcast;
 156}
 157
 158static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
 159{
 160	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 161	struct rb_node *n = priv->multicast_tree.rb_node;
 162
 163	while (n) {
 164		struct ipoib_mcast *mcast;
 165		int ret;
 166
 167		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
 168
 169		ret = memcmp(mgid, mcast->mcmember.mgid.raw,
 170			     sizeof (union ib_gid));
 171		if (ret < 0)
 172			n = n->rb_left;
 173		else if (ret > 0)
 174			n = n->rb_right;
 175		else
 176			return mcast;
 177	}
 178
 179	return NULL;
 180}
 181
 182static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
 183{
 184	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 185	struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
 186
 187	while (*n) {
 188		struct ipoib_mcast *tmcast;
 189		int ret;
 190
 191		pn = *n;
 192		tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
 193
 194		ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
 195			     sizeof (union ib_gid));
 196		if (ret < 0)
 197			n = &pn->rb_left;
 198		else if (ret > 0)
 199			n = &pn->rb_right;
 200		else
 201			return -EEXIST;
 202	}
 203
 204	rb_link_node(&mcast->rb_node, pn, n);
 205	rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
 206
 207	return 0;
 208}
 209
 210static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
 211				   struct ib_sa_mcmember_rec *mcmember)
 212{
 213	struct net_device *dev = mcast->dev;
 214	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 215	struct rdma_netdev *rn = netdev_priv(dev);
 216	struct ipoib_ah *ah;
 217	struct rdma_ah_attr av;
 218	int ret;
 219	int set_qkey = 0;
 220	int mtu;
 221
 222	mcast->mcmember = *mcmember;
 223
 224	/* Set the multicast MTU and cached Q_Key before we attach if it's
 225	 * the broadcast group.
 226	 */
 227	if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 228		    sizeof (union ib_gid))) {
 229		spin_lock_irq(&priv->lock);
 230		if (!priv->broadcast) {
 231			spin_unlock_irq(&priv->lock);
 232			return -EAGAIN;
 233		}
 234		/*update priv member according to the new mcast*/
 235		priv->broadcast->mcmember.qkey = mcmember->qkey;
 236		priv->broadcast->mcmember.mtu = mcmember->mtu;
 237		priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
 238		priv->broadcast->mcmember.rate = mcmember->rate;
 239		priv->broadcast->mcmember.sl = mcmember->sl;
 240		priv->broadcast->mcmember.flow_label = mcmember->flow_label;
 241		priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
 242		/* assume if the admin and the mcast are the same both can be changed */
 243		mtu = rdma_mtu_enum_to_int(priv->ca,  priv->port,
 244					   priv->broadcast->mcmember.mtu);
 245		if (priv->mcast_mtu == priv->admin_mtu)
 246			priv->admin_mtu = IPOIB_UD_MTU(mtu);
 247		priv->mcast_mtu = IPOIB_UD_MTU(mtu);
 248		rn->mtu = priv->mcast_mtu;
 249
 250		priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
 251		spin_unlock_irq(&priv->lock);
 252		priv->tx_wr.remote_qkey = priv->qkey;
 253		set_qkey = 1;
 254	}
 255
 256	if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 257		if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 258			ipoib_warn(priv, "multicast group %pI6 already attached\n",
 259				   mcast->mcmember.mgid.raw);
 260
 261			return 0;
 262		}
 263
 264		ret = rn->attach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
 265				       be16_to_cpu(mcast->mcmember.mlid),
 266				       set_qkey, priv->qkey);
 267		if (ret < 0) {
 268			ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
 269				   mcast->mcmember.mgid.raw);
 270
 271			clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
 272			return ret;
 273		}
 274	}
 275
 276	memset(&av, 0, sizeof(av));
 277	av.type = rdma_ah_find_type(priv->ca, priv->port);
 278	rdma_ah_set_dlid(&av, be16_to_cpu(mcast->mcmember.mlid));
 279	rdma_ah_set_port_num(&av, priv->port);
 280	rdma_ah_set_sl(&av, mcast->mcmember.sl);
 281	rdma_ah_set_static_rate(&av, mcast->mcmember.rate);
 282
 283	rdma_ah_set_grh(&av, &mcast->mcmember.mgid,
 284			be32_to_cpu(mcast->mcmember.flow_label),
 285			0, mcast->mcmember.hop_limit,
 286			mcast->mcmember.traffic_class);
 287
 288	ah = ipoib_create_ah(dev, priv->pd, &av);
 289	if (IS_ERR(ah)) {
 290		ipoib_warn(priv, "ib_address_create failed %pe\n", ah);
 291		/* use original error */
 292		return PTR_ERR(ah);
 293	}
 294	spin_lock_irq(&priv->lock);
 295	mcast->ah = ah;
 296	spin_unlock_irq(&priv->lock);
 297
 298	ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
 299			mcast->mcmember.mgid.raw,
 300			mcast->ah->ah,
 301			be16_to_cpu(mcast->mcmember.mlid),
 302			mcast->mcmember.sl);
 303
 304	/* actually send any queued packets */
 305	netif_tx_lock_bh(dev);
 306	while (!skb_queue_empty(&mcast->pkt_queue)) {
 307		struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
 308
 309		netif_tx_unlock_bh(dev);
 310
 311		skb->dev = dev;
 312
 313		ret = dev_queue_xmit(skb);
 314		if (ret)
 315			ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
 316				   __func__, ret);
 317		netif_tx_lock_bh(dev);
 318	}
 319	netif_tx_unlock_bh(dev);
 320
 321	return 0;
 322}
 323
 324void ipoib_mcast_carrier_on_task(struct work_struct *work)
 325{
 326	struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
 327						   carrier_on_task);
 328	struct ib_port_attr attr;
 329
 330	if (ib_query_port(priv->ca, priv->port, &attr) ||
 331	    attr.state != IB_PORT_ACTIVE) {
 332		ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
 333		return;
 334	}
 335	/*
 336	 * Take rtnl_lock to avoid racing with ipoib_stop() and
 337	 * turning the carrier back on while a device is being
 338	 * removed.  However, ipoib_stop() will attempt to flush
 339	 * the workqueue while holding the rtnl lock, so loop
 340	 * on trylock until either we get the lock or we see
 341	 * FLAG_OPER_UP go away as that signals that we are bailing
 342	 * and can safely ignore the carrier on work.
 343	 */
 344	while (!rtnl_trylock()) {
 345		if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 346			return;
 347		else
 348			msleep(20);
 349	}
 350	if (!ipoib_cm_admin_enabled(priv->dev))
 351		dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
 352	netif_carrier_on(priv->dev);
 353	rtnl_unlock();
 354}
 355
 356static int ipoib_mcast_join_complete(int status,
 357				     struct ib_sa_multicast *multicast)
 358{
 359	struct ipoib_mcast *mcast = multicast->context;
 360	struct net_device *dev = mcast->dev;
 361	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 362
 363	ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
 364			test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
 365			"sendonly " : "",
 366			mcast->mcmember.mgid.raw, status);
 367
 368	/* We trap for port events ourselves. */
 369	if (status == -ENETRESET) {
 370		status = 0;
 371		goto out;
 372	}
 373
 374	if (!status)
 375		status = ipoib_mcast_join_finish(mcast, &multicast->rec);
 376
 377	if (!status) {
 378		mcast->backoff = 1;
 379		mcast->delay_until = jiffies;
 380
 381		/*
 382		 * Defer carrier on work to priv->wq to avoid a
 383		 * deadlock on rtnl_lock here.  Requeue our multicast
 384		 * work too, which will end up happening right after
 385		 * our carrier on task work and will allow us to
 386		 * send out all of the non-broadcast joins
 387		 */
 388		if (mcast == priv->broadcast) {
 389			spin_lock_irq(&priv->lock);
 390			queue_work(priv->wq, &priv->carrier_on_task);
 391			__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 392			goto out_locked;
 393		}
 394	} else {
 395		bool silent_fail =
 396		    test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 397		    status == -EINVAL;
 398
 399		if (mcast->logcount < 20) {
 400			if (status == -ETIMEDOUT || status == -EAGAIN ||
 401			    silent_fail) {
 402				ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
 403						test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 404						mcast->mcmember.mgid.raw, status);
 405			} else {
 406				ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
 407						test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
 408					   mcast->mcmember.mgid.raw, status);
 409			}
 410
 411			if (!silent_fail)
 412				mcast->logcount++;
 413		}
 414
 415		if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
 416		    mcast->backoff >= 2) {
 417			/*
 418			 * We only retry sendonly joins once before we drop
 419			 * the packet and quit trying to deal with the
 420			 * group.  However, we leave the group in the
 421			 * mcast list as an unjoined group.  If we want to
 422			 * try joining again, we simply queue up a packet
 423			 * and restart the join thread.  The empty queue
 424			 * is why the join thread ignores this group.
 425			 */
 426			mcast->backoff = 1;
 427			netif_tx_lock_bh(dev);
 428			while (!skb_queue_empty(&mcast->pkt_queue)) {
 429				++dev->stats.tx_dropped;
 430				dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
 431			}
 432			netif_tx_unlock_bh(dev);
 433		} else {
 434			spin_lock_irq(&priv->lock);
 435			/* Requeue this join task with a backoff delay */
 436			__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 437			goto out_locked;
 438		}
 439	}
 440out:
 441	spin_lock_irq(&priv->lock);
 442out_locked:
 443	/*
 444	 * Make sure to set mcast->mc before we clear the busy flag to avoid
 445	 * racing with code that checks for BUSY before checking mcast->mc
 446	 */
 447	if (status)
 448		mcast->mc = NULL;
 449	else
 450		mcast->mc = multicast;
 451	clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 452	spin_unlock_irq(&priv->lock);
 453	complete(&mcast->done);
 454
 455	return status;
 456}
 457
 458/*
 459 * Caller must hold 'priv->lock'
 460 */
 461static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
 462{
 463	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 464	struct ib_sa_multicast *multicast;
 465	struct ib_sa_mcmember_rec rec = {
 466		.join_state = 1
 467	};
 468	ib_sa_comp_mask comp_mask;
 469	int ret = 0;
 470
 471	if (!priv->broadcast ||
 472	    !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 473		return -EINVAL;
 474
 475	init_completion(&mcast->done);
 476	set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 477
 478	ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 479
 480	rec.mgid     = mcast->mcmember.mgid;
 481	rec.port_gid = priv->local_gid;
 482	rec.pkey     = cpu_to_be16(priv->pkey);
 483
 484	comp_mask =
 485		IB_SA_MCMEMBER_REC_MGID		|
 486		IB_SA_MCMEMBER_REC_PORT_GID	|
 487		IB_SA_MCMEMBER_REC_PKEY		|
 488		IB_SA_MCMEMBER_REC_JOIN_STATE;
 489
 490	if (mcast != priv->broadcast) {
 491		/*
 492		 * RFC 4391:
 493		 *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
 494		 *  and HopLimit as those used in the broadcast-GID.  The rest
 495		 *  of attributes SHOULD follow the values used in the
 496		 *  broadcast-GID as well.
 497		 */
 498		comp_mask |=
 499			IB_SA_MCMEMBER_REC_QKEY			|
 500			IB_SA_MCMEMBER_REC_MTU_SELECTOR		|
 501			IB_SA_MCMEMBER_REC_MTU			|
 502			IB_SA_MCMEMBER_REC_TRAFFIC_CLASS	|
 503			IB_SA_MCMEMBER_REC_RATE_SELECTOR	|
 504			IB_SA_MCMEMBER_REC_RATE			|
 505			IB_SA_MCMEMBER_REC_SL			|
 506			IB_SA_MCMEMBER_REC_FLOW_LABEL		|
 507			IB_SA_MCMEMBER_REC_HOP_LIMIT;
 508
 509		rec.qkey	  = priv->broadcast->mcmember.qkey;
 510		rec.mtu_selector  = IB_SA_EQ;
 511		rec.mtu		  = priv->broadcast->mcmember.mtu;
 512		rec.traffic_class = priv->broadcast->mcmember.traffic_class;
 513		rec.rate_selector = IB_SA_EQ;
 514		rec.rate	  = priv->broadcast->mcmember.rate;
 515		rec.sl		  = priv->broadcast->mcmember.sl;
 516		rec.flow_label	  = priv->broadcast->mcmember.flow_label;
 517		rec.hop_limit	  = priv->broadcast->mcmember.hop_limit;
 518
 519		/*
 520		 * Send-only IB Multicast joins work at the core IB layer but
 521		 * require specific SM support.
 522		 * We can use such joins here only if the current SM supports that feature.
 523		 * However, if not, we emulate an Ethernet multicast send,
 524		 * which does not require a multicast subscription and will
 525		 * still send properly. The most appropriate thing to
 526		 * do is to create the group if it doesn't exist as that
 527		 * most closely emulates the behavior, from a user space
 528		 * application perspective, of Ethernet multicast operation.
 529		 */
 530		if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 531			rec.join_state = SENDONLY_FULLMEMBER_JOIN;
 532	}
 533
 534	multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
 535					 &rec, comp_mask, GFP_ATOMIC,
 536					 ipoib_mcast_join_complete, mcast);
 537	if (IS_ERR(multicast)) {
 538		ret = PTR_ERR(multicast);
 539		ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
 540		/* Requeue this join task with a backoff delay */
 541		__ipoib_mcast_schedule_join_thread(priv, mcast, 1);
 542		clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
 543		complete(&mcast->done);
 544		return ret;
 545	}
 546	return 0;
 547}
 548
 549void ipoib_mcast_join_task(struct work_struct *work)
 550{
 551	struct ipoib_dev_priv *priv =
 552		container_of(work, struct ipoib_dev_priv, mcast_task.work);
 553	struct net_device *dev = priv->dev;
 554	struct ib_port_attr port_attr;
 555	unsigned long delay_until = 0;
 556	struct ipoib_mcast *mcast = NULL;
 557
 558	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 559		return;
 560
 561	if (ib_query_port(priv->ca, priv->port, &port_attr)) {
 562		ipoib_dbg(priv, "ib_query_port() failed\n");
 563		return;
 564	}
 565	if (port_attr.state != IB_PORT_ACTIVE) {
 566		ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
 567			  port_attr.state);
 568		return;
 569	}
 570	priv->local_lid = port_attr.lid;
 571	netif_addr_lock_bh(dev);
 572
 573	if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
 574		netif_addr_unlock_bh(dev);
 575		return;
 576	}
 577	netif_addr_unlock_bh(dev);
 578
 579	spin_lock_irq(&priv->lock);
 580	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 581		goto out;
 582
 583	if (!priv->broadcast) {
 584		struct ipoib_mcast *broadcast;
 585
 586		broadcast = ipoib_mcast_alloc(dev);
 587		if (!broadcast) {
 588			ipoib_warn(priv, "failed to allocate broadcast group\n");
 589			/*
 590			 * Restart us after a 1 second delay to retry
 591			 * creating our broadcast group and attaching to
 592			 * it.  Until this succeeds, this ipoib dev is
 593			 * completely stalled (multicast wise).
 594			 */
 595			__ipoib_mcast_schedule_join_thread(priv, NULL, 1);
 596			goto out;
 597		}
 598
 599		memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
 600		       sizeof (union ib_gid));
 601		priv->broadcast = broadcast;
 602
 603		__ipoib_mcast_add(dev, priv->broadcast);
 604	}
 605
 606	if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 607		if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
 608		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
 609			mcast = priv->broadcast;
 610			if (mcast->backoff > 1 &&
 611			    time_before(jiffies, mcast->delay_until)) {
 612				delay_until = mcast->delay_until;
 613				mcast = NULL;
 614			}
 615		}
 616		goto out;
 617	}
 618
 619	/*
 620	 * We'll never get here until the broadcast group is both allocated
 621	 * and attached
 622	 */
 623	list_for_each_entry(mcast, &priv->multicast_list, list) {
 624		if (IS_ERR_OR_NULL(mcast->mc) &&
 625		    !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
 626		    (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
 627		     !skb_queue_empty(&mcast->pkt_queue))) {
 628			if (mcast->backoff == 1 ||
 629			    time_after_eq(jiffies, mcast->delay_until)) {
 630				/* Found the next unjoined group */
 631				if (ipoib_mcast_join(dev, mcast)) {
 632					spin_unlock_irq(&priv->lock);
 633					return;
 634				}
 635			} else if (!delay_until ||
 636				 time_before(mcast->delay_until, delay_until))
 637				delay_until = mcast->delay_until;
 638		}
 639	}
 640
 641	mcast = NULL;
 642	ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
 643
 644out:
 645	if (delay_until) {
 646		cancel_delayed_work(&priv->mcast_task);
 647		queue_delayed_work(priv->wq, &priv->mcast_task,
 648				   delay_until - jiffies);
 649	}
 650	if (mcast)
 651		ipoib_mcast_join(dev, mcast);
 652
 653	spin_unlock_irq(&priv->lock);
 654}
 655
 656void ipoib_mcast_start_thread(struct net_device *dev)
 657{
 658	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 659	unsigned long flags;
 660
 661	ipoib_dbg_mcast(priv, "starting multicast thread\n");
 662
 663	spin_lock_irqsave(&priv->lock, flags);
 664	__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 665	spin_unlock_irqrestore(&priv->lock, flags);
 666}
 667
 668void ipoib_mcast_stop_thread(struct net_device *dev)
 669{
 670	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 671
 672	ipoib_dbg_mcast(priv, "stopping multicast thread\n");
 673
 674	cancel_delayed_work_sync(&priv->mcast_task);
 675}
 676
 677static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
 678{
 679	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 680	struct rdma_netdev *rn = netdev_priv(dev);
 681	int ret = 0;
 682
 683	if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 684		ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
 685
 686	if (!IS_ERR_OR_NULL(mcast->mc))
 687		ib_sa_free_multicast(mcast->mc);
 688
 689	if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
 690		ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
 691				mcast->mcmember.mgid.raw);
 692
 693		/* Remove ourselves from the multicast group */
 694		ret = rn->detach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
 695				       be16_to_cpu(mcast->mcmember.mlid));
 696		if (ret)
 697			ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
 698	} else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
 699		ipoib_dbg(priv, "leaving with no mcmember but not a "
 700			  "SENDONLY join\n");
 701
 702	return 0;
 703}
 704
 705/*
 706 * Check if the multicast group is sendonly. If so remove it from the maps
 707 * and add to the remove list
 708 */
 709void ipoib_check_and_add_mcast_sendonly(struct ipoib_dev_priv *priv, u8 *mgid,
 710				struct list_head *remove_list)
 711{
 712	/* Is this multicast ? */
 713	if (*mgid == 0xff) {
 714		struct ipoib_mcast *mcast = __ipoib_mcast_find(priv->dev, mgid);
 715
 716		if (mcast && test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 717			list_del(&mcast->list);
 718			rb_erase(&mcast->rb_node, &priv->multicast_tree);
 719			list_add_tail(&mcast->list, remove_list);
 720		}
 721	}
 722}
 723
 724void ipoib_mcast_remove_list(struct list_head *remove_list)
 725{
 726	struct ipoib_mcast *mcast, *tmcast;
 727
 728	/*
 729	 * make sure the in-flight joins have finished before we attempt
 730	 * to leave
 731	 */
 732	list_for_each_entry_safe(mcast, tmcast, remove_list, list)
 733		if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
 734			wait_for_completion(&mcast->done);
 735
 736	list_for_each_entry_safe(mcast, tmcast, remove_list, list) {
 737		ipoib_mcast_leave(mcast->dev, mcast);
 738		ipoib_mcast_free(mcast);
 739	}
 740}
 741
 742void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
 743{
 744	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 745	struct rdma_netdev *rn = netdev_priv(dev);
 746	struct ipoib_mcast *mcast;
 747	unsigned long flags;
 748	void *mgid = daddr + 4;
 749
 750	spin_lock_irqsave(&priv->lock, flags);
 751
 752	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)		||
 753	    !priv->broadcast					||
 754	    !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
 755		++dev->stats.tx_dropped;
 756		dev_kfree_skb_any(skb);
 757		goto unlock;
 758	}
 759
 760	mcast = __ipoib_mcast_find(dev, mgid);
 761	if (!mcast || !mcast->ah) {
 762		if (!mcast) {
 763			/* Let's create a new send only group now */
 764			ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
 765					mgid);
 766
 767			mcast = ipoib_mcast_alloc(dev);
 768			if (!mcast) {
 769				ipoib_warn(priv, "unable to allocate memory "
 770					   "for multicast structure\n");
 771				++dev->stats.tx_dropped;
 772				dev_kfree_skb_any(skb);
 773				goto unlock;
 774			}
 775
 776			set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
 777			memcpy(mcast->mcmember.mgid.raw, mgid,
 778			       sizeof (union ib_gid));
 779			__ipoib_mcast_add(dev, mcast);
 780			list_add_tail(&mcast->list, &priv->multicast_list);
 781		}
 782		if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
 783			/* put pseudoheader back on for next time */
 784			skb_push(skb, sizeof(struct ipoib_pseudo_header));
 785			skb_queue_tail(&mcast->pkt_queue, skb);
 786		} else {
 787			++dev->stats.tx_dropped;
 788			dev_kfree_skb_any(skb);
 789		}
 790		if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
 791			__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 792		}
 793	} else {
 794		struct ipoib_neigh *neigh;
 795
 796		spin_unlock_irqrestore(&priv->lock, flags);
 797		neigh = ipoib_neigh_get(dev, daddr);
 798		spin_lock_irqsave(&priv->lock, flags);
 799		if (!neigh) {
 800			neigh = ipoib_neigh_alloc(daddr, dev);
 801			/* Make sure that the neigh will be added only
 802			 * once to mcast list.
 803			 */
 804			if (neigh && list_empty(&neigh->list)) {
 805				kref_get(&mcast->ah->ref);
 806				neigh->ah	= mcast->ah;
 807				neigh->ah->valid = 1;
 808				list_add_tail(&neigh->list, &mcast->neigh_list);
 809			}
 810		}
 811		spin_unlock_irqrestore(&priv->lock, flags);
 812		mcast->ah->last_send = rn->send(dev, skb, mcast->ah->ah,
 813						IB_MULTICAST_QPN);
 814		if (neigh)
 815			ipoib_neigh_put(neigh);
 816		return;
 817	}
 818
 819unlock:
 820	spin_unlock_irqrestore(&priv->lock, flags);
 821}
 822
 823void ipoib_mcast_dev_flush(struct net_device *dev)
 824{
 825	struct ipoib_dev_priv *priv = ipoib_priv(dev);
 826	LIST_HEAD(remove_list);
 827	struct ipoib_mcast *mcast, *tmcast;
 828	unsigned long flags;
 829
 830	mutex_lock(&priv->mcast_mutex);
 831	ipoib_dbg_mcast(priv, "flushing multicast list\n");
 832
 833	spin_lock_irqsave(&priv->lock, flags);
 834
 835	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 836		list_del(&mcast->list);
 837		rb_erase(&mcast->rb_node, &priv->multicast_tree);
 838		list_add_tail(&mcast->list, &remove_list);
 839	}
 840
 841	if (priv->broadcast) {
 842		rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
 843		list_add_tail(&priv->broadcast->list, &remove_list);
 844		priv->broadcast = NULL;
 845	}
 846
 847	spin_unlock_irqrestore(&priv->lock, flags);
 848
 849	ipoib_mcast_remove_list(&remove_list);
 850	mutex_unlock(&priv->mcast_mutex);
 851}
 852
 853static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
 854{
 855	/* reserved QPN, prefix, scope */
 856	if (memcmp(addr, broadcast, 6))
 857		return 0;
 858	/* signature lower, pkey */
 859	if (memcmp(addr + 7, broadcast + 7, 3))
 860		return 0;
 861	return 1;
 862}
 863
 864void ipoib_mcast_restart_task(struct work_struct *work)
 865{
 866	struct ipoib_dev_priv *priv =
 867		container_of(work, struct ipoib_dev_priv, restart_task);
 868	struct net_device *dev = priv->dev;
 869	struct netdev_hw_addr *ha;
 870	struct ipoib_mcast *mcast, *tmcast;
 871	LIST_HEAD(remove_list);
 872	struct ib_sa_mcmember_rec rec;
 873
 874	if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
 875		/*
 876		 * shortcut...on shutdown flush is called next, just
 877		 * let it do all the work
 878		 */
 879		return;
 880
 881	ipoib_dbg_mcast(priv, "restarting multicast task\n");
 882
 883	netif_addr_lock_bh(dev);
 884	spin_lock_irq(&priv->lock);
 885
 886	/*
 887	 * Unfortunately, the networking core only gives us a list of all of
 888	 * the multicast hardware addresses. We need to figure out which ones
 889	 * are new and which ones have been removed
 890	 */
 891
 892	/* Clear out the found flag */
 893	list_for_each_entry(mcast, &priv->multicast_list, list)
 894		clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 895
 896	/* Mark all of the entries that are found or don't exist */
 897	netdev_for_each_mc_addr(ha, dev) {
 898		union ib_gid mgid;
 899
 900		if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
 901			continue;
 902
 903		memcpy(mgid.raw, ha->addr + 4, sizeof(mgid));
 904
 905		mcast = __ipoib_mcast_find(dev, &mgid);
 906		if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 907			struct ipoib_mcast *nmcast;
 908
 909			/* ignore group which is directly joined by userspace */
 910			if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
 911			    !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
 912				ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
 913						mgid.raw);
 914				continue;
 915			}
 916
 917			/* Not found or send-only group, let's add a new entry */
 918			ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
 919					mgid.raw);
 920
 921			nmcast = ipoib_mcast_alloc(dev);
 922			if (!nmcast) {
 923				ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
 924				continue;
 925			}
 926
 927			set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
 928
 929			nmcast->mcmember.mgid = mgid;
 930
 931			if (mcast) {
 932				/* Destroy the send only entry */
 933				list_move_tail(&mcast->list, &remove_list);
 934
 935				rb_replace_node(&mcast->rb_node,
 936						&nmcast->rb_node,
 937						&priv->multicast_tree);
 938			} else
 939				__ipoib_mcast_add(dev, nmcast);
 940
 941			list_add_tail(&nmcast->list, &priv->multicast_list);
 942		}
 943
 944		if (mcast)
 945			set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
 946	}
 947
 948	/* Remove all of the entries don't exist anymore */
 949	list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
 950		if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
 951		    !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
 952			ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
 953					mcast->mcmember.mgid.raw);
 954
 955			rb_erase(&mcast->rb_node, &priv->multicast_tree);
 956
 957			/* Move to the remove list */
 958			list_move_tail(&mcast->list, &remove_list);
 959		}
 960	}
 961
 962	spin_unlock_irq(&priv->lock);
 963	netif_addr_unlock_bh(dev);
 964
 965	ipoib_mcast_remove_list(&remove_list);
 966
 967	/*
 968	 * Double check that we are still up
 969	 */
 970	if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
 971		spin_lock_irq(&priv->lock);
 972		__ipoib_mcast_schedule_join_thread(priv, NULL, 0);
 973		spin_unlock_irq(&priv->lock);
 974	}
 975}
 976
 977#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 978
 979struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
 980{
 981	struct ipoib_mcast_iter *iter;
 982
 983	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
 984	if (!iter)
 985		return NULL;
 986
 987	iter->dev = dev;
 988	memset(iter->mgid.raw, 0, 16);
 989
 990	if (ipoib_mcast_iter_next(iter)) {
 991		kfree(iter);
 992		return NULL;
 993	}
 994
 995	return iter;
 996}
 997
 998int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
 999{
1000	struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
1001	struct rb_node *n;
1002	struct ipoib_mcast *mcast;
1003	int ret = 1;
1004
1005	spin_lock_irq(&priv->lock);
1006
1007	n = rb_first(&priv->multicast_tree);
1008
1009	while (n) {
1010		mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1011
1012		if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1013			   sizeof (union ib_gid)) < 0) {
1014			iter->mgid      = mcast->mcmember.mgid;
1015			iter->created   = mcast->created;
1016			iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1017			iter->complete  = !!mcast->ah;
1018			iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1019
1020			ret = 0;
1021
1022			break;
1023		}
1024
1025		n = rb_next(n);
1026	}
1027
1028	spin_unlock_irq(&priv->lock);
1029
1030	return ret;
1031}
1032
1033void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1034			   union ib_gid *mgid,
1035			   unsigned long *created,
1036			   unsigned int *queuelen,
1037			   unsigned int *complete,
1038			   unsigned int *send_only)
1039{
1040	*mgid      = iter->mgid;
1041	*created   = iter->created;
1042	*queuelen  = iter->queuelen;
1043	*complete  = iter->complete;
1044	*send_only = iter->send_only;
1045}
1046
1047#endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */