ulp/ipoib/ipoib_multicast.c

0001 /*
0002  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
0003  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
0004  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
0005  *
0006  * This software is available to you under a choice of one of two
0007  * licenses.  You may choose to be licensed under the terms of the GNU
0008  * General Public License (GPL) Version 2, available from the file
0009  * COPYING in the main directory of this source tree, or the
0010  * OpenIB.org BSD license below:
0011  *
0012  *     Redistribution and use in source and binary forms, with or
0013  *     without modification, are permitted provided that the following
0014  *     conditions are met:
0015  *
0016  *      - Redistributions of source code must retain the above
0017  *        copyright notice, this list of conditions and the following
0018  *        disclaimer.
0019  *
0020  *      - Redistributions in binary form must reproduce the above
0021  *        copyright notice, this list of conditions and the following
0022  *        disclaimer in the documentation and/or other materials
0023  *        provided with the distribution.
0024  *
0025  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0026  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0027  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0028  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0029  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0030  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0031  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0032  * SOFTWARE.
0033  */
0034
0035 #include <linux/skbuff.h>
0036 #include <linux/rtnetlink.h>
0037 #include <linux/moduleparam.h>
0038 #include <linux/ip.h>
0039 #include <linux/in.h>
0040 #include <linux/igmp.h>
0041 #include <linux/inetdevice.h>
0042 #include <linux/delay.h>
0043 #include <linux/completion.h>
0044 #include <linux/slab.h>
0045
0046 #include <net/dst.h>
0047
0048 #include "ipoib.h"
0049
0050 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
0051 static int mcast_debug_level;
0052
0053 module_param(mcast_debug_level, int, 0644);
0054 MODULE_PARM_DESC(mcast_debug_level,
0055          "Enable multicast debug tracing if > 0");
0056 #endif
0057
0058 struct ipoib_mcast_iter {
0059     struct net_device *dev;
0060     union ib_gid       mgid;
0061     unsigned long      created;
0062     unsigned int       queuelen;
0063     unsigned int       complete;
0064     unsigned int       send_only;
0065 };
0066
0067 /* join state that allows creating mcg with sendonly member request */
0068 #define SENDONLY_FULLMEMBER_JOIN    8
0069
0070 /*
0071  * This should be called with the priv->lock held
0072  */
0073 static void __ipoib_mcast_schedule_join_thread(struct ipoib_dev_priv *priv,
0074                            struct ipoib_mcast *mcast,
0075                            bool delay)
0076 {
0077     if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
0078         return;
0079
0080     /*
0081      * We will be scheduling *something*, so cancel whatever is
0082      * currently scheduled first
0083      */
0084     cancel_delayed_work(&priv->mcast_task);
0085     if (mcast && delay) {
0086         /*
0087          * We had a failure and want to schedule a retry later
0088          */
0089         mcast->backoff *= 2;
0090         if (mcast->backoff > IPOIB_MAX_BACKOFF_SECONDS)
0091             mcast->backoff = IPOIB_MAX_BACKOFF_SECONDS;
0092         mcast->delay_until = jiffies + (mcast->backoff * HZ);
0093         /*
0094          * Mark this mcast for its delay, but restart the
0095          * task immediately.  The join task will make sure to
0096          * clear out all entries without delays, and then
0097          * schedule itself to run again when the earliest
0098          * delay expires
0099          */
0100         queue_delayed_work(priv->wq, &priv->mcast_task, 0);
0101     } else if (delay) {
0102         /*
0103          * Special case of retrying after a failure to
0104          * allocate the broadcast multicast group, wait
0105          * 1 second and try again
0106          */
0107         queue_delayed_work(priv->wq, &priv->mcast_task, HZ);
0108     } else
0109         queue_delayed_work(priv->wq, &priv->mcast_task, 0);
0110 }
0111
0112 static void ipoib_mcast_free(struct ipoib_mcast *mcast)
0113 {
0114     struct net_device *dev = mcast->dev;
0115     int tx_dropped = 0;
0116
0117     ipoib_dbg_mcast(ipoib_priv(dev), "deleting multicast group %pI6\n",
0118             mcast->mcmember.mgid.raw);
0119
0120     /* remove all neigh connected to this mcast */
0121     ipoib_del_neighs_by_gid(dev, mcast->mcmember.mgid.raw);
0122
0123     if (mcast->ah)
0124         ipoib_put_ah(mcast->ah);
0125
0126     while (!skb_queue_empty(&mcast->pkt_queue)) {
0127         ++tx_dropped;
0128         dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
0129     }
0130
0131     netif_tx_lock_bh(dev);
0132     dev->stats.tx_dropped += tx_dropped;
0133     netif_tx_unlock_bh(dev);
0134
0135     kfree(mcast);
0136 }
0137
0138 static struct ipoib_mcast *ipoib_mcast_alloc(struct net_device *dev)
0139 {
0140     struct ipoib_mcast *mcast;
0141
0142     mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
0143     if (!mcast)
0144         return NULL;
0145
0146     mcast->dev = dev;
0147     mcast->created = jiffies;
0148     mcast->delay_until = jiffies;
0149     mcast->backoff = 1;
0150
0151     INIT_LIST_HEAD(&mcast->list);
0152     INIT_LIST_HEAD(&mcast->neigh_list);
0153     skb_queue_head_init(&mcast->pkt_queue);
0154
0155     return mcast;
0156 }
0157
0158 static struct ipoib_mcast *__ipoib_mcast_find(struct net_device *dev, void *mgid)
0159 {
0160     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0161     struct rb_node *n = priv->multicast_tree.rb_node;
0162
0163     while (n) {
0164         struct ipoib_mcast *mcast;
0165         int ret;
0166
0167         mcast = rb_entry(n, struct ipoib_mcast, rb_node);
0168
0169         ret = memcmp(mgid, mcast->mcmember.mgid.raw,
0170                  sizeof (union ib_gid));
0171         if (ret < 0)
0172             n = n->rb_left;
0173         else if (ret > 0)
0174             n = n->rb_right;
0175         else
0176             return mcast;
0177     }
0178
0179     return NULL;
0180 }
0181
0182 static int __ipoib_mcast_add(struct net_device *dev, struct ipoib_mcast *mcast)
0183 {
0184     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0185     struct rb_node **n = &priv->multicast_tree.rb_node, *pn = NULL;
0186
0187     while (*n) {
0188         struct ipoib_mcast *tmcast;
0189         int ret;
0190
0191         pn = *n;
0192         tmcast = rb_entry(pn, struct ipoib_mcast, rb_node);
0193
0194         ret = memcmp(mcast->mcmember.mgid.raw, tmcast->mcmember.mgid.raw,
0195                  sizeof (union ib_gid));
0196         if (ret < 0)
0197             n = &pn->rb_left;
0198         else if (ret > 0)
0199             n = &pn->rb_right;
0200         else
0201             return -EEXIST;
0202     }
0203
0204     rb_link_node(&mcast->rb_node, pn, n);
0205     rb_insert_color(&mcast->rb_node, &priv->multicast_tree);
0206
0207     return 0;
0208 }
0209
0210 static int ipoib_mcast_join_finish(struct ipoib_mcast *mcast,
0211                    struct ib_sa_mcmember_rec *mcmember)
0212 {
0213     struct net_device *dev = mcast->dev;
0214     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0215     struct rdma_netdev *rn = netdev_priv(dev);
0216     struct ipoib_ah *ah;
0217     struct rdma_ah_attr av;
0218     int ret;
0219     int set_qkey = 0;
0220     int mtu;
0221
0222     mcast->mcmember = *mcmember;
0223
0224     /* Set the multicast MTU and cached Q_Key before we attach if it's
0225      * the broadcast group.
0226      */
0227     if (!memcmp(mcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
0228             sizeof (union ib_gid))) {
0229         spin_lock_irq(&priv->lock);
0230         if (!priv->broadcast) {
0231             spin_unlock_irq(&priv->lock);
0232             return -EAGAIN;
0233         }
0234         /*update priv member according to the new mcast*/
0235         priv->broadcast->mcmember.qkey = mcmember->qkey;
0236         priv->broadcast->mcmember.mtu = mcmember->mtu;
0237         priv->broadcast->mcmember.traffic_class = mcmember->traffic_class;
0238         priv->broadcast->mcmember.rate = mcmember->rate;
0239         priv->broadcast->mcmember.sl = mcmember->sl;
0240         priv->broadcast->mcmember.flow_label = mcmember->flow_label;
0241         priv->broadcast->mcmember.hop_limit = mcmember->hop_limit;
0242         /* assume if the admin and the mcast are the same both can be changed */
0243         mtu = rdma_mtu_enum_to_int(priv->ca,  priv->port,
0244                        priv->broadcast->mcmember.mtu);
0245         if (priv->mcast_mtu == priv->admin_mtu)
0246             priv->admin_mtu = IPOIB_UD_MTU(mtu);
0247         priv->mcast_mtu = IPOIB_UD_MTU(mtu);
0248         rn->mtu = priv->mcast_mtu;
0249
0250         priv->qkey = be32_to_cpu(priv->broadcast->mcmember.qkey);
0251         spin_unlock_irq(&priv->lock);
0252         priv->tx_wr.remote_qkey = priv->qkey;
0253         set_qkey = 1;
0254     }
0255
0256     if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
0257         if (test_and_set_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
0258             ipoib_warn(priv, "multicast group %pI6 already attached\n",
0259                    mcast->mcmember.mgid.raw);
0260
0261             return 0;
0262         }
0263
0264         ret = rn->attach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
0265                        be16_to_cpu(mcast->mcmember.mlid),
0266                        set_qkey, priv->qkey);
0267         if (ret < 0) {
0268             ipoib_warn(priv, "couldn't attach QP to multicast group %pI6\n",
0269                    mcast->mcmember.mgid.raw);
0270
0271             clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags);
0272             return ret;
0273         }
0274     }
0275
0276     memset(&av, 0, sizeof(av));
0277     av.type = rdma_ah_find_type(priv->ca, priv->port);
0278     rdma_ah_set_dlid(&av, be16_to_cpu(mcast->mcmember.mlid));
0279     rdma_ah_set_port_num(&av, priv->port);
0280     rdma_ah_set_sl(&av, mcast->mcmember.sl);
0281     rdma_ah_set_static_rate(&av, mcast->mcmember.rate);
0282
0283     rdma_ah_set_grh(&av, &mcast->mcmember.mgid,
0284             be32_to_cpu(mcast->mcmember.flow_label),
0285             0, mcast->mcmember.hop_limit,
0286             mcast->mcmember.traffic_class);
0287
0288     ah = ipoib_create_ah(dev, priv->pd, &av);
0289     if (IS_ERR(ah)) {
0290         ipoib_warn(priv, "ib_address_create failed %ld\n",
0291                -PTR_ERR(ah));
0292         /* use original error */
0293         return PTR_ERR(ah);
0294     }
0295     spin_lock_irq(&priv->lock);
0296     mcast->ah = ah;
0297     spin_unlock_irq(&priv->lock);
0298
0299     ipoib_dbg_mcast(priv, "MGID %pI6 AV %p, LID 0x%04x, SL %d\n",
0300             mcast->mcmember.mgid.raw,
0301             mcast->ah->ah,
0302             be16_to_cpu(mcast->mcmember.mlid),
0303             mcast->mcmember.sl);
0304
0305     /* actually send any queued packets */
0306     netif_tx_lock_bh(dev);
0307     while (!skb_queue_empty(&mcast->pkt_queue)) {
0308         struct sk_buff *skb = skb_dequeue(&mcast->pkt_queue);
0309
0310         netif_tx_unlock_bh(dev);
0311
0312         skb->dev = dev;
0313
0314         ret = dev_queue_xmit(skb);
0315         if (ret)
0316             ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
0317                    __func__, ret);
0318         netif_tx_lock_bh(dev);
0319     }
0320     netif_tx_unlock_bh(dev);
0321
0322     return 0;
0323 }
0324
0325 void ipoib_mcast_carrier_on_task(struct work_struct *work)
0326 {
0327     struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
0328                            carrier_on_task);
0329     struct ib_port_attr attr;
0330
0331     if (ib_query_port(priv->ca, priv->port, &attr) ||
0332         attr.state != IB_PORT_ACTIVE) {
0333         ipoib_dbg(priv, "Keeping carrier off until IB port is active\n");
0334         return;
0335     }
0336     /*
0337      * Take rtnl_lock to avoid racing with ipoib_stop() and
0338      * turning the carrier back on while a device is being
0339      * removed.  However, ipoib_stop() will attempt to flush
0340      * the workqueue while holding the rtnl lock, so loop
0341      * on trylock until either we get the lock or we see
0342      * FLAG_OPER_UP go away as that signals that we are bailing
0343      * and can safely ignore the carrier on work.
0344      */
0345     while (!rtnl_trylock()) {
0346         if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
0347             return;
0348         else
0349             msleep(20);
0350     }
0351     if (!ipoib_cm_admin_enabled(priv->dev))
0352         dev_set_mtu(priv->dev, min(priv->mcast_mtu, priv->admin_mtu));
0353     netif_carrier_on(priv->dev);
0354     rtnl_unlock();
0355 }
0356
0357 static int ipoib_mcast_join_complete(int status,
0358                      struct ib_sa_multicast *multicast)
0359 {
0360     struct ipoib_mcast *mcast = multicast->context;
0361     struct net_device *dev = mcast->dev;
0362     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0363
0364     ipoib_dbg_mcast(priv, "%sjoin completion for %pI6 (status %d)\n",
0365             test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ?
0366             "sendonly " : "",
0367             mcast->mcmember.mgid.raw, status);
0368
0369     /* We trap for port events ourselves. */
0370     if (status == -ENETRESET) {
0371         status = 0;
0372         goto out;
0373     }
0374
0375     if (!status)
0376         status = ipoib_mcast_join_finish(mcast, &multicast->rec);
0377
0378     if (!status) {
0379         mcast->backoff = 1;
0380         mcast->delay_until = jiffies;
0381
0382         /*
0383          * Defer carrier on work to priv->wq to avoid a
0384          * deadlock on rtnl_lock here.  Requeue our multicast
0385          * work too, which will end up happening right after
0386          * our carrier on task work and will allow us to
0387          * send out all of the non-broadcast joins
0388          */
0389         if (mcast == priv->broadcast) {
0390             spin_lock_irq(&priv->lock);
0391             queue_work(priv->wq, &priv->carrier_on_task);
0392             __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
0393             goto out_locked;
0394         }
0395     } else {
0396         bool silent_fail =
0397             test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
0398             status == -EINVAL;
0399
0400         if (mcast->logcount < 20) {
0401             if (status == -ETIMEDOUT || status == -EAGAIN ||
0402                 silent_fail) {
0403                 ipoib_dbg_mcast(priv, "%smulticast join failed for %pI6, status %d\n",
0404                         test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
0405                         mcast->mcmember.mgid.raw, status);
0406             } else {
0407                 ipoib_warn(priv, "%smulticast join failed for %pI6, status %d\n",
0408                         test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ? "sendonly " : "",
0409                        mcast->mcmember.mgid.raw, status);
0410             }
0411
0412             if (!silent_fail)
0413                 mcast->logcount++;
0414         }
0415
0416         if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) &&
0417             mcast->backoff >= 2) {
0418             /*
0419              * We only retry sendonly joins once before we drop
0420              * the packet and quit trying to deal with the
0421              * group.  However, we leave the group in the
0422              * mcast list as an unjoined group.  If we want to
0423              * try joining again, we simply queue up a packet
0424              * and restart the join thread.  The empty queue
0425              * is why the join thread ignores this group.
0426              */
0427             mcast->backoff = 1;
0428             netif_tx_lock_bh(dev);
0429             while (!skb_queue_empty(&mcast->pkt_queue)) {
0430                 ++dev->stats.tx_dropped;
0431                 dev_kfree_skb_any(skb_dequeue(&mcast->pkt_queue));
0432             }
0433             netif_tx_unlock_bh(dev);
0434         } else {
0435             spin_lock_irq(&priv->lock);
0436             /* Requeue this join task with a backoff delay */
0437             __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
0438             goto out_locked;
0439         }
0440     }
0441 out:
0442     spin_lock_irq(&priv->lock);
0443 out_locked:
0444     /*
0445      * Make sure to set mcast->mc before we clear the busy flag to avoid
0446      * racing with code that checks for BUSY before checking mcast->mc
0447      */
0448     if (status)
0449         mcast->mc = NULL;
0450     else
0451         mcast->mc = multicast;
0452     clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
0453     spin_unlock_irq(&priv->lock);
0454     complete(&mcast->done);
0455
0456     return status;
0457 }
0458
0459 /*
0460  * Caller must hold 'priv->lock'
0461  */
0462 static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
0463 {
0464     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0465     struct ib_sa_multicast *multicast;
0466     struct ib_sa_mcmember_rec rec = {
0467         .join_state = 1
0468     };
0469     ib_sa_comp_mask comp_mask;
0470     int ret = 0;
0471
0472     if (!priv->broadcast ||
0473         !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
0474         return -EINVAL;
0475
0476     init_completion(&mcast->done);
0477     set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
0478
0479     ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
0480
0481     rec.mgid     = mcast->mcmember.mgid;
0482     rec.port_gid = priv->local_gid;
0483     rec.pkey     = cpu_to_be16(priv->pkey);
0484
0485     comp_mask =
0486         IB_SA_MCMEMBER_REC_MGID     |
0487         IB_SA_MCMEMBER_REC_PORT_GID |
0488         IB_SA_MCMEMBER_REC_PKEY     |
0489         IB_SA_MCMEMBER_REC_JOIN_STATE;
0490
0491     if (mcast != priv->broadcast) {
0492         /*
0493          * RFC 4391:
0494          *  The MGID MUST use the same P_Key, Q_Key, SL, MTU,
0495          *  and HopLimit as those used in the broadcast-GID.  The rest
0496          *  of attributes SHOULD follow the values used in the
0497          *  broadcast-GID as well.
0498          */
0499         comp_mask |=
0500             IB_SA_MCMEMBER_REC_QKEY         |
0501             IB_SA_MCMEMBER_REC_MTU_SELECTOR     |
0502             IB_SA_MCMEMBER_REC_MTU          |
0503             IB_SA_MCMEMBER_REC_TRAFFIC_CLASS    |
0504             IB_SA_MCMEMBER_REC_RATE_SELECTOR    |
0505             IB_SA_MCMEMBER_REC_RATE         |
0506             IB_SA_MCMEMBER_REC_SL           |
0507             IB_SA_MCMEMBER_REC_FLOW_LABEL       |
0508             IB_SA_MCMEMBER_REC_HOP_LIMIT;
0509
0510         rec.qkey      = priv->broadcast->mcmember.qkey;
0511         rec.mtu_selector  = IB_SA_EQ;
0512         rec.mtu       = priv->broadcast->mcmember.mtu;
0513         rec.traffic_class = priv->broadcast->mcmember.traffic_class;
0514         rec.rate_selector = IB_SA_EQ;
0515         rec.rate      = priv->broadcast->mcmember.rate;
0516         rec.sl        = priv->broadcast->mcmember.sl;
0517         rec.flow_label    = priv->broadcast->mcmember.flow_label;
0518         rec.hop_limit     = priv->broadcast->mcmember.hop_limit;
0519
0520         /*
0521          * Send-only IB Multicast joins work at the core IB layer but
0522          * require specific SM support.
0523          * We can use such joins here only if the current SM supports that feature.
0524          * However, if not, we emulate an Ethernet multicast send,
0525          * which does not require a multicast subscription and will
0526          * still send properly. The most appropriate thing to
0527          * do is to create the group if it doesn't exist as that
0528          * most closely emulates the behavior, from a user space
0529          * application perspective, of Ethernet multicast operation.
0530          */
0531         if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
0532             rec.join_state = SENDONLY_FULLMEMBER_JOIN;
0533     }
0534     spin_unlock_irq(&priv->lock);
0535
0536     multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
0537                      &rec, comp_mask, GFP_KERNEL,
0538                      ipoib_mcast_join_complete, mcast);
0539     spin_lock_irq(&priv->lock);
0540     if (IS_ERR(multicast)) {
0541         ret = PTR_ERR(multicast);
0542         ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
0543         /* Requeue this join task with a backoff delay */
0544         __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
0545         clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
0546         spin_unlock_irq(&priv->lock);
0547         complete(&mcast->done);
0548         spin_lock_irq(&priv->lock);
0549     }
0550     return 0;
0551 }
0552
0553 void ipoib_mcast_join_task(struct work_struct *work)
0554 {
0555     struct ipoib_dev_priv *priv =
0556         container_of(work, struct ipoib_dev_priv, mcast_task.work);
0557     struct net_device *dev = priv->dev;
0558     struct ib_port_attr port_attr;
0559     unsigned long delay_until = 0;
0560     struct ipoib_mcast *mcast = NULL;
0561
0562     if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
0563         return;
0564
0565     if (ib_query_port(priv->ca, priv->port, &port_attr)) {
0566         ipoib_dbg(priv, "ib_query_port() failed\n");
0567         return;
0568     }
0569     if (port_attr.state != IB_PORT_ACTIVE) {
0570         ipoib_dbg(priv, "port state is not ACTIVE (state = %d) suspending join task\n",
0571               port_attr.state);
0572         return;
0573     }
0574     priv->local_lid = port_attr.lid;
0575     netif_addr_lock_bh(dev);
0576
0577     if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
0578         netif_addr_unlock_bh(dev);
0579         return;
0580     }
0581     netif_addr_unlock_bh(dev);
0582
0583     spin_lock_irq(&priv->lock);
0584     if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
0585         goto out;
0586
0587     if (!priv->broadcast) {
0588         struct ipoib_mcast *broadcast;
0589
0590         broadcast = ipoib_mcast_alloc(dev);
0591         if (!broadcast) {
0592             ipoib_warn(priv, "failed to allocate broadcast group\n");
0593             /*
0594              * Restart us after a 1 second delay to retry
0595              * creating our broadcast group and attaching to
0596              * it.  Until this succeeds, this ipoib dev is
0597              * completely stalled (multicast wise).
0598              */
0599             __ipoib_mcast_schedule_join_thread(priv, NULL, 1);
0600             goto out;
0601         }
0602
0603         memcpy(broadcast->mcmember.mgid.raw, priv->dev->broadcast + 4,
0604                sizeof (union ib_gid));
0605         priv->broadcast = broadcast;
0606
0607         __ipoib_mcast_add(dev, priv->broadcast);
0608     }
0609
0610     if (!test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
0611         if (IS_ERR_OR_NULL(priv->broadcast->mc) &&
0612             !test_bit(IPOIB_MCAST_FLAG_BUSY, &priv->broadcast->flags)) {
0613             mcast = priv->broadcast;
0614             if (mcast->backoff > 1 &&
0615                 time_before(jiffies, mcast->delay_until)) {
0616                 delay_until = mcast->delay_until;
0617                 mcast = NULL;
0618             }
0619         }
0620         goto out;
0621     }
0622
0623     /*
0624      * We'll never get here until the broadcast group is both allocated
0625      * and attached
0626      */
0627     list_for_each_entry(mcast, &priv->multicast_list, list) {
0628         if (IS_ERR_OR_NULL(mcast->mc) &&
0629             !test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags) &&
0630             (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags) ||
0631              !skb_queue_empty(&mcast->pkt_queue))) {
0632             if (mcast->backoff == 1 ||
0633                 time_after_eq(jiffies, mcast->delay_until)) {
0634                 /* Found the next unjoined group */
0635                 if (ipoib_mcast_join(dev, mcast)) {
0636                     spin_unlock_irq(&priv->lock);
0637                     return;
0638                 }
0639             } else if (!delay_until ||
0640                  time_before(mcast->delay_until, delay_until))
0641                 delay_until = mcast->delay_until;
0642         }
0643     }
0644
0645     mcast = NULL;
0646     ipoib_dbg_mcast(priv, "successfully started all multicast joins\n");
0647
0648 out:
0649     if (delay_until) {
0650         cancel_delayed_work(&priv->mcast_task);
0651         queue_delayed_work(priv->wq, &priv->mcast_task,
0652                    delay_until - jiffies);
0653     }
0654     if (mcast)
0655         ipoib_mcast_join(dev, mcast);
0656
0657     spin_unlock_irq(&priv->lock);
0658 }
0659
0660 void ipoib_mcast_start_thread(struct net_device *dev)
0661 {
0662     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0663     unsigned long flags;
0664
0665     ipoib_dbg_mcast(priv, "starting multicast thread\n");
0666
0667     spin_lock_irqsave(&priv->lock, flags);
0668     __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
0669     spin_unlock_irqrestore(&priv->lock, flags);
0670 }
0671
0672 void ipoib_mcast_stop_thread(struct net_device *dev)
0673 {
0674     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0675
0676     ipoib_dbg_mcast(priv, "stopping multicast thread\n");
0677
0678     cancel_delayed_work_sync(&priv->mcast_task);
0679 }
0680
0681 static int ipoib_mcast_leave(struct net_device *dev, struct ipoib_mcast *mcast)
0682 {
0683     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0684     struct rdma_netdev *rn = netdev_priv(dev);
0685     int ret = 0;
0686
0687     if (test_and_clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
0688         ipoib_warn(priv, "ipoib_mcast_leave on an in-flight join\n");
0689
0690     if (!IS_ERR_OR_NULL(mcast->mc))
0691         ib_sa_free_multicast(mcast->mc);
0692
0693     if (test_and_clear_bit(IPOIB_MCAST_FLAG_ATTACHED, &mcast->flags)) {
0694         ipoib_dbg_mcast(priv, "leaving MGID %pI6\n",
0695                 mcast->mcmember.mgid.raw);
0696
0697         /* Remove ourselves from the multicast group */
0698         ret = rn->detach_mcast(dev, priv->ca, &mcast->mcmember.mgid,
0699                        be16_to_cpu(mcast->mcmember.mlid));
0700         if (ret)
0701             ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
0702     } else if (!test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
0703         ipoib_dbg(priv, "leaving with no mcmember but not a "
0704               "SENDONLY join\n");
0705
0706     return 0;
0707 }
0708
0709 /*
0710  * Check if the multicast group is sendonly. If so remove it from the maps
0711  * and add to the remove list
0712  */
0713 void ipoib_check_and_add_mcast_sendonly(struct ipoib_dev_priv *priv, u8 *mgid,
0714                 struct list_head *remove_list)
0715 {
0716     /* Is this multicast ? */
0717     if (*mgid == 0xff) {
0718         struct ipoib_mcast *mcast = __ipoib_mcast_find(priv->dev, mgid);
0719
0720         if (mcast && test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
0721             list_del(&mcast->list);
0722             rb_erase(&mcast->rb_node, &priv->multicast_tree);
0723             list_add_tail(&mcast->list, remove_list);
0724         }
0725     }
0726 }
0727
0728 void ipoib_mcast_remove_list(struct list_head *remove_list)
0729 {
0730     struct ipoib_mcast *mcast, *tmcast;
0731
0732     /*
0733      * make sure the in-flight joins have finished before we attempt
0734      * to leave
0735      */
0736     list_for_each_entry_safe(mcast, tmcast, remove_list, list)
0737         if (test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags))
0738             wait_for_completion(&mcast->done);
0739
0740     list_for_each_entry_safe(mcast, tmcast, remove_list, list) {
0741         ipoib_mcast_leave(mcast->dev, mcast);
0742         ipoib_mcast_free(mcast);
0743     }
0744 }
0745
0746 void ipoib_mcast_send(struct net_device *dev, u8 *daddr, struct sk_buff *skb)
0747 {
0748     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0749     struct rdma_netdev *rn = netdev_priv(dev);
0750     struct ipoib_mcast *mcast;
0751     unsigned long flags;
0752     void *mgid = daddr + 4;
0753
0754     spin_lock_irqsave(&priv->lock, flags);
0755
0756     if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)     ||
0757         !priv->broadcast                    ||
0758         !test_bit(IPOIB_MCAST_FLAG_ATTACHED, &priv->broadcast->flags)) {
0759         ++dev->stats.tx_dropped;
0760         dev_kfree_skb_any(skb);
0761         goto unlock;
0762     }
0763
0764     mcast = __ipoib_mcast_find(dev, mgid);
0765     if (!mcast || !mcast->ah) {
0766         if (!mcast) {
0767             /* Let's create a new send only group now */
0768             ipoib_dbg_mcast(priv, "setting up send only multicast group for %pI6\n",
0769                     mgid);
0770
0771             mcast = ipoib_mcast_alloc(dev);
0772             if (!mcast) {
0773                 ipoib_warn(priv, "unable to allocate memory "
0774                        "for multicast structure\n");
0775                 ++dev->stats.tx_dropped;
0776                 dev_kfree_skb_any(skb);
0777                 goto unlock;
0778             }
0779
0780             set_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags);
0781             memcpy(mcast->mcmember.mgid.raw, mgid,
0782                    sizeof (union ib_gid));
0783             __ipoib_mcast_add(dev, mcast);
0784             list_add_tail(&mcast->list, &priv->multicast_list);
0785         }
0786         if (skb_queue_len(&mcast->pkt_queue) < IPOIB_MAX_MCAST_QUEUE) {
0787             /* put pseudoheader back on for next time */
0788             skb_push(skb, sizeof(struct ipoib_pseudo_header));
0789             skb_queue_tail(&mcast->pkt_queue, skb);
0790         } else {
0791             ++dev->stats.tx_dropped;
0792             dev_kfree_skb_any(skb);
0793         }
0794         if (!test_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags)) {
0795             __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
0796         }
0797     } else {
0798         struct ipoib_neigh *neigh;
0799
0800         spin_unlock_irqrestore(&priv->lock, flags);
0801         neigh = ipoib_neigh_get(dev, daddr);
0802         spin_lock_irqsave(&priv->lock, flags);
0803         if (!neigh) {
0804             neigh = ipoib_neigh_alloc(daddr, dev);
0805             /* Make sure that the neigh will be added only
0806              * once to mcast list.
0807              */
0808             if (neigh && list_empty(&neigh->list)) {
0809                 kref_get(&mcast->ah->ref);
0810                 neigh->ah   = mcast->ah;
0811                 neigh->ah->valid = 1;
0812                 list_add_tail(&neigh->list, &mcast->neigh_list);
0813             }
0814         }
0815         spin_unlock_irqrestore(&priv->lock, flags);
0816         mcast->ah->last_send = rn->send(dev, skb, mcast->ah->ah,
0817                         IB_MULTICAST_QPN);
0818         if (neigh)
0819             ipoib_neigh_put(neigh);
0820         return;
0821     }
0822
0823 unlock:
0824     spin_unlock_irqrestore(&priv->lock, flags);
0825 }
0826
0827 void ipoib_mcast_dev_flush(struct net_device *dev)
0828 {
0829     struct ipoib_dev_priv *priv = ipoib_priv(dev);
0830     LIST_HEAD(remove_list);
0831     struct ipoib_mcast *mcast, *tmcast;
0832     unsigned long flags;
0833
0834     mutex_lock(&priv->mcast_mutex);
0835     ipoib_dbg_mcast(priv, "flushing multicast list\n");
0836
0837     spin_lock_irqsave(&priv->lock, flags);
0838
0839     list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
0840         list_del(&mcast->list);
0841         rb_erase(&mcast->rb_node, &priv->multicast_tree);
0842         list_add_tail(&mcast->list, &remove_list);
0843     }
0844
0845     if (priv->broadcast) {
0846         rb_erase(&priv->broadcast->rb_node, &priv->multicast_tree);
0847         list_add_tail(&priv->broadcast->list, &remove_list);
0848         priv->broadcast = NULL;
0849     }
0850
0851     spin_unlock_irqrestore(&priv->lock, flags);
0852
0853     ipoib_mcast_remove_list(&remove_list);
0854     mutex_unlock(&priv->mcast_mutex);
0855 }
0856
0857 static int ipoib_mcast_addr_is_valid(const u8 *addr, const u8 *broadcast)
0858 {
0859     /* reserved QPN, prefix, scope */
0860     if (memcmp(addr, broadcast, 6))
0861         return 0;
0862     /* signature lower, pkey */
0863     if (memcmp(addr + 7, broadcast + 7, 3))
0864         return 0;
0865     return 1;
0866 }
0867
0868 void ipoib_mcast_restart_task(struct work_struct *work)
0869 {
0870     struct ipoib_dev_priv *priv =
0871         container_of(work, struct ipoib_dev_priv, restart_task);
0872     struct net_device *dev = priv->dev;
0873     struct netdev_hw_addr *ha;
0874     struct ipoib_mcast *mcast, *tmcast;
0875     LIST_HEAD(remove_list);
0876     struct ib_sa_mcmember_rec rec;
0877
0878     if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
0879         /*
0880          * shortcut...on shutdown flush is called next, just
0881          * let it do all the work
0882          */
0883         return;
0884
0885     ipoib_dbg_mcast(priv, "restarting multicast task\n");
0886
0887     netif_addr_lock_bh(dev);
0888     spin_lock_irq(&priv->lock);
0889
0890     /*
0891      * Unfortunately, the networking core only gives us a list of all of
0892      * the multicast hardware addresses. We need to figure out which ones
0893      * are new and which ones have been removed
0894      */
0895
0896     /* Clear out the found flag */
0897     list_for_each_entry(mcast, &priv->multicast_list, list)
0898         clear_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
0899
0900     /* Mark all of the entries that are found or don't exist */
0901     netdev_for_each_mc_addr(ha, dev) {
0902         union ib_gid mgid;
0903
0904         if (!ipoib_mcast_addr_is_valid(ha->addr, dev->broadcast))
0905             continue;
0906
0907         memcpy(mgid.raw, ha->addr + 4, sizeof(mgid));
0908
0909         mcast = __ipoib_mcast_find(dev, &mgid);
0910         if (!mcast || test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
0911             struct ipoib_mcast *nmcast;
0912
0913             /* ignore group which is directly joined by userspace */
0914             if (test_bit(IPOIB_FLAG_UMCAST, &priv->flags) &&
0915                 !ib_sa_get_mcmember_rec(priv->ca, priv->port, &mgid, &rec)) {
0916                 ipoib_dbg_mcast(priv, "ignoring multicast entry for mgid %pI6\n",
0917                         mgid.raw);
0918                 continue;
0919             }
0920
0921             /* Not found or send-only group, let's add a new entry */
0922             ipoib_dbg_mcast(priv, "adding multicast entry for mgid %pI6\n",
0923                     mgid.raw);
0924
0925             nmcast = ipoib_mcast_alloc(dev);
0926             if (!nmcast) {
0927                 ipoib_warn(priv, "unable to allocate memory for multicast structure\n");
0928                 continue;
0929             }
0930
0931             set_bit(IPOIB_MCAST_FLAG_FOUND, &nmcast->flags);
0932
0933             nmcast->mcmember.mgid = mgid;
0934
0935             if (mcast) {
0936                 /* Destroy the send only entry */
0937                 list_move_tail(&mcast->list, &remove_list);
0938
0939                 rb_replace_node(&mcast->rb_node,
0940                         &nmcast->rb_node,
0941                         &priv->multicast_tree);
0942             } else
0943                 __ipoib_mcast_add(dev, nmcast);
0944
0945             list_add_tail(&nmcast->list, &priv->multicast_list);
0946         }
0947
0948         if (mcast)
0949             set_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags);
0950     }
0951
0952     /* Remove all of the entries don't exist anymore */
0953     list_for_each_entry_safe(mcast, tmcast, &priv->multicast_list, list) {
0954         if (!test_bit(IPOIB_MCAST_FLAG_FOUND, &mcast->flags) &&
0955             !test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags)) {
0956             ipoib_dbg_mcast(priv, "deleting multicast group %pI6\n",
0957                     mcast->mcmember.mgid.raw);
0958
0959             rb_erase(&mcast->rb_node, &priv->multicast_tree);
0960
0961             /* Move to the remove list */
0962             list_move_tail(&mcast->list, &remove_list);
0963         }
0964     }
0965
0966     spin_unlock_irq(&priv->lock);
0967     netif_addr_unlock_bh(dev);
0968
0969     ipoib_mcast_remove_list(&remove_list);
0970
0971     /*
0972      * Double check that we are still up
0973      */
0974     if (test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
0975         spin_lock_irq(&priv->lock);
0976         __ipoib_mcast_schedule_join_thread(priv, NULL, 0);
0977         spin_unlock_irq(&priv->lock);
0978     }
0979 }
0980
0981 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
0982
0983 struct ipoib_mcast_iter *ipoib_mcast_iter_init(struct net_device *dev)
0984 {
0985     struct ipoib_mcast_iter *iter;
0986
0987     iter = kmalloc(sizeof(*iter), GFP_KERNEL);
0988     if (!iter)
0989         return NULL;
0990
0991     iter->dev = dev;
0992     memset(iter->mgid.raw, 0, 16);
0993
0994     if (ipoib_mcast_iter_next(iter)) {
0995         kfree(iter);
0996         return NULL;
0997     }
0998
0999     return iter;
1000 }
1001
1002 int ipoib_mcast_iter_next(struct ipoib_mcast_iter *iter)
1003 {
1004     struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
1005     struct rb_node *n;
1006     struct ipoib_mcast *mcast;
1007     int ret = 1;
1008
1009     spin_lock_irq(&priv->lock);
1010
1011     n = rb_first(&priv->multicast_tree);
1012
1013     while (n) {
1014         mcast = rb_entry(n, struct ipoib_mcast, rb_node);
1015
1016         if (memcmp(iter->mgid.raw, mcast->mcmember.mgid.raw,
1017                sizeof (union ib_gid)) < 0) {
1018             iter->mgid      = mcast->mcmember.mgid;
1019             iter->created   = mcast->created;
1020             iter->queuelen  = skb_queue_len(&mcast->pkt_queue);
1021             iter->complete  = !!mcast->ah;
1022             iter->send_only = !!(mcast->flags & (1 << IPOIB_MCAST_FLAG_SENDONLY));
1023
1024             ret = 0;
1025
1026             break;
1027         }
1028
1029         n = rb_next(n);
1030     }
1031
1032     spin_unlock_irq(&priv->lock);
1033
1034     return ret;
1035 }
1036
1037 void ipoib_mcast_iter_read(struct ipoib_mcast_iter *iter,
1038                union ib_gid *mgid,
1039                unsigned long *created,
1040                unsigned int *queuelen,
1041                unsigned int *complete,
1042                unsigned int *send_only)
1043 {
1044     *mgid      = iter->mgid;
1045     *created   = iter->created;
1046     *queuelen  = iter->queuelen;
1047     *complete  = iter->complete;
1048     *send_only = iter->send_only;
1049 }
1050
1051 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */