IB/ipoib: fix for rare multicast join race condition
authorAlex Estrin <alex.estrin@intel.com>
Thu, 11 Feb 2016 21:30:51 +0000 (16:30 -0500)
committerDoug Ledford <dledford@redhat.com>
Fri, 12 Feb 2016 19:53:22 +0000 (14:53 -0500)
A narrow window for race condition still exist between
multicast join thread and *dev_flush workers.
A kernel crash caused by prolong erratic link state changes
was observed (most likely a faulty cabling):

[167275.656270] BUG: unable to handle kernel NULL pointer dereference at
0000000000000020
[167275.665973] IP: [<ffffffffa05f8f2e>] ipoib_mcast_join+0xae/0x1d0 [ib_ipoib]
[167275.674443] PGD 0
[167275.677373] Oops: 0000 [#1] SMP
...
[167275.977530] Call Trace:
[167275.982225]  [<ffffffffa05f92f0>] ? ipoib_mcast_free+0x200/0x200 [ib_ipoib]
[167275.992024]  [<ffffffffa05fa1b7>] ipoib_mcast_join_task+0x2a7/0x490
[ib_ipoib]
[167276.002149]  [<ffffffff8109d5fb>] process_one_work+0x17b/0x470
[167276.010754]  [<ffffffff8109e3cb>] worker_thread+0x11b/0x400
[167276.019088]  [<ffffffff8109e2b0>] ? rescuer_thread+0x400/0x400
[167276.027737]  [<ffffffff810a5aef>] kthread+0xcf/0xe0
Here was a hit spot:
ipoib_mcast_join() {
..............
      rec.qkey      = priv->broadcast->mcmember.qkey;
                                       ^^^^^^^
.....
 }
Proposed patch should prevent multicast join task to continue
if link state change is detected.

Signed-off-by: Alex Estrin <alex.estrin@intel.com>
Changes from v4:
- as suggested by Doug Ledford, optimized spinlock usage,
i.e. ipoib_mcast_join() is called with lock held.
Changes from v3:
- sync with priv->lock before flag check.
Chages from v2:
- Move check for OPER_UP flag state to mcast_join() to
ensure no event worker is in progress.
- minor style fixes.
Changes from v1:
- No need to lock again if error detected.
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/ulp/ipoib/ipoib_multicast.c

index 050dfa175d169dd3f77ee83767f1988062989891..25889311b1e9c8db412f784a77de5aa8ef0a29cb 100644 (file)
@@ -456,7 +456,10 @@ out_locked:
        return status;
 }
 
-static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
+/*
+ * Caller must hold 'priv->lock'
+ */
+static int ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
 {
        struct ipoib_dev_priv *priv = netdev_priv(dev);
        struct ib_sa_multicast *multicast;
@@ -466,6 +469,10 @@ static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
        ib_sa_comp_mask comp_mask;
        int ret = 0;
 
+       if (!priv->broadcast ||
+           !test_bit(IPOIB_FLAG_OPER_UP, &priv->flags))
+               return -EINVAL;
+
        ipoib_dbg_mcast(priv, "joining MGID %pI6\n", mcast->mcmember.mgid.raw);
 
        rec.mgid     = mcast->mcmember.mgid;
@@ -525,20 +532,23 @@ static void ipoib_mcast_join(struct net_device *dev, struct ipoib_mcast *mcast)
                        rec.join_state = 4;
 #endif
        }
+       spin_unlock_irq(&priv->lock);
 
        multicast = ib_sa_join_multicast(&ipoib_sa_client, priv->ca, priv->port,
                                         &rec, comp_mask, GFP_KERNEL,
                                         ipoib_mcast_join_complete, mcast);
+       spin_lock_irq(&priv->lock);
        if (IS_ERR(multicast)) {
                ret = PTR_ERR(multicast);
                ipoib_warn(priv, "ib_sa_join_multicast failed, status %d\n", ret);
-               spin_lock_irq(&priv->lock);
                /* Requeue this join task with a backoff delay */
                __ipoib_mcast_schedule_join_thread(priv, mcast, 1);
                clear_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
                spin_unlock_irq(&priv->lock);
                complete(&mcast->done);
+               spin_lock_irq(&priv->lock);
        }
+       return 0;
 }
 
 void ipoib_mcast_join_task(struct work_struct *work)
@@ -620,9 +630,10 @@ void ipoib_mcast_join_task(struct work_struct *work)
                                /* Found the next unjoined group */
                                init_completion(&mcast->done);
                                set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
-                               spin_unlock_irq(&priv->lock);
-                               ipoib_mcast_join(dev, mcast);
-                               spin_lock_irq(&priv->lock);
+                               if (ipoib_mcast_join(dev, mcast)) {
+                                       spin_unlock_irq(&priv->lock);
+                                       return;
+                               }
                        } else if (!delay_until ||
                                 time_before(mcast->delay_until, delay_until))
                                delay_until = mcast->delay_until;
@@ -641,10 +652,9 @@ out:
        if (mcast) {
                init_completion(&mcast->done);
                set_bit(IPOIB_MCAST_FLAG_BUSY, &mcast->flags);
+               ipoib_mcast_join(dev, mcast);
        }
        spin_unlock_irq(&priv->lock);
-       if (mcast)
-               ipoib_mcast_join(dev, mcast);
 }
 
 int ipoib_mcast_start_thread(struct net_device *dev)