Permalink
Browse files

IB/ipoib: Don't allow MC joins during light MC flush

This fix solves a race between light flush and on the fly joins.
Light flush doesn't set the device to down and unset IPOIB_OPER_UP
flag, this means that if while flushing we have a MC join in progress
and the QP was attached to BC MGID we can have a mismatches when
re-attaching a QP to the BC MGID.

The light flush would set the broadcast group to NULL causing an on
the fly join to rejoin and reattach to the BC MCG as well as adding
the BC MGID to the multicast list. The flush process would later on
remove the BC MGID and detach it from the QP. On the next flush
the BC MGID is present in the multicast list but not found when trying
to detach it because of the previous double attach and single detach.

[18332.714265] ------------[ cut here ]------------
[18332.717775] WARNING: CPU: 6 PID: 3767 at drivers/infiniband/core/verbs.c:280 ib_dealloc_pd+0xff/0x120 [ib_core]
...
[18332.775198] Hardware name: Red Hat KVM, BIOS Bochs 01/01/2011
[18332.779411]  0000000000000000 ffff8800b50dfbb0 ffffffff813fed47 0000000000000000
[18332.784960]  0000000000000000 ffff8800b50dfbf0 ffffffff8109add1 0000011832f58300
[18332.790547]  ffff880226a596c0 ffff880032482000 ffff880032482830 ffff880226a59280
[18332.796199] Call Trace:
[18332.798015]  [<ffffffff813fed47>] dump_stack+0x63/0x8c
[18332.801831]  [<ffffffff8109add1>] __warn+0xd1/0xf0
[18332.805403]  [<ffffffff8109aebd>] warn_slowpath_null+0x1d/0x20
[18332.809706]  [<ffffffffa025d90f>] ib_dealloc_pd+0xff/0x120 [ib_core]
[18332.814384]  [<ffffffffa04f3d7c>] ipoib_transport_dev_cleanup+0xfc/0x1d0 [ib_ipoib]
[18332.820031]  [<ffffffffa04ed648>] ipoib_ib_dev_cleanup+0x98/0x110 [ib_ipoib]
[18332.825220]  [<ffffffffa04e62c8>] ipoib_dev_cleanup+0x2d8/0x550 [ib_ipoib]
[18332.830290]  [<ffffffffa04e656f>] ipoib_uninit+0x2f/0x40 [ib_ipoib]
[18332.834911]  [<ffffffff81772a8a>] rollback_registered_many+0x1aa/0x2c0
[18332.839741]  [<ffffffff81772bd1>] rollback_registered+0x31/0x40
[18332.844091]  [<ffffffff81773b18>] unregister_netdevice_queue+0x48/0x80
[18332.848880]  [<ffffffffa04f489b>] ipoib_vlan_delete+0x1fb/0x290 [ib_ipoib]
[18332.853848]  [<ffffffffa04df1cd>] delete_child+0x7d/0xf0 [ib_ipoib]
[18332.858474]  [<ffffffff81520c08>] dev_attr_store+0x18/0x30
[18332.862510]  [<ffffffff8127fe4a>] sysfs_kf_write+0x3a/0x50
[18332.866349]  [<ffffffff8127f4e0>] kernfs_fop_write+0x120/0x170
[18332.870471]  [<ffffffff81207198>] __vfs_write+0x28/0xe0
[18332.874152]  [<ffffffff810e09bf>] ? percpu_down_read+0x1f/0x50
[18332.878274]  [<ffffffff81208062>] vfs_write+0xa2/0x1a0
[18332.881896]  [<ffffffff812093a6>] SyS_write+0x46/0xa0
[18332.885632]  [<ffffffff810039b7>] do_syscall_64+0x57/0xb0
[18332.889709]  [<ffffffff81883321>] entry_SYSCALL64_slow_path+0x25/0x25
[18332.894727] ---[ end trace 09ebbe31f831ef17 ]---

Fixes: ee1e2c8 ("IPoIB: Refresh paths instead of flushing them on SM change events")
Signed-off-by: Alex Vesker <valex@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
  • Loading branch information...
1 parent 5e102b3 commit 344bacca8cd811809fc33a249f2738ab757d327f Alex Vesker committed with dledford Sep 12, 2016
Showing with 9 additions and 0 deletions.
  1. +9 −0 drivers/infiniband/ulp/ipoib/ipoib_ib.c
@@ -1161,8 +1161,17 @@ static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,
}
if (level == IPOIB_FLUSH_LIGHT) {
+ int oper_up;
ipoib_mark_paths_invalid(dev);
+ /* Set IPoIB operation as down to prevent races between:
+ * the flush flow which leaves MCG and on the fly joins
+ * which can happen during that time. mcast restart task
+ * should deal with join requests we missed.
+ */
+ oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
ipoib_mcast_dev_flush(dev);
+ if (oper_up)
+ set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);
ipoib_flush_ah(dev);
}

0 comments on commit 344bacc

Please sign in to comment.