IB/mlx4: Fix possible deadlock on sm_lock spinlock
authorJack Morgenstein <jackm@dev.mellanox.co.il>
Fri, 3 Aug 2012 08:26:45 +0000 (08:26 +0000)
committerRoland Dreier <roland@purestorage.com>
Fri, 10 Aug 2012 20:02:24 +0000 (13:02 -0700)
The sm_lock spinlock is taken in the process context by
mlx4_ib_modify_device, and in the interrupt context by update_sm_ah,
so we need to take that spinlock with irqsave, and release it with
irqrestore.

Lockdeps reports this as follows:

    [ INFO: inconsistent lock state ]
    3.5.0+ #20 Not tainted
    inconsistent {HARDIRQ-ON-W} -> {IN-HARDIRQ-W} usage.
    swapper/0/0 [HC1[1]:SC0[0]:HE0:SE1] takes:
    (&(&ibdev->sm_lock)->rlock){?.+...}, at: [<ffffffffa028af1d>] update_sm_ah+0xad/0x100 [mlx4_ib]
    {HARDIRQ-ON-W} state was registered at:
      [<ffffffff810b84a0>] mark_irqflags+0x120/0x190
      [<ffffffff810b9ce7>] __lock_acquire+0x307/0x4c0
      [<ffffffff810b9f51>] lock_acquire+0xb1/0x150
      [<ffffffff815523b1>] _raw_spin_lock+0x41/0x50
      [<ffffffffa028d563>] mlx4_ib_modify_device+0x63/0x240 [mlx4_ib]
      [<ffffffffa026d1fc>] ib_modify_device+0x1c/0x20 [ib_core]
      [<ffffffffa026c353>] set_node_desc+0x83/0xc0 [ib_core]
      [<ffffffff8136a150>] dev_attr_store+0x20/0x30
      [<ffffffff81201fd6>] sysfs_write_file+0xe6/0x170
      [<ffffffff8118da38>] vfs_write+0xc8/0x190
      [<ffffffff8118dc01>] sys_write+0x51/0x90
      [<ffffffff8155b869>] system_call_fastpath+0x16/0x1b

    ...
    *** DEADLOCK ***

    1 lock held by swapper/0/0:

    stack backtrace:
    Pid: 0, comm: swapper/0 Not tainted 3.5.0+ #20
    Call Trace:
    <IRQ>  [<ffffffff810b7bea>] print_usage_bug+0x18a/0x190
    [<ffffffff810b7370>] ? print_irq_inversion_bug+0x210/0x210
    [<ffffffff810b7fb2>] mark_lock_irq+0xf2/0x280
    [<ffffffff810b8290>] mark_lock+0x150/0x240
    [<ffffffff810b84ef>] mark_irqflags+0x16f/0x190
    [<ffffffff810b9ce7>] __lock_acquire+0x307/0x4c0
    [<ffffffffa028af1d>] ? update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffff810b9f51>] lock_acquire+0xb1/0x150
    [<ffffffffa028af1d>] ? update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffff815523b1>] _raw_spin_lock+0x41/0x50
    [<ffffffffa028af1d>] ? update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffffa026b2fa>] ? ib_create_ah+0x1a/0x40 [ib_core]
    [<ffffffffa028af1d>] update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffff810c27c3>] ? is_module_address+0x23/0x30
    [<ffffffffa028b05b>] handle_port_mgmt_change_event+0xeb/0x150 [mlx4_ib]
    [<ffffffffa028c177>] mlx4_ib_event+0x117/0x160 [mlx4_ib]
    [<ffffffff81552501>] ? _raw_spin_lock_irqsave+0x61/0x70
    [<ffffffffa022718c>] mlx4_dispatch_event+0x6c/0x90 [mlx4_core]
    [<ffffffffa0221b40>] mlx4_eq_int+0x500/0x950 [mlx4_core]

Reported by: Or Gerlitz <ogerlitz@mellanox.com>
Tested-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <roland@purestorage.com>
drivers/infiniband/hw/mlx4/mad.c
drivers/infiniband/hw/mlx4/main.c

index c27141f..9c2ae7e 100644 (file)
@@ -125,6 +125,7 @@ static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
 {
        struct ib_ah *new_ah;
        struct ib_ah_attr ah_attr;
+       unsigned long flags;
 
        if (!dev->send_agent[port_num - 1][0])
                return;
@@ -139,11 +140,11 @@ static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
        if (IS_ERR(new_ah))
                return;
 
-       spin_lock(&dev->sm_lock);
+       spin_lock_irqsave(&dev->sm_lock, flags);
        if (dev->sm_ah[port_num - 1])
                ib_destroy_ah(dev->sm_ah[port_num - 1]);
        dev->sm_ah[port_num - 1] = new_ah;
-       spin_unlock(&dev->sm_lock);
+       spin_unlock_irqrestore(&dev->sm_lock, flags);
 }
 
 /*
@@ -197,13 +198,15 @@ static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
 static void node_desc_override(struct ib_device *dev,
                               struct ib_mad *mad)
 {
+       unsigned long flags;
+
        if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
             mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
            mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
            mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
-               spin_lock(&to_mdev(dev)->sm_lock);
+               spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
                memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
-               spin_unlock(&to_mdev(dev)->sm_lock);
+               spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
        }
 }
 
@@ -213,6 +216,7 @@ static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *ma
        struct ib_mad_send_buf *send_buf;
        struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
        int ret;
+       unsigned long flags;
 
        if (agent) {
                send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
@@ -225,13 +229,13 @@ static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *ma
                 * wrong following the IB spec strictly, but we know
                 * it's OK for our devices).
                 */
-               spin_lock(&dev->sm_lock);
+               spin_lock_irqsave(&dev->sm_lock, flags);
                memcpy(send_buf->mad, mad, sizeof *mad);
                if ((send_buf->ah = dev->sm_ah[port_num - 1]))
                        ret = ib_post_send_mad(send_buf, NULL);
                else
                        ret = -EINVAL;
-               spin_unlock(&dev->sm_lock);
+               spin_unlock_irqrestore(&dev->sm_lock, flags);
 
                if (ret)
                        ib_free_send_mad(send_buf);
index fe2088c..cc05579 100644 (file)
@@ -423,6 +423,7 @@ static int mlx4_ib_modify_device(struct ib_device *ibdev, int mask,
                                 struct ib_device_modify *props)
 {
        struct mlx4_cmd_mailbox *mailbox;
+       unsigned long flags;
 
        if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
                return -EOPNOTSUPP;
@@ -430,9 +431,9 @@ static int mlx4_ib_modify_device(struct ib_device *ibdev, int mask,
        if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
                return 0;
 
-       spin_lock(&to_mdev(ibdev)->sm_lock);
+       spin_lock_irqsave(&to_mdev(ibdev)->sm_lock, flags);
        memcpy(ibdev->node_desc, props->node_desc, 64);
-       spin_unlock(&to_mdev(ibdev)->sm_lock);
+       spin_unlock_irqrestore(&to_mdev(ibdev)->sm_lock, flags);
 
        /*
         * If possible, pass node desc to FW, so it can generate