ixgbevf: fix softirq-safe to unsafe splat on internal mbx_lock
authorJohn Fastabend <john.r.fastabend@intel.com>
Mon, 1 Oct 2012 14:52:20 +0000 (14:52 +0000)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Tue, 23 Oct 2012 08:14:04 +0000 (01:14 -0700)
The lockdep splat below identifies a case where irq safe to unsafe
lock order is detected. Resolved by making mbx_lock bh.

======================================================
[ INFO: SOFTIRQ-safe -> SOFTIRQ-unsafe lock order detected ]
3.6.0-rc5jk-net-next+ #119 Not tainted
------------------------------------------------------
ip/2608 [HC0[0]:SC0[2]:HE1:SE0] is trying to acquire:
 (&(&adapter->mbx_lock)->rlock){+.+...}, at: [<ffffffffa008114e>] ixgbevf_set_rx_mode+0x36/0xd2 [ixgbevf]

and this task is already holding:
 (_xmit_ETHER){+.....}, at: [<ffffffff814097c8>] dev_set_rx_mode+0x1e/0x33
which would create a new lock dependency:
 (_xmit_ETHER){+.....} -> (&(&adapter->mbx_lock)->rlock){+.+...}

but this new dependency connects a SOFTIRQ-irq-safe lock:
 (&(&mc->mca_lock)->rlock){+.-...}
... which became SOFTIRQ-irq-safe at:
  [<ffffffff81092ee5>] __lock_acquire+0x2f2/0xdf3
  [<ffffffff81093b11>] lock_acquire+0x12b/0x158
  [<ffffffff814bdbcd>] _raw_spin_lock_bh+0x4a/0x7d
  [<ffffffffa011a740>] mld_ifc_timer_expire+0x1b2/0x282 [ipv6]
  [<ffffffff81054580>] run_timer_softirq+0x2a2/0x3ee
  [<ffffffff8104cc42>] __do_softirq+0x161/0x2b9
  [<ffffffff814c6a7c>] call_softirq+0x1c/0x30
  [<ffffffff81011bc7>] do_softirq+0x4b/0xa3
  [<ffffffff8104c8d5>] irq_exit+0x53/0xd7
  [<ffffffff814c734d>] do_IRQ+0x9d/0xb4
  [<ffffffff814be56f>] ret_from_intr+0x0/0x1a
  [<ffffffff813de21c>] cpuidle_enter+0x12/0x14
  [<ffffffff813de235>] cpuidle_enter_state+0x17/0x3f
  [<ffffffff813deb6c>] cpuidle_idle_call+0x140/0x21c
  [<ffffffff8101764c>] cpu_idle+0x79/0xcd
  [<ffffffff814a59f5>] rest_init+0x149/0x150
  [<ffffffff81ca2cbc>] start_kernel+0x37c/0x389
  [<ffffffff81ca22dd>] x86_64_start_reservations+0xb8/0xbd
  [<ffffffff81ca23e3>] x86_64_start_kernel+0x101/0x110

to a SOFTIRQ-irq-unsafe lock:
 (&(&adapter->mbx_lock)->rlock){+.+...}
... which became SOFTIRQ-irq-unsafe at:
...  [<ffffffff81092f59>] __lock_acquire+0x366/0xdf3
  [<ffffffff81093b11>] lock_acquire+0x12b/0x158
  [<ffffffff814bd862>] _raw_spin_lock+0x45/0x7a
  [<ffffffffa0080fde>] ixgbevf_negotiate_api+0x3d/0x6d [ixgbevf]
  [<ffffffffa008404b>] ixgbevf_open+0x6c/0x43e [ixgbevf]
  [<ffffffff8140b2c1>] __dev_open+0xa0/0xe6
  [<ffffffff814099b6>] __dev_change_flags+0xbe/0x142
  [<ffffffff8140b1eb>] dev_change_flags+0x21/0x57
  [<ffffffff8141a523>] do_setlink+0x2e2/0x7f4
  [<ffffffff8141ad8c>] rtnl_newlink+0x277/0x4bb
  [<ffffffff81419c08>] rtnetlink_rcv_msg+0x236/0x253
  [<ffffffff8142f92d>] netlink_rcv_skb+0x43/0x94
  [<ffffffff814199cb>] rtnetlink_rcv+0x26/0x2d
  [<ffffffff8142f6dc>] netlink_unicast+0xee/0x174
  [<ffffffff8142ff12>] netlink_sendmsg+0x26a/0x288
  [<ffffffff813f5a0d>] __sock_sendmsg_nosec+0x58/0x61
  [<ffffffff813f7d57>] __sock_sendmsg+0x3d/0x48
  [<ffffffff813f7ed9>] sock_sendmsg+0x6e/0x87
  [<ffffffff813f93d4>] __sys_sendmsg+0x206/0x288
  [<ffffffff813f95ce>] sys_sendmsg+0x42/0x60
  [<ffffffff814c57a9>] system_call_fastpath+0x16/0x1b

other info that might help us debug this:

Chain exists of:
  &(&mc->mca_lock)->rlock --> _xmit_ETHER --> &(&adapter->mbx_lock)->rlock

 Possible interrupt unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&(&adapter->mbx_lock)->rlock);
                               local_irq_disable();
                               lock(&(&mc->mca_lock)->rlock);
                               lock(_xmit_ETHER);
  <Interrupt>
    lock(&(&mc->mca_lock)->rlock);

 *** DEADLOCK ***

Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Greg Rose <gregory.v.rose@intel.com>
Tested-by: Sibai Li <sibai.li@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/ixgbevf/ixgbevf_main.c

index 805e02c..846ac51 100644 (file)
@@ -1138,12 +1138,12 @@ static int ixgbevf_vlan_rx_add_vid(struct net_device *netdev, u16 vid)
        if (!hw->mac.ops.set_vfta)
                return -EOPNOTSUPP;
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        /* add VID to filter table */
        err = hw->mac.ops.set_vfta(hw, vid, 0, true);
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 
        /* translate error return types so error makes sense */
        if (err == IXGBE_ERR_MBX)
@@ -1163,13 +1163,13 @@ static int ixgbevf_vlan_rx_kill_vid(struct net_device *netdev, u16 vid)
        struct ixgbe_hw *hw = &adapter->hw;
        int err = -EOPNOTSUPP;
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        /* remove VID from filter table */
        if (hw->mac.ops.set_vfta)
                err = hw->mac.ops.set_vfta(hw, vid, 0, false);
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 
        clear_bit(vid, adapter->active_vlans);
 
@@ -1225,7 +1225,7 @@ static void ixgbevf_set_rx_mode(struct net_device *netdev)
        struct ixgbevf_adapter *adapter = netdev_priv(netdev);
        struct ixgbe_hw *hw = &adapter->hw;
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        /* reprogram multicast list */
        if (hw->mac.ops.update_mc_addr_list)
@@ -1233,7 +1233,7 @@ static void ixgbevf_set_rx_mode(struct net_device *netdev)
 
        ixgbevf_write_uc_addr_list(netdev);
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 }
 
 static void ixgbevf_napi_enable_all(struct ixgbevf_adapter *adapter)
@@ -1347,7 +1347,7 @@ static void ixgbevf_negotiate_api(struct ixgbevf_adapter *adapter)
                      ixgbe_mbox_api_unknown };
        int err = 0, idx = 0;
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        while (api[idx] != ixgbe_mbox_api_unknown) {
                err = ixgbevf_negotiate_api_version(hw, api[idx]);
@@ -1356,7 +1356,7 @@ static void ixgbevf_negotiate_api(struct ixgbevf_adapter *adapter)
                idx++;
        }
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 }
 
 static void ixgbevf_up_complete(struct ixgbevf_adapter *adapter)
@@ -1397,7 +1397,7 @@ static void ixgbevf_up_complete(struct ixgbevf_adapter *adapter)
 
        ixgbevf_configure_msix(adapter);
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        if (hw->mac.ops.set_rar) {
                if (is_valid_ether_addr(hw->mac.addr))
@@ -1406,7 +1406,7 @@ static void ixgbevf_up_complete(struct ixgbevf_adapter *adapter)
                        hw->mac.ops.set_rar(hw, 0, hw->mac.perm_addr, 0);
        }
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 
        clear_bit(__IXGBEVF_DOWN, &adapter->state);
        ixgbevf_napi_enable_all(adapter);
@@ -1430,12 +1430,12 @@ static int ixgbevf_reset_queues(struct ixgbevf_adapter *adapter)
        unsigned int num_rx_queues = 1;
        int err, i;
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        /* fetch queue configuration from the PF */
        err = ixgbevf_get_queues(hw, &num_tcs, &def_q);
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 
        if (err)
                return err;
@@ -1694,14 +1694,14 @@ void ixgbevf_reset(struct ixgbevf_adapter *adapter)
        struct ixgbe_hw *hw = &adapter->hw;
        struct net_device *netdev = adapter->netdev;
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        if (hw->mac.ops.reset_hw(hw))
                hw_dbg(hw, "PF still resetting\n");
        else
                hw->mac.ops.init_hw(hw);
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 
        if (is_valid_ether_addr(adapter->hw.mac.addr)) {
                memcpy(netdev->dev_addr, adapter->hw.mac.addr,
@@ -2195,12 +2195,12 @@ static void ixgbevf_watchdog_task(struct work_struct *work)
        if (hw->mac.ops.check_link) {
                s32 need_reset;
 
-               spin_lock(&adapter->mbx_lock);
+               spin_lock_bh(&adapter->mbx_lock);
 
                need_reset = hw->mac.ops.check_link(hw, &link_speed,
                                                    &link_up, false);
 
-               spin_unlock(&adapter->mbx_lock);
+               spin_unlock_bh(&adapter->mbx_lock);
 
                if (need_reset) {
                        adapter->link_up = link_up;
@@ -2468,12 +2468,12 @@ static int ixgbevf_setup_queues(struct ixgbevf_adapter *adapter)
        unsigned int num_rx_queues = 1;
        int err, i;
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        /* fetch queue configuration from the PF */
        err = ixgbevf_get_queues(hw, &num_tcs, &def_q);
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 
        if (err)
                return err;
@@ -3047,12 +3047,12 @@ static int ixgbevf_set_mac(struct net_device *netdev, void *p)
        memcpy(netdev->dev_addr, addr->sa_data, netdev->addr_len);
        memcpy(hw->mac.addr, addr->sa_data, netdev->addr_len);
 
-       spin_lock(&adapter->mbx_lock);
+       spin_lock_bh(&adapter->mbx_lock);
 
        if (hw->mac.ops.set_rar)
                hw->mac.ops.set_rar(hw, 0, hw->mac.addr, 0);
 
-       spin_unlock(&adapter->mbx_lock);
+       spin_unlock_bh(&adapter->mbx_lock);
 
        return 0;
 }