ice: Fix race condition during interface enslave

Commit 5dbbbd01 ("ice: Avoid RTNL lock when re-creating
auxiliary device") changes a process of re-creation of aux device
so ice_plug_aux_dev() is called from ice_service_task() context.
This unfortunately opens a race window that can result in dead-lock
when interface has left LAG and immediately enters LAG again.

Reproducer:
```
#!/bin/sh

ip link add lag0 type bond mode 1 miimon 100
ip link set lag0

for n in {1..10}; do
        echo Cycle: $n
        ip link set ens7f0 master lag0
        sleep 1
        ip link set ens7f0 nomaster
done
```

This results in:
[20976.208697] Workqueue: ice ice_service_task [ice]
[20976.213422] Call Trace:
[20976.215871]  __schedule+0x2d1/0x830
[20976.219364]  schedule+0x35/0xa0
[20976.222510]  schedule_preempt_disabled+0xa/0x10
[20976.227043]  __mutex_lock.isra.7+0x310/0x420
[20976.235071]  enum_all_gids_of_dev_cb+0x1c/0x100 [ib_core]
[20976.251215]  ib_enum_roce_netdev+0xa4/0xe0 [ib_core]
[20976.256192]  ib_cache_setup_one+0x33/0xa0 [ib_core]
[20976.261079]  ib_register_device+0x40d/0x580 [ib_core]
[20976.266139]  irdma_ib_register_device+0x129/0x250 [irdma]
[20976.281409]  irdma_probe+0x2c1/0x360 [irdma]
[20976.285691]  auxiliary_bus_probe+0x45/0x70
[20976.289790]  really_probe+0x1f2/0x480
[20976.298509]  driver_probe_device+0x49/0xc0
[20976.302609]  bus_for_each_drv+0x79/0xc0
[20976.306448]  __device_attach+0xdc/0x160
[20976.310286]  bus_probe_device+0x9d/0xb0
[20976.314128]  device_add+0x43c/0x890
[20976.321287]  __auxiliary_device_add+0x43/0x60
[20976.325644]  ice_plug_aux_dev+0xb2/0x100 [ice]
[20976.330109]  ice_service_task+0xd0c/0xed0 [ice]
[20976.342591]  process_one_work+0x1a7/0x360
[20976.350536]  worker_thread+0x30/0x390
[20976.358128]  kthread+0x10a/0x120
[20976.365547]  ret_from_fork+0x1f/0x40
...
[20976.438030] task:ip              state:D stack:    0 pid:213658 ppid:213627 flags:0x00004084
[20976.446469] Call Trace:
[20976.448921]  __schedule+0x2d1/0x830
[20976.452414]  schedule+0x35/0xa0
[20976.455559]  schedule_preempt_disabled+0xa/0x10
[20976.460090]  __mutex_lock.isra.7+0x310/0x420
[20976.464364]  device_del+0x36/0x3c0
[20976.467772]  ice_unplug_aux_dev+0x1a/0x40 [ice]
[20976.472313]  ice_lag_event_handler+0x2a2/0x520 [ice]
[20976.477288]  notifier_call_chain+0x47/0x70
[20976.481386]  __netdev_upper_dev_link+0x18b/0x280
[20976.489845]  bond_enslave+0xe05/0x1790 [bonding]
[20976.494475]  do_setlink+0x336/0xf50
[20976.502517]  __rtnl_newlink+0x529/0x8b0
[20976.543441]  rtnl_newlink+0x43/0x60
[20976.546934]  rtnetlink_rcv_msg+0x2b1/0x360
[20976.559238]  netlink_rcv_skb+0x4c/0x120
[20976.563079]  netlink_unicast+0x196/0x230
[20976.567005]  netlink_sendmsg+0x204/0x3d0
[20976.570930]  sock_sendmsg+0x4c/0x50
[20976.574423]  ____sys_sendmsg+0x1eb/0x250
[20976.586807]  ___sys_sendmsg+0x7c/0xc0
[20976.606353]  __sys_sendmsg+0x57/0xa0
[20976.609930]  do_syscall_64+0x5b/0x1a0
[20976.613598]  entry_SYSCALL_64_after_hwframe+0x65/0xca

1. Command 'ip link ... set nomaster' causes that ice_plug_aux_dev()
   is called from ice_service_task() context, aux device is created
   and associated device->lock is taken.
2. Command 'ip link ... set master...' calls ice's notifier under
   RTNL lock and that notifier calls ice_unplug_aux_dev(). That
   function tries to take aux device->lock but this is already taken
   by ice_plug_aux_dev() in step 1
3. Later ice_plug_aux_dev() tries to take RTNL lock but this is already
   taken in step 2
4. Dead-lock

The patch fixes this issue by following changes:
- Bit ICE_FLAG_PLUG_AUX_DEV is kept to be set during ice_plug_aux_dev()
  call in ice_service_task()
- The bit is checked in ice_clear_rdma_cap() and only if it is not set
  then ice_unplug_aux_dev() is called. If it is set (in other words
  plugging of aux device was requested and ice_plug_aux_dev() is
  potentially running) then the function only clears the bit
- Once ice_plug_aux_dev() call (in ice_service_task) is finished
  the bit ICE_FLAG_PLUG_AUX_DEV is cleared but it is also checked
  whether it was already cleared by ice_clear_rdma_cap(). If so then
  aux device is unplugged.
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Co-developed-by: Petr Oros <poros@redhat.com>
Signed-off-by: Petr Oros <poros@redhat.com>
Reviewed-by: Dave Ertman <david.m.ertman@intel.com>
Link: https://lore.kernel.org/r/20220310171641.3863659-1-ivecera@redhat.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

drivers/net/ethernet/intel/ice/ice.h

@@ -898,7 +898,16 @@ static inline void ice_set_rdma_cap(struct ice_pf *pf)
  */
 static inline void ice_clear_rdma_cap(struct ice_pf *pf)
 {
-	ice_unplug_aux_dev(pf);
+	/* We can directly unplug aux device here only if the flag bit
+	 * ICE_FLAG_PLUG_AUX_DEV is not set because ice_unplug_aux_dev()
+	 * could race with ice_plug_aux_dev() called from
+	 * ice_service_task(). In this case we only clear that bit now and
+	 * aux device will be unplugged later once ice_plug_aux_device()
+	 * called from ice_service_task() finishes (see ice_service_task()).
+	 */
+	if (!test_and_clear_bit(ICE_FLAG_PLUG_AUX_DEV, pf->flags))
+		ice_unplug_aux_dev(pf);
+
 	clear_bit(ICE_FLAG_RDMA_ENA, pf->flags);
 	clear_bit(ICE_FLAG_AUX_ENA, pf->flags);
 }

drivers/net/ethernet/intel/ice/ice_main.c

@@ -2255,9 +2255,19 @@ static void ice_service_task(struct work_struct *work)
 		return;
 	}
 
-	if (test_and_clear_bit(ICE_FLAG_PLUG_AUX_DEV, pf->flags))
+	if (test_bit(ICE_FLAG_PLUG_AUX_DEV, pf->flags)) {
+		/* Plug aux device per request */
 		ice_plug_aux_dev(pf);
 
+		/* Mark plugging as done but check whether unplug was
+		 * requested during ice_plug_aux_dev() call
+		 * (e.g. from ice_clear_rdma_cap()) and if so then
+		 * plug aux device.
+		 */
+		if (!test_and_clear_bit(ICE_FLAG_PLUG_AUX_DEV, pf->flags))
+			ice_unplug_aux_dev(pf);
+	}
+
 	if (test_and_clear_bit(ICE_FLAG_MTU_CHANGED, pf->flags)) {
 		struct iidc_event *event;