ks8851: Fix mutex deadlock in ks8851_net_stop()
authorStephen Boyd <sboyd@codeaurora.org>
Wed, 18 Apr 2012 17:25:58 +0000 (17:25 +0000)
committerDavid S. Miller <davem@davemloft.net>
Sat, 21 Apr 2012 19:33:14 +0000 (15:33 -0400)
commitc5a99937a9cf74a623384023201a7d98b51e7e3b
tree48f114446b6e0c6dde76a6b0bd0aa7fae9af76bf
parent3adadc08cc1e2cbcc15a640d639297ef5fcb17f5
ks8851: Fix mutex deadlock in ks8851_net_stop()

There is a potential deadlock scenario when the ks8851 driver
is removed. The interrupt handler schedules a workqueue which
acquires a mutex that ks8851_net_stop() also acquires before
flushing the workqueue. Previously lockdep wouldn't be able
to find this problem but now that it has the support we can
trigger this lockdep warning by rmmoding the driver after
an ifconfig up.

Fix the possible deadlock by disabling the interrupts in
the chip and then release the lock across the workqueue
flushing. The mutex is only there to proect the registers
anyway so this should be ok.

=======================================================
[ INFO: possible circular locking dependency detected ]
3.0.21-00021-g8b33780-dirty #2911
-------------------------------------------------------
rmmod/125 is trying to acquire lock:
 ((&ks->irq_work)){+.+...}, at: [<c019e0b8>] flush_work+0x0/0xac

but task is already holding lock:
 (&ks->lock){+.+...}, at: [<bf00b850>] ks8851_net_stop+0x64/0x138 [ks8851]

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 (&ks->lock){+.+...}:
       [<c01b89c8>] __lock_acquire+0x940/0x9f8
       [<c01b9058>] lock_acquire+0x10c/0x130
       [<c083dbec>] mutex_lock_nested+0x68/0x3dc
       [<bf00bd48>] ks8851_irq_work+0x24/0x46c [ks8851]
       [<c019c580>] process_one_work+0x2d8/0x518
       [<c019cb98>] worker_thread+0x220/0x3a0
       [<c01a2ad4>] kthread+0x88/0x94
       [<c0107008>] kernel_thread_exit+0x0/0x8

-> #0 ((&ks->irq_work)){+.+...}:
       [<c01b7984>] validate_chain+0x914/0x1018
       [<c01b89c8>] __lock_acquire+0x940/0x9f8
       [<c01b9058>] lock_acquire+0x10c/0x130
       [<c019e104>] flush_work+0x4c/0xac
       [<bf00b858>] ks8851_net_stop+0x6c/0x138 [ks8851]
       [<c06b209c>] __dev_close_many+0x98/0xcc
       [<c06b2174>] dev_close_many+0x68/0xd0
       [<c06b22ec>] rollback_registered_many+0xcc/0x2b8
       [<c06b2554>] rollback_registered+0x28/0x34
       [<c06b25b8>] unregister_netdevice_queue+0x58/0x7c
       [<c06b25f4>] unregister_netdev+0x18/0x20
       [<bf00c1f4>] ks8851_remove+0x64/0xb4 [ks8851]
       [<c049ddf0>] spi_drv_remove+0x18/0x1c
       [<c0468e98>] __device_release_driver+0x7c/0xbc
       [<c0468f64>] driver_detach+0x8c/0xb4
       [<c0467f00>] bus_remove_driver+0xb8/0xe8
       [<c01c1d20>] sys_delete_module+0x1e8/0x27c
       [<c0105ec0>] ret_fast_syscall+0x0/0x3c

other info that might help us debug this:

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&ks->lock);
                               lock((&ks->irq_work));
                               lock(&ks->lock);
  lock((&ks->irq_work));

 *** DEADLOCK ***

4 locks held by rmmod/125:
 #0:  (&__lockdep_no_validate__){+.+.+.}, at: [<c0468f44>] driver_detach+0x6c/0xb4
 #1:  (&__lockdep_no_validate__){+.+.+.}, at: [<c0468f50>] driver_detach+0x78/0xb4
 #2:  (rtnl_mutex){+.+.+.}, at: [<c06b25e8>] unregister_netdev+0xc/0x20
 #3:  (&ks->lock){+.+...}, at: [<bf00b850>] ks8851_net_stop+0x64/0x138 [ks8851]

Cc: Ben Dooks <ben-linux@fluff.org>
Signed-off-by: Stephen Boyd <sboyd@codeaurora.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/micrel/ks8851.c