driver core: fix potential deadlock in __driver_attach
authorZhang Wensheng <zhangwensheng5@huawei.com>
Wed, 22 Jun 2022 07:43:27 +0000 (15:43 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 27 Jun 2022 14:43:51 +0000 (16:43 +0200)
commit70fe758352cafdee72a7b13bf9db065f9613ced8
tree8039f0567d8638a8829fbe8dfe20e2192d8aef6d
parent1d248d2302da5b96e06281a52056eeafec0d2e11
driver core: fix potential deadlock in __driver_attach

In __driver_attach function, There are also AA deadlock problem,
like the commit b232b02bf3c2 ("driver core: fix deadlock in
__device_attach").

stack like commit b232b02bf3c2 ("driver core: fix deadlock in
__device_attach").
list below:
    In __driver_attach function, The lock holding logic is as follows:
    ...
    __driver_attach
    if (driver_allows_async_probing(drv))
      device_lock(dev)      // get lock dev
        async_schedule_dev(__driver_attach_async_helper, dev); // func
          async_schedule_node
            async_schedule_node_domain(func)
              entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC);
              /* when fail or work limit, sync to execute func, but
                 __driver_attach_async_helper will get lock dev as
                 will, which will lead to A-A deadlock.  */
              if (!entry || atomic_read(&entry_count) > MAX_WORK) {
                func;
              else
                queue_work_node(node, system_unbound_wq, &entry->work)
      device_unlock(dev)

    As above show, when it is allowed to do async probes, because of
    out of memory or work limit, async work is not be allowed, to do
    sync execute instead. it will lead to A-A deadlock because of
    __driver_attach_async_helper getting lock dev.

Reproduce:
and it can be reproduce by make the condition
(if (!entry || atomic_read(&entry_count) > MAX_WORK)) untenable, like
below:

[  370.785650] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables
this message.
[  370.787154] task:swapper/0       state:D stack:    0 pid:    1 ppid:
0 flags:0x00004000
[  370.788865] Call Trace:
[  370.789374]  <TASK>
[  370.789841]  __schedule+0x482/0x1050
[  370.790613]  schedule+0x92/0x1a0
[  370.791290]  schedule_preempt_disabled+0x2c/0x50
[  370.792256]  __mutex_lock.isra.0+0x757/0xec0
[  370.793158]  __mutex_lock_slowpath+0x1f/0x30
[  370.794079]  mutex_lock+0x50/0x60
[  370.794795]  __device_driver_lock+0x2f/0x70
[  370.795677]  ? driver_probe_device+0xd0/0xd0
[  370.796576]  __driver_attach_async_helper+0x1d/0xd0
[  370.797318]  ? driver_probe_device+0xd0/0xd0
[  370.797957]  async_schedule_node_domain+0xa5/0xc0
[  370.798652]  async_schedule_node+0x19/0x30
[  370.799243]  __driver_attach+0x246/0x290
[  370.799828]  ? driver_allows_async_probing+0xa0/0xa0
[  370.800548]  bus_for_each_dev+0x9d/0x130
[  370.801132]  driver_attach+0x22/0x30
[  370.801666]  bus_add_driver+0x290/0x340
[  370.802246]  driver_register+0x88/0x140
[  370.802817]  ? virtio_scsi_init+0x116/0x116
[  370.803425]  scsi_register_driver+0x1a/0x30
[  370.804057]  init_sd+0x184/0x226
[  370.804533]  do_one_initcall+0x71/0x3a0
[  370.805107]  kernel_init_freeable+0x39a/0x43a
[  370.805759]  ? rest_init+0x150/0x150
[  370.806283]  kernel_init+0x26/0x230
[  370.806799]  ret_from_fork+0x1f/0x30

To fix the deadlock, move the async_schedule_dev outside device_lock,
as we can see, in async_schedule_node_domain, the parameter of
queue_work_node is system_unbound_wq, so it can accept concurrent
operations. which will also not change the code logic, and will
not lead to deadlock.

Fixes: ef0ff68351be ("driver core: Probe devices asynchronously instead of the driver")
Signed-off-by: Zhang Wensheng <zhangwensheng5@huawei.com>
Link: https://lore.kernel.org/r/20220622074327.497102-1-zhangwensheng5@huawei.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/base/dd.c