smpboot: make cleanup to mirror setup
authorFrederic Weisbecker <fweisbec@gmail.com>
Fri, 4 Sep 2015 22:45:03 +0000 (15:45 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 4 Sep 2015 23:54:41 +0000 (16:54 -0700)
commit3dd08c0c918f9bf058572ddbf26e7d6fb5674a5c
treedb82da2bf13515643d5b36cf6f2f0584f1355b97
parent5869b5064b0950afc447610c0f5d4134b71d9e94
smpboot: make cleanup to mirror setup

The per-cpu kthread cleanup() callback is the mirror of the setup()
callback.  When the per-cpu kthread is started, it first calls setup()
to initialize the resources which are then released by cleanup() when
the kthread exits.

Now since the introduction of a per-cpu kthread cpumask, the kthreads
excluded by the cpumask on boot may happen to be parked immediately
after their creation without taking the setup() stage, waiting to be
asked to unpark to do so.  Then when smpboot_unregister_percpu_thread()
is later called, the kthread is stopped without having ever called
setup().

But this triggers a bug as the kthread unconditionally calls cleanup()
on exit but this doesn't mirror any setup().  Thus the kernel crashes
because we try to free resources that haven't been initialized, as in
the watchdog case:

    WATCHDOG disable 0
    WATCHDOG disable 1
    WATCHDOG disable 2
    BUG: unable to handle kernel NULL pointer dereference at           (null)
    IP: hrtimer_active+0x26/0x60
    [...]
    Call Trace:
      hrtimer_try_to_cancel+0x1c/0x280
      hrtimer_cancel+0x1d/0x30
      watchdog_disable+0x56/0x70
      watchdog_cleanup+0xe/0x10
      smpboot_thread_fn+0x23c/0x2c0
      kthread+0xf8/0x110
      ret_from_fork+0x3f/0x70

This bug is currently masked with explicit kthread unparking before
kthread_stop() on smpboot_destroy_threads(). This forces a call to
setup() and then unpark().

We could fix this by unconditionally calling setup() on kthread entry.
But setup() isn't always cheap.  In the case of watchdog it launches
hrtimer, perf events, etc...  So we may as well like to skip it if there
are chances the kthread will never be used, as in a reduced cpumask value.

So let's simply do a state machine check before calling cleanup() that
makes sure setup() has been called before mirroring it.

And remove the nasty hack workaround.

Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Chris Metcalf <cmetcalf@ezchip.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ulrich Obergfell <uobergfe@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
kernel/smpboot.c