implement in-kernel gendisk events handling

Currently, media presence polling for removeable block devices is done
from userland.  There are several issues with this.

* Polling is done by periodically opening the device.  For SCSI
  devices, the command sequence generated by such action involves a
  few different commands including TEST_UNIT_READY.  This behavior,
  while perfectly legal, is different from Windows which only issues
  single command, GET_EVENT_STATUS_NOTIFICATION.  Unfortunately, some
  ATAPI devices lock up after being periodically queried such command
  sequences.

* There is no reliable and unintrusive way for a userland program to
  tell whether the target device is safe for media presence polling.
  For example, polling for media presence during an on-going burning
  session can make it fail.  The polling program can avoid this by
  opening the device with O_EXCL but then it risks making a valid
  exclusive user of the device fail w/ -EBUSY.

* Userland polling is unnecessarily heavy and in-kernel implementation
  is lighter and better coordinated (workqueue, timer slack).

This patch implements framework for in-kernel disk event handling,
which includes media presence polling.

* bdops->check_events() is added, which supercedes ->media_changed().
  It should check whether there's any pending event and return if so.
  Currently, two events are defined - DISK_EVENT_MEDIA_CHANGE and
  DISK_EVENT_EJECT_REQUEST.  ->check_events() is guaranteed not to be
  called parallelly.

* gendisk->events and ->async_events are added.  These should be
  initialized by block driver before passing the device to add_disk().
  The former contains the mask of all supported events and the latter
  the mask of all events which the device can report without polling.
  /sys/block/*/events[_async] export these to userland.

* Kernel parameter block.events_dfl_poll_msecs controls the system
  polling interval (default is 0 which means disable) and
  /sys/block/*/events_poll_msecs control polling intervals for
  individual devices (default is -1 meaning use system setting).  Note
  that if a device can report all supported events asynchronously and
  its polling interval isn't explicitly set, the device won't be
  polled regardless of the system polling interval.

* If a device is opened exclusively with write access, event checking
  is automatically disabled until all write exclusive accesses are
  released.

* There are event 'clearing' events.  For example, both of currently
  defined events are cleared after the device has been successfully
  opened.  This information is passed to ->check_events() callback
  using @clearing argument as a hint.

* Event checking is always performed from system_nrt_wq and timer
  slack is set to 25% for polling.

* Nothing changes for drivers which implement ->media_changed() but
  not ->check_events().  Going forward, all drivers will be converted
  to ->check_events() and ->media_change() will be dropped.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Kay Sievers <kay.sievers@vrfy.org>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>