Re: [RFC 0/4] Fix hci_dev ref-counts

On Mon, Oct 31, 2011 at 12:58 PM, Marcel Holtmann [off-list ref] wrote:

Hi David,

quoted

We currently have two reference counters for each hci_dev. This patchset tries
to reduce this to one ref-count and fix several bugs with them. The two
available ref-counts currently are:

1) hci_dev->dev internal
This ref-count is increased by hci_alloc_dev() and decreased inside
hci_free_dev(). No other code currently uses this refcount. It can be set with
get_device() and put_device().
When this ref-count drops zero, then the hci_dev structure is deallocated (see
bt_link_release inside hci_sysfs.c), hence, this ref-count is the most important
one. However, we currently do not use it correctly.

2) hci_dev->refcnt
This ref-count is used inside hci_dev_hold/put() and practically used as if it
would protect the hci_dev structure. However, if this ref-count drops zero, then
the ->destruct callback is called which is used by all drivers *exclusively* to
deallocate driver structures and *not* to deallocate or destroy the hci_dev.

actually I think we should convert the tasklets into one workqueue first
and then integrate deeper into sysfs before trying to hack around this
problem.

I am currently playing around with several patches for that but I
still need to get away from that single-threaded workqueue as it is
quite ugly.

Our core driver registration code is from the 2.4 series and needs major
updating. The sysfs code is glued on top of it and not really well
integrated. So fixing that one properly might make some of these race
conditions go away.

Ok. I'll will reconsider sending a new RFC when we fixed the workqueue
stuff. As our system is a simple bus-system, it shouldn't be too hard
to do that.

Regards

Marcel

Regards
David