On Thu, 14 Nov 2013 10:51:05 +0100, Pantelis Antoniou <panto@xxxxxxxxxxxxxxxxxxxxxxx> wrote:
> Hi Grant,
>
> On Nov 14, 2013, at 2:44 AM, Grant Likely wrote:
>
> > On Wed, 13 Nov 2013 10:03:37 +0100, Pantelis Antoniou <panto@xxxxxxxxxxxxxxxxxxxxxxx> wrote:
> >> On Nov 13, 2013, at 2:34 AM, Grant Likely wrote:
> >>> On Tue, 12 Nov 2013 11:39:08 +0100, Pantelis Antoniou <panto@xxxxxxxxxxxxxxxxxxxxxxx> wrote:
> >>>>> On Tue, 5 Nov 2013 19:50:16 +0200, Pantelis Antoniou <panto@xxxxxxxxxxxxxxxxxxxxxxx> wrote:
> >>>>>> + } else {
> >>>>>> + pr_warn("%s: node %p cannot be freed; memory is gone\n",
> >>>>>> + __func__, node);
> >>>>>> + }
> >>>>>> +}
> >>>>>
> >>>>> All of the above is potentially dangerous. There is no way to determine
> >>>>> if anything still holds a reference to a node. The proper way to handle
> >>>>> removal of properties is to have a release method when the last
> >>>>> of_node_put is called.
> >>>>>
> >>>>
> >>>> This is safe, and expected to be called only on a dynamically created tree,
> >>>> that's what all the checks against OF_DYNAMIC guard against.
> >>>>
> >>>> It is not ever meant to be called on an arbitrary tree, created by unflattening
> >>>> a blob.
> >>>
> >>> I am talking about when being used on a dynamic tree. The problem is
> >>> when a driver or other code holds a reference to a dynamic nodes, but
> >>> doesn't release it correctly. The memory must not be freed until all of
> >>> the references are relased. OF_DYNAMIC doesn't actually help in that
> >>> case, and it is the reason for of_node_get()/of_node_put()
> >>>
> >>
> >> I know, but even that is not enough. of_node_get()/of_node_put() handles the
> >> case of references to the nodes, but not what happens with references to
> >> properties. deadprops is mitigating the problem somewhat, but if we're going
> >> to go to all the trouble of kobjectification let's do the props as well.
> >>
> >> of_get_property could be modified to return a devm_kmalloced copy of the real
> >> property and that would deal with most of the callers. Of course for
> >> the small sized scalar data we can avoid the copy.
> >>
> >> By using the devm_* interface we also avoid having to mess too much with the callers.
> >>
> >> I.e. what about something like devm_of_get_property()?
> >
> > Reference counting is already a horrible pain to keep correct. I don't
> > see a better way to handle it in the dynamic case, so we're stuck with
> > it, but I don't want to make it any harder. Adding ref counting to
> > properties will make it harder than it already is to get the code right.
> > I'm absolutely fine with a little bit of wasted memory in the form of
> > deadprops when the alternative is so horrible. References at the node
> > level is enough granularity.
> >
> > I don't think kduping the property is the solution either. I strongly
> > suspect that will be far more expensive than the deadprop solution.
> >
>
> As long as we can live with deadprops all is fine. Perhaps a devm_of_get_property()
> makes sense for new drivers though? What do you think? Perhaps copying to a
> user supplied buffer as well?
I still don't think it is necessary. The device lifetime should always
be shorter than the node lifetime.
> It's a kind of drag. That means you get handed a device_node pointer you are not
> able to free it without having the blob along with the container/accessor of it.
> I.e. For the normal case where the blob comes from a request_firmware() call
> You have to keep the firmware structure around.
>
> Depending on what other method you're going to use tends to make the code a little
> bit messier.
Understood. Stick with keeping the blob around for now. It can be
reworkd in the future if necessary since there are no associated
userspace ABI issues.
g.
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
