On 09/09/2013 12:24 PM, Laurent Pinchart wrote: > Hi Hans, > > On Monday 09 September 2013 12:17:34 Hans Verkuil wrote: >> On 09/09/2013 12:10 PM, Laurent Pinchart wrote: >>> On Monday 09 September 2013 12:07:18 Hans Verkuil wrote: >>>> On 09/09/2013 12:00 PM, Laurent Pinchart wrote: >>>>> On Monday 09 September 2013 11:07:43 Hans Verkuil wrote: >>>>>> On 09/06/2013 12:33 AM, Sylwester Nawrocki wrote: >>>>> [snip] >>>>> >>>>>>> The main issue as I see it is that we need to track both driver >>>>>>> remove() and struct device .release() calls and free resources only >>>>>>> when last of them executes. Data structures which are referenced in >>>>>>> fops must not be freed in remove() and we cannot use dev_get_drvdata() >>>>>>> in fops, e.g. not protected with device_lock(). >>>>>> >>>>>> You can do all that by returning 0 if probe() was partially successful >>>>>> (i.e. one or more, but not all, nodes were created successfully) by >>>>>> doing what I described above. I don't see another way that doesn't >>>>>> introduce a race condition. >>>>> >>>>> But isn't this just plain wrong ? If probing fails, I don't see how >>>>> returning success could be a good idea. >>>> >>>> Well, the nodes that are created are working fine. So it's partially OK >>>> :-) >>>> >>>> That said, yes it would be better if it could safely clean up and return >>>> an error. But it is better than returning an error and introducing a race >>>> condition. >>>> >>>>>> That doesn't mean that there isn't one, it's just that I don't know of >>>>>> a better way of doing this. >>>>> >>>>> We might need support from the device core. >>>> >>>> I do come back to my main question: has anyone actually experienced this >>>> error in a realistic scenario? Other than in very low-memory situations I >>>> cannot imagine this happening. >>> >>> What about running out of minors, which could very well happen with subdev >>> nodes in complex SoCs ? >> >> Is that really realistic? What's the worst-case SoC we have in terms of >> device nodes? Frankly, if this might happen then we should allow for more >> minors or make the minor allocation completely dynamic. > > For the 4 VSP1 instances on the R-Car H2, I need 33 video nodes and 65 (if I'm > not mistaken) subdev nodes. That doesn't include the camera interface. So that leaves 158 free minors. That's a lot of webcams that can be attached :-) > On a side note, this seems to indicate that the subdev API should probably > move to the /dev/media device node. That's something else to discuss. I have 70 tty nodes in /dev. Just because there are a lot of them doesn't mean that we have to merge them somehow. There may be other arguments for changing how we handle them, but 'there are a lot of them' isn't a good argument, IMHO. Regards, Hans >> If you run into this situation then you have bigger problems than a >> potential race condition. > -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
