Re: [PATCH v6] uio: Fix uio driver to refcount device

From: Brian Russell
Date: Thu Mar 19 2015 - 12:19:43 EST




On 19/03/15 15:14, Greg Kroah-Hartman wrote:
> On Thu, Mar 19, 2015 at 02:49:21PM +0000, Brian Russell wrote:
>> On 19/03/15 14:23, Greg Kroah-Hartman wrote:
>>> On Thu, Mar 19, 2015 at 02:11:19PM +0000, Brian Russell wrote:
>>>> if (info->irq && (info->irq != UIO_IRQ_CUSTOM)) {
>>>> - ret = devm_request_irq(idev->dev, info->irq, uio_interrupt,
>>>> + ret = request_irq(info->irq, uio_interrupt,
>>>> info->irq_flags, info->name, idev);
>>>
>>> Why move this away from the device lifecycle?
>>>
>>
>> I ran into a problem here: after device unregister the parent module (in my case igb_uio in DPDK) can call pci_disable_msi. From the PCI MSI how to:
>>
>> "Before calling this function, a device driver must always call free_irq()
>> on any interrupt for which it previously called request_irq().
>> Failure to do so results in a BUG_ON(), leaving the device with
>> MSI enabled and thus leaking its vector."
>>
>> So I need the free_irq, but the device may still have open fds and therefore be around a bit longer waiting for them to be released.
>
> Note, please wrap your email lines...
>
> Anyway, ok, that makes sense. But put in a big comment in here about
> this so that someone doesn't try to convert it to devm_* in the future.
>
> Can you do this in a separate patch first as well? Burrying it in a
> pointer conversion patch isn't good.
>

I've separated out the irq bits into a separate patch, which depends
on patching on top of the first patch. I'm a bit unsure how to go about
sending the update now the patches are broken out and only one featured
in the previous versions. Does this seek ok?

3 separate mails:

[PATCH v8 0/2] uio: Fix uio hotplug issues
summary of patches

[PATCH v8 1/2] uio: Fix uio driver to refcount device
original patch minus irq and include version changes

[PATCH v8 2/2] uio: Request/free irq separate from dev lifecycle
broken out irq bits, no other version info


... or should the v8 only go on the first patch?

Thanks,

Brian


>>>> err_device_create:
>>>> uio_free_minor(idev);
>>>> return ret;
>>>> @@ -871,8 +892,10 @@ void uio_unregister_device(struct uio_info *info)
>>>>
>>>> uio_dev_del_attributes(idev);
>>>>
>>>> - device_destroy(&uio_class, MKDEV(uio_major, idev->minor));
>>>> + device_unregister(&idev->dev);
>>>> + free_irq(idev->info->irq, idev);
>>>>
>>>> + idev->info = NULL;
>>>
>>> Why? If this was the last reference count, isn't idev now gone? You
>>> shouldn't be referencing it anymore.
>>>
>>
>> No, the device can be unregistered here but still have outstanding refcount because of open fds. The count can get to zero in the release function, so idev could still be around after this but the calling module could free info.
>
> Yes, but if it was the last reference, idev is now a stale pointer and
> you just derefrenced it. Twice. :(
>
> Once you do a put, and don't have a reference on the pointer, you can't
> touch that pointer again.
>
> thanks,
>
> greg k-h
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/