On 29/09/14 16:40, Konrad Rzeszutek Wilk wrote:
> On Mon, Sep 29, 2014 at 03:17:10PM +0100, David Vrabel wrote:
>> On 29/09/14 15:02, Konrad Rzeszutek Wilk wrote:
>>> On Sat, Sep 27, 2014 at 12:36:42AM +0800, Chen Gang wrote:
>>>> When xenbus_switch_state() fails, it will call xenbus_switch_fatal()
>>>
>>> Only on the first depth, not on the subsequent ones (as in if
>>> the first xenbus_switch_fail fails, it won't try to call
>>> xenbus_switch_state again and again).
>>>
>>>> internally, so need not return any status value, then use 'void' instead
>>>> of 'int' for xenbus_switch_state() and __xenbus_switch_state().
>>>
>>> When that switch occurs (to XenbusStateConnected) won't the watches
>>> fire - meaning we MUST make sure that the watch functions - if they
>>> use the xenbus_switch_state() they MUST not hold any locks - because
>>> they could be executed once more?
>>>
>>> Oh wait, we don't have to worry about that right now as the callbacks
>>> that pick up the messages from the XenBus are all gated on one mutex
>>> anyhow.
>>>
>>> Hm, anyhow, I would add this extra piece of information to the patch:
>>>
>>>
>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>> index c214daa..f7399fd 100644
>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>> @@ -661,6 +661,12 @@ static void xen_pcibk_be_watch(struct xenbus_watch *watch,
>>>
>>> switch (xenbus_read_driver_state(pdev->xdev->nodename)) {
>>> case XenbusStateInitWait:
>>> + /*
>>> + * xenbus_switch_state can call xenbus_switch_fatal which will
>>> + * immediately set the state to XenbusStateClosing which
>>> + * means if we were reading for it here we MUST drop any
>>> + * locks so that we don't dead-lock.
>>> + */
>>
>> Watches are asynchronous and serialised by the xenwatch thread. I can't
>> see what deadlock you're talking about here. Particularly since the
>> backend doesn't watch its own state node (it watches the frontend one).
>>
>>> xen_pcibk_setup_backend(pdev);
>>> break;
>>>
>>>>
>>>> Also need be sure that all callers which check the return value must let
>>>> 'err' be 0.
>>>
>>> I am bit uncomfortable with that, that is due to:
>>>
>>>
>>> .. snip..
>>>> diff --git a/drivers/net/xen-netback/xenbus.c b/drivers/net/xen-netback/xenbus.c
>>>> index 9c47b89..b5c3d47 100644
>>>> --- a/drivers/net/xen-netback/xenbus.c
>>>> +++ b/drivers/net/xen-netback/xenbus.c
>>>> @@ -337,10 +337,7 @@ static int netback_probe(struct xenbus_device *dev,
>>>> if (err)
>>>> pr_debug("Error writing multi-queue-max-queues\n");
>>>>
>>>> - err = xenbus_switch_state(dev, XenbusStateInitWait);
>>>> - if (err)
>>>> - goto fail;
>>>> -
>>>> + xenbus_switch_state(dev, XenbusStateInitWait);
>>>
>>> Which if it fails it won't call:
>>>
>>> 354 fail:
>>> 355 pr_debug("failed\n");
>>> 356 netback_remove(dev);
>>> 357 return err;
>>>
>>>
>>> And since there is no watch on the backend state to go in Closing it won't
>>> ever call those and we leak memory.
>>
>> It's not leaking the memory. All resources will be recovered when the
>> device is removed.
>
> I presume you mean when the XenBus entries are torn down? It does look
> like it would call the .remove functionality. That should take care of that.
>
> In which case we can just remove all of the 'netback_remove()' and also
> remove some of the labels.
No. If the final xenbus_switch_state() fails then at least the device
is in a consistent state, waiting for the other end to notice.
We don't want to return success from a probe with a half-setup device.
David
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
