[dpdk-dev] [PATCH v8 04/19] ethdev: introduce device lock

Thomas Monjalon thomas at monjalon.net
Wed Jul 4 09:27:02 CEST 2018


04/07/2018 03:47, Zhang, Qi Z:
> From: Thomas Monjalon [mailto:thomas at monjalon.net]
> > 03/07/2018 17:08, Zhang, Qi Z:
> > > From: Thomas Monjalon [mailto:thomas at monjalon.net]
> > > > 02/07/2018 07:44, Qi Zhang:
> > > > > Introduce API rte_eth_dev_lock and rte_eth_dev_unlock to let
> > > > > application lock or unlock on specific ethdev, a locked device
> > > > > can't be detached, this help applicaiton to prevent unexpected
> > > > > device detaching, especially in multi-process envrionment.
> > > >
> > > > Trying to understand: a process of an application could try to
> > > > detach a port while another process is against this decision.
> > > > Why an application needs to be protected against itself?
> > >
> > > I think we can regard this as a help function, it help application to simplified
> > the situation when one process want to detach a device while another one is
> > still using it.
> > > Application can register a callback which can do to necessary clean up (like
> > stop traffic, release memory ...) before device be detached.
> > 
> > Yes I agree such hook can be a good idea.
> > 
> > 
> > > > I guess it is only an application inter-process management.
> > > > If we really want to provide such helper in DPDK, it should not be
> > > > limited to ethdev.
> > >
> > > Once we move to eal layer, we will have rte_eal_dev_lock/unlock(devname,
> > busname).
> > > But its also better we keep rte_eth_dev_lock/unlock to make ethdev API
> > > more completed (any port be locked means underline rte_device also be
> > locked?) and this is same for other device family.
> > 
> > No. Again, a port is not exactly a device.
> > There can be several ports per device.
> 
> Yes, I know that.
> what I mean is, we should assume lock any port of that rte_device will prevent the device be detached.
> 
> > 
> > I think the right place for this hook is in port-level API (ethdev, crypto, etc). And
> > as we improve only ethdev currently, without any common genericity for other
> > device classes, it is probably fine in ethdev for now.
> > >
> > > > (for info, see class implementation:
> > > > https://patches.dpdk.org/patch/41605/)
> > > >
> > > > What about hardware unplug?
> > > > Can we detach the locked ports associated to the unplugged device?
> > >
> > > NO, we can't.
> > > But do you think, we need to introduce a "force detach" version, which will
> > ignore all locks.
> > 
> > No, I don't think so.
> > I am just trying to show that you cannot really "lock" a port.
> > Maybe you should just rename those functions.
> > After all, it is just a pre-detach hook.
> 
> > Wait, how is it different of RTE_ETH_EVENT_DESTROY callback?
> > Perhaps we just need to improve the handling of the DESTROY event?
> 
> I have thought about this before.
> Not like RTE_ETH_EVENT_DESTROY and other event hook, the hook here need to give feedback, pass or fail will impact the following behavior, this make it special, so I separate it from all exist rte_eth_event_type handle mechanism. 

Look at _rte_eth_dev_callback_process, there is a "ret_param".

> The alternative solution is 
> we just introduce a new event type like RTE_ETH_EVENT_PRE_DETACH and reuse all exist API 
> rte_eth_dev_callback_register/rte_eth_dev_callback_unregister.

I don't think we need a new event.
Let's try to use RTE_ETH_EVENT_DESTROY.

> But in _rte_eth_dev_callback_process we need to add a code branch for PRE_DETACH handle.
> 
> If (event = RTE_ETH_EVENT_PRE_DETACH)
> 	<...>.
> else {
> 	<....>
> }
> 
> And we may also need to keep rte_eth_dev_lock/unlock which will register a default callback for PRE_DETACH.

The default callback can be registered by the application.

> What do you think about?





More information about the dev mailing list