[dpdk-dev] [PATCH v3 7/7] app/testpmd: adjust ethdev port ownership

[Matan Azrad]

Hi

From: Gaëtan Rivet [mailto:gaetan.rivet at 6wind.com]
<snip>
> > > > > > Look,
> > > > > > > Testpmd initiates some of its internal databases depends on
> > > > > > > specific port iteration, In some time someone may take
> > > > > > > ownership of Testpmd ports and testpmd will continue to touch
> them.
> > > > >
> > > > > But if someone will take the ownership (assign new owner_id)
> > > > > that port will not appear in RTE_ETH_FOREACH_DEV() any more.
> > > > >
> > > >
> > > > Yes, but testpmd sometimes depends on previous iteration using
> internal database.
> > > > So it uses internal database that was updated by old iteration.
> > >
> > > That sounds like just a bug in testpmd that need to be fixed, no?
> >
> > If Testpmd already took ownership for these ports(like I did), it is ok.
> >
> 
> Have you tested using the default iterator (NO_OWNER)?
> It worked until now with the bare minimal device tagging using
> DEV_DEFERRED. Testpmd did not seem to mind having to skip this port.
> 
> I'm sure there were places where this was overlooked, but overall, I'd think
> everything should be fixable using only the NO_OWNER iteration.

I don't think so.

> Can you point to a specific scenario (command line, chain of event) that
> would lead to a problem?
>

I didn't construct a race test to catch testpmd issue, but I think without this patch, there is a lot of issues.
Go to the testpmd code (before ownership) and find usage of the old iterator(after the first iteration in main),
Ask yourself what should happen if exactly in this time, a new port is created by fail-safe(plug in event).
 
> > > Any particular places where outdated device info is used?
> >
> > For example, look for the stream management in testpmd(I think I saw it
> there).
> >
> 
> The stream management is certainly shaky, but it happens after the EAL
> initial port creation, and is not able to update itself for new hotplugged ports
> (unless something changed).
> 

Yes, but conceptually someone in the future may take the port(because it ownerless).

> > > > > > If I look back on the fail-safe, its sole purpose is to have
> > > > > > seamless hotplug with existing applications.
> > > > > >
> > > > > > Port ownership is a genericization of some functions
> > > > > > introduced by the fail-safe, that could structure DPDK
> > > > > > further. It should allow applications to have a seamless
> > > > > > integration with subsystems using port ownership. Without this,
> port ownership cannot be used.
> > > > > >
> > > > > > Testpmd should be fixed, but follow the most common design
> > > > > > patterns of DPDK applications. Going with port ownership seems
> > > > > > like a paradigm shift.
> > > > > >
> > > > > > > In addition
> > > > > > > Using the old iterator in some places in testpmd will cause
> > > > > > > a race for run-
> > > > > time new ports(can be created by failsafe or any hotplug code):
> > > > > > > - testpmd finds an ownerless port(just now created) by the
> > > > > > > old iterator and start traffic there,
> 
> How does testpmd start traffic there? Testpmd has only a callback for
> displaying that it received an event for a new port. It has no concept of
> hotplugging beyond that.
> 

Yes, so no traffic just some control command.

> Testpmd will not start using any new port probed using the hotplug API on its
> own, again, unless something has drastically changed.
> 

Every iterator using in testpmd is exposed to race.

> > > > > > > - failsafe takes ownership of this new port and start traffic there.
> > > > > > > Problem!
> > > > >
> > > > > Could you shed a bit more light here - it would be race
> > > > > condition between whom and whom?
> > > >
> > > > Sure.
> > > >
> > > > > As I remember in testpmd all control ops are done within one
> > > > > thread (main lcore).
> > > >
> > > > But other dpdk entity can use another thread, for example:
> > > > Failsafe uses the host thread(using alarm callback) to create a
> > > > new port and
> > > to take ownership of a port.
> > >
> > > Hm, and you create new ports inside failsafe PMD, right and then set
> > > new owner_id for it?
> >
> > Yes.
> >
> > > And all this in alarm in interrupt thread?
> >
> > Yes.
> >
> > > If so I wonder how you can guarantee that no-one else will set
> > > different owner_id between
> > > rte_eth_dev_allocate() and rte_eth_dev_owner_set()?
> >
> > I check it (see failsafe patch to this series - V5).
> > Function: fs_bus_init.
> >
> > > Could you point me to that place (I am not really familiar with
> > > familiar with failsafe code)?
> > >
> > > >
> > > > The race:
> > > > Testpmd iterates over all ports by the master thread.
> > > > Failsafe takes ownership of a port by the host thread and start using it.
> > > > => The two dpdk entities may use the device at same time!
> > >
> 
> When can this happen? Fail-safe creates its initial pool of ports during EAL
> init, before testpmd scans eth_dev ports and configure its streams.
> At that point, it has taken ownership, from the master lcore context.
> 
> After this point, new ports could be detected and hotplugged by fail-safe.
> However, even if testpmd had a callback to capture those new ports and
> reconfigure its streams, it would be executed from within the intr-thread,
> same as failsafe. If the thread was interrupted, by a dataplane-lcore for
> example, streams would not have been reconfigured.
> The fail-safe would execute its callback and set the owner-id before the
> callback chains goes to the application.
>

Some iterator may be invoked in plug out process by other thread in testpmd and causes to control command 
 
> And that would only be if testpmd had any callback for hotplugging ports and
> reconfiguring its streams, which it hasn't, as far as I know.
>

We don't need to implement it in testpmd.
 
> > > Ok, if failsafe really assigns its owner_id(s) to ports that are
> > > already in use by the app, then how such scheme supposed to work at
> all?
> >
> > If the app works well (with the new rules) it already took ownership and
> failsafe will see it and will wait until the application release it.
> > Every dpdk entity should know which port it wants to manage, If 2
> > entities want to manage the same device -  it can be ok and port ownership
> can synchronize the usage.
> >
> > Probably, application which will run fail-safe wants to manage only the fail-
> safe port and therefor to take ownership only for it.
> >
> > > I.E. application has a port - it assigns some owner_id != 0 to it,
> > > then PMD tries to set its owner_id tot the same port.
> > > Obviously failsafe's set_owner() will always fail in such case.
> > >
> > Yes, and will try again after some time.
> >
> > > From what I hear we need to introduce a concept of 'default owner id'.
> > > I.E. when failsafe PMD is created - user assigns some owner_id to it
> (default).
> > > Then failsafe PMD generates it's own owner_id and assigns it only to
> > > the ports whose current owner_id is equal either 0 or 'default' owner_id.
> > >
> >
> > It is a suggestion and we need to think about it more (I'm talking about it
> with Gaetan in another thread).
> > Actually I think, if we want a generic solution to the generic problem the
> current solution is ok.
> >
> 
> We could as well conclude this other thread there.
> 
> The only solution would be to have a default relationship between owners,
> something that goes beyond the scope assigned by Thomas to your
> evolution, but would be necessary for this API to be properly used by
> existing applications.
> 
> I think it's the only way to have a sane default behavior with your API, but I
> also think this goes beyong the scope of the DPDK altogether.
> 
> But even with those considerations that could be ironed out later (API is still
> experimental anyway), in the meantime, I think we should strive not to
> break "userland" as much as possible. Meaning that unless you have a
> specific situation creating a bug, you shouldn't have to modify testpmd, and if
> an issues arises, you need to try to improve your API before resorting to
> changing the resource management model of all existing applications.
> 

I understand it.
Suggestion:

2 system owners.
APP_OWNER - 0.
NO_OWNER - 1.

And allowing for more owners as now.

1. Every port creation will set the owner for NO_OWNER (as now).
2. There is option for all dpdk entities to take owner of  NO_OWNER ports all the time(as now).
3. In some point in the end of EAL init: set all the NO_OWNER to APP_OWNER(for V6).
4. Change the old iterator to iterate over APP_OWNER ports(for V6).

What do you think?

<snip>