Re: [dpdk-dev] [PATCH] app/testpmd: change port detach interface

[Nithin Dabilpuram <nithind1988@gmail.com>]

On Tue, Jul 02, 2019 at 04:58:17PM +0100, Yigit, Ferruh wrote:
> On 5/29/2019 9:16 AM, Nithin Dabilpuram wrote:
> > Hi Thomas,
> > 
> > On Mon, May 20, 2019 at 06:20:53PM +0530, Nithin Dabilpuram wrote:
> >> On Fri, May 17, 2019 at 10:59:38AM +0200, Thomas Monjalon wrote:
> >>> 17/05/2019 10:55, Nithin Dabilpuram:
> >>>> On Wed, May 15, 2019 at 09:27:22AM +0200, Thomas Monjalon wrote:
> >>>>> 15/05/2019 08:52, Nithin Dabilpuram:
> >>>>>> Hi Thomas,
> >>>>>> On Tue, May 14, 2019 at 05:39:30PM +0200, Thomas Monjalon wrote:
> >>>>>>> Hi,
> >>>>>>>
> >>>>>>> 13/05/2019 13:21, Nithin Dabilpuram:
> >>>>>>>> With the latest published interface of
> >>>>>>>> rte_eal_hotplug_[add,remove](), and rte_eth_dev_close(),
> >>>>>>>> rte_eth_dev_close() would cleanup all the data structures of
> >>>>>>>> port's eth dev leaving the device common resource intact
> >>>>>>>> if RTE_ETH_DEV_CLOSE_REMOVE is set in dev flags.
> >>>>>>>> So "port detach" (~hotplug remove) should be able to work,
> >>>>>>>> with device identifier like "port attach" as eth_dev could have
> >>>>>>>> been closed already and rte_eth_devices[port_id] reused.
> >>>>>>>
> >>>>>>> "port attach" uses devargs as identifier because there
> >>>>>>> is no port id before creating it. But "detach port" uses
> >>>>>>> logically the port id to close.
> >>>>>>
> >>>>>> But if "port close" was already called on that port,
> >>>>>> eth_dev->state would be set as RTE_ETH_DEV_UNUSED and
> >>>>>> that port id could be reused.
> >>>>>> So after "port close" if we call "port detach", isn't it
> >>>>>> incorrect to use the same port id ?
> >>>>>
> >>>>> Yes it is incorrect to close a port which is already closed :)
> >>>>>
> >>>>>>>> This change alters "port detach" cmdline interface to
> >>>>>>>> work with device identifier like "port attach".
> >>>>>>>
> >>>>>>> The word "port" means an ethdev port, so it should be
> >>>>>>> referenced with a port id.
> >>>>>>> If you want to close an EAL rte_device, then you should
> >>>>>>> rename the command.
> >>>>>>> But testpmd purpose should be to work with ethdev ports only.
> >>>>>>
> >>>>>> Renaming the command to "detach <identifier>" ?
> >>>>>
> >>>>> Yes something like that.
> >>>>> But why do you want to manage rte_device in testpmd?
> >>>>> Being able to close ports in not enough?
> >>>>> Please describe a scenario.
> >>>>>
> >>>>
> >>>> We just want to support testing hotplug detach along with
> >>>> hotplug attach from testpmd. Currently there is no way to detach
> >>>> if we close the port first.
> >>>
> >>> OK
> >> So can I send next revision with command renamed to "detach <identifier>" ?
> > 
> > Any info on this ? I can even add it as another cmd without disturbing existing
> > command if needed.
> 
> This sounds better option to me. I see the need to remove device via
> 'identifier' but also still it is easier to use 'port_id' for removal when
> applicable, so I am for keeping it.
Thanks. Will send out a patch for the same.
> 
> What do you think adding a new command:
> 'device detach'
> 
> Also testpmd doesn't dead with 'device' much, it mainly works in port level,
> because of this does it make sense to add another command something like:
> "show device info all"
Sure. Will add it.
> 
> > 
> >>>
> >>>> Another reason is that in our new PMD, for detaching one specific port,
> >>>> we need more than one try as the PMD might return -EAGAIN.
> >>>> So with the current "port detach" implementation, after closing the port,
> >>>> if PMD returns -EAGAIN for rte_dev_remove() call, there is no way to
> >>>> try it again.
> >>>
> >>> This is a bug.
> >>> Should we catch -EAGAIN somewhere?
> >>
> >> It is already caught in local_dev_remove() and
> >> rte_dev_remove() fails. Only problem as I said below is
> >> in testpmd if first call to detach_port_device() i.e handler of "port detach", 
> >> rte_dev_remove() returns -EAGAIN and PMD cleaned up the resources partially like eth_dev
> >> resources, the second time call cannot work port_id will not be valid anymore.
> >>
> >>>
> >>>
>