[dpdk-dev] [PATCH] app/testpmd: fix testpmd exit using ctrl+d

Dumitrescu, Cristian cristian.dumitrescu at intel.com
Wed Jul 25 15:21:10 CEST 2018



> -----Original Message-----
> From: Van Haaren, Harry
> Sent: Wednesday, July 25, 2018 1:27 PM
> To: Ananyev, Konstantin <konstantin.ananyev at intel.com>; Dumitrescu,
> Cristian <cristian.dumitrescu at intel.com>; Mordechay Haimovsky
> <motih at mellanox.com>; Thomas Monjalon <thomas at monjalon.net>;
> Singh, Jasvinder <jasvinder.singh at intel.com>
> Cc: dev at dpdk.org; Iremonger, Bernard <bernard.iremonger at intel.com>;
> Pattan, Reshma <reshma.pattan at intel.com>; olivier.matz at 6wind.com
> Subject: RE: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit using ctrl+d
> 
> > -----Original Message-----
> > From: Ananyev, Konstantin
> > Sent: Wednesday, July 25, 2018 1:18 PM
> > To: Dumitrescu, Cristian <cristian.dumitrescu at intel.com>; Mordechay
> Haimovsky
> > <motih at mellanox.com>; Thomas Monjalon <thomas at monjalon.net>;
> Singh, Jasvinder
> > <jasvinder.singh at intel.com>
> > Cc: dev at dpdk.org; Iremonger, Bernard <bernard.iremonger at intel.com>;
> Pattan,
> > Reshma <reshma.pattan at intel.com>; olivier.matz at 6wind.com; Van
> Haaren, Harry
> > <harry.van.haaren at intel.com>
> > Subject: RE: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit using ctrl+d
> >
> >
> >
> > > -----Original Message-----
> > > From: Dumitrescu, Cristian
> > > Sent: Wednesday, July 25, 2018 12:57 PM
> > > To: Ananyev, Konstantin <konstantin.ananyev at intel.com>; Mordechay
> Haimovsky
> > <motih at mellanox.com>; Thomas Monjalon
> > > <thomas at monjalon.net>; Singh, Jasvinder <jasvinder.singh at intel.com>
> > > Cc: dev at dpdk.org; Iremonger, Bernard <bernard.iremonger at intel.com>;
> Pattan,
> > Reshma <reshma.pattan at intel.com>;
> > > olivier.matz at 6wind.com; Van Haaren, Harry
> <harry.van.haaren at intel.com>
> > > Subject: RE: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit using
> ctrl+d
> > >
> > >
> > >
> > > > -----Original Message-----
> > > > From: Ananyev, Konstantin
> > > > Sent: Wednesday, July 25, 2018 12:39 PM
> > > > To: Dumitrescu, Cristian <cristian.dumitrescu at intel.com>; Mordechay
> > > > Haimovsky <motih at mellanox.com>; Thomas Monjalon
> > > > <thomas at monjalon.net>; Singh, Jasvinder
> <jasvinder.singh at intel.com>
> > > > Cc: dev at dpdk.org; Iremonger, Bernard
> <bernard.iremonger at intel.com>;
> > > > Pattan, Reshma <reshma.pattan at intel.com>;
> olivier.matz at 6wind.com; Van
> > > > Haaren, Harry <harry.van.haaren at intel.com>
> > > > Subject: RE: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit using
> ctrl+d
> > > >
> > > >
> > > >
> > > > > -----Original Message-----
> > > > > From: Dumitrescu, Cristian
> > > > > Sent: Wednesday, July 25, 2018 10:36 AM
> > > > > To: Ananyev, Konstantin <konstantin.ananyev at intel.com>;
> Mordechay
> > > > Haimovsky <motih at mellanox.com>; Thomas Monjalon
> > > > > <thomas at monjalon.net>; Singh, Jasvinder
> <jasvinder.singh at intel.com>
> > > > > Cc: dev at dpdk.org; Iremonger, Bernard
> <bernard.iremonger at intel.com>;
> > > > Pattan, Reshma <reshma.pattan at intel.com>;
> > > > > olivier.matz at 6wind.com
> > > > > Subject: RE: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit using
> > ctrl+d
> > > > >
> > > > >
> > > > >
> > > > > > -----Original Message-----
> > > > > > From: Ananyev, Konstantin
> > > > > > Sent: Wednesday, July 25, 2018 10:04 AM
> > > > > > To: Dumitrescu, Cristian <cristian.dumitrescu at intel.com>;
> Mordechay
> > > > > > Haimovsky <motih at mellanox.com>; Thomas Monjalon
> > > > > > <thomas at monjalon.net>; Singh, Jasvinder
> <jasvinder.singh at intel.com>
> > > > > > Cc: dev at dpdk.org; Iremonger, Bernard
> <bernard.iremonger at intel.com>;
> > > > > > Pattan, Reshma <reshma.pattan at intel.com>;
> olivier.matz at 6wind.com
> > > > > > Subject: RE: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit
> using
> > > > ctrl+d
> > > > > >
> > > > > >
> > > > > >
> > > > > > > -----Original Message-----
> > > > > > > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of
> Dumitrescu,
> > > > > > Cristian
> > > > > > > Sent: Tuesday, July 24, 2018 5:59 PM
> > > > > > > To: Mordechay Haimovsky <motih at mellanox.com>; Thomas
> Monjalon
> > > > > > <thomas at monjalon.net>; Singh, Jasvinder
> > > > > > > <jasvinder.singh at intel.com>
> > > > > > > Cc: dev at dpdk.org; Iremonger, Bernard
> > > > <bernard.iremonger at intel.com>;
> > > > > > Pattan, Reshma <reshma.pattan at intel.com>;
> > > > > > > olivier.matz at 6wind.com
> > > > > > > Subject: Re: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit
> using
> > > > ctrl+d
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > > -----Original Message-----
> > > > > > > > From: Mordechay Haimovsky [mailto:motih at mellanox.com]
> > > > > > > > Sent: Tuesday, July 24, 2018 3:37 PM
> > > > > > > > To: Thomas Monjalon <thomas at monjalon.net>; Singh,
> Jasvinder
> > > > > > > > <jasvinder.singh at intel.com>
> > > > > > > > Cc: dev at dpdk.org; Iremonger, Bernard
> > > > <bernard.iremonger at intel.com>;
> > > > > > > > Pattan, Reshma <reshma.pattan at intel.com>;
> > > > olivier.matz at 6wind.com;
> > > > > > > > Dumitrescu, Cristian <cristian.dumitrescu at intel.com>
> > > > > > > > Subject: RE: [dpdk-dev] [PATCH] app/testpmd: fix testpmd exit
> > using
> > > > > > ctrl+d
> > > > > > > >
> > > > > > > > Even after this fix we still have setups that use netvsc  for
> > example,
> > > > on
> > > > > > > > which testpmd exits with rte_panic right after loading it even
> > > > without
> > > > > > > > touching the KBD.
> > > > > > > >
> > > > > > > > I recommend returning the previous prompt routine in test-
> > > > > > pmd/cmdline.c
> > > > > > > > and rework the SOFTNIC section there, preferably moving its
> poll
> > > > section
> > > > > > to
> > > > > > > > use rte_service in a separate file cleaning the CLI files from
> > PMD-
> > > > specific
> > > > > > > > implementation.
> > > > > > > >
> > > > > > > > > -----Original Message-----
> > > > > > > > > From: Thomas Monjalon [mailto:thomas at monjalon.net]
> > > > > > > > > Sent: Tuesday, July 24, 2018 3:34 PM
> > > > > > > > > To: Jasvinder Singh <jasvinder.singh at intel.com>
> > > > > > > > > Cc: dev at dpdk.org; bernard.iremonger at intel.com;
> > > > > > > > > reshma.pattan at intel.com; Mordechay Haimovsky
> > > > > > > > <motih at mellanox.com>;
> > > > > > > > > olivier.matz at 6wind.com; cristian.dumitrescu at intel.com
> > > > > > > > > Subject: Re: [dpdk-dev] [PATCH] app/testpmd: fix testpmd
> exit
> > > > using
> > > > > > ctrl+d
> > > > > > > > >
> > > > > > > > > Important note:
> > > > > > > > > 	testpmd is currently really broken.
> > > > > > > > > 	We cannot have a RC2 until it is fixed.
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > 24/07/2018 13:25, Thomas Monjalon:
> > > > > > > > > > 23/07/2018 12:44, Jasvinder Singh:
> > > > > > > > > > > Fix testpmd app exit by pressing ctrl+d, End-of-
> Transmission
> > > > > > > > > > > character (EOT) on the empty command line.
> > > > > > > > > >
> > > > > > > > > > Please describe what is the issue and what is the cause.
> > > > > > > > > >
> > > > > > > > > > > Fixes: 0ad778b398c6 ("app/testpmd: rework softnic
> forward
> > > > > > mode")
> > > > > > > > > > >
> > > > > > > > > > > Reported-by: Mordechay Haimovsky
> <motih at mellanox.com>
> > > > > > > > > > > Signed-off-by: Jasvinder Singh
> <jasvinder.singh at intel.com>
> > > > > > > > > > > ---
> > > > > > > > > > >  app/test-pmd/cmdline.c       | 10 ++++++----
> > > > > > > > > > >  lib/librte_cmdline/cmdline.c |  2 +-
> > > > > > > > > >
> > > > > > > > > > It is very suspicious to change the cmdline library.
> > > > > > > > > > If there is a bug in this library, it deserves a separate fix.
> > > > > > > > >
> > > > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > First, testpmd is not really broken, as only thing that changed is
> > the Ctrl
> > > > +
> > > > > > D behavior. I agree this is an issue that we need to fix, as
> > > > > > > it looks that it is breaking some automation scripts for some
> > people.
> > > > > > >
> > > > > > > The change in behavior for Ctrl + D exit is caused by replacing the
> > call
> > > > to
> > > > > > cmdline_interact() with calling cmdline_poll() in a loop.
> > > > > > > These two approaches should be identical in behavior, but it looks
> > like
> > > > they
> > > > > > are not due to some issue in the cmdline library.
> > > > > > > Jasvinder proposed a quick patch, but it looks like something else
> > needs
> > > > to
> > > > > > be fixed in cmdline library in order to bring
> > > > > > > cmdline_poll() on parity with cmdline_interact(). Any advice from
> > > > Olivier
> > > > > > would be very much appreciated!
> > > > > > >
> > > > > > > It is really a bad practice to use cmdline_interact() in testpmd, as
> > it is a
> > > > > > blocking call that prohibits doing other things on the same
> > > > > > > thread that runs the CLI. Sometimes we need to run other things
> on
> > the
> > > > > > same core, e.g. the slow softnic_manage() function.
> > > > > >
> > > > > > Curious why not use a service core for softnic background stuff, and
> > leave
> > > > CLI
> > > > > > one for CLI?
> > > > > > Konstantin
> > > > >
> > > > > I guess for a test application you can do anything you want, but in
> real
> > life
> > > > CPU cores are really expensive and dedicating a CPU core
> > > > > just for CLI is a colossal waste.
> > > >
> > > > Ok, but let's application developer to decide how to use (waste) the
> cores
> > he
> > > > owns :)
> > > > What I am saying: there is a special thing (developed by Harry) service
> > cores.
> > > > As I understand why of it's the purpose - allow PMD(s) to allocate CPU
> > > > resources for
> > > > there background tasks in a unified and transparent way.
> > > > From the description above - that seems to fit your needs (softnic
> > > > background processing), no?
> > > > Konstantin
> > > >
> > >
> > > Then why not put the testpmd CLI itself on a service core? Are you
> > volunteering for a patch on this? :)
> >
> > It was not me who broke testpmd at first place.
> > I'd better live without softnic support in testpmd :)
> > Though I still don't understand why do you feel that service cores are not
> > good enough for you?
> > From my understanding they were introduced for similar purposes
> > (Harry please feel free to correct me here).
> > So if you think they not fit for your case -
> > at least would be good to understand why.
> > Konstantin
> 
> 
> Service-cores is a mechanism to abstract SW fallback CPU core
> requirements;
> Eg: A nic provides RSS, 0 cores required
>     A NIC does not provide RSS, use 1 service core to provide that RSS
> function
> 
> Result: the application can rely on RSS, as we have a transparent fallback:
> no application controlled lcore is performing any different than before with
> HW provided functionality
> 
> Service cores is not intended to be used for "background" tasks, as
> the service-core consumes an Lcore and polls 100% of the time, regardless
> of what services it is performing.
> 
> 
> If there are multiple things that we can run on service-cores, that would
> make
> sense. Eg: timer_manage(), CLI (non-blocking..), SoftNIC background tasks...
> 
> This design makes sense as the application can also register its own
> "background"
> services, and run them all on a single service-core.
> 
> 

Thanks, Harry, we might have to revert to service core if we are not able to reproduce & fix this issue quickly enough.

> > >
> > >
> > > > >
> > > > > We did use the non-blocking cmdline_poll() function instead of the
> > > > blocking cmdline_interact() in the past without any issues. The
> > > > > issues reported by Moti come as a surprise. It is probably good to see
> > what
> > > > this is about and see if we can quickly fix the issue in
> > > > > cmdline library. Otherwise, we can revert the usage of cmdline_poll()
> > with
> > > > cmdline_interact().
> > > > >
> > > > >
> > > > > >
> > > > > > >
> > > > > > > Worst case scenario: We can revert the cmdline_poll() back to
> > > > > > cmdline_interact(), this is a small change, but not the proper way
> of
> > > > > > > doing things, as this is simply hiding the issue in cmdline library.
> > It
> > > > would
> > > > > > also prevent us from testing some Soft NIC functionality.



More information about the dev mailing list