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

Mordechay Haimovsky motih at mellanox.com
Wed Jul 25 13:55:50 CEST 2018


This is how we reproduce the issue :
app/testpmd -n 4   --vdev="net_tap0,iface=tap0,remote=eth1"  -- --burst=64 --mbcache=512 -i   --rxq=2 --txq=2 --txd=512 --rxd=512  --port-topology=chained --forward-mode=rxonly

> -----Original Message-----
> From: Ananyev, Konstantin [mailto:konstantin.ananyev at intel.com]
> Sent: Wednesday, July 25, 2018 2:39 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
> 
> >
> > 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