[dpdk-dev] [EXT] Re: [PATCH] eal: change init macro as exec environment specific

[Thomas Monjalon]

12/03/2019 20:25, Jerin Jacob Kollanukkaran:
> On Fri, 2019-03-01 at 18:28 +0100, Thomas Monjalon wrote:
> > External Email
> > 
> > -------------------------------------------------------------------
> > ---
> > 01/03/2019 18:05, Ferruh Yigit:
> > > On 10/11/2017 3:33 PM, jerin.jacob at caviumnetworks.com (Jerin
> > > Jacob) wrote:
> > > > From: Thomas Monjalon <thomas at monjalon.net>
> > > > > 07/08/2017 14:04, Jerin Jacob:
> > > > > > baremetal execution environments may have a different
> > > > > > method to enable RTE_INIT instead of using compiler
> > > > > > constructor scheme. Move RTE_INIT* definition under
> > > > > > exec-env to support different execution environments.
> > > > > > 
> > > > > > Signed-off-by: Jerin Jacob <jerin.jacob at
> > > > > > caviumnetworks.com>
> > > > > > ---
> > > > > >  app/test-eventdev/evt_test.h                       |  2 +-
> > > > > >  lib/librte_eal/bsdapp/eal/Makefile                 |  2 +-
> > > > > >  .../bsdapp/eal/include/exec-env/rte_eal.h          | 51
> > > > > > ++++++++++++++++++++++
> > > > > >  lib/librte_eal/common/eal_common_log.c             |  2 +
> > > > > >  lib/librte_eal/common/include/rte_bus.h            |  2 +
> > > > > >  lib/librte_eal/common/include/rte_eal.h            |  6 ---
> > > > > >  lib/librte_eal/common/include/rte_tailq.h          |  2 +
> > > > > >  lib/librte_eal/linuxapp/eal/Makefile               |  2 +-
> > > > > >  .../linuxapp/eal/include/exec-env/rte_eal.h        | 51
> > > > > > ++++++++++++++++++++++
> > > > > >  9 files changed, 111 insertions(+), 9 deletions(-)
> > > > > >  create mode 100644 lib/librte_eal/bsdapp/eal/include/exec-
> > > > > > env/rte_eal.h
> > > > > >  create mode 100644 lib/librte_eal/linuxapp/eal/include/exec-
> > > > > > env/rte_eal.h
> > > > > 
> > > > > I am not a big fan of duplicating code for Linux and BSD.
> > > > > 
> > > > > Maybe we should have different splits and include a common file
> > > > > in Linux and BSD?
> > > > 
> > > > OK. This is doable.
> > > > 
> > > > > I feel it would be easier to think about the split when adding
> > > > > a new environment.
> > > > > It is also an open question whether we want to support (again)
> > > > > some
> > > > > bare metal environments.
> > > > 
> > > > IMO, A factor could be, how much we are OK to change?
> > > > 
> > > > Our internal prototype implementation for a bare metal
> > > > environment
> > > > shows things are already in place and may need minor changes like
> > > > this to
> > > > accommodate a bare metal execution environment(accounting the
> > > > latest
> > > > changes of moving pci to driver/pci/..)
> > > > 
> > > > If no one care about need for such abstraction then we could drop
> > > > this
> > > > patch. We can always keep local copy of such patches in our
> > > > internal
> > > > tree. I thought to upstream it as it may be useful for someone
> > > > else and
> > > > it is easy for us maintain if changes are in
> > > > lib/librte_eal/<new environment>/eal/ and drivers/*/
> > >  Hi Jerin, Thomas,
> > > 
> > > This is an old patch, the abstraction seems good idea but it comes
> > > with a
> > > duplication.
> > > 
> > > Is there an intention to continue the work? Are we waiting for any
> > > decision?
> > > Any objection to mark it as rejected?
> > 
> > I am not sure there is a real desire to make DPDK
> > ready for bare-metal (back again).
> > If any of you are aware of a real use-case, we can re-consider.
> 
> Some of the usecases:
> 
> # PCIe endpoint mode aka Smart NIC(Where DPDK runs on PCIe card), May
> not need to waste one core for Linux. Specially Smart NIC market has
> less number of cores.
> On the endpoint side, It treats as FW so customer may not have access
> to so nobdoy cares it is Linux or baremetal so may need to waste one
> core for Linux
> 
> # VM case, it possible to have bare metal guest just to save one a
> logical core for Linux
> 
> # Some of the RTOS like Zephyr already provide TCP/IP stack and good
> subsystems for specific usecases.
> 
> # We are using DPDK for pre silicon validation for SoC mode. Bringing
> up linux on emulator takes ages, Baremetal can be used for Harware
> verification too.
> 
> 
> IMO, As long as it not limiting the a feature of Linux app, Why not to
> allow baremetal? I agree with code duplication. I think, it can be
> fixed easily, Other than that, Is there any concern?

The concern is about the effort required.
Which libc to use? Which dependency is acceptable?