> -----Original Message-----
> From: Ruifeng Wang (Arm Technology China) [mailto:ruifeng.w...@arm.com]
> Sent: Wednesday, September 11, 2019 7:58 AM
> To: Honnappa Nagarahalli <honnappa.nagaraha...@arm.com>; Ananyev, Konstantin
> <konstantin.anan...@intel.com>; Kantecki, Tomasz
> <tomasz.kante...@intel.com>
> Cc: dev@dpdk.org; Gavin Hu (Arm Technology China) <gavin...@arm.com>; nd
> <n...@arm.com>; nd <n...@arm.com>; nd <n...@arm.com>
> Subject: RE: [dpdk-dev] [PATCH 0/2] add lock-free mode for l3fwd
>
> > -----Original Message-----
> > From: Honnappa Nagarahalli <honnappa.nagaraha...@arm.com>
> > Sent: Wednesday, September 11, 2019 13:38
> > To: Ruifeng Wang (Arm Technology China) <ruifeng.w...@arm.com>;
> > Ananyev, Konstantin <konstantin.anan...@intel.com>; Kantecki, Tomasz
> > <tomasz.kante...@intel.com>
> > Cc: dev@dpdk.org; Gavin Hu (Arm Technology China) <gavin...@arm.com>;
> > Honnappa Nagarahalli <honnappa.nagaraha...@arm.com>; nd
> > <n...@arm.com>; nd <n...@arm.com>
> > Subject: RE: [dpdk-dev] [PATCH 0/2] add lock-free mode for l3fwd
> >
> > <snip>
> > > >
> > > >
> > > >
> > > > > >
> > > > > > >
> > > > > > > Lock-free mode is supported by hash library and LPM library.
> > > > > > > Now we add an option for l3fwd example to enable the lock-free
> > > mode.
> > > > > > > Necessary preparation steps are added to use lock-free LPM mode.
> > > > > >
> > > > > > Can I ask about the purpose of these changes?
> > > > > > Right now in l3fwd both lpm and hash tables are static and hard-
> > coded.
> > > > > > we initialize them at startup and then just do read from them.
> > > > > > Do you plan to enhance l3fwd with ability to dynamically update
> > > > > > tables contents?
> > > > > > Though fir that we first have to get rid of hard-coded values
> > > > > > (config file or
> > > > so).
> > > > > > Konstantin
> > > > > >
> > > > > Thanks for your questions.
> > > > > Currently, we have no plan to enhance l3fwd with ability to
> > > > > dynamically
> > > > update table contents.
> > > > > Lock-free method is being integrated into Hash library and LPM
> > > > > library. Lock-free algorithms are not only about control plane
> > > > > (adding or deleting routes), they affect the data path performance as
> > well.
> > > > > Since l3fwd application is showcasing data path performance, we
> > > > > need to show the impact of including the quiescent state reporting
> > > > > on data
> > > path.
> > > > > This change also serves as an example of using the RCU APIs.
> > > >
> > > >
> > > > But what you suggest doesn't provide the complete picture.
> > > > With dynamic updates in place (via control path) the data-path
> > > > impact might be completely different then without.
> > > > Again without dynamic updates how can you test that your data-path
> > > > lock- free approach does work as expected?
> > > > Also it can't even be used as a reference implementation for users,
> > > > as half of the functionality they need to implement is simply missing.
> > > > My opinion - we either need to leave l3fwd as it is (static routes),
> > > > or implement a proper control path with ability to dynamically
> > > > update routes before starting to introduce some synchronization
> > > > schemes (RCU or whatever).
> > > >
> > > > Konstantin
> > > >
> > >
> > > Agree that dynamic control path updates should be included for a whole
> > > picture.
> > > I will add dynamic update to l3fwd and reroll the patch series.
> > > Thanks.
> > I think we should have an agreement on what exactly we mean by
> > 'dynamically update routes'.
> > IMO, we should not disturb the existing static routes as there might be
> > automated tests running in the labs. I suggest that we should add/delete
> > new routes/hash entries which are different from the existing routes/hash
> > entries. This should be sufficient to showcase the functionality as well as
> > measure the impact.
> >
> Yes, existing static routes should be kept intact.
> To perform regular route/hash entries add/delete, a dedicated lcore will be
> needed.
> An interactive prompt is not an option since we need automatic add/delete.
> We can skip master core for data path main loop. And perform unrelated
> route/hash entries add/delete regularly on master core.
> The impact is that command lines used in tests will need update since master
> core will no longer do data path work.
Not sure why it has to be master core?
Why interrupt thread wouldn't do?
I think what we need to:
1. introduce reading routes from config file instead of having them hard-coded
within the app.
2. add ability to update routes dynamically.
Probably the easiest (and commonly used way) re-read conf file and update
routes on the signal (SIGUSR1 or so).
Konstantin
> > >
> > > > >
> > > > > > >
> > > > > > > Patch 2/2 has dependency on RCU QSBR integration with LPM
> > library:
> > > > > > > http://patches.dpdk.org/project/dpdk/list/?series=6288
> > > > > > >
> > > > > > >
> > > > > > > Ruifeng Wang (2):
> > > > > > > examples/l3fwd: add lock-free option for l3fwd
> > > > > > > examples/l3fwd: integrate RCU QSBR for LPM mode
> > > > > > >
> > > > > > > doc/guides/sample_app_ug/l3_forward.rst | 3 ++
> > > > > > > examples/l3fwd/Makefile | 1 +
> > > > > > > examples/l3fwd/l3fwd.h | 4 +-
> > > > > > > examples/l3fwd/l3fwd_em.c | 10 +++-
> > > > > > > examples/l3fwd/l3fwd_lpm.c | 72
> > > > +++++++++++++++++++++++--
> > > > > > > examples/l3fwd/main.c | 27 ++++++++--
> > > > > > > examples/l3fwd/meson.build | 1 +
> > > > > > > 7 files changed, 108 insertions(+), 10 deletions(-)
> > > > > > >
> > > > > > > --
> > > > > > > 2.17.1
> > >
> >
