[dpdk-dev] [PATCH] eal/ppc: remove fix of memory barrier for IBM POWER

Thomas Monjalon thomas at monjalon.net
Wed Mar 27 10:19:03 CET 2019


Pradeep, Chao,

Do we have more news?
We must merge this patch for DPDK 19.05-rc1.

I understand you want to try improving performance
by using lightweight sync for SMP barrier,
and this change can be done later.
First priority is to fix the bug of the general barrier.
That's why I should push this patch before the end of the week.



> Shahaf Shuler <shahafs at mellanox.com<mailto:shahafs at mellanox.com>> wrote on 03/23/2019 11:37:42 PM:
> > From: Shahaf Shuler <shahafs at mellanox.com<mailto:shahafs at mellanox.com>>
> > Pradeep Satyanarayana wrote on Saturday, March 23, 2019 12:58 AM
> > >Thomas Monjalon <thomas at monjalon.net<mailto:thomas at monjalon.net>> wrote on 03/22/2019 10:51:17 AM:
> > >> 22/03/2019 16:30, Pradeep Satyanarayana:
> > >> > Thomas Monjalon <thomas at monjalon.net<mailto:thomas at monjalon.net>> wrote on 03/22/2019 01:49:03 AM:
> > >> > > 22/03/2019 02:40, Pradeep Satyanarayana:
> > >> > > > - rte_[rw]mb (general memory barrier) --> should be lwsync
> > >> > >
> > >> > > This is what may be discussed.
> > >> > > The assumption is that the general memory barrier should cover
> > >> > > all cases (CPU caches, SMP and I/O).
> > >> > > That's why we think it should "sync" for Power.
> > >> >
> > >> > In that case, at a minimum we must de-link rte_smp_[rw]mb from rte_[rw]mb
> > >> > and retain it as lwsync. Agreed?
> > >>
> > >> I have no clue about what is needed for SMP barrier in Power.
> > >> As long as it works as expected, no problem.
> > >>
> > >
> > >We will try that out and report back here, later next week
> >
> > Till then, i think there are 2 orthogonal issues:
> > 1. ppc rte_wmb is incorrect
> > 2. ppc rte_smp_[rw]mb may be improved.
> >
> > for #1 the current patch from Dekel seems to be OK. do you agree?
> > for #2 i guess you will check and come back w/ patch/answer?
> 
> That has been the line of thinking. However, we need to do some extensive testing
> to confirm that it all holds up.





More information about the dev mailing list