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

[Thomas Monjalon <thomas@monjalon.net>]

We need to agree on the definitions.
Please see below,

22/03/2019 02:40, Pradeep Satyanarayana:
> Shahaf Shuler <shahafs@mellanox.com> wrote on 03/21/2019 01:49:39 AM:
> > Pradeep Satyanarayana <pradeep@us.ibm.com> wrote on Thu 3/21/2019 12:41
> AM:
> > >> > So far, when not running on power, we used the rte_wmb for that.
> > >> On x86 and ARM systems it provided the needed guarantees.
> > >> > It is also mentioned in the barrier doxygen on ARM arch:
> > >> > "
> > >> > Write memory barrier.
> > >> >
> > >> > Guarantees that the STORE operations generated before the barrier
> > >> > occur before the STORE operations generated after.
> > >> > "
> > >> >
> > >> > It doesn't restrict to store to system memory only.
> > >> > w/ power is on somewhat different and in fact rte_mb is required.
> > >> It obviously miss the point of those barrier if we will need to use
> > >> a different barrier based on the system arch.
> > >> >
> > >> > We need to align the definition of the different barriers in DPDK:
> > >> > 1. need a clear documentation of each. this should be global and
> > >> not part of the specific implementation on each arch.
> > >
> > >A single approach may not work for all architectures. Power is different
> > >from others, so we need to be able to accommodate that. More comments
> below.
> >
> > it don't get this claim.
> > It is ok to have some differences between the different arch, but
> > here you implement a well-defined barrier - rte_wmb.
> > if you see a need we can discuss to define a **new** barrier which
> > sync STORE only to system memory, and will be able to utilize the
> > lwsync command.
> >
> > >
> > >> The global definition is in lib/librte_eal/common/include/
> > generic/rte_atomic.h
> > >>
> > >> There are some copy/paste in Arm32 and PPC that I will remove.
> > >>
> > >> > 2. either modify ppc rte_wmb to match ARM and x86 ones or to
> > >> define a new type of barrier which will sync between both I/O and
> > >> stores to systems memory.
> > >>
> > >> The basic memory barrier of DPDK does not mention
> > >> a difference between I/O and system memory.
> > >
> > >In the case of Power, sync will cater to both I/O and system
> > memory. However, that
> > >is too big a hammer in all cases.
> >
> > rte_wmb requires such sync. you propose to have the wrong barrier in
> > favor of performance.
> > to mitigate this you can take my suggestion above and define a new,
> > more lightweight one.
> >
> > >
> > >> It is not explicit (yet) but I assume it is protecting both.
> > >> So, in my opinion, we need to make it explicit in the doc,
> > >> and fix the PPC implementation to comply with this definition.
> > >>
> > >> Anyway, I don't see any significant effort from IBM to move from
> > >> the alpha support stage to a real Open Source support.
> > >> PS: sending a mail every two months, to promise improvements, is
> > not enough!
> > >
> >
> > […]
> >
> > >
> > >We should retain lwsync, should not be removed as discussed in here:
> > >
> > >http://mails.dpdk.org/archives/dev/2019-March/126746.html
> >
> > i don't agree.
> > it is very clear the rte_wmb implementation in power is broken and
> > we need to fix this right away before other customers will hit the
> > same issue.
> 
> 
> In the DPDK source I see a couple of different classes of memory barriers.
> I am
> not clear on the usage of these in the drivers, but I would think the
> guidelines
> to be as shown below (for Power):
> 
> - 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.

> - rte_smp_[rw]mb (SMP memory barrier) -->should be lwsync
> - rte_io_[rw]mb (I/O memory barrier)  --> should be sync
> - rte_cio_[rw]mb (coherent I/O memory barrier) -->should be sync
> 
> lwsync is appropriate for cases where CPUs are accessing cacheable memory
> (i.e. Memory Coherence Required) while the sync instruction should be used
> in all other cases.
> 
> With the patch in:
> http://mails.dpdk.org/archives/dev/2019-March/126746.html
> 
> It converts even the rte_smp_[rw]mb into sync. That is not what the
> rte_smp*() should
> be implementing as per the guidelines above.
> 
> static __rte_always_inline void
> mlx5_tx_dbrec_cond_wmb(struct mlx5_txq_data *txq, volatile struct mlx5_wqe
> *wqe,
>                        int cond)
> {
>         uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
>         volatile uint64_t *src = ((volatile uint64_t *)wqe);
> 
>         rte_cio_wmb(); --> would rte_cio_rmb() be more appropriate?
>         *txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
>         /* Ensure ordering between DB record and BF copy. */
>         rte_wmb(); --> what has been established is that for Power we need
> "sync" instead of lwsync
> 		We are dealing with device memory -should we be using an
> rte_io_wmb() here?
> 
>         mlx5_uar_write64_relaxed(*src, dst, txq->uar_lock);
>         if (cond)
>                 rte_wmb(); --> what about here? Are there conditions when
> we need sync?
> }