Re: [net-next 01/14] net/mlx5e: RX, Add a prefetch command for small L1_CACHE_BYTES

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Tue, 23 Apr 2019 10:27:32 -0700
Alexander Duyck <alexander.duyck@gmail.com> wrote:

> On Tue, Apr 23, 2019 at 9:42 AM Saeed Mahameed <saeedm@mellanox.com> wrote:
> >
> > On Tue, 2019-04-23 at 08:21 -0700, Alexander Duyck wrote:  
> > > On Tue, Apr 23, 2019 at 6:23 AM Jesper Dangaard Brouer
> > > <brouer@redhat.com> wrote:  
> > > > On Mon, 22 Apr 2019 19:46:47 -0700
> > > > Jakub Kicinski <jakub.kicinski@netronome.com> wrote:
> > > >  
> > > > > On Mon, 22 Apr 2019 15:32:53 -0700, Saeed Mahameed wrote:  
> > > > > > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en.h
> > > > > > b/drivers/net/ethernet/mellanox/mlx5/core/en.h
> > > > > > index 51e109fdeec1..6147be23a9b9 100644
> > > > > > --- a/drivers/net/ethernet/mellanox/mlx5/core/en.h
> > > > > > +++ b/drivers/net/ethernet/mellanox/mlx5/core/en.h
> > > > > > @@ -50,6 +50,7 @@
> > > > > >  #include <net/xdp.h>
> > > > > >  #include <linux/net_dim.h>
> > > > > >  #include <linux/bits.h>
> > > > > > +#include <linux/prefetch.h>
> > > > > >  #include "wq.h"
> > > > > >  #include "mlx5_core.h"
> > > > > >  #include "en_stats.h"
> > > > > > @@ -986,6 +987,22 @@ static inline void mlx5e_cq_arm(struct
> > > > > > mlx5e_cq *cq)
> > > > > >     mlx5_cq_arm(mcq, MLX5_CQ_DB_REQ_NOT, mcq->uar->map, cq-  
> > > > > > >wq.cc);  
> > > > > >  }
> > > > > >
> > > > > > +static inline void mlx5e_prefetch(void *p)
> > > > > > +{
> > > > > > +   prefetch(p);
> > > > > > +#if L1_CACHE_BYTES < 128
> > > > > > +   prefetch(p + L1_CACHE_BYTES);
> > > > > > +#endif
> > > > > > +}
> > > > > > +
> > > > > > +static inline void mlx5e_prefetchw(void *p)
> > > > > > +{
> > > > > > +   prefetchw(p);
> > > > > > +#if L1_CACHE_BYTES < 128
> > > > > > +   prefetchw(p + L1_CACHE_BYTES);
> > > > > > +#endif
> > > > > > +}  
> > > > >
> > > > > All Intel drivers do the exact same thing, perhaps it's time to
> > > > > add a
> > > > > helper fot this?
> > > > >
> > > > > net_prefetch_headers()
> > > > >
> > > > > or some such?  
> > > >
> > > > I wonder if Tariq measured any effect from doing this?
> > > >
> > > > Because Intel CPUs will usually already prefetch the next cache-
> > > > line,
> > > > as described in [1], you can even read (and modify) this MSR 0x1A4
> > > > e.g. via tools in [2].  Maybe Intel guys added it before this was
> > > > done
> > > > in HW, and never cleaned it up?
> > > >
> > > > [1]
> > > > https://software.intel.com/en-us/articles/disclosure-of-hw-prefetcher-control-on-some-intel-processors  
> > >
> > > The issue is the adjacent cache line prefetcher can be on or off and
> > > a
> > > network driver shouldn't really be going through and twiddling those
> > > sort of bits. In some cases having it on can result in more memory
> > > being consumed then is needed. The reason why I enabled the
> > > additional
> > > cacheline prefetch for the Intel NICs is because most TCP packets are
> > > at a minimum 68 bytes for just the headers so there was an advantage
> > > for TCP traffic to make certain we prefetched at least enough for us
> > > to process the headers.
> > >  
> >
> > So if L2 adjacent cache line prefetcher bit is enabled then this

Nitpick: is it the DCU prefetcher bit that "Fetches the next cache line
into L1-D cache" in the link[1].

> > additional prefetch step is redundant ? what is the performance cost in
> > this case ?  
> 
> I don't recall. I don't think it would be anything too significant though.

I tried to measure this (approx 1 year ago), a prefetch that is not
needed, and AFAICR the overhead was below 1 nanosec, approx 0.333 ns.
(but anyone claiming to be able to measure below 2 ns variation
accuracy should be questioned...)

> > > As far as Jakub comment about combining the functions I would be okay
> > > with that. We just need to make it a static inline function available
> > > to all the network drivers.
> > >  
> >
> > Agreed, will drop this patch for now and Tariq will address, in next
> > version.

I don't mind the patch, and Alex provided a good argument why is still
makes sense.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer