[dpdk-dev] [PATCH v5 00/23] Add DLB2 PMD

[Thomas Monjalon]

30/10/2020 17:02, McDaniel, Timothy:
> From: Thomas Monjalon <thomas at monjalon.net>
> > 30/10/2020 16:35, McDaniel, Timothy:
> > > From: Thomas Monjalon <thomas at monjalon.net>
> > > > 30/10/2020 12:58, McDaniel, Timothy:
> > > > > From: McDaniel, Timothy
> > > > > > From: Thomas Monjalon <thomas at monjalon.net>
> > > > > > > 30/10/2020 10:43, Timothy McDaniel:
> > > > > > > > - note that the code still uses its private byte-encoded versions of
> > > > > > > >   umonitor/umwait, rather than the new functions in the power
> > > > > > > >   patch that are built on top of those intrinsics. This is intentional.
> > > > > > >
> > > > > > > Why? Now these intrinsics are available in the main branch.
> > > > > > > We should avoid duplicating such code.
> > > > > > >
> > > > > > >
> > > > > >
> > > > > > I had asked that the low level intrinsics (UMWAIT/UMONITOR) be split
> > out so
> > > > > > that DLB/DLB2 could use them instead of its own private byte-encoded
> > > > versions,
> > > > > > but instead we have these wrappers that call the low level intrinsics.
> > Those
> > > > > > wrappers
> > > > > > introduce additional overhead that is not required for DLB/DLB2. I have a
> > > > > > meeting with Ma Liang on Monday to discuss.
> > > > >
> > > > > I thought the ask of DLB was to just substitute the low level
> > umwait/umonitor
> > > > byte
> > > > > encoded instructions DLB has defined privately with similar byte-encoded
> > > > instructions defined in the power
> > > > > patch. The power patch does not directly expose those, which is why I did
> > not
> > > > update DLB/DLB2.
> > > > > The power patch does have the advantage of centralizing the race
> > avoidance
> > > > > logic, which is a good thing for any PMD that wishes to take advantage of
> > > > umwait/umonitor.
> > > >
> > > > So you mean the overhead is a good thing?
> > > >
> > > > > Sorry for the confusion. I just misunderstood what was being asked of DLB
> > in
> > > > regard to switching over..  That being said,
> > > > > I am willing to convert DLB/DLB2 to use  rte_power_monitor(...) in a future
> > > > patch-set.
> > > >
> > > > Why not now?
> > > >
> > > > Indeed there is a confusion and it looks like a lot of novlang
> > > > to exit from the situation.
> > > > We'll wait a clear decision with facts.
> > > >
> > >
> > > Hi Thomas,
> > >
> > > I have updated DLB and DLB2 to use rte_power_monitor(...), and those
> > patches are
> > > ready if you are willing to accept them and the 3 power patches.
> > >
> > > For the sake of consistency, I see the benefit of using the power patch, even if
> > it is
> > > slightly less efficient that the DLB-specific implementation that I currently
> > have.
> > > We have already encountered an empty queue, so this is no longer fast path
> > for the PMD.
> > 
> > I am really concerned that the API in EAL is not the most efficient.
> > Why is that? Can we improve the EAL API?
> > 
> 
> From an efficiency perspective, I only noticed 2 things.
> 1) The size check and associated logic. This could be avoided if we had _8, _16, _32, _64 variants instead of 1 function that handled all possibilities, but even that may be a wash
> with branch prediction 
> 2) The spinlock - but this observation was my mistake, and was flawed. I was looking at the rte_power_monitor_sync(...), and not rte_power_monitor(...), the latter of which does not take a spinlock.
> 
> In summary, I am no longer concerned about efficiency and suitability for DLB/DLB2.

OK, so let's go ahead with a DLB PMD using this API.