RE: [RFC PATCH 1/8] x86/cpufeatures: Enumerate MOVDIRI instruction

[David Laight]

From: Fenghua Yu
> Sent: 19 June 2018 22:37
> To: Thomas Gleixner
> Cc: Fenghua Yu; Ingo Molnar; H Peter Anvin; Ashok Raj; Alan Cox; Ravi V Shankar; linux-kernel; x86
> Subject: Re: [RFC PATCH 1/8] x86/cpufeatures: Enumerate MOVDIRI instruction
> 
> On Tue, Jun 19, 2018 at 10:57:44AM +0200, Thomas Gleixner wrote:
> > On Fri, 15 Jun 2018, Fenghua Yu wrote:
> >
> > > MOVDIRI moves doubleword or quadword from register to memory through
> > > direct store which is implemented by using write combining (WC) for
> > > writing data directly into memory without caching the data.
> >
> > And that is useful for what?
> 
> Programmable agents can handle streaming offload (e.g. high speed packet
> processing in network). Hardware implements a doorbell (tail pointer)
> register that is updated by software when adding new work-elements to
> the streaming offload work-queue.
> 
> MOVDIRI can be used as the doorbell write which is a 4-byte or 8-byte
> uncachable write to MMIO. MOVDIRI has lower overhead than other ways
> to write the doorbell.

I'd have thought that it wouldn't make any significant difference for
uncached accesses to device registers.

> In low latency offload (e.g. Non-Volatile Memory, etc), MOVDIR64B writes
> work descriptors (and data in some cases) to device-hosted work-queues
> with atomicity.

More likely it is useful more writing to memory without polluting the
data cache.
This might be because the programmer knows the data won't be read for
a long time (at least by the cpu in question).
It might also be useful to avoid a lot of cache snooping on data that
will be accessed by hardware - especially if the hardware is also
likely to be writing to the same cache line.

I can also just about imagine MOVDIR64B being useful for generating
64 byte PCIe TLP to optimise memcpy_to/fromio() without needing
an AVX512 register.

	David