Re: CVS commit: src/sys/arch/arm/include

[Taylor R Campbell <riastradh%NetBSD.org@localhost>]

> Module Name:    src
> Committed By:   skrll
> Date:           Fri Jun  5 08:03:38 UTC 2026
> 
> Modified Files:
>         src/sys/arch/arm/include: cpufunc.h
> 
> Log Message:
> arm: relax coherent DMA ordering barriers from DSB to DMB
> 
> Use DMB instead of DSB for the ARM coherent DMA ordering macros dma_*_*()
> 
> The previous definitions used DSB, which enforces completion semantics and
> is heavier than needed for coherent device DMA ordering. DMB provides ordering
> of memory operations without requiring full completion, making it the
> appropriate barrier for these coherent-only CPU/device DMA paths.
> 
> Tested on Fusion VM, Orion O6, and Thunderx.
> 
> There is an approximate 1% performance improvement for the Fusion VM, but
> less for Orion O6 and Thunderx.

Thanks!

I was preparing to commit this with the following message but you beat
me to it.


arm: Relax coherent case of bus_dmamap_sync to use DMB, not DSB.

=> DMB is needed to order memory operations with respect to other
   memory operations, such as filling an outgoing DMA buffer and
   writing to a doorbell register.  We definitely need this.

=> DSB implies DMB, and is needed to order TLB and cache maintenance
   instructions with with respect to memory operations.  But the
   coherent case is the case where we don't need to issue cache
   maintenance instructions.  So there is no need to use DSB here,
   and I don't think there ever was one -- not even in earlier
   revisions of the architecture before some parts of the memory
   model were strengthened (when Arm decided to guarantee
   other-multicopy-atomicity, which is not relevant to DSB vs DMB).

On some cores (e.g., Arm Cortex-A9 r4 Software Developers Errata
Notice, ARM UAN 0009D ID32315, Erratum 794072: A short loop including
a DMB instruction might cause a denial of service on another
processor which executes a CP15 broadcast operation
<https://documentation-service.arm.com/static/608118675e70d934bc69f143>)
there are bugs requiring the use of DSB instead of DMB -- but that's
a systemwide issue, not related to DMA I/O.  And, without rewriting
every instruction stream to replace DMB by DSB, it can _only_ be
properly fixed by a CPU state bit causing DMB to act like DSB.

If this causes issues, let's get them documented and justified!
Because right now, there is no Arm documentation to my knowledge --
nor has there ever been -- justifying the use of DSB here.