On Saturday 31 October 2015 06:52 PM, Noam Camus wrote:
> From: Noam Camus <[email protected]>
>
> NPS device got 256 cores and each got 16 HW threads (SMT).
> We use EZchip dedicated ISA to trigger HW scheduler of the
> core that current HW thread belongs to.
> This scheduling makes sure that data beyond barrier is available
> to all HW threads in core and by that to all in device (4K).
>
> Signed-off-by: Noam Camus <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> ---
> arch/arc/include/asm/barrier.h | 8 ++++++++
> 1 files changed, 8 insertions(+), 0 deletions(-)
>
> diff --git a/arch/arc/include/asm/barrier.h b/arch/arc/include/asm/barrier.h
> index a720998..c8229b8 100644
> --- a/arch/arc/include/asm/barrier.h
> +++ b/arch/arc/include/asm/barrier.h
> @@ -34,6 +34,12 @@
>
> #ifdef CONFIG_ISA_ARCOMPACT
>
> +#ifdef CONFIG_ARC_PLAT_EZNPS
> +#include <plat/ctop.h>
> +#define mb() asm volatile (".word %0" : : "i"(CTOP_INST_SCHD_RW) : "memory")
> +#define rmb() asm volatile (".word %0" : : "i"(CTOP_INST_SCHD_RD) : "memory")
Do u need this even for mandatory barriers whose semantics are not related to SMP
at all ? I think you need them only for smb_*
Following is a good introduction to difference between the two !
https://community.arm.com/groups/processors/blog/2011/04/11/memory-access-ordering-part-2--barriers-and-the-linux-kernel
> +#else
> +
> /*
> * ARCompact based cores (ARC700) only have SYNC instruction which is super
> * heavy weight as it flushes the pipeline as well.
> @@ -41,6 +47,8 @@
> */
>
> #define mb() asm volatile("sync\n" : : : "memory")
> +#endif /* CONFIG_ARC_PLAT_EZNPS */
> +
> #endif
>
> #include <asm-generic/barrier.h>
From: Vineet Gupta [mailto:[email protected]]
Sent: Monday, November 02, 2015 10:03 AM
> Do u need this even for mandatory barriers whose semantics are not related to SMP at all ? I think you need them only for smb_*
Yes I do.
For example it is used to proceed access to our Global Interrupt Manager (GIM) when we want to make sure that driver of some device committed all read/writes before we are acknowledging the GIM.
We are not falling to use "sync" since we prefer to have HW thread schedule in the meantime we wait for load/store to be done.
- Noam