[ofa-general] Re: [PATCH/RFC] Remove IB_DEVICE_SEND_W_INV from 2.6.26

[Steve Wise]

Sounds ok to me.


Roland Dreier wrote:
> Given that we should have full support for memory management extensions
> pending for 2.6.27, and the support we have for send w/ invalidate in
> 2.6.26 is incomplete (no provision for returning STag/L_Key in receive
> completion and no implementation of that in amso1100 for one thing), I
> think it makes sense to simply remove the IB_DEVICE_SEND_W_INV
> capability flag rather than moving it to a new bit position.
>
> Then when we add all the memory management extension support in 2.6.27,
> we can just use bit 21 for IB_DEVICE_MEM_MGT_EXTENSIONS and avoid having
> such fine grained distinctions, and avoid having all sorts of strange
> code to monkey around with the SEND_W_INV bit in libibverbs and
> userspace driver libraries.
>
> Thoughts pro or con?
>
>  - R.
>
> diff --git a/drivers/infiniband/hw/amso1100/c2_rnic.c b/drivers/infiniband/hw/amso1100/c2_rnic.c
> index 9a054c6..b1441ae 100644
> --- a/drivers/infiniband/hw/amso1100/c2_rnic.c
> +++ b/drivers/infiniband/hw/amso1100/c2_rnic.c
> @@ -455,8 +455,7 @@ int __devinit c2_rnic_init(struct c2_dev *c2dev)
>  	     IB_DEVICE_CURR_QP_STATE_MOD |
>  	     IB_DEVICE_SYS_IMAGE_GUID |
>  	     IB_DEVICE_ZERO_STAG |
> -	     IB_DEVICE_MEM_WINDOW |
> -	     IB_DEVICE_SEND_W_INV);
> +	     IB_DEVICE_MEM_WINDOW);
>  
>  	/* Allocate the qptr_array */
>  	c2dev->qptr_array = vmalloc(C2_MAX_CQS * sizeof(void *));
> diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h
> index 911a661..31d30b1 100644
> --- a/include/rdma/ib_verbs.h
> +++ b/include/rdma/ib_verbs.h
> @@ -105,7 +105,6 @@ enum ib_device_cap_flags {
>  	 */
>  	IB_DEVICE_UD_IP_CSUM		= (1<<18),
>  	IB_DEVICE_UD_TSO		= (1<<19),
> -	IB_DEVICE_SEND_W_INV		= (1<<21),
>  };
>  
>  enum ib_atomic_cap {
>