[Skiboot] [PATCH v2 1/2] astbmc: Add NPU slot type and use it in Garrison

[Alistair Popple]

On Tue, 21 Jun 2016 18:04:47 Russell Currey wrote:
> NPU links are grouped into pairs, signifying that they both correlate to
> the same physical GPU.  These pairs are presented in the Garrison slot
> table as typical PCI devices, creating what are essentially redundant
> entries.  Add a new slot type specifically for NPUs, and subsequently
> perform bdfn correlation using it.
> 
> Signed-off-by: Russell Currey <ruscur at russell.cc>
> ---
> V2: New change to facilitate bdfn allocation without using pbcq
> ---
>  platforms/astbmc/astbmc.h   |  2 ++
>  platforms/astbmc/garrison.c | 36 ++++++++----------------------------
>  platforms/astbmc/slots.c    |  7 +++++++
>  3 files changed, 17 insertions(+), 28 deletions(-)

<snip>
 
> diff --git a/platforms/astbmc/slots.c b/platforms/astbmc/slots.c
> index 36547e1..992f68b 100644
> --- a/platforms/astbmc/slots.c
> +++ b/platforms/astbmc/slots.c
> @@ -70,6 +70,13 @@ static const struct slot_table_entry 
*match_slot_dev_entry(struct phb *phb,
>  			prerror("SLOT: Bad PHB entry type in table !\n");
>  			continue;
>  		}
> +
> +		if (ent->etype == st_npu_slot) {
> +			/* NPU groups are at device level, so ignore function 
*/
> +			if (ent->location == ((pd->bdfn & 0xff) >> 3))

Is this reliable? ent->location is a ST_LOC_NPU_GROUP code but what guarantees 
that the pd->bdfn device number matches the NPU_GROUP number?

Do we need something like "if (ent->location == pd->npu_group)" instead?

- Alistair

> +				return ent;
> +		}
> +
>  		if (ent->location == (pd->bdfn & 0xff))
>  			return ent;
>  	}
>