[ofa-general] Re: [PATCH] mlx4: Fix the value of the pkey_index in the completion

[Dotan Barak]

Roland Dreier wrote:
>  > Without this fix, incoming GSI packets on port 2 gets invalid pkey index in the completion,
>  > which prevent from the mad layer to send back a response.
>
> Does this explain the problems that someone saw when using all port 2
> of a connectx setup?
>   
Yes. without this mask the value of the pkey_index was 0x80. In the 
function build_mlx_header
there is a call to ib_get_cached_pkey which failed because the pkey 
index was out of range
(i sent you a patch that check the return value). This failure caused to 
uninitialized pkey value to
be sent in the MAD response ....

> Why does it only affect port 2?
>
>  > -		wc->pkey_index     = be32_to_cpu(cqe->immed_rss_invalid) >> 16;
>  > +		wc->pkey_index     = (u16)(be32_to_cpu(cqe->immed_rss_invalid) & 0x7f);
>
> It seems the (u16) cast here is doubly useless, since wc->pkey_index
> is already u16, and we're masking with 0x7f anyway.
>   
If you want i can send a new patch without the casting ...
> By the way, will it always work to mask with 0x7f?  Or is it possible
> that the P_Key table might have more than 128 entries?
>   
Yes, 0x7f should be enough for the ConnectX HCA because it supports 128 
pkey entries per port.
(this line will be written in the next version of the ConnectX PRM)

Dotan