[ofa-general] Re: [PATCH 1/11] IB/ipoib: high dma support

[Eli Cohen]

On Sun, 2007-09-30 at 11:13 +0200, Or Gerlitz wrote: 
> Roland Dreier wrote:
> >  > So after all we do need a flag?
> > 
> > No, after thinking about it I don't think there's any reason to use
> > __GFP_HIGHMEM... it would use less kernel memory on highmem systems
> > but I don't think it really helps in the end.
> 
> I see that none of the drivers for NICs which set the NETIF_F_HIGHDMA 
> bit use some _HIGHMEM GFP directive for their skb allocation, so maybe 
> the two flags need not be correlated. So the NETIF_F_HIGHDMA just comes 
> to say that if from some reason an skb which was not allocated by the 
> driver points to buffer in high-memory, its fine.
> 
> Anyway, I am fine with the patch once the change-log comment is changed 
> AND the ipath issue is resolved with agreement of the parties (Roland, 
> Ralph), if no device capability flag nor implementing high-mem support 
> in the ipath driver is needed, let it be.
> 
> Or.
> 
If I understand correctly, there were two kinds of flags discussed in
this thread. One of them is using __GFP_HIGHMEM for allocating buffers
to be utilized in the receive flow of ipoib cm which it was agreed that
will hurt performance.
The second one is using a capability flag which devices use to declare
that they support high dma - this one is not required too but implies
that ipath will not work on 32 bit platforms (e.g. i386 with PAE).

Saying that, Or, can you clarify how would you like the change-long
comment to be changed?