[ofa-general] [PATCH] infiniband: add userspace support for invalidate stag

[Steve Wise]

On Fri, 2007-05-04 at 19:32 -0500, Steve Wise wrote:
> On Fri, 2007-05-04 at 14:50 -0700, Roland Dreier wrote:
> > A few general things:
> >  - please always submit patches with a changelog entry and
> >    Signed-off-by: line
> >  - please send patches in logical chunks.  Usually I'm complaining
> >    about people combining unrelated things into one patch, but in this
> >    case I think you divided the patch up too much -- rather than 5
> >    patches, this should probably be one kernel patch and one userspace
> >    patch.
> >  - please make libibverbs patches apply to the libibverbs git tree
> >    with -p1.  You seem to have generated patches against an OFED package.
> > 
> > OK, with that out of the way, I think there are still some issues to
> > sort out with how to handle send with invalidate from userspace.
> > These patches don't address the case of new userspace with
> > send-with-invalidate support talking to an unpatched kernel -- it
> > seems that send-with-invalidate would be silently turned into a plain
> > send request, which is not a very good failure mode.
> > 
> > I don't know what the right solution is yet -- a kernel ABI bump for
> > this one case (send with invalidate support for userspace drivers that
> > don't do kernel bypass == amso1100) is ugly.  Maybe we also need a
> > device capabilities bit that says whether send-with-invalidate is
> > supported?
> > 
> 
> There already exists a SEND-INV capabilities flag.
> 
> <snip>
>         IB_DEVICE_SEND_W_INV            = (1<<16),
> 
> I think with the capabilities flag, we shouldn't worry about changing
> the ABI.  But the drivers will need to set this flag.  Amso currently
> does...

Actually, since Amso has set this flag since day one, it doesn't really
solve the ABI issue Roland describes.


Steve.