On Mon, 20 Aug 2012 21:56:08 +0200, Daniel Vetter <daniel at ffwll.ch> wrote: > On Sat, Aug 11, 2012 at 03:41:21PM +0100, Chris Wilson wrote: > > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk> > > What about putting kmap/unmap abstractions into obj->ops (like the dma_buf > interface already has)? Since the pwrite/pread code is already rather > branch heave I hope we don't see the overhead of the indirect call even > in microbenchmarks (haven't checked). And this way we would also neatly > wrap up dma_bufs for pwrite (if anyone ever really wants that ...). > > The kmap(_atomic) for stolen mem backed objects would boil down to doing > the pointer arithmetic, kunmap would be just a noop. Tried doing so. The lack of struct page for the stolen makes it more cumbersome than it is worth, and worse confusing. -Chris -- Chris Wilson, Intel Open Source Technology Centre
