The PAT work on x86 has finally made pgprot_writecombine() a usable API for modular drivers. As the comment indicates, this is exactly what we want to use in mlx4_ib to map BlueFlame pages up to userspace, since using WC for these pages improves small message latency significantly.
Signed-off-by: Roland Dreier <[email protected]> --- Could someone test this patch on top of 2.6.29, and see if it provides the right latency benefit? If it looks good, I'll send it upstream for 2.6.30. Thanks, Roland diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c index 2ccb9d3..ae3d759 100644 --- a/drivers/infiniband/hw/mlx4/main.c +++ b/drivers/infiniband/hw/mlx4/main.c @@ -394,8 +394,7 @@ static int mlx4_ib_mmap(struct ib_ucontext *context, struct vm_area_struct *vma) PAGE_SIZE, vma->vm_page_prot)) return -EAGAIN; } else if (vma->vm_pgoff == 1 && dev->dev->caps.bf_reg_size != 0) { - /* FIXME want pgprot_writecombine() for BlueFlame pages */ - vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); + vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot); if (io_remap_pfn_range(vma, vma->vm_start, to_mucontext(context)->uar.pfn + _______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
