On Tue, 2013-11-12 at 09:38 +0200, Sagi Grimberg wrote: > On 11/12/2013 9:33 AM, Or Gerlitz wrote: > > On 12/11/2013 09:21, Nicholas A. Bellinger wrote: > >> The point is that if FMR is only required for future T10 PI support, why > >> not just disable it by default, instead of just avoiding it for the > >> small block count == 1 case..? > > b/c for larger IOs (e.g that involves > 1 pages) it gives better > > performance, since you don't need to use SGEs in for RDMA or > > multiple RDMAs, just one RDMA and one SGE > > > > Or. > > Or is correct, > > In the general case of large scatterlists it is better to perform fast > registration and use a single RDMA, but for small IOs (single entry) > skipping fast registration and using dma MR is faster as still you use a > single RDMA. > You are correct that in the future, this optimization will consider DIF > support as even for single entry scatterlists you need to register for > specific protection attributes. > Or & Sagi, thanks for the explanation. In that case, applying to for-next, with a CC' stable for v3.12.y to address this particular performance regression. --nab -- To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
