On Tue, Apr 22, 2025 at 06:39:56AM +0200, Christoph Hellwig wrote: > On Fri, Apr 18, 2025 at 09:47:54AM +0300, Leon Romanovsky wrote: > > From: Kanchan Joshi <joshi.k@xxxxxxxxxxx> > > > > blk_rq_dma_map API is costly for single-segment requests. > > Avoid using it and map the bio_vec directly. > > This needs to be folded into the earlier patches or split prep patches > instead of undoing work done earlier, preferably combined with a bit > of code movement so that the new nvme_try_setup_prp_simple stays in > the same place as before and the diff shows it reusing code. > > E.g. change > > "nvme-pci: use a better encoding for small prp pool allocations" to > already use the flags instead of my boolean, and maybe include > abort in the flags instead of using a separate bool so that we > don't increase hte iod size. > > Slot in a new patch after that that dropping the single SGL segment > fastpath if we think we don't need that, although if we need the PRP > one I suspect that one would still be very helpful as well. > > Add a patch if we want the try_ version of, although when keeping > the optimization for SGLs as well that are will look a bit different. I uploaded new tag dma-split-Apr-22 with the changes. At the end, I decided to keep SGL optimized path too, but didn't create separate patch for try_* variants as they require block iterator, which is introduced in conversion patch only. > > I'm happy to give away my patch authorship credits if that helps with > the folding. I kept you as an author. Please advise if I need to change it to be someone else. Thanks
