Christoph, On 6/24/25 05:46, Christoph Hellwig wrote: > On Mon, Jun 23, 2025 at 09:27:30AM -0600, Keith Busch wrote: >>> - if (nvme_pci_use_sgls(dev, req, iod->sgt.nents)) >>> + if (use_sgl == SGL_FORCED || >>> + (use_sgl == SGL_SUPPORTED && >>> + (!sgl_threshold || nvme_pci_avg_seg_size(req) >= sgl_threshold))) >>> ret = nvme_pci_setup_sgls(nvmeq, req, &cmnd->rw); >> We historically interpreted sgl_threshold set to 0 to mean disable SGL >> usage, maybe because the controller is broken or something. It might be >> okay to have 0 mean to not consider segment sizes, but I just wanted to >> point out this is a different interpretation of the user parameter. > This is still a leftover from the threshold mess in the first version. > I'll fix it for the next one. Are you planning to send a V3 with a fix sg_threshold parameter ? or you are planning to send out a separate patch to fix this change in the behavior ? -ck
