On 6/25/19 10:45 PM, Michal Hocko wrote: > On Tue 25-06-19 20:15:28, John Hubbard wrote: >> On 6/19/19 12:27 PM, Jason Gunthorpe wrote: >>> On Thu, Jun 13, 2019 at 06:23:04PM -0700, John Hubbard wrote: >>>> On 6/13/19 5:43 PM, Ira Weiny wrote: >>>>> On Thu, Jun 13, 2019 at 07:58:29PM +0000, Jason Gunthorpe wrote: >>>>>> On Thu, Jun 13, 2019 at 12:53:02PM -0700, Ralph Campbell wrote: >>>>>>> >>>> ... >>>>> So I think it is ok. Frankly I was wondering if we should remove the public >>>>> type altogether but conceptually it seems ok. But I don't see any users of it >>>>> so... should we get rid of it in the code rather than turning the config off? >>>>> >>>>> Ira >>>> >>>> That seems reasonable. I recall that the hope was for those IBM Power 9 >>>> systems to use _PUBLIC, as they have hardware-based coherent device (GPU) >>>> memory, and so the memory really is visible to the CPU. And the IBM team >>>> was thinking of taking advantage of it. But I haven't seen anything on >>>> that front for a while. >>> >>> Does anyone know who those people are and can we encourage them to >>> send some patches? :) >>> >> >> I asked about this, and it seems that the idea was: DEVICE_PUBLIC was there >> in order to provide an alternative way to do things (such as migrate memory >> to and from a device), in case the combination of existing and near-future >> NUMA APIs was insufficient. This probably came as a follow-up to the early >> 2017-ish conversations about NUMA, in which the linux-mm recommendation was >> "try using HMM mechanisms, and if those are inadequate, then maybe we can >> look at enhancing NUMA so that it has better handling of advanced (GPU-like) >> devices". > > Yes that was the original idea. It sounds so much better to use a common > framework rather than awkward special cased cpuless NUMA nodes with > a weird semantic. User of the neither of the two has shown up so I guess > that the envisioned HW just didn't materialized. Or has there been a > completely different approach chosen? The HW showed up, alright: it's the IBM Power 9, which provides HW-based memory coherency between its CPUs and GPUs. So on this system, the CPU is allowed to access GPU memory, which *could* be modeled as DEVICE_PUBLIC. However, what happened was that the system worked well enough with a combination of the device driver, plus NUMA APIs, plus heaven knows what sort of /proc tuning might have also gone on. :) No one saw the need to reach for the DEVICE_PUBLIC functionality. > >> In the end, however, _PUBLIC was never used, nor does anyone in the local >> (NVIDIA + IBM) kernel vicinity seem to have plans to use it. So it really >> does seem safe to remove, although of course it's good to start with >> BROKEN and see if anyone pops up and complains. > > Well, I do not really see much of a difference. Preserving an unused > code which doesn't have any user in sight just adds a maintenance burden > whether the code depends on BROKEN or not. We can always revert patches > which remove the code once a real user shows up. Sure, I don't see much difference either. Either way seems fine. thanks, -- John Hubbard NVIDIA _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
