On 23/11/2023 19:46, Peter Xu wrote: > On Thu, Nov 23, 2023 at 07:11:19PM +0000, Ryan Roberts wrote: >> Hi, >> >> I'm not sure I've 100% understood the crossover between this series and my work >> to support arm64's contpte mappings generally for anonymous and file-backed memory. > > No worry, there's no confliction. If you worked on that it's only be > something nice on top. Also, I'm curious if you have performance numbers, I have perf numbers for high level use cases (kernel compilation and Speedometer Java Script benchmarks) at https://lore.kernel.org/linux-arm-kernel/20230622144210.2623299-1-ryan.roberts@xxxxxxx/ I don't have any micro-benchmarks for GUP though, if that's your question. Is there an easy-to-use test I can run to get some numbers? I'd be happy to try it out. > because I'm going to do some test for hugetlb cont_ptes (which is only the > current plan), and if you got those it'll be a great baseline for me, > because it should be similar in you case even though the goal is slightly > different. > >> >> My approach is to transparently use contpte mappings when core-mm request pte >> mappings that meet the requirements; and its all based around intercepting the >> normal (non-hugetlb) helpers (e.g. set_ptes(), ptep_get() and friends). There is >> no semantic change to the core-mm. See [1]. It relies on 1) the page cache using >> large folios and 2) my "small-sized THP" series which starts using arbitrary >> sized large folios for anonymous memory [2]. >> >> If I've understood this conversation correctly there is an object called hugepd, >> which today is only supported by powerpc, but which could allow the core-mm to >> control the mapping granularity? I can see some value in exposing that control >> to core-mm in the (very) long term. > > For me it's needed immediately, because hugetlb_follow_page_mask() will be > gone after the last patch. > >> >> [1] https://lore.kernel.org/all/20231115163018.1303287-1-ryan.roberts@xxxxxxx/ >> [2] https://lore.kernel.org/linux-mm/20231115132734.931023-1-ryan.roberts@xxxxxxx/ > > AFAICT you haven't yet worked on gup then, after I glimpsed the above > series. No, I haven't touched GUP at all. The approach is fully inside the arm64 arch code (except 1 patch to core-mm which enables an optimization). So as far as GUP and the rest of the core-mm is concerned, there are still only page-sized ptes and they can all be iterated over and accessed as normal. > > It's a matter of whether one follow_page_mask() call can fetch more than > one page* for a cont_pte entry on aarch64 for a large non-hugetlb folio > (and if this series lands, it'll be the same to hugetlb or non-hugetlb). > Now the current code can only fetch one page I think. > > Thanks, >
