On Wed, Dec 9, 2020 at 1:32 AM Joao Martins <joao.m.martins@xxxxxxxxxx> wrote: > > Hey, > > This small series, attempts at minimizing 'struct page' overhead by > pursuing a similar approach as Muchun Song series "Free some vmemmap > pages of hugetlb page"[0] but applied to devmap/ZONE_DEVICE. > > [0] https://lore.kernel.org/linux-mm/20201130151838.11208-1-songmuchun@xxxxxxxxxxxxx/ > Oh, well. It looks like you agree with my optimization approach and have fully understood. Also, welcome help me review that series if you have time. :) > The link above describes it quite nicely, but the idea is to reuse tail > page vmemmap areas, particular the area which only describes tail pages. > So a vmemmap page describes 64 struct pages, and the first page for a given > ZONE_DEVICE vmemmap would contain the head page and 63 tail pages. The second > vmemmap page would contain only tail pages, and that's what gets reused across > the rest of the subsection/section. The bigger the page size, the bigger the > savings (2M hpage -> save 6 vmemmap pages; 1G hpage -> save 4094 vmemmap pages). > > In terms of savings, per 1Tb of memory, the struct page cost would go down > with compound pagemap: > > * with 2M pages we lose 4G instead of 16G (0.39% instead of 1.5% of total memory) > * with 1G pages we lose 8MB instead of 16G (0.0007% instead of 1.5% of total memory) > > Along the way I've extended it past 'struct page' overhead *trying* to address a > few performance issues we knew about for pmem, specifically on the > {pin,get}_user_pages* function family with device-dax vmas which are really > slow even of the fast variants. THP is great on -fast variants but all except > hugetlbfs perform rather poorly on non-fast gup. > > So to summarize what the series does: > > Patches 1-5: Much like Muchun series, we reuse tail page areas across a given > page size (namely @align was referred by remaining memremap/dax code) and > enabling of memremap to initialize the ZONE_DEVICE pages as compound pages or a > given @align order. The main difference though, is that contrary to the hugetlbfs > series, there's no vmemmap for the area, because we are onlining it. IOW no > freeing of pages of already initialized vmemmap like the case for hugetlbfs, > which simplifies the logic (besides not being arch-specific). After these, > there's quite visible region bootstrap of pmem memmap given that we would > initialize fewer struct pages depending on the page size. > > NVDIMM namespace bootstrap improves from ~750ms to ~190ms/<=1ms on emulated NVDIMMs > with 2M and 1G respectivally. The net gain in improvement is similarly observed > in proportion when running on actual NVDIMMs. > > Patch 6 - 8: Optimize grabbing/release a page refcount changes given that we > are working with compound pages i.e. we do 1 increment/decrement to the head > page for a given set of N subpages compared as opposed to N individual writes. > {get,pin}_user_pages_fast() for zone_device with compound pagemap consequently > improves considerably, and unpin_user_pages() improves as well when passed a > set of consecutive pages: > > before after > (get_user_pages_fast 1G;2M page size) ~75k us -> ~3.2k ; ~5.2k us > (pin_user_pages_fast 1G;2M page size) ~125k us -> ~3.4k ; ~5.5k us > > The RDMA patch (patch 8/9) is to demonstrate the improvement for an existing > user. For unpin_user_pages() we have an additional test to demonstrate the > improvement. The test performs MR reg/unreg continuously and measuring its > rate for a given period. So essentially ib_mem_get and ib_mem_release being > stress tested which at the end of day means: pin_user_pages_longterm() and > unpin_user_pages() for a scatterlist: > > Before: > 159 rounds in 5.027 sec: 31617.923 usec / round (device-dax) > 466 rounds in 5.009 sec: 10748.456 usec / round (hugetlbfs) > > After: > 305 rounds in 5.010 sec: 16426.047 usec / round (device-dax) > 1073 rounds in 5.004 sec: 4663.622 usec / round (hugetlbfs) > > Patch 9: Improves {pin,get}_user_pages() and its longterm counterpart. It > is very experimental, and I imported most of follow_hugetlb_page(), except > that we do the same trick as gup-fast. In doing the patch I feel this batching > should live in follow_page_mask() and having that being changed to return a set > of pages/something-else when walking over PMD/PUDs for THP / devmap pages. This > patch then brings the previous test of mr reg/unreg (above) on parity > between device-dax and hugetlbfs. > > Some of the patches are a little fresh/WIP (specially patch 3 and 9) and we are > still running tests. Hence the RFC, asking for comments and general direction > of the work before continuing. > > Patches apply on top of linux-next tag next-20201208 (commit a9e26cb5f261). > > Comments and suggestions very much appreciated! > > Thanks, > Joao > > Joao Martins (9): > memremap: add ZONE_DEVICE support for compound pages > sparse-vmemmap: Consolidate arguments in vmemmap section populate > sparse-vmemmap: Reuse vmemmap areas for a given page size > mm/page_alloc: Reuse tail struct pages for compound pagemaps > device-dax: Compound pagemap support > mm/gup: Grab head page refcount once for group of subpages > mm/gup: Decrement head page once for group of subpages > RDMA/umem: batch page unpin in __ib_mem_release() > mm: Add follow_devmap_page() for devdax vmas > > drivers/dax/device.c | 54 ++++++--- > drivers/infiniband/core/umem.c | 25 +++- > include/linux/huge_mm.h | 4 + > include/linux/memory_hotplug.h | 16 ++- > include/linux/memremap.h | 2 + > include/linux/mm.h | 6 +- > mm/gup.c | 130 ++++++++++++++++----- > mm/huge_memory.c | 202 +++++++++++++++++++++++++++++++++ > mm/memory_hotplug.c | 13 ++- > mm/memremap.c | 13 ++- > mm/page_alloc.c | 28 ++++- > mm/sparse-vmemmap.c | 97 +++++++++++++--- > mm/sparse.c | 16 +-- > 13 files changed, 531 insertions(+), 75 deletions(-) > > -- > 2.17.1 > -- Yours, Muchun
