| Message ID | 20201026145114.59424-1-songmuchun@bytedance.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Free some vmemmap pages of hugetlb page | expand |
On Mon, Oct 26, 2020 at 10:50:55PM +0800, Muchun Song wrote:
> For tail pages, the value of compound_dtor is the same. So we can reuse
compound_dtor is only set on the first tail page. compound_head is
what you mean here, I think.
On Mon, Oct 26, 2020 at 11:53 PM Matthew Wilcox <willy@infradead.org> wrote: > > On Mon, Oct 26, 2020 at 10:50:55PM +0800, Muchun Song wrote: > > For tail pages, the value of compound_dtor is the same. So we can reuse > > compound_dtor is only set on the first tail page. compound_head is > what you mean here, I think. > Yes, that's right. Sorry for the confusion. Thanks.
On Mon 26-10-20 22:50:55, Muchun Song wrote: > If we uses the 1G hugetlbpage, we can save 4095 pages. This is a very > substantial gain. On our server, run some SPDK/QEMU applications which > will use 1000GB hugetlbpage. With this feature enabled, we can save > ~16GB(1G hugepage)/~11GB(2MB hugepage) memory. [...] > 15 files changed, 1091 insertions(+), 165 deletions(-) > create mode 100644 include/linux/bootmem_info.h > create mode 100644 mm/bootmem_info.c This is a neat idea but the code footprint is really non trivial. To a very tricky code which hugetlb is unfortunately. Saving 1,6% of memory is definitely interesting especially for 1GB pages which tend to be more static and where the savings are more visible. Anyway, I haven't seen any runtime overhead analysis here. What is the price to modify the vmemmap page tables and make them pte rather than pmd based (especially for 2MB hugetlb). Also, how expensive is the vmemmap page tables reconstruction on the freeing path? Thanks!
On Fri, Oct 30, 2020 at 5:14 PM Michal Hocko <mhocko@suse.com> wrote: > > On Mon 26-10-20 22:50:55, Muchun Song wrote: > > If we uses the 1G hugetlbpage, we can save 4095 pages. This is a very > > substantial gain. On our server, run some SPDK/QEMU applications which > > will use 1000GB hugetlbpage. With this feature enabled, we can save > > ~16GB(1G hugepage)/~11GB(2MB hugepage) memory. > [...] > > 15 files changed, 1091 insertions(+), 165 deletions(-) > > create mode 100644 include/linux/bootmem_info.h > > create mode 100644 mm/bootmem_info.c > > This is a neat idea but the code footprint is really non trivial. To a > very tricky code which hugetlb is unfortunately. > > Saving 1,6% of memory is definitely interesting especially for 1GB pages > which tend to be more static and where the savings are more visible. > > Anyway, I haven't seen any runtime overhead analysis here. What is the > price to modify the vmemmap page tables and make them pte rather than > pmd based (especially for 2MB hugetlb). Also, how expensive is the > vmemmap page tables reconstruction on the freeing path? Yeah, I haven't tested the remapping overhead of reserving a hugetlb page. I can do that. But the overhead is not on the allocation/freeing of each hugetlb page, it is only once when we reserve some hugetlb pages through /proc/sys/vm/nr_hugepages. Once the reservation is successful, the subsequent allocation, freeing and using are the same as before (not patched). So I think that the overhead is acceptable. Thanks. > > Thanks! > -- > Michal Hocko > SUSE Labs
On Fri 30-10-20 18:24:25, Muchun Song wrote: > On Fri, Oct 30, 2020 at 5:14 PM Michal Hocko <mhocko@suse.com> wrote: > > > > On Mon 26-10-20 22:50:55, Muchun Song wrote: > > > If we uses the 1G hugetlbpage, we can save 4095 pages. This is a very > > > substantial gain. On our server, run some SPDK/QEMU applications which > > > will use 1000GB hugetlbpage. With this feature enabled, we can save > > > ~16GB(1G hugepage)/~11GB(2MB hugepage) memory. > > [...] > > > 15 files changed, 1091 insertions(+), 165 deletions(-) > > > create mode 100644 include/linux/bootmem_info.h > > > create mode 100644 mm/bootmem_info.c > > > > This is a neat idea but the code footprint is really non trivial. To a > > very tricky code which hugetlb is unfortunately. > > > > Saving 1,6% of memory is definitely interesting especially for 1GB pages > > which tend to be more static and where the savings are more visible. > > > > Anyway, I haven't seen any runtime overhead analysis here. What is the > > price to modify the vmemmap page tables and make them pte rather than > > pmd based (especially for 2MB hugetlb). Also, how expensive is the > > vmemmap page tables reconstruction on the freeing path? > > Yeah, I haven't tested the remapping overhead of reserving a hugetlb > page. I can do that. But the overhead is not on the allocation/freeing of > each hugetlb page, it is only once when we reserve some hugetlb pages > through /proc/sys/vm/nr_hugepages. Once the reservation is successful, > the subsequent allocation, freeing and using are the same as before > (not patched). Yes, that is quite clear. Except for the hugetlb overcommit and migration if the pool is depeleted. Maybe few other cases. > So I think that the overhead is acceptable. Having some numbers for a such a large feature is really needed.
Hi all, This patch series will free some vmemmap pages(struct page structures) associated with each hugetlbpage when preallocated to save memory. Nowadays we track the status of physical page frames using `struct page` arranged in one or more arrays. And here exists one-to-one mapping between the physical page frame and the corresponding `struct page`. The hugetlbpage support is built on top of multiple page size support that is provided by most modern architectures. For example, x86 CPUs normally support 4K and 2M (1G if architecturally supported) page sizes. Every hugetlbpage has more than one `struct page`. The 2M hugetlbpage has 512 `struct page` and 1G hugetlbpage has 4096 `struct page`. But in the core of hugetlbpage only uses the first 4 `struct page` to store metadata associated with each hugetlbpage. The rest of the `struct page` are usually read the compound_head field which are all the same value. If we can free some struct page memory to buddy system so that we can save a lot of memory. When the system boot up, every 2M hugetlbpage has 512 `struct page` which is 8 pages(sizeof(struct page) * 512 / PAGE_SIZE). hugetlbpage struct pages(8 pages) page frame(8 pages) +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ | | | 0 | -------------> | 0 | | | | 1 | -------------> | 1 | | | | 2 | -------------> | 2 | | | | 3 | -------------> | 3 | | | | 4 | -------------> | 4 | | 2M | | 5 | -------------> | 5 | | | | 6 | -------------> | 6 | | | | 7 | -------------> | 7 | | | +-----------+ +-----------+ | | | | +-----------+ When a hugetlbpage is preallocated, we can change the mapping from above to bellow. hugetlbpage struct pages(8 pages) page frame(8 pages) +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ | | | 0 | -------------> | 0 | | | | 1 | -------------> | 1 | | | | 2 | -------------> +-----------+ | | | 3 | -----------------^ ^ ^ ^ ^ | | | 4 | -------------------+ | | | | 2M | | 5 | ---------------------+ | | | | | 6 | -----------------------+ | | | | 7 | -------------------------+ | | +-----------+ | | | | +-----------+ For tail pages, the value of compound_dtor is the same. So we can reuse first page of tail page structs. We map the virtual addresses of the remaining 6 pages of tail page structs to the first tail page struct, and then free these 6 pages. Therefore, we need to reserve at least 2 pages as vmemmap areas. When a hugetlbpage is freed to the buddy system, we should allocate 6 pages for vmemmap pages and restore the previous mapping relationship. If we uses the 1G hugetlbpage, we can save 4095 pages. This is a very substantial gain. On our server, run some SPDK/QEMU applications which will use 1000GB hugetlbpage. With this feature enabled, we can save ~16GB(1G hugepage)/~11GB(2MB hugepage) memory. changelog in v2: 1. Fix do not call dissolve_compound_page in alloc_huge_page_vmemmap(). 2. Fix some typo and code style problems. 3. Remove unused handle_vmemmap_fault(). 4. Merge some commits to one commit suggested by Mike. Muchun Song (19): mm/memory_hotplug: Move bootmem info registration API to bootmem_info.c mm/memory_hotplug: Move {get,put}_page_bootmem() to bootmem_info.c mm/hugetlb: Introduce a new config HUGETLB_PAGE_FREE_VMEMMAP mm/hugetlb: Introduce nr_free_vmemmap_pages in the struct hstate mm/hugetlb: Introduce pgtable allocation/freeing helpers mm/bootmem_info: Introduce {free,prepare}_vmemmap_page() mm/hugetlb: Free the vmemmap pages associated with each hugetlb page mm/hugetlb: Defer freeing of hugetlb pages mm/hugetlb: Allocate the vmemmap pages associated with each hugetlb page mm/hugetlb: Introduce remap_huge_page_pmd_vmemmap helper mm/hugetlb: Use PG_slab to indicate split pmd mm/hugetlb: Support freeing vmemmap pages of gigantic page mm/hugetlb: Add a BUILD_BUG_ON to check if struct page size is a power of two mm/hugetlb: Clear PageHWPoison on the non-error memory page mm/hugetlb: Flush work when dissolving hugetlb page mm/hugetlb: Add a kernel parameter hugetlb_free_vmemmap mm/hugetlb: Merge pte to huge pmd only for gigantic page mm/hugetlb: Gather discrete indexes of tail page mm/hugetlb: Add BUILD_BUG_ON to catch invalid usage of tail struct page .../admin-guide/kernel-parameters.txt | 9 + Documentation/admin-guide/mm/hugetlbpage.rst | 3 + arch/x86/include/asm/hugetlb.h | 20 + arch/x86/include/asm/pgtable_64_types.h | 8 + arch/x86/mm/init_64.c | 5 +- fs/Kconfig | 16 + include/linux/bootmem_info.h | 65 ++ include/linux/hugetlb.h | 50 ++ include/linux/hugetlb_cgroup.h | 15 +- include/linux/memory_hotplug.h | 27 - mm/Makefile | 1 + mm/bootmem_info.c | 125 +++ mm/hugetlb.c | 795 +++++++++++++++++- mm/memory_hotplug.c | 116 --- mm/sparse.c | 1 + 15 files changed, 1091 insertions(+), 165 deletions(-) create mode 100644 include/linux/bootmem_info.h create mode 100644 mm/bootmem_info.c