1.. _transhuge: 2 3============================ 4Transparent Hugepage Support 5============================ 6 7This document describes design principles Transparent Hugepage (THP) 8Support and its interaction with other parts of the memory management. 9 10Design principles 11================= 12 13- "graceful fallback": mm components which don't have transparent hugepage 14 knowledge fall back to breaking huge pmd mapping into table of ptes and, 15 if necessary, split a transparent hugepage. Therefore these components 16 can continue working on the regular pages or regular pte mappings. 17 18- if a hugepage allocation fails because of memory fragmentation, 19 regular pages should be gracefully allocated instead and mixed in 20 the same vma without any failure or significant delay and without 21 userland noticing 22 23- if some task quits and more hugepages become available (either 24 immediately in the buddy or through the VM), guest physical memory 25 backed by regular pages should be relocated on hugepages 26 automatically (with khugepaged) 27 28- it doesn't require memory reservation and in turn it uses hugepages 29 whenever possible (the only possible reservation here is kernelcore= 30 to avoid unmovable pages to fragment all the memory but such a tweak 31 is not specific to transparent hugepage support and it's a generic 32 feature that applies to all dynamic high order allocations in the 33 kernel) 34 35get_user_pages and follow_page 36============================== 37 38get_user_pages and follow_page if run on a hugepage, will return the 39head or tail pages as usual (exactly as they would do on 40hugetlbfs). Most gup users will only care about the actual physical 41address of the page and its temporary pinning to release after the I/O 42is complete, so they won't ever notice the fact the page is huge. But 43if any driver is going to mangle over the page structure of the tail 44page (like for checking page->mapping or other bits that are relevant 45for the head page and not the tail page), it should be updated to jump 46to check head page instead. Taking reference on any head/tail page would 47prevent page from being split by anyone. 48 49.. note:: 50 these aren't new constraints to the GUP API, and they match the 51 same constrains that applies to hugetlbfs too, so any driver capable 52 of handling GUP on hugetlbfs will also work fine on transparent 53 hugepage backed mappings. 54 55In case you can't handle compound pages if they're returned by 56follow_page, the FOLL_SPLIT bit can be specified as parameter to 57follow_page, so that it will split the hugepages before returning 58them. Migration for example passes FOLL_SPLIT as parameter to 59follow_page because it's not hugepage aware and in fact it can't work 60at all on hugetlbfs (but it instead works fine on transparent 61hugepages thanks to FOLL_SPLIT). migration simply can't deal with 62hugepages being returned (as it's not only checking the pfn of the 63page and pinning it during the copy but it pretends to migrate the 64memory in regular page sizes and with regular pte/pmd mappings). 65 66Graceful fallback 67================= 68 69Code walking pagetables but unaware about huge pmds can simply call 70split_huge_pmd(vma, pmd, addr) where the pmd is the one returned by 71pmd_offset. It's trivial to make the code transparent hugepage aware 72by just grepping for "pmd_offset" and adding split_huge_pmd where 73missing after pmd_offset returns the pmd. Thanks to the graceful 74fallback design, with a one liner change, you can avoid to write 75hundred if not thousand of lines of complex code to make your code 76hugepage aware. 77 78If you're not walking pagetables but you run into a physical hugepage 79but you can't handle it natively in your code, you can split it by 80calling split_huge_page(page). This is what the Linux VM does before 81it tries to swapout the hugepage for example. split_huge_page() can fail 82if the page is pinned and you must handle this correctly. 83 84Example to make mremap.c transparent hugepage aware with a one liner 85change:: 86 87 diff --git a/mm/mremap.c b/mm/mremap.c 88 --- a/mm/mremap.c 89 +++ b/mm/mremap.c 90 @@ -41,6 +41,7 @@ static pmd_t *get_old_pmd(struct mm_stru 91 return NULL; 92 93 pmd = pmd_offset(pud, addr); 94 + split_huge_pmd(vma, pmd, addr); 95 if (pmd_none_or_clear_bad(pmd)) 96 return NULL; 97 98Locking in hugepage aware code 99============================== 100 101We want as much code as possible hugepage aware, as calling 102split_huge_page() or split_huge_pmd() has a cost. 103 104To make pagetable walks huge pmd aware, all you need to do is to call 105pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the 106mmap_sem in read (or write) mode to be sure an huge pmd cannot be 107created from under you by khugepaged (khugepaged collapse_huge_page 108takes the mmap_sem in write mode in addition to the anon_vma lock). If 109pmd_trans_huge returns false, you just fallback in the old code 110paths. If instead pmd_trans_huge returns true, you have to take the 111page table lock (pmd_lock()) and re-run pmd_trans_huge. Taking the 112page table lock will prevent the huge pmd to be converted into a 113regular pmd from under you (split_huge_pmd can run in parallel to the 114pagetable walk). If the second pmd_trans_huge returns false, you 115should just drop the page table lock and fallback to the old code as 116before. Otherwise you can proceed to process the huge pmd and the 117hugepage natively. Once finished you can drop the page table lock. 118 119Refcounts and transparent huge pages 120==================================== 121 122Refcounting on THP is mostly consistent with refcounting on other compound 123pages: 124 125 - get_page()/put_page() and GUP operate in head page's ->_refcount. 126 127 - ->_refcount in tail pages is always zero: get_page_unless_zero() never 128 succeed on tail pages. 129 130 - map/unmap of the pages with PTE entry increment/decrement ->_mapcount 131 on relevant sub-page of the compound page. 132 133 - map/unmap of the whole compound page accounted in compound_mapcount 134 (stored in first tail page). For file huge pages, we also increment 135 ->_mapcount of all sub-pages in order to have race-free detection of 136 last unmap of subpages. 137 138PageDoubleMap() indicates that the page is *possibly* mapped with PTEs. 139 140For anonymous pages PageDoubleMap() also indicates ->_mapcount in all 141subpages is offset up by one. This additional reference is required to 142get race-free detection of unmap of subpages when we have them mapped with 143both PMDs and PTEs. 144 145This is optimization required to lower overhead of per-subpage mapcount 146tracking. The alternative is alter ->_mapcount in all subpages on each 147map/unmap of the whole compound page. 148 149For anonymous pages, we set PG_double_map when a PMD of the page got split 150for the first time, but still have PMD mapping. The additional references 151go away with last compound_mapcount. 152 153File pages get PG_double_map set on first map of the page with PTE and 154goes away when the page gets evicted from page cache. 155 156split_huge_page internally has to distribute the refcounts in the head 157page to the tail pages before clearing all PG_head/tail bits from the page 158structures. It can be done easily for refcounts taken by page table 159entries. But we don't have enough information on how to distribute any 160additional pins (i.e. from get_user_pages). split_huge_page() fails any 161requests to split pinned huge page: it expects page count to be equal to 162sum of mapcount of all sub-pages plus one (split_huge_page caller must 163have reference for head page). 164 165split_huge_page uses migration entries to stabilize page->_refcount and 166page->_mapcount of anonymous pages. File pages just got unmapped. 167 168We safe against physical memory scanners too: the only legitimate way 169scanner can get reference to a page is get_page_unless_zero(). 170 171All tail pages have zero ->_refcount until atomic_add(). This prevents the 172scanner from getting a reference to the tail page up to that point. After the 173atomic_add() we don't care about the ->_refcount value. We already known how 174many references should be uncharged from the head page. 175 176For head page get_page_unless_zero() will succeed and we don't mind. It's 177clear where reference should go after split: it will stay on head page. 178 179Note that split_huge_pmd() doesn't have any limitation on refcounting: 180pmd can be split at any point and never fails. 181 182Partial unmap and deferred_split_huge_page() 183============================================ 184 185Unmapping part of THP (with munmap() or other way) is not going to free 186memory immediately. Instead, we detect that a subpage of THP is not in use 187in page_remove_rmap() and queue the THP for splitting if memory pressure 188comes. Splitting will free up unused subpages. 189 190Splitting the page right away is not an option due to locking context in 191the place where we can detect partial unmap. It's also might be 192counterproductive since in many cases partial unmap happens during exit(2) if 193a THP crosses a VMA boundary. 194 195Function deferred_split_huge_page() is used to queue page for splitting. 196The splitting itself will happen when we get memory pressure via shrinker 197interface. 198