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