3 ============================
4 Transparent Hugepage Support
5 ============================
7 This document describes design principles for Transparent Hugepage (THP)
8 support and its interaction with other parts of the memory management
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.
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
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)
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
36 get_user_pages and follow_page
37 ==============================
39 get_user_pages and follow_page if run on a hugepage, will return the
40 head or tail pages as usual (exactly as they would do on
41 hugetlbfs). Most GUP users will only care about the actual physical
42 address of the page and its temporary pinning to release after the I/O
43 is complete, so they won't ever notice the fact the page is huge. But
44 if any driver is going to mangle over the page structure of the tail
45 page (like for checking page->mapping or other bits that are relevant
46 for the head page and not the tail page), it should be updated to jump
47 to check head page instead. Taking a reference on any head/tail page would
48 prevent the page from being split by anyone.
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.
56 In case you can't handle compound pages if they're returned by
57 follow_page, the FOLL_SPLIT bit can be specified as a parameter to
58 follow_page, so that it will split the hugepages before returning
64 Code walking pagetables but unaware about huge pmds can simply call
65 split_huge_pmd(vma, pmd, addr) where the pmd is the one returned by
66 pmd_offset. It's trivial to make the code transparent hugepage aware
67 by just grepping for "pmd_offset" and adding split_huge_pmd where
68 missing after pmd_offset returns the pmd. Thanks to the graceful
69 fallback design, with a one liner change, you can avoid to write
70 hundreds if not thousands of lines of complex code to make your code
73 If you're not walking pagetables but you run into a physical hugepage
74 that you can't handle natively in your code, you can split it by
75 calling split_huge_page(page). This is what the Linux VM does before
76 it tries to swapout the hugepage for example. split_huge_page() can fail
77 if the page is pinned and you must handle this correctly.
79 Example to make mremap.c transparent hugepage aware with a one liner
82 diff --git a/mm/mremap.c b/mm/mremap.c
85 @@ -41,6 +41,7 @@ static pmd_t *get_old_pmd(struct mm_stru
88 pmd = pmd_offset(pud, addr);
89 + split_huge_pmd(vma, pmd, addr);
90 if (pmd_none_or_clear_bad(pmd))
93 Locking in hugepage aware code
94 ==============================
96 We want as much code as possible hugepage aware, as calling
97 split_huge_page() or split_huge_pmd() has a cost.
99 To make pagetable walks huge pmd aware, all you need to do is to call
100 pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the
101 mmap_sem in read (or write) mode to be sure a huge pmd cannot be
102 created from under you by khugepaged (khugepaged collapse_huge_page
103 takes the mmap_sem in write mode in addition to the anon_vma lock). If
104 pmd_trans_huge returns false, you just fallback in the old code
105 paths. If instead pmd_trans_huge returns true, you have to take the
106 page table lock (pmd_lock()) and re-run pmd_trans_huge. Taking the
107 page table lock will prevent the huge pmd being converted into a
108 regular pmd from under you (split_huge_pmd can run in parallel to the
109 pagetable walk). If the second pmd_trans_huge returns false, you
110 should just drop the page table lock and fallback to the old code as
111 before. Otherwise, you can proceed to process the huge pmd and the
112 hugepage natively. Once finished, you can drop the page table lock.
114 Refcounts and transparent huge pages
115 ====================================
117 Refcounting on THP is mostly consistent with refcounting on other compound
120 - get_page()/put_page() and GUP operate on head page's ->_refcount.
122 - ->_refcount in tail pages is always zero: get_page_unless_zero() never
123 succeeds on tail pages.
125 - map/unmap of the pages with PTE entry increment/decrement ->_mapcount
126 on relevant sub-page of the compound page.
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.
133 PageDoubleMap() indicates that the page is *possibly* mapped with PTEs.
135 For anonymous pages, PageDoubleMap() also indicates ->_mapcount in all
136 subpages is offset up by one. This additional reference is required to
137 get race-free detection of unmap of subpages when we have them mapped with
140 This optimization is required to lower the overhead of per-subpage mapcount
141 tracking. The alternative is to alter ->_mapcount in all subpages on each
142 map/unmap of the whole compound page.
144 For anonymous pages, we set PG_double_map when a PMD of the page is split
145 for the first time, but still have a PMD mapping. The additional references
146 go away with the last compound_mapcount.
148 File pages get PG_double_map set on the first map of the page with PTE and
149 goes away when the page gets evicted from the page cache.
151 split_huge_page internally has to distribute the refcounts in the head
152 page to the tail pages before clearing all PG_head/tail bits from the page
153 structures. It can be done easily for refcounts taken by page table
154 entries, but we don't have enough information on how to distribute any
155 additional pins (i.e. from get_user_pages). split_huge_page() fails any
156 requests to split pinned huge pages: it expects page count to be equal to
157 the sum of mapcount of all sub-pages plus one (split_huge_page caller must
158 have a reference to the head page).
160 split_huge_page uses migration entries to stabilize page->_refcount and
161 page->_mapcount of anonymous pages. File pages just get unmapped.
163 We are safe against physical memory scanners too: the only legitimate way
164 a scanner can get a reference to a page is get_page_unless_zero().
166 All tail pages have zero ->_refcount until atomic_add(). This prevents the
167 scanner from getting a reference to the tail page up to that point. After the
168 atomic_add() we don't care about the ->_refcount value. We already know how
169 many references should be uncharged from the head page.
171 For head page get_page_unless_zero() will succeed and we don't mind. It's
172 clear where references should go after split: it will stay on the head page.
174 Note that split_huge_pmd() doesn't have any limitations on refcounting:
175 pmd can be split at any point and never fails.
177 Partial unmap and deferred_split_huge_page()
178 ============================================
180 Unmapping part of THP (with munmap() or other way) is not going to free
181 memory immediately. Instead, we detect that a subpage of THP is not in use
182 in page_remove_rmap() and queue the THP for splitting if memory pressure
183 comes. Splitting will free up unused subpages.
185 Splitting the page right away is not an option due to locking context in
186 the place where we can detect partial unmap. It also might be
187 counterproductive since in many cases partial unmap happens during exit(2) if
188 a THP crosses a VMA boundary.
190 The function deferred_split_huge_page() is used to queue a page for splitting.
191 The splitting itself will happen when we get memory pressure via shrinker