1 2 .. SPDX-License-Identifier: GPL-2.0 3 4 ========================================= 5 A vmemmap diet for HugeTLB and Device DAX 6 ========================================= 7 8 HugeTLB 9 ======= 10 11 This section is to explain how HugeTLB Vmemmap Optimization (HVO) works. 12 13 The ``struct page`` structures are used to describe a physical page frame. By 14 default, there is a one-to-one mapping from a page frame to its corresponding 15 ``struct page``. 16 17 HugeTLB pages consist of multiple base page size pages and is supported by many 18 architectures. See Documentation/admin-guide/mm/hugetlbpage.rst for more 19 details. On the x86-64 architecture, HugeTLB pages of size 2MB and 1GB are 20 currently supported. Since the base page size on x86 is 4KB, a 2MB HugeTLB page 21 consists of 512 base pages and a 1GB HugeTLB page consists of 262144 base pages. 22 For each base page, there is a corresponding ``struct page``. 23 24 Within the HugeTLB subsystem, only the first 4 ``struct page`` are used to 25 contain unique information about a HugeTLB page. ``__NR_USED_SUBPAGE`` provides 26 this upper limit. The only 'useful' information in the remaining ``struct page`` 27 is the compound_head field, and this field is the same for all tail pages. 28 29 By removing redundant ``struct page`` for HugeTLB pages, memory can be returned 30 to the buddy allocator for other uses. 31 32 Different architectures support different HugeTLB pages. For example, the 33 following table is the HugeTLB page size supported by x86 and arm64 34 architectures. Because arm64 supports 4k, 16k, and 64k base pages and 35 supports contiguous entries, so it supports many kinds of sizes of HugeTLB 36 page. 37 38 +--------------+-----------+-----------------------------------------------+ 39 | Architecture | Page Size | HugeTLB Page Size | 40 +--------------+-----------+-----------+-----------+-----------+-----------+ 41 | x86-64 | 4KB | 2MB | 1GB | | | 42 +--------------+-----------+-----------+-----------+-----------+-----------+ 43 | | 4KB | 64KB | 2MB | 32MB | 1GB | 44 | +-----------+-----------+-----------+-----------+-----------+ 45 | arm64 | 16KB | 2MB | 32MB | 1GB | | 46 | +-----------+-----------+-----------+-----------+-----------+ 47 | | 64KB | 2MB | 512MB | 16GB | | 48 +--------------+-----------+-----------+-----------+-----------+-----------+ 49 50 When the system boot up, every HugeTLB page has more than one ``struct page`` 51 structs which size is (unit: pages):: 52 53 struct_size = HugeTLB_Size / PAGE_SIZE * sizeof(struct page) / PAGE_SIZE 54 55 Where HugeTLB_Size is the size of the HugeTLB page. We know that the size 56 of the HugeTLB page is always n times PAGE_SIZE. So we can get the following 57 relationship:: 58 59 HugeTLB_Size = n * PAGE_SIZE 60 61 Then:: 62 63 struct_size = n * PAGE_SIZE / PAGE_SIZE * sizeof(struct page) / PAGE_SIZE 64 = n * sizeof(struct page) / PAGE_SIZE 65 66 We can use huge mapping at the pud/pmd level for the HugeTLB page. 67 68 For the HugeTLB page of the pmd level mapping, then:: 69 70 struct_size = n * sizeof(struct page) / PAGE_SIZE 71 = PAGE_SIZE / sizeof(pte_t) * sizeof(struct page) / PAGE_SIZE 72 = sizeof(struct page) / sizeof(pte_t) 73 = 64 / 8 74 = 8 (pages) 75 76 Where n is how many pte entries which one page can contains. So the value of 77 n is (PAGE_SIZE / sizeof(pte_t)). 78 79 This optimization only supports 64-bit system, so the value of sizeof(pte_t) 80 is 8. And this optimization also applicable only when the size of ``struct page`` 81 is a power of two. In most cases, the size of ``struct page`` is 64 bytes (e.g. 82 x86-64 and arm64). So if we use pmd level mapping for a HugeTLB page, the 83 size of ``struct page`` structs of it is 8 page frames which size depends on the 84 size of the base page. 85 86 For the HugeTLB page of the pud level mapping, then:: 87 88 struct_size = PAGE_SIZE / sizeof(pmd_t) * struct_size(pmd) 89 = PAGE_SIZE / 8 * 8 (pages) 90 = PAGE_SIZE (pages) 91 92 Where the struct_size(pmd) is the size of the ``struct page`` structs of a 93 HugeTLB page of the pmd level mapping. 94 95 E.g.: A 2MB HugeTLB page on x86_64 consists in 8 page frames while 1GB 96 HugeTLB page consists in 4096. 97 98 Next, we take the pmd level mapping of the HugeTLB page as an example to 99 show the internal implementation of this optimization. There are 8 pages 100 ``struct page`` structs associated with a HugeTLB page which is pmd mapped. 101 102 Here is how things look before optimization:: 103 104 HugeTLB struct pages(8 pages) page frame(8 pages) 105 +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ 106 | | | 0 | -------------> | 0 | 107 | | +-----------+ +-----------+ 108 | | | 1 | -------------> | 1 | 109 | | +-----------+ +-----------+ 110 | | | 2 | -------------> | 2 | 111 | | +-----------+ +-----------+ 112 | | | 3 | -------------> | 3 | 113 | | +-----------+ +-----------+ 114 | | | 4 | -------------> | 4 | 115 | PMD | +-----------+ +-----------+ 116 | level | | 5 | -------------> | 5 | 117 | mapping | +-----------+ +-----------+ 118 | | | 6 | -------------> | 6 | 119 | | +-----------+ +-----------+ 120 | | | 7 | -------------> | 7 | 121 | | +-----------+ +-----------+ 122 | | 123 | | 124 | | 125 +-----------+ 126 127 The value of page->compound_head is the same for all tail pages. The first 128 page of ``struct page`` (page 0) associated with the HugeTLB page contains the 4 129 ``struct page`` necessary to describe the HugeTLB. The only use of the remaining 130 pages of ``struct page`` (page 1 to page 7) is to point to page->compound_head. 131 Therefore, we can remap pages 1 to 7 to page 0. Only 1 page of ``struct page`` 132 will be used for each HugeTLB page. This will allow us to free the remaining 133 7 pages to the buddy allocator. 134 135 Here is how things look after remapping:: 136 137 HugeTLB struct pages(8 pages) page frame(8 pages) 138 +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ 139 | | | 0 | -------------> | 0 | 140 | | +-----------+ +-----------+ 141 | | | 1 | ---------------^ ^ ^ ^ ^ ^ ^ 142 | | +-----------+ | | | | | | 143 | | | 2 | -----------------+ | | | | | 144 | | +-----------+ | | | | | 145 | | | 3 | -------------------+ | | | | 146 | | +-----------+ | | | | 147 | | | 4 | ---------------------+ | | | 148 | PMD | +-----------+ | | | 149 | level | | 5 | -----------------------+ | | 150 | mapping | +-----------+ | | 151 | | | 6 | -------------------------+ | 152 | | +-----------+ | 153 | | | 7 | ---------------------------+ 154 | | +-----------+ 155 | | 156 | | 157 | | 158 +-----------+ 159 160 When a HugeTLB is freed to the buddy system, we should allocate 7 pages for 161 vmemmap pages and restore the previous mapping relationship. 162 163 For the HugeTLB page of the pud level mapping. It is similar to the former. 164 We also can use this approach to free (PAGE_SIZE - 1) vmemmap pages. 165 166 Apart from the HugeTLB page of the pmd/pud level mapping, some architectures 167 (e.g. aarch64) provides a contiguous bit in the translation table entries 168 that hints to the MMU to indicate that it is one of a contiguous set of 169 entries that can be cached in a single TLB entry. 170 171 The contiguous bit is used to increase the mapping size at the pmd and pte 172 (last) level. So this type of HugeTLB page can be optimized only when its 173 size of the ``struct page`` structs is greater than **1** page. 174 175 Notice: The head vmemmap page is not freed to the buddy allocator and all 176 tail vmemmap pages are mapped to the head vmemmap page frame. So we can see 177 more than one ``struct page`` struct with ``PG_head`` (e.g. 8 per 2 MB HugeTLB 178 page) associated with each HugeTLB page. The ``compound_head()`` can handle 179 this correctly. There is only **one** head ``struct page``, the tail 180 ``struct page`` with ``PG_head`` are fake head ``struct page``. We need an 181 approach to distinguish between those two different types of ``struct page`` so 182 that ``compound_head()`` can return the real head ``struct page`` when the 183 parameter is the tail ``struct page`` but with ``PG_head``. 184 185 Device DAX 186 ========== 187 188 The device-dax interface uses the same tail deduplication technique explained 189 in the previous chapter, except when used with the vmemmap in 190 the device (altmap). 191 192 The following page sizes are supported in DAX: PAGE_SIZE (4K on x86_64), 193 PMD_SIZE (2M on x86_64) and PUD_SIZE (1G on x86_64). 194 For powerpc equivalent details see Documentation/arch/powerpc/vmemmap_dedup.rst 195 196 The differences with HugeTLB are relatively minor. 197 198 It only use 3 ``struct page`` for storing all information as opposed 199 to 4 on HugeTLB pages. 200 201 There's no remapping of vmemmap given that device-dax memory is not part of 202 System RAM ranges initialized at boot. Thus the tail page deduplication 203 happens at a later stage when we populate the sections. HugeTLB reuses the 204 the head vmemmap page representing, whereas device-dax reuses the tail 205 vmemmap page. This results in only half of the savings compared to HugeTLB. 206 207 Deduplicated tail pages are not mapped read-only. 208 209 Here's how things look like on device-dax after the sections are populated:: 210 211 +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ 212 | | | 0 | -------------> | 0 | 213 | | +-----------+ +-----------+ 214 | | | 1 | -------------> | 1 | 215 | | +-----------+ +-----------+ 216 | | | 2 | ----------------^ ^ ^ ^ ^ ^ 217 | | +-----------+ | | | | | 218 | | | 3 | ------------------+ | | | | 219 | | +-----------+ | | | | 220 | | | 4 | --------------------+ | | | 221 | PMD | +-----------+ | | | 222 | level | | 5 | ----------------------+ | | 223 | mapping | +-----------+ | | 224 | | | 6 | ------------------------+ | 225 | | +-----------+ | 226 | | | 7 | --------------------------+ 227 | | +-----------+ 228 | | 229 | | 230 | | 231 +-----------+
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.