1.. SPDX-License-Identifier: GPL-2.0 2 3========================================= 4A vmemmap diet for HugeTLB and Device DAX 5========================================= 6 7HugeTLB 8======= 9 10The struct page structures (page structs) are used to describe a physical 11page frame. By default, there is a one-to-one mapping from a page frame to 12it's corresponding page struct. 13 14HugeTLB pages consist of multiple base page size pages and is supported by many 15architectures. See Documentation/admin-guide/mm/hugetlbpage.rst for more 16details. On the x86-64 architecture, HugeTLB pages of size 2MB and 1GB are 17currently supported. Since the base page size on x86 is 4KB, a 2MB HugeTLB page 18consists of 512 base pages and a 1GB HugeTLB page consists of 4096 base pages. 19For each base page, there is a corresponding page struct. 20 21Within the HugeTLB subsystem, only the first 4 page structs are used to 22contain unique information about a HugeTLB page. __NR_USED_SUBPAGE provides 23this upper limit. The only 'useful' information in the remaining page structs 24is the compound_head field, and this field is the same for all tail pages. 25 26By removing redundant page structs for HugeTLB pages, memory can be returned 27to the buddy allocator for other uses. 28 29Different architectures support different HugeTLB pages. For example, the 30following table is the HugeTLB page size supported by x86 and arm64 31architectures. Because arm64 supports 4k, 16k, and 64k base pages and 32supports contiguous entries, so it supports many kinds of sizes of HugeTLB 33page. 34 35+--------------+-----------+-----------------------------------------------+ 36| Architecture | Page Size | HugeTLB Page Size | 37+--------------+-----------+-----------+-----------+-----------+-----------+ 38| x86-64 | 4KB | 2MB | 1GB | | | 39+--------------+-----------+-----------+-----------+-----------+-----------+ 40| | 4KB | 64KB | 2MB | 32MB | 1GB | 41| +-----------+-----------+-----------+-----------+-----------+ 42| arm64 | 16KB | 2MB | 32MB | 1GB | | 43| +-----------+-----------+-----------+-----------+-----------+ 44| | 64KB | 2MB | 512MB | 16GB | | 45+--------------+-----------+-----------+-----------+-----------+-----------+ 46 47When the system boot up, every HugeTLB page has more than one struct page 48structs which size is (unit: pages):: 49 50 struct_size = HugeTLB_Size / PAGE_SIZE * sizeof(struct page) / PAGE_SIZE 51 52Where HugeTLB_Size is the size of the HugeTLB page. We know that the size 53of the HugeTLB page is always n times PAGE_SIZE. So we can get the following 54relationship:: 55 56 HugeTLB_Size = n * PAGE_SIZE 57 58Then:: 59 60 struct_size = n * PAGE_SIZE / PAGE_SIZE * sizeof(struct page) / PAGE_SIZE 61 = n * sizeof(struct page) / PAGE_SIZE 62 63We can use huge mapping at the pud/pmd level for the HugeTLB page. 64 65For the HugeTLB page of the pmd level mapping, then:: 66 67 struct_size = n * sizeof(struct page) / PAGE_SIZE 68 = PAGE_SIZE / sizeof(pte_t) * sizeof(struct page) / PAGE_SIZE 69 = sizeof(struct page) / sizeof(pte_t) 70 = 64 / 8 71 = 8 (pages) 72 73Where n is how many pte entries which one page can contains. So the value of 74n is (PAGE_SIZE / sizeof(pte_t)). 75 76This optimization only supports 64-bit system, so the value of sizeof(pte_t) 77is 8. And this optimization also applicable only when the size of struct page 78is a power of two. In most cases, the size of struct page is 64 bytes (e.g. 79x86-64 and arm64). So if we use pmd level mapping for a HugeTLB page, the 80size of struct page structs of it is 8 page frames which size depends on the 81size of the base page. 82 83For the HugeTLB page of the pud level mapping, then:: 84 85 struct_size = PAGE_SIZE / sizeof(pmd_t) * struct_size(pmd) 86 = PAGE_SIZE / 8 * 8 (pages) 87 = PAGE_SIZE (pages) 88 89Where the struct_size(pmd) is the size of the struct page structs of a 90HugeTLB page of the pmd level mapping. 91 92E.g.: A 2MB HugeTLB page on x86_64 consists in 8 page frames while 1GB 93HugeTLB page consists in 4096. 94 95Next, we take the pmd level mapping of the HugeTLB page as an example to 96show the internal implementation of this optimization. There are 8 pages 97struct page structs associated with a HugeTLB page which is pmd mapped. 98 99Here is how things look before optimization:: 100 101 HugeTLB struct pages(8 pages) page frame(8 pages) 102 +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ 103 | | | 0 | -------------> | 0 | 104 | | +-----------+ +-----------+ 105 | | | 1 | -------------> | 1 | 106 | | +-----------+ +-----------+ 107 | | | 2 | -------------> | 2 | 108 | | +-----------+ +-----------+ 109 | | | 3 | -------------> | 3 | 110 | | +-----------+ +-----------+ 111 | | | 4 | -------------> | 4 | 112 | PMD | +-----------+ +-----------+ 113 | level | | 5 | -------------> | 5 | 114 | mapping | +-----------+ +-----------+ 115 | | | 6 | -------------> | 6 | 116 | | +-----------+ +-----------+ 117 | | | 7 | -------------> | 7 | 118 | | +-----------+ +-----------+ 119 | | 120 | | 121 | | 122 +-----------+ 123 124The value of page->compound_head is the same for all tail pages. The first 125page of page structs (page 0) associated with the HugeTLB page contains the 4 126page structs necessary to describe the HugeTLB. The only use of the remaining 127pages of page structs (page 1 to page 7) is to point to page->compound_head. 128Therefore, we can remap pages 1 to 7 to page 0. Only 1 page of page structs 129will be used for each HugeTLB page. This will allow us to free the remaining 1307 pages to the buddy allocator. 131 132Here is how things look after remapping:: 133 134 HugeTLB struct pages(8 pages) page frame(8 pages) 135 +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ 136 | | | 0 | -------------> | 0 | 137 | | +-----------+ +-----------+ 138 | | | 1 | ---------------^ ^ ^ ^ ^ ^ ^ 139 | | +-----------+ | | | | | | 140 | | | 2 | -----------------+ | | | | | 141 | | +-----------+ | | | | | 142 | | | 3 | -------------------+ | | | | 143 | | +-----------+ | | | | 144 | | | 4 | ---------------------+ | | | 145 | PMD | +-----------+ | | | 146 | level | | 5 | -----------------------+ | | 147 | mapping | +-----------+ | | 148 | | | 6 | -------------------------+ | 149 | | +-----------+ | 150 | | | 7 | ---------------------------+ 151 | | +-----------+ 152 | | 153 | | 154 | | 155 +-----------+ 156 157When a HugeTLB is freed to the buddy system, we should allocate 7 pages for 158vmemmap pages and restore the previous mapping relationship. 159 160For the HugeTLB page of the pud level mapping. It is similar to the former. 161We also can use this approach to free (PAGE_SIZE - 1) vmemmap pages. 162 163Apart from the HugeTLB page of the pmd/pud level mapping, some architectures 164(e.g. aarch64) provides a contiguous bit in the translation table entries 165that hints to the MMU to indicate that it is one of a contiguous set of 166entries that can be cached in a single TLB entry. 167 168The contiguous bit is used to increase the mapping size at the pmd and pte 169(last) level. So this type of HugeTLB page can be optimized only when its 170size of the struct page structs is greater than 1 page. 171 172Notice: The head vmemmap page is not freed to the buddy allocator and all 173tail vmemmap pages are mapped to the head vmemmap page frame. So we can see 174more than one struct page struct with PG_head (e.g. 8 per 2 MB HugeTLB page) 175associated with each HugeTLB page. The compound_head() can handle this 176correctly (more details refer to the comment above compound_head()). 177 178Device DAX 179========== 180 181The device-dax interface uses the same tail deduplication technique explained 182in the previous chapter, except when used with the vmemmap in 183the device (altmap). 184 185The following page sizes are supported in DAX: PAGE_SIZE (4K on x86_64), 186PMD_SIZE (2M on x86_64) and PUD_SIZE (1G on x86_64). 187 188The differences with HugeTLB are relatively minor. 189 190It only use 3 page structs for storing all information as opposed 191to 4 on HugeTLB pages. 192 193There's no remapping of vmemmap given that device-dax memory is not part of 194System RAM ranges initialized at boot. Thus the tail page deduplication 195happens at a later stage when we populate the sections. HugeTLB reuses the 196the head vmemmap page representing, whereas device-dax reuses the tail 197vmemmap page. This results in only half of the savings compared to HugeTLB. 198 199Deduplicated tail pages are not mapped read-only. 200 201Here's how things look like on device-dax after the sections are populated:: 202 203 +-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+ 204 | | | 0 | -------------> | 0 | 205 | | +-----------+ +-----------+ 206 | | | 1 | -------------> | 1 | 207 | | +-----------+ +-----------+ 208 | | | 2 | ----------------^ ^ ^ ^ ^ ^ 209 | | +-----------+ | | | | | 210 | | | 3 | ------------------+ | | | | 211 | | +-----------+ | | | | 212 | | | 4 | --------------------+ | | | 213 | PMD | +-----------+ | | | 214 | level | | 5 | ----------------------+ | | 215 | mapping | +-----------+ | | 216 | | | 6 | ------------------------+ | 217 | | +-----------+ | 218 | | | 7 | --------------------------+ 219 | | +-----------+ 220 | | 221 | | 222 | | 223 +-----------+ 224