Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 1 | = Transparent Hugepage Support = |
| 2 | |
| 3 | == Objective == |
| 4 | |
| 5 | Performance critical computing applications dealing with large memory |
| 6 | working sets are already running on top of libhugetlbfs and in turn |
| 7 | hugetlbfs. Transparent Hugepage Support is an alternative means of |
| 8 | using huge pages for the backing of virtual memory with huge pages |
| 9 | that supports the automatic promotion and demotion of page sizes and |
| 10 | without the shortcomings of hugetlbfs. |
| 11 | |
| 12 | Currently it only works for anonymous memory mappings but in the |
| 13 | future it can expand over the pagecache layer starting with tmpfs. |
| 14 | |
| 15 | The reason applications are running faster is because of two |
| 16 | factors. The first factor is almost completely irrelevant and it's not |
| 17 | of significant interest because it'll also have the downside of |
| 18 | requiring larger clear-page copy-page in page faults which is a |
| 19 | potentially negative effect. The first factor consists in taking a |
| 20 | single page fault for each 2M virtual region touched by userland (so |
| 21 | reducing the enter/exit kernel frequency by a 512 times factor). This |
| 22 | only matters the first time the memory is accessed for the lifetime of |
| 23 | a memory mapping. The second long lasting and much more important |
| 24 | factor will affect all subsequent accesses to the memory for the whole |
| 25 | runtime of the application. The second factor consist of two |
| 26 | components: 1) the TLB miss will run faster (especially with |
| 27 | virtualization using nested pagetables but almost always also on bare |
| 28 | metal without virtualization) and 2) a single TLB entry will be |
| 29 | mapping a much larger amount of virtual memory in turn reducing the |
| 30 | number of TLB misses. With virtualization and nested pagetables the |
| 31 | TLB can be mapped of larger size only if both KVM and the Linux guest |
| 32 | are using hugepages but a significant speedup already happens if only |
| 33 | one of the two is using hugepages just because of the fact the TLB |
| 34 | miss is going to run faster. |
| 35 | |
| 36 | == Design == |
| 37 | |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 38 | - "graceful fallback": mm components which don't have transparent hugepage |
| 39 | knowledge fall back to breaking huge pmd mapping into table of ptes and, |
| 40 | if necessary, split a transparent hugepage. Therefore these components |
| 41 | can continue working on the regular pages or regular pte mappings. |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 42 | |
| 43 | - if a hugepage allocation fails because of memory fragmentation, |
| 44 | regular pages should be gracefully allocated instead and mixed in |
| 45 | the same vma without any failure or significant delay and without |
| 46 | userland noticing |
| 47 | |
| 48 | - if some task quits and more hugepages become available (either |
| 49 | immediately in the buddy or through the VM), guest physical memory |
| 50 | backed by regular pages should be relocated on hugepages |
| 51 | automatically (with khugepaged) |
| 52 | |
| 53 | - it doesn't require memory reservation and in turn it uses hugepages |
| 54 | whenever possible (the only possible reservation here is kernelcore= |
| 55 | to avoid unmovable pages to fragment all the memory but such a tweak |
| 56 | is not specific to transparent hugepage support and it's a generic |
| 57 | feature that applies to all dynamic high order allocations in the |
| 58 | kernel) |
| 59 | |
| 60 | - this initial support only offers the feature in the anonymous memory |
| 61 | regions but it'd be ideal to move it to tmpfs and the pagecache |
| 62 | later |
| 63 | |
| 64 | Transparent Hugepage Support maximizes the usefulness of free memory |
| 65 | if compared to the reservation approach of hugetlbfs by allowing all |
| 66 | unused memory to be used as cache or other movable (or even unmovable |
| 67 | entities). It doesn't require reservation to prevent hugepage |
| 68 | allocation failures to be noticeable from userland. It allows paging |
| 69 | and all other advanced VM features to be available on the |
| 70 | hugepages. It requires no modifications for applications to take |
| 71 | advantage of it. |
| 72 | |
| 73 | Applications however can be further optimized to take advantage of |
| 74 | this feature, like for example they've been optimized before to avoid |
| 75 | a flood of mmap system calls for every malloc(4k). Optimizing userland |
| 76 | is by far not mandatory and khugepaged already can take care of long |
| 77 | lived page allocations even for hugepage unaware applications that |
| 78 | deals with large amounts of memory. |
| 79 | |
| 80 | In certain cases when hugepages are enabled system wide, application |
| 81 | may end up allocating more memory resources. An application may mmap a |
| 82 | large region but only touch 1 byte of it, in that case a 2M page might |
| 83 | be allocated instead of a 4k page for no good. This is why it's |
| 84 | possible to disable hugepages system-wide and to only have them inside |
| 85 | MADV_HUGEPAGE madvise regions. |
| 86 | |
| 87 | Embedded systems should enable hugepages only inside madvise regions |
| 88 | to eliminate any risk of wasting any precious byte of memory and to |
| 89 | only run faster. |
| 90 | |
| 91 | Applications that gets a lot of benefit from hugepages and that don't |
| 92 | risk to lose memory by using hugepages, should use |
| 93 | madvise(MADV_HUGEPAGE) on their critical mmapped regions. |
| 94 | |
| 95 | == sysfs == |
| 96 | |
| 97 | Transparent Hugepage Support can be entirely disabled (mostly for |
| 98 | debugging purposes) or only enabled inside MADV_HUGEPAGE regions (to |
| 99 | avoid the risk of consuming more memory resources) or enabled system |
| 100 | wide. This can be achieved with one of: |
| 101 | |
| 102 | echo always >/sys/kernel/mm/transparent_hugepage/enabled |
| 103 | echo madvise >/sys/kernel/mm/transparent_hugepage/enabled |
| 104 | echo never >/sys/kernel/mm/transparent_hugepage/enabled |
| 105 | |
| 106 | It's also possible to limit defrag efforts in the VM to generate |
| 107 | hugepages in case they're not immediately free to madvise regions or |
| 108 | to never try to defrag memory and simply fallback to regular pages |
| 109 | unless hugepages are immediately available. Clearly if we spend CPU |
| 110 | time to defrag memory, we would expect to gain even more by the fact |
| 111 | we use hugepages later instead of regular pages. This isn't always |
| 112 | guaranteed, but it may be more likely in case the allocation is for a |
| 113 | MADV_HUGEPAGE region. |
| 114 | |
| 115 | echo always >/sys/kernel/mm/transparent_hugepage/defrag |
| 116 | echo madvise >/sys/kernel/mm/transparent_hugepage/defrag |
| 117 | echo never >/sys/kernel/mm/transparent_hugepage/defrag |
| 118 | |
Kirill A. Shutemov | 79da540 | 2012-12-12 13:51:12 -0800 | [diff] [blame] | 119 | By default kernel tries to use huge zero page on read page fault. |
| 120 | It's possible to disable huge zero page by writing 0 or enable it |
| 121 | back by writing 1: |
| 122 | |
Wanpeng Li | f49cbdd | 2013-07-08 16:00:16 -0700 | [diff] [blame] | 123 | echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page |
| 124 | echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page |
Kirill A. Shutemov | 79da540 | 2012-12-12 13:51:12 -0800 | [diff] [blame] | 125 | |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 126 | khugepaged will be automatically started when |
| 127 | transparent_hugepage/enabled is set to "always" or "madvise, and it'll |
| 128 | be automatically shutdown if it's set to "never". |
| 129 | |
| 130 | khugepaged runs usually at low frequency so while one may not want to |
| 131 | invoke defrag algorithms synchronously during the page faults, it |
| 132 | should be worth invoking defrag at least in khugepaged. However it's |
David Rientjes | e369fde | 2011-09-22 14:11:38 -0700 | [diff] [blame] | 133 | also possible to disable defrag in khugepaged by writing 0 or enable |
| 134 | defrag in khugepaged by writing 1: |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 135 | |
David Rientjes | e369fde | 2011-09-22 14:11:38 -0700 | [diff] [blame] | 136 | echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag |
| 137 | echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 138 | |
| 139 | You can also control how many pages khugepaged should scan at each |
| 140 | pass: |
| 141 | |
| 142 | /sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan |
| 143 | |
| 144 | and how many milliseconds to wait in khugepaged between each pass (you |
| 145 | can set this to 0 to run khugepaged at 100% utilization of one core): |
| 146 | |
| 147 | /sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs |
| 148 | |
| 149 | and how many milliseconds to wait in khugepaged if there's an hugepage |
| 150 | allocation failure to throttle the next allocation attempt. |
| 151 | |
| 152 | /sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs |
| 153 | |
| 154 | The khugepaged progress can be seen in the number of pages collapsed: |
| 155 | |
| 156 | /sys/kernel/mm/transparent_hugepage/khugepaged/pages_collapsed |
| 157 | |
| 158 | for each pass: |
| 159 | |
| 160 | /sys/kernel/mm/transparent_hugepage/khugepaged/full_scans |
| 161 | |
Ebru Akagunduz | 9ddfa69 | 2015-02-26 23:34:36 +0200 | [diff] [blame] | 162 | max_ptes_none specifies how many extra small pages (that are |
| 163 | not already mapped) can be allocated when collapsing a group |
| 164 | of small pages into one large page. |
| 165 | |
| 166 | /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none |
| 167 | |
| 168 | A higher value leads to use additional memory for programs. |
| 169 | A lower value leads to gain less thp performance. Value of |
| 170 | max_ptes_none can waste cpu time very little, you can |
| 171 | ignore it. |
| 172 | |
Ebru Akagunduz | 80f73b4 | 2015-11-05 18:47:32 -0800 | [diff] [blame] | 173 | max_ptes_swap specifies how many pages can be brought in from |
| 174 | swap when collapsing a group of pages into a transparent huge page. |
| 175 | |
| 176 | /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_swap |
| 177 | |
| 178 | A higher value can cause excessive swap IO and waste |
| 179 | memory. A lower value can prevent THPs from being |
| 180 | collapsed, resulting fewer pages being collapsed into |
| 181 | THPs, and lower memory access performance. |
| 182 | |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 183 | == Boot parameter == |
| 184 | |
| 185 | You can change the sysfs boot time defaults of Transparent Hugepage |
| 186 | Support by passing the parameter "transparent_hugepage=always" or |
| 187 | "transparent_hugepage=madvise" or "transparent_hugepage=never" |
| 188 | (without "") to the kernel command line. |
| 189 | |
| 190 | == Need of application restart == |
| 191 | |
| 192 | The transparent_hugepage/enabled values only affect future |
| 193 | behavior. So to make them effective you need to restart any |
| 194 | application that could have been using hugepages. This also applies to |
| 195 | the regions registered in khugepaged. |
| 196 | |
Mel Gorman | 6925699 | 2012-05-29 15:06:45 -0700 | [diff] [blame] | 197 | == Monitoring usage == |
| 198 | |
| 199 | The number of transparent huge pages currently used by the system is |
| 200 | available by reading the AnonHugePages field in /proc/meminfo. To |
| 201 | identify what applications are using transparent huge pages, it is |
| 202 | necessary to read /proc/PID/smaps and count the AnonHugePages fields |
| 203 | for each mapping. Note that reading the smaps file is expensive and |
| 204 | reading it frequently will incur overhead. |
| 205 | |
| 206 | There are a number of counters in /proc/vmstat that may be used to |
| 207 | monitor how successfully the system is providing huge pages for use. |
| 208 | |
| 209 | thp_fault_alloc is incremented every time a huge page is successfully |
| 210 | allocated to handle a page fault. This applies to both the |
| 211 | first time a page is faulted and for COW faults. |
| 212 | |
| 213 | thp_collapse_alloc is incremented by khugepaged when it has found |
| 214 | a range of pages to collapse into one huge page and has |
| 215 | successfully allocated a new huge page to store the data. |
| 216 | |
| 217 | thp_fault_fallback is incremented if a page fault fails to allocate |
| 218 | a huge page and instead falls back to using small pages. |
| 219 | |
| 220 | thp_collapse_alloc_failed is incremented if khugepaged found a range |
| 221 | of pages that should be collapsed into one huge page but failed |
| 222 | the allocation. |
| 223 | |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 224 | thp_split_page is incremented every time a huge page is split into base |
Mel Gorman | 6925699 | 2012-05-29 15:06:45 -0700 | [diff] [blame] | 225 | pages. This can happen for a variety of reasons but a common |
| 226 | reason is that a huge page is old and is being reclaimed. |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 227 | This action implies splitting all PMD the page mapped with. |
| 228 | |
| 229 | thp_split_page_failed is is incremented if kernel fails to split huge |
| 230 | page. This can happen if the page was pinned by somebody. |
| 231 | |
| 232 | thp_split_pmd is incremented every time a PMD split into table of PTEs. |
| 233 | This can happen, for instance, when application calls mprotect() or |
| 234 | munmap() on part of huge page. It doesn't split huge page, only |
| 235 | page table entry. |
Mel Gorman | 6925699 | 2012-05-29 15:06:45 -0700 | [diff] [blame] | 236 | |
Kirill A. Shutemov | d8a8e1f | 2012-12-12 13:51:09 -0800 | [diff] [blame] | 237 | thp_zero_page_alloc is incremented every time a huge zero page is |
| 238 | successfully allocated. It includes allocations which where |
| 239 | dropped due race with other allocation. Note, it doesn't count |
| 240 | every map of the huge zero page, only its allocation. |
| 241 | |
| 242 | thp_zero_page_alloc_failed is incremented if kernel fails to allocate |
| 243 | huge zero page and falls back to using small pages. |
| 244 | |
Mel Gorman | 6925699 | 2012-05-29 15:06:45 -0700 | [diff] [blame] | 245 | As the system ages, allocating huge pages may be expensive as the |
| 246 | system uses memory compaction to copy data around memory to free a |
| 247 | huge page for use. There are some counters in /proc/vmstat to help |
| 248 | monitor this overhead. |
| 249 | |
| 250 | compact_stall is incremented every time a process stalls to run |
| 251 | memory compaction so that a huge page is free for use. |
| 252 | |
| 253 | compact_success is incremented if the system compacted memory and |
| 254 | freed a huge page for use. |
| 255 | |
| 256 | compact_fail is incremented if the system tries to compact memory |
| 257 | but failed. |
| 258 | |
| 259 | compact_pages_moved is incremented each time a page is moved. If |
| 260 | this value is increasing rapidly, it implies that the system |
| 261 | is copying a lot of data to satisfy the huge page allocation. |
| 262 | It is possible that the cost of copying exceeds any savings |
| 263 | from reduced TLB misses. |
| 264 | |
| 265 | compact_pagemigrate_failed is incremented when the underlying mechanism |
| 266 | for moving a page failed. |
| 267 | |
| 268 | compact_blocks_moved is incremented each time memory compaction examines |
| 269 | a huge page aligned range of pages. |
| 270 | |
| 271 | It is possible to establish how long the stalls were using the function |
| 272 | tracer to record how long was spent in __alloc_pages_nodemask and |
| 273 | using the mm_page_alloc tracepoint to identify which allocations were |
| 274 | for huge pages. |
| 275 | |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 276 | == get_user_pages and follow_page == |
| 277 | |
| 278 | get_user_pages and follow_page if run on a hugepage, will return the |
| 279 | head or tail pages as usual (exactly as they would do on |
| 280 | hugetlbfs). Most gup users will only care about the actual physical |
| 281 | address of the page and its temporary pinning to release after the I/O |
| 282 | is complete, so they won't ever notice the fact the page is huge. But |
| 283 | if any driver is going to mangle over the page structure of the tail |
| 284 | page (like for checking page->mapping or other bits that are relevant |
| 285 | for the head page and not the tail page), it should be updated to jump |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 286 | to check head page instead. Taking reference on any head/tail page would |
| 287 | prevent page from being split by anyone. |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 288 | |
| 289 | NOTE: these aren't new constraints to the GUP API, and they match the |
| 290 | same constrains that applies to hugetlbfs too, so any driver capable |
| 291 | of handling GUP on hugetlbfs will also work fine on transparent |
| 292 | hugepage backed mappings. |
| 293 | |
| 294 | In case you can't handle compound pages if they're returned by |
| 295 | follow_page, the FOLL_SPLIT bit can be specified as parameter to |
| 296 | follow_page, so that it will split the hugepages before returning |
| 297 | them. Migration for example passes FOLL_SPLIT as parameter to |
| 298 | follow_page because it's not hugepage aware and in fact it can't work |
| 299 | at all on hugetlbfs (but it instead works fine on transparent |
| 300 | hugepages thanks to FOLL_SPLIT). migration simply can't deal with |
| 301 | hugepages being returned (as it's not only checking the pfn of the |
| 302 | page and pinning it during the copy but it pretends to migrate the |
| 303 | memory in regular page sizes and with regular pte/pmd mappings). |
| 304 | |
| 305 | == Optimizing the applications == |
| 306 | |
| 307 | To be guaranteed that the kernel will map a 2M page immediately in any |
| 308 | memory region, the mmap region has to be hugepage naturally |
| 309 | aligned. posix_memalign() can provide that guarantee. |
| 310 | |
| 311 | == Hugetlbfs == |
| 312 | |
| 313 | You can use hugetlbfs on a kernel that has transparent hugepage |
| 314 | support enabled just fine as always. No difference can be noted in |
| 315 | hugetlbfs other than there will be less overall fragmentation. All |
| 316 | usual features belonging to hugetlbfs are preserved and |
| 317 | unaffected. libhugetlbfs will also work fine as usual. |
| 318 | |
| 319 | == Graceful fallback == |
| 320 | |
| 321 | Code walking pagetables but unware about huge pmds can simply call |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 322 | split_huge_pmd(vma, pmd, addr) where the pmd is the one returned by |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 323 | pmd_offset. It's trivial to make the code transparent hugepage aware |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 324 | by just grepping for "pmd_offset" and adding split_huge_pmd where |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 325 | missing after pmd_offset returns the pmd. Thanks to the graceful |
| 326 | fallback design, with a one liner change, you can avoid to write |
| 327 | hundred if not thousand of lines of complex code to make your code |
| 328 | hugepage aware. |
| 329 | |
| 330 | If you're not walking pagetables but you run into a physical hugepage |
| 331 | but you can't handle it natively in your code, you can split it by |
| 332 | calling split_huge_page(page). This is what the Linux VM does before |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 333 | it tries to swapout the hugepage for example. split_huge_page() can fail |
| 334 | if the page is pinned and you must handle this correctly. |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 335 | |
| 336 | Example to make mremap.c transparent hugepage aware with a one liner |
| 337 | change: |
| 338 | |
| 339 | diff --git a/mm/mremap.c b/mm/mremap.c |
| 340 | --- a/mm/mremap.c |
| 341 | +++ b/mm/mremap.c |
| 342 | @@ -41,6 +41,7 @@ static pmd_t *get_old_pmd(struct mm_stru |
| 343 | return NULL; |
| 344 | |
| 345 | pmd = pmd_offset(pud, addr); |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 346 | + split_huge_pmd(vma, pmd, addr); |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 347 | if (pmd_none_or_clear_bad(pmd)) |
| 348 | return NULL; |
| 349 | |
| 350 | == Locking in hugepage aware code == |
| 351 | |
| 352 | We want as much code as possible hugepage aware, as calling |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 353 | split_huge_page() or split_huge_pmd() has a cost. |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 354 | |
| 355 | To make pagetable walks huge pmd aware, all you need to do is to call |
| 356 | pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the |
| 357 | mmap_sem in read (or write) mode to be sure an huge pmd cannot be |
| 358 | created from under you by khugepaged (khugepaged collapse_huge_page |
| 359 | takes the mmap_sem in write mode in addition to the anon_vma lock). If |
| 360 | pmd_trans_huge returns false, you just fallback in the old code |
| 361 | paths. If instead pmd_trans_huge returns true, you have to take the |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 362 | page table lock (pmd_lock()) and re-run pmd_trans_huge. Taking the |
| 363 | page table lock will prevent the huge pmd to be converted into a |
| 364 | regular pmd from under you (split_huge_pmd can run in parallel to the |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 365 | pagetable walk). If the second pmd_trans_huge returns false, you |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 366 | should just drop the page table lock and fallback to the old code as |
| 367 | before. Otherwise you can proceed to process the huge pmd and the |
| 368 | hugepage natively. Once finished you can drop the page table lock. |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 369 | |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 370 | == Refcounts and transparent huge pages == |
| 371 | |
| 372 | Refcounting on THP is mostly consistent with refcounting on other compound |
| 373 | pages: |
| 374 | |
| 375 | - get_page()/put_page() and GUP operate in head page's ->_count. |
| 376 | |
| 377 | - ->_count in tail pages is always zero: get_page_unless_zero() never |
| 378 | succeed on tail pages. |
| 379 | |
| 380 | - map/unmap of the pages with PTE entry increment/decrement ->_mapcount |
| 381 | on relevant sub-page of the compound page. |
| 382 | |
| 383 | - map/unmap of the whole compound page accounted in compound_mapcount |
| 384 | (stored in first tail page). |
| 385 | |
| 386 | PageDoubleMap() indicates that ->_mapcount in all subpages is offset up by one. |
| 387 | This additional reference is required to get race-free detection of unmap of |
| 388 | subpages when we have them mapped with both PMDs and PTEs. |
| 389 | |
| 390 | This is optimization required to lower overhead of per-subpage mapcount |
| 391 | tracking. The alternative is alter ->_mapcount in all subpages on each |
| 392 | map/unmap of the whole compound page. |
| 393 | |
| 394 | We set PG_double_map when a PMD of the page got split for the first time, |
| 395 | but still have PMD mapping. The addtional references go away with last |
| 396 | compound_mapcount. |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 397 | |
| 398 | split_huge_page internally has to distribute the refcounts in the head |
Kirill A. Shutemov | a46e637 | 2016-01-15 16:54:30 -0800 | [diff] [blame] | 399 | page to the tail pages before clearing all PG_head/tail bits from the page |
| 400 | structures. It can be done easily for refcounts taken by page table |
| 401 | entries. But we don't have enough information on how to distribute any |
| 402 | additional pins (i.e. from get_user_pages). split_huge_page() fails any |
| 403 | requests to split pinned huge page: it expects page count to be equal to |
| 404 | sum of mapcount of all sub-pages plus one (split_huge_page caller must |
| 405 | have reference for head page). |
| 406 | |
| 407 | split_huge_page uses migration entries to stabilize page->_count and |
| 408 | page->_mapcount. |
| 409 | |
| 410 | We safe against physical memory scanners too: the only legitimate way |
| 411 | scanner can get reference to a page is get_page_unless_zero(). |
| 412 | |
| 413 | All tail pages has zero ->_count until atomic_add(). It prevent scanner |
| 414 | from geting reference to tail page up to the point. After the atomic_add() |
| 415 | we don't care about ->_count value. We already known how many references |
| 416 | with should uncharge from head page. |
| 417 | |
| 418 | For head page get_page_unless_zero() will succeed and we don't mind. It's |
| 419 | clear where reference should go after split: it will stay on head page. |
| 420 | |
| 421 | Note that split_huge_pmd() doesn't have any limitation on refcounting: |
| 422 | pmd can be split at any point and never fails. |
| 423 | |
| 424 | == Partial unmap and deferred_split_huge_page() == |
| 425 | |
| 426 | Unmapping part of THP (with munmap() or other way) is not going to free |
| 427 | memory immediately. Instead, we detect that a subpage of THP is not in use |
| 428 | in page_remove_rmap() and queue the THP for splitting if memory pressure |
| 429 | comes. Splitting will free up unused subpages. |
| 430 | |
| 431 | Splitting the page right away is not an option due to locking context in |
| 432 | the place where we can detect partial unmap. It's also might be |
| 433 | counterproductive since in many cases partial unmap unmap happens during |
| 434 | exit(2) if an THP crosses VMA boundary. |
| 435 | |
| 436 | Function deferred_split_huge_page() is used to queue page for splitting. |
| 437 | The splitting itself will happen when we get memory pressure via shrinker |
| 438 | interface. |