Based on kernel version 4.9. Page generated on 2016-12-21 14:37 EST.
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 and tmpfs/shmem. 13 But in the future it can expand to other filesystems. 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 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. 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 Transparent Hugepage Support maximizes the usefulness of free memory 61 if compared to the reservation approach of hugetlbfs by allowing all 62 unused memory to be used as cache or other movable (or even unmovable 63 entities). It doesn't require reservation to prevent hugepage 64 allocation failures to be noticeable from userland. It allows paging 65 and all other advanced VM features to be available on the 66 hugepages. It requires no modifications for applications to take 67 advantage of it. 68 69 Applications however can be further optimized to take advantage of 70 this feature, like for example they've been optimized before to avoid 71 a flood of mmap system calls for every malloc(4k). Optimizing userland 72 is by far not mandatory and khugepaged already can take care of long 73 lived page allocations even for hugepage unaware applications that 74 deals with large amounts of memory. 75 76 In certain cases when hugepages are enabled system wide, application 77 may end up allocating more memory resources. An application may mmap a 78 large region but only touch 1 byte of it, in that case a 2M page might 79 be allocated instead of a 4k page for no good. This is why it's 80 possible to disable hugepages system-wide and to only have them inside 81 MADV_HUGEPAGE madvise regions. 82 83 Embedded systems should enable hugepages only inside madvise regions 84 to eliminate any risk of wasting any precious byte of memory and to 85 only run faster. 86 87 Applications that gets a lot of benefit from hugepages and that don't 88 risk to lose memory by using hugepages, should use 89 madvise(MADV_HUGEPAGE) on their critical mmapped regions. 90 91 == sysfs == 92 93 Transparent Hugepage Support for anonymous memory can be entirely disabled 94 (mostly for debugging purposes) or only enabled inside MADV_HUGEPAGE 95 regions (to avoid the risk of consuming more memory resources) or enabled 96 system wide. This can be achieved with one of: 97 98 echo always >/sys/kernel/mm/transparent_hugepage/enabled 99 echo madvise >/sys/kernel/mm/transparent_hugepage/enabled 100 echo never >/sys/kernel/mm/transparent_hugepage/enabled 101 102 It's also possible to limit defrag efforts in the VM to generate 103 anonymous hugepages in case they're not immediately free to madvise 104 regions or to never try to defrag memory and simply fallback to regular 105 pages unless hugepages are immediately available. Clearly if we spend CPU 106 time to defrag memory, we would expect to gain even more by the fact we 107 use hugepages later instead of regular pages. This isn't always 108 guaranteed, but it may be more likely in case the allocation is for a 109 MADV_HUGEPAGE region. 110 111 echo always >/sys/kernel/mm/transparent_hugepage/defrag 112 echo defer >/sys/kernel/mm/transparent_hugepage/defrag 113 echo madvise >/sys/kernel/mm/transparent_hugepage/defrag 114 echo never >/sys/kernel/mm/transparent_hugepage/defrag 115 116 "always" means that an application requesting THP will stall on allocation 117 failure and directly reclaim pages and compact memory in an effort to 118 allocate a THP immediately. This may be desirable for virtual machines 119 that benefit heavily from THP use and are willing to delay the VM start 120 to utilise them. 121 122 "defer" means that an application will wake kswapd in the background 123 to reclaim pages and wake kcompact to compact memory so that THP is 124 available in the near future. It's the responsibility of khugepaged 125 to then install the THP pages later. 126 127 "madvise" will enter direct reclaim like "always" but only for regions 128 that are have used madvise(MADV_HUGEPAGE). This is the default behaviour. 129 130 "never" should be self-explanatory. 131 132 By default kernel tries to use huge zero page on read page fault to 133 anonymous mapping. It's possible to disable huge zero page by writing 0 134 or enable it back by writing 1: 135 136 echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page 137 echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page 138 139 khugepaged will be automatically started when 140 transparent_hugepage/enabled is set to "always" or "madvise, and it'll 141 be automatically shutdown if it's set to "never". 142 143 khugepaged runs usually at low frequency so while one may not want to 144 invoke defrag algorithms synchronously during the page faults, it 145 should be worth invoking defrag at least in khugepaged. However it's 146 also possible to disable defrag in khugepaged by writing 0 or enable 147 defrag in khugepaged by writing 1: 148 149 echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag 150 echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag 151 152 You can also control how many pages khugepaged should scan at each 153 pass: 154 155 /sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan 156 157 and how many milliseconds to wait in khugepaged between each pass (you 158 can set this to 0 to run khugepaged at 100% utilization of one core): 159 160 /sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs 161 162 and how many milliseconds to wait in khugepaged if there's an hugepage 163 allocation failure to throttle the next allocation attempt. 164 165 /sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs 166 167 The khugepaged progress can be seen in the number of pages collapsed: 168 169 /sys/kernel/mm/transparent_hugepage/khugepaged/pages_collapsed 170 171 for each pass: 172 173 /sys/kernel/mm/transparent_hugepage/khugepaged/full_scans 174 175 max_ptes_none specifies how many extra small pages (that are 176 not already mapped) can be allocated when collapsing a group 177 of small pages into one large page. 178 179 /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none 180 181 A higher value leads to use additional memory for programs. 182 A lower value leads to gain less thp performance. Value of 183 max_ptes_none can waste cpu time very little, you can 184 ignore it. 185 186 max_ptes_swap specifies how many pages can be brought in from 187 swap when collapsing a group of pages into a transparent huge page. 188 189 /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_swap 190 191 A higher value can cause excessive swap IO and waste 192 memory. A lower value can prevent THPs from being 193 collapsed, resulting fewer pages being collapsed into 194 THPs, and lower memory access performance. 195 196 == Boot parameter == 197 198 You can change the sysfs boot time defaults of Transparent Hugepage 199 Support by passing the parameter "transparent_hugepage=always" or 200 "transparent_hugepage=madvise" or "transparent_hugepage=never" 201 (without "") to the kernel command line. 202 203 == Hugepages in tmpfs/shmem == 204 205 You can control hugepage allocation policy in tmpfs with mount option 206 "huge=". It can have following values: 207 208 - "always": 209 Attempt to allocate huge pages every time we need a new page; 210 211 - "never": 212 Do not allocate huge pages; 213 214 - "within_size": 215 Only allocate huge page if it will be fully within i_size. 216 Also respect fadvise()/madvise() hints; 217 218 - "advise: 219 Only allocate huge pages if requested with fadvise()/madvise(); 220 221 The default policy is "never". 222 223 "mount -o remount,huge= /mountpoint" works fine after mount: remounting 224 huge=never will not attempt to break up huge pages at all, just stop more 225 from being allocated. 226 227 There's also sysfs knob to control hugepage allocation policy for internal 228 shmem mount: /sys/kernel/mm/transparent_hugepage/shmem_enabled. The mount 229 is used for SysV SHM, memfds, shared anonymous mmaps (of /dev/zero or 230 MAP_ANONYMOUS), GPU drivers' DRM objects, Ashmem. 231 232 In addition to policies listed above, shmem_enabled allows two further 233 values: 234 235 - "deny": 236 For use in emergencies, to force the huge option off from 237 all mounts; 238 - "force": 239 Force the huge option on for all - very useful for testing; 240 241 == Need of application restart == 242 243 The transparent_hugepage/enabled values and tmpfs mount option only affect 244 future behavior. So to make them effective you need to restart any 245 application that could have been using hugepages. This also applies to the 246 regions registered in khugepaged. 247 248 == Monitoring usage == 249 250 The number of anonymous transparent huge pages currently used by the 251 system is available by reading the AnonHugePages field in /proc/meminfo. 252 To identify what applications are using anonymous transparent huge pages, 253 it is necessary to read /proc/PID/smaps and count the AnonHugePages fields 254 for each mapping. 255 256 The number of file transparent huge pages mapped to userspace is available 257 by reading ShmemPmdMapped and ShmemHugePages fields in /proc/meminfo. 258 To identify what applications are mapping file transparent huge pages, it 259 is necessary to read /proc/PID/smaps and count the FileHugeMapped fields 260 for each mapping. 261 262 Note that reading the smaps file is expensive and reading it 263 frequently will incur overhead. 264 265 There are a number of counters in /proc/vmstat that may be used to 266 monitor how successfully the system is providing huge pages for use. 267 268 thp_fault_alloc is incremented every time a huge page is successfully 269 allocated to handle a page fault. This applies to both the 270 first time a page is faulted and for COW faults. 271 272 thp_collapse_alloc is incremented by khugepaged when it has found 273 a range of pages to collapse into one huge page and has 274 successfully allocated a new huge page to store the data. 275 276 thp_fault_fallback is incremented if a page fault fails to allocate 277 a huge page and instead falls back to using small pages. 278 279 thp_collapse_alloc_failed is incremented if khugepaged found a range 280 of pages that should be collapsed into one huge page but failed 281 the allocation. 282 283 thp_file_alloc is incremented every time a file huge page is successfully 284 i allocated. 285 286 thp_file_mapped is incremented every time a file huge page is mapped into 287 user address space. 288 289 thp_split_page is incremented every time a huge page is split into base 290 pages. This can happen for a variety of reasons but a common 291 reason is that a huge page is old and is being reclaimed. 292 This action implies splitting all PMD the page mapped with. 293 294 thp_split_page_failed is is incremented if kernel fails to split huge 295 page. This can happen if the page was pinned by somebody. 296 297 thp_deferred_split_page is incremented when a huge page is put onto split 298 queue. This happens when a huge page is partially unmapped and 299 splitting it would free up some memory. Pages on split queue are 300 going to be split under memory pressure. 301 302 thp_split_pmd is incremented every time a PMD split into table of PTEs. 303 This can happen, for instance, when application calls mprotect() or 304 munmap() on part of huge page. It doesn't split huge page, only 305 page table entry. 306 307 thp_zero_page_alloc is incremented every time a huge zero page is 308 successfully allocated. It includes allocations which where 309 dropped due race with other allocation. Note, it doesn't count 310 every map of the huge zero page, only its allocation. 311 312 thp_zero_page_alloc_failed is incremented if kernel fails to allocate 313 huge zero page and falls back to using small pages. 314 315 As the system ages, allocating huge pages may be expensive as the 316 system uses memory compaction to copy data around memory to free a 317 huge page for use. There are some counters in /proc/vmstat to help 318 monitor this overhead. 319 320 compact_stall is incremented every time a process stalls to run 321 memory compaction so that a huge page is free for use. 322 323 compact_success is incremented if the system compacted memory and 324 freed a huge page for use. 325 326 compact_fail is incremented if the system tries to compact memory 327 but failed. 328 329 compact_pages_moved is incremented each time a page is moved. If 330 this value is increasing rapidly, it implies that the system 331 is copying a lot of data to satisfy the huge page allocation. 332 It is possible that the cost of copying exceeds any savings 333 from reduced TLB misses. 334 335 compact_pagemigrate_failed is incremented when the underlying mechanism 336 for moving a page failed. 337 338 compact_blocks_moved is incremented each time memory compaction examines 339 a huge page aligned range of pages. 340 341 It is possible to establish how long the stalls were using the function 342 tracer to record how long was spent in __alloc_pages_nodemask and 343 using the mm_page_alloc tracepoint to identify which allocations were 344 for huge pages. 345 346 == get_user_pages and follow_page == 347 348 get_user_pages and follow_page if run on a hugepage, will return the 349 head or tail pages as usual (exactly as they would do on 350 hugetlbfs). Most gup users will only care about the actual physical 351 address of the page and its temporary pinning to release after the I/O 352 is complete, so they won't ever notice the fact the page is huge. But 353 if any driver is going to mangle over the page structure of the tail 354 page (like for checking page->mapping or other bits that are relevant 355 for the head page and not the tail page), it should be updated to jump 356 to check head page instead. Taking reference on any head/tail page would 357 prevent page from being split by anyone. 358 359 NOTE: these aren't new constraints to the GUP API, and they match the 360 same constrains that applies to hugetlbfs too, so any driver capable 361 of handling GUP on hugetlbfs will also work fine on transparent 362 hugepage backed mappings. 363 364 In case you can't handle compound pages if they're returned by 365 follow_page, the FOLL_SPLIT bit can be specified as parameter to 366 follow_page, so that it will split the hugepages before returning 367 them. Migration for example passes FOLL_SPLIT as parameter to 368 follow_page because it's not hugepage aware and in fact it can't work 369 at all on hugetlbfs (but it instead works fine on transparent 370 hugepages thanks to FOLL_SPLIT). migration simply can't deal with 371 hugepages being returned (as it's not only checking the pfn of the 372 page and pinning it during the copy but it pretends to migrate the 373 memory in regular page sizes and with regular pte/pmd mappings). 374 375 == Optimizing the applications == 376 377 To be guaranteed that the kernel will map a 2M page immediately in any 378 memory region, the mmap region has to be hugepage naturally 379 aligned. posix_memalign() can provide that guarantee. 380 381 == Hugetlbfs == 382 383 You can use hugetlbfs on a kernel that has transparent hugepage 384 support enabled just fine as always. No difference can be noted in 385 hugetlbfs other than there will be less overall fragmentation. All 386 usual features belonging to hugetlbfs are preserved and 387 unaffected. libhugetlbfs will also work fine as usual. 388 389 == Graceful fallback == 390 391 Code walking pagetables but unaware about huge pmds can simply call 392 split_huge_pmd(vma, pmd, addr) where the pmd is the one returned by 393 pmd_offset. It's trivial to make the code transparent hugepage aware 394 by just grepping for "pmd_offset" and adding split_huge_pmd where 395 missing after pmd_offset returns the pmd. Thanks to the graceful 396 fallback design, with a one liner change, you can avoid to write 397 hundred if not thousand of lines of complex code to make your code 398 hugepage aware. 399 400 If you're not walking pagetables but you run into a physical hugepage 401 but you can't handle it natively in your code, you can split it by 402 calling split_huge_page(page). This is what the Linux VM does before 403 it tries to swapout the hugepage for example. split_huge_page() can fail 404 if the page is pinned and you must handle this correctly. 405 406 Example to make mremap.c transparent hugepage aware with a one liner 407 change: 408 409 diff --git a/mm/mremap.c b/mm/mremap.c 410 --- a/mm/mremap.c 411 +++ b/mm/mremap.c 412 @@ -41,6 +41,7 @@ static pmd_t *get_old_pmd(struct mm_stru 413 return NULL; 414 415 pmd = pmd_offset(pud, addr); 416 + split_huge_pmd(vma, pmd, addr); 417 if (pmd_none_or_clear_bad(pmd)) 418 return NULL; 419 420 == Locking in hugepage aware code == 421 422 We want as much code as possible hugepage aware, as calling 423 split_huge_page() or split_huge_pmd() has a cost. 424 425 To make pagetable walks huge pmd aware, all you need to do is to call 426 pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the 427 mmap_sem in read (or write) mode to be sure an huge pmd cannot be 428 created from under you by khugepaged (khugepaged collapse_huge_page 429 takes the mmap_sem in write mode in addition to the anon_vma lock). If 430 pmd_trans_huge returns false, you just fallback in the old code 431 paths. If instead pmd_trans_huge returns true, you have to take the 432 page table lock (pmd_lock()) and re-run pmd_trans_huge. Taking the 433 page table lock will prevent the huge pmd to be converted into a 434 regular pmd from under you (split_huge_pmd can run in parallel to the 435 pagetable walk). If the second pmd_trans_huge returns false, you 436 should just drop the page table lock and fallback to the old code as 437 before. Otherwise you can proceed to process the huge pmd and the 438 hugepage natively. Once finished you can drop the page table lock. 439 440 == Refcounts and transparent huge pages == 441 442 Refcounting on THP is mostly consistent with refcounting on other compound 443 pages: 444 445 - get_page()/put_page() and GUP operate in head page's ->_refcount. 446 447 - ->_refcount in tail pages is always zero: get_page_unless_zero() never 448 succeed on tail pages. 449 450 - map/unmap of the pages with PTE entry increment/decrement ->_mapcount 451 on relevant sub-page of the compound page. 452 453 - map/unmap of the whole compound page accounted in compound_mapcount 454 (stored in first tail page). For file huge pages, we also increment 455 ->_mapcount of all sub-pages in order to have race-free detection of 456 last unmap of subpages. 457 458 PageDoubleMap() indicates that the page is *possibly* mapped with PTEs. 459 460 For anonymous pages PageDoubleMap() also indicates ->_mapcount in all 461 subpages is offset up by one. This additional reference is required to 462 get race-free detection of unmap of subpages when we have them mapped with 463 both PMDs and PTEs. 464 465 This is optimization required to lower overhead of per-subpage mapcount 466 tracking. The alternative is alter ->_mapcount in all subpages on each 467 map/unmap of the whole compound page. 468 469 For anonymous pages, we set PG_double_map when a PMD of the page got split 470 for the first time, but still have PMD mapping. The additional references 471 go away with last compound_mapcount. 472 473 File pages get PG_double_map set on first map of the page with PTE and 474 goes away when the page gets evicted from page cache. 475 476 split_huge_page internally has to distribute the refcounts in the head 477 page to the tail pages before clearing all PG_head/tail bits from the page 478 structures. It can be done easily for refcounts taken by page table 479 entries. But we don't have enough information on how to distribute any 480 additional pins (i.e. from get_user_pages). split_huge_page() fails any 481 requests to split pinned huge page: it expects page count to be equal to 482 sum of mapcount of all sub-pages plus one (split_huge_page caller must 483 have reference for head page). 484 485 split_huge_page uses migration entries to stabilize page->_refcount and 486 page->_mapcount of anonymous pages. File pages just got unmapped. 487 488 We safe against physical memory scanners too: the only legitimate way 489 scanner can get reference to a page is get_page_unless_zero(). 490 491 All tail pages have zero ->_refcount until atomic_add(). This prevents the 492 scanner from getting a reference to the tail page up to that point. After the 493 atomic_add() we don't care about the ->_refcount value. We already known how 494 many references should be uncharged from the head page. 495 496 For head page get_page_unless_zero() will succeed and we don't mind. It's 497 clear where reference should go after split: it will stay on head page. 498 499 Note that split_huge_pmd() doesn't have any limitation on refcounting: 500 pmd can be split at any point and never fails. 501 502 == Partial unmap and deferred_split_huge_page() == 503 504 Unmapping part of THP (with munmap() or other way) is not going to free 505 memory immediately. Instead, we detect that a subpage of THP is not in use 506 in page_remove_rmap() and queue the THP for splitting if memory pressure 507 comes. Splitting will free up unused subpages. 508 509 Splitting the page right away is not an option due to locking context in 510 the place where we can detect partial unmap. It's also might be 511 counterproductive since in many cases partial unmap unmap happens during 512 exit(2) if an THP crosses VMA boundary. 513 514 Function deferred_split_huge_page() is used to queue page for splitting. 515 The splitting itself will happen when we get memory pressure via shrinker 516 interface.