Based on kernel version 4.16.1. Page generated on 2018-04-09 11:53 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 defer+madvise >/sys/kernel/mm/transparent_hugepage/defrag 114 echo madvise >/sys/kernel/mm/transparent_hugepage/defrag 115 echo never >/sys/kernel/mm/transparent_hugepage/defrag 116 117 "always" means that an application requesting THP will stall on allocation 118 failure and directly reclaim pages and compact memory in an effort to 119 allocate a THP immediately. This may be desirable for virtual machines 120 that benefit heavily from THP use and are willing to delay the VM start 121 to utilise them. 122 123 "defer" means that an application will wake kswapd in the background 124 to reclaim pages and wake kcompactd to compact memory so that THP is 125 available in the near future. It's the responsibility of khugepaged 126 to then install the THP pages later. 127 128 "defer+madvise" will enter direct reclaim and compaction like "always", but 129 only for regions that have used madvise(MADV_HUGEPAGE); all other regions 130 will wake kswapd in the background to reclaim pages and wake kcompactd to 131 compact memory so that THP is available in the near future. 132 133 "madvise" will enter direct reclaim like "always" but only for regions 134 that are have used madvise(MADV_HUGEPAGE). This is the default behaviour. 135 136 "never" should be self-explanatory. 137 138 By default kernel tries to use huge zero page on read page fault to 139 anonymous mapping. It's possible to disable huge zero page by writing 0 140 or enable it back by writing 1: 141 142 echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page 143 echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page 144 145 Some userspace (such as a test program, or an optimized memory allocation 146 library) may want to know the size (in bytes) of a transparent hugepage: 147 148 cat /sys/kernel/mm/transparent_hugepage/hpage_pmd_size 149 150 khugepaged will be automatically started when 151 transparent_hugepage/enabled is set to "always" or "madvise, and it'll 152 be automatically shutdown if it's set to "never". 153 154 khugepaged runs usually at low frequency so while one may not want to 155 invoke defrag algorithms synchronously during the page faults, it 156 should be worth invoking defrag at least in khugepaged. However it's 157 also possible to disable defrag in khugepaged by writing 0 or enable 158 defrag in khugepaged by writing 1: 159 160 echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag 161 echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag 162 163 You can also control how many pages khugepaged should scan at each 164 pass: 165 166 /sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan 167 168 and how many milliseconds to wait in khugepaged between each pass (you 169 can set this to 0 to run khugepaged at 100% utilization of one core): 170 171 /sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs 172 173 and how many milliseconds to wait in khugepaged if there's an hugepage 174 allocation failure to throttle the next allocation attempt. 175 176 /sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs 177 178 The khugepaged progress can be seen in the number of pages collapsed: 179 180 /sys/kernel/mm/transparent_hugepage/khugepaged/pages_collapsed 181 182 for each pass: 183 184 /sys/kernel/mm/transparent_hugepage/khugepaged/full_scans 185 186 max_ptes_none specifies how many extra small pages (that are 187 not already mapped) can be allocated when collapsing a group 188 of small pages into one large page. 189 190 /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none 191 192 A higher value leads to use additional memory for programs. 193 A lower value leads to gain less thp performance. Value of 194 max_ptes_none can waste cpu time very little, you can 195 ignore it. 196 197 max_ptes_swap specifies how many pages can be brought in from 198 swap when collapsing a group of pages into a transparent huge page. 199 200 /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_swap 201 202 A higher value can cause excessive swap IO and waste 203 memory. A lower value can prevent THPs from being 204 collapsed, resulting fewer pages being collapsed into 205 THPs, and lower memory access performance. 206 207 == Boot parameter == 208 209 You can change the sysfs boot time defaults of Transparent Hugepage 210 Support by passing the parameter "transparent_hugepage=always" or 211 "transparent_hugepage=madvise" or "transparent_hugepage=never" 212 (without "") to the kernel command line. 213 214 == Hugepages in tmpfs/shmem == 215 216 You can control hugepage allocation policy in tmpfs with mount option 217 "huge=". It can have following values: 218 219 - "always": 220 Attempt to allocate huge pages every time we need a new page; 221 222 - "never": 223 Do not allocate huge pages; 224 225 - "within_size": 226 Only allocate huge page if it will be fully within i_size. 227 Also respect fadvise()/madvise() hints; 228 229 - "advise: 230 Only allocate huge pages if requested with fadvise()/madvise(); 231 232 The default policy is "never". 233 234 "mount -o remount,huge= /mountpoint" works fine after mount: remounting 235 huge=never will not attempt to break up huge pages at all, just stop more 236 from being allocated. 237 238 There's also sysfs knob to control hugepage allocation policy for internal 239 shmem mount: /sys/kernel/mm/transparent_hugepage/shmem_enabled. The mount 240 is used for SysV SHM, memfds, shared anonymous mmaps (of /dev/zero or 241 MAP_ANONYMOUS), GPU drivers' DRM objects, Ashmem. 242 243 In addition to policies listed above, shmem_enabled allows two further 244 values: 245 246 - "deny": 247 For use in emergencies, to force the huge option off from 248 all mounts; 249 - "force": 250 Force the huge option on for all - very useful for testing; 251 252 == Need of application restart == 253 254 The transparent_hugepage/enabled values and tmpfs mount option only affect 255 future behavior. So to make them effective you need to restart any 256 application that could have been using hugepages. This also applies to the 257 regions registered in khugepaged. 258 259 == Monitoring usage == 260 261 The number of anonymous transparent huge pages currently used by the 262 system is available by reading the AnonHugePages field in /proc/meminfo. 263 To identify what applications are using anonymous transparent huge pages, 264 it is necessary to read /proc/PID/smaps and count the AnonHugePages fields 265 for each mapping. 266 267 The number of file transparent huge pages mapped to userspace is available 268 by reading ShmemPmdMapped and ShmemHugePages fields in /proc/meminfo. 269 To identify what applications are mapping file transparent huge pages, it 270 is necessary to read /proc/PID/smaps and count the FileHugeMapped fields 271 for each mapping. 272 273 Note that reading the smaps file is expensive and reading it 274 frequently will incur overhead. 275 276 There are a number of counters in /proc/vmstat that may be used to 277 monitor how successfully the system is providing huge pages for use. 278 279 thp_fault_alloc is incremented every time a huge page is successfully 280 allocated to handle a page fault. This applies to both the 281 first time a page is faulted and for COW faults. 282 283 thp_collapse_alloc is incremented by khugepaged when it has found 284 a range of pages to collapse into one huge page and has 285 successfully allocated a new huge page to store the data. 286 287 thp_fault_fallback is incremented if a page fault fails to allocate 288 a huge page and instead falls back to using small pages. 289 290 thp_collapse_alloc_failed is incremented if khugepaged found a range 291 of pages that should be collapsed into one huge page but failed 292 the allocation. 293 294 thp_file_alloc is incremented every time a file huge page is successfully 295 allocated. 296 297 thp_file_mapped is incremented every time a file huge page is mapped into 298 user address space. 299 300 thp_split_page is incremented every time a huge page is split into base 301 pages. This can happen for a variety of reasons but a common 302 reason is that a huge page is old and is being reclaimed. 303 This action implies splitting all PMD the page mapped with. 304 305 thp_split_page_failed is incremented if kernel fails to split huge 306 page. This can happen if the page was pinned by somebody. 307 308 thp_deferred_split_page is incremented when a huge page is put onto split 309 queue. This happens when a huge page is partially unmapped and 310 splitting it would free up some memory. Pages on split queue are 311 going to be split under memory pressure. 312 313 thp_split_pmd is incremented every time a PMD split into table of PTEs. 314 This can happen, for instance, when application calls mprotect() or 315 munmap() on part of huge page. It doesn't split huge page, only 316 page table entry. 317 318 thp_zero_page_alloc is incremented every time a huge zero page is 319 successfully allocated. It includes allocations which where 320 dropped due race with other allocation. Note, it doesn't count 321 every map of the huge zero page, only its allocation. 322 323 thp_zero_page_alloc_failed is incremented if kernel fails to allocate 324 huge zero page and falls back to using small pages. 325 326 As the system ages, allocating huge pages may be expensive as the 327 system uses memory compaction to copy data around memory to free a 328 huge page for use. There are some counters in /proc/vmstat to help 329 monitor this overhead. 330 331 compact_stall is incremented every time a process stalls to run 332 memory compaction so that a huge page is free for use. 333 334 compact_success is incremented if the system compacted memory and 335 freed a huge page for use. 336 337 compact_fail is incremented if the system tries to compact memory 338 but failed. 339 340 compact_pages_moved is incremented each time a page is moved. If 341 this value is increasing rapidly, it implies that the system 342 is copying a lot of data to satisfy the huge page allocation. 343 It is possible that the cost of copying exceeds any savings 344 from reduced TLB misses. 345 346 compact_pagemigrate_failed is incremented when the underlying mechanism 347 for moving a page failed. 348 349 compact_blocks_moved is incremented each time memory compaction examines 350 a huge page aligned range of pages. 351 352 It is possible to establish how long the stalls were using the function 353 tracer to record how long was spent in __alloc_pages_nodemask and 354 using the mm_page_alloc tracepoint to identify which allocations were 355 for huge pages. 356 357 == get_user_pages and follow_page == 358 359 get_user_pages and follow_page if run on a hugepage, will return the 360 head or tail pages as usual (exactly as they would do on 361 hugetlbfs). Most gup users will only care about the actual physical 362 address of the page and its temporary pinning to release after the I/O 363 is complete, so they won't ever notice the fact the page is huge. But 364 if any driver is going to mangle over the page structure of the tail 365 page (like for checking page->mapping or other bits that are relevant 366 for the head page and not the tail page), it should be updated to jump 367 to check head page instead. Taking reference on any head/tail page would 368 prevent page from being split by anyone. 369 370 NOTE: these aren't new constraints to the GUP API, and they match the 371 same constrains that applies to hugetlbfs too, so any driver capable 372 of handling GUP on hugetlbfs will also work fine on transparent 373 hugepage backed mappings. 374 375 In case you can't handle compound pages if they're returned by 376 follow_page, the FOLL_SPLIT bit can be specified as parameter to 377 follow_page, so that it will split the hugepages before returning 378 them. Migration for example passes FOLL_SPLIT as parameter to 379 follow_page because it's not hugepage aware and in fact it can't work 380 at all on hugetlbfs (but it instead works fine on transparent 381 hugepages thanks to FOLL_SPLIT). migration simply can't deal with 382 hugepages being returned (as it's not only checking the pfn of the 383 page and pinning it during the copy but it pretends to migrate the 384 memory in regular page sizes and with regular pte/pmd mappings). 385 386 == Optimizing the applications == 387 388 To be guaranteed that the kernel will map a 2M page immediately in any 389 memory region, the mmap region has to be hugepage naturally 390 aligned. posix_memalign() can provide that guarantee. 391 392 == Hugetlbfs == 393 394 You can use hugetlbfs on a kernel that has transparent hugepage 395 support enabled just fine as always. No difference can be noted in 396 hugetlbfs other than there will be less overall fragmentation. All 397 usual features belonging to hugetlbfs are preserved and 398 unaffected. libhugetlbfs will also work fine as usual. 399 400 == Graceful fallback == 401 402 Code walking pagetables but unaware about huge pmds can simply call 403 split_huge_pmd(vma, pmd, addr) where the pmd is the one returned by 404 pmd_offset. It's trivial to make the code transparent hugepage aware 405 by just grepping for "pmd_offset" and adding split_huge_pmd where 406 missing after pmd_offset returns the pmd. Thanks to the graceful 407 fallback design, with a one liner change, you can avoid to write 408 hundred if not thousand of lines of complex code to make your code 409 hugepage aware. 410 411 If you're not walking pagetables but you run into a physical hugepage 412 but you can't handle it natively in your code, you can split it by 413 calling split_huge_page(page). This is what the Linux VM does before 414 it tries to swapout the hugepage for example. split_huge_page() can fail 415 if the page is pinned and you must handle this correctly. 416 417 Example to make mremap.c transparent hugepage aware with a one liner 418 change: 419 420 diff --git a/mm/mremap.c b/mm/mremap.c 421 --- a/mm/mremap.c 422 +++ b/mm/mremap.c 423 @@ -41,6 +41,7 @@ static pmd_t *get_old_pmd(struct mm_stru 424 return NULL; 425 426 pmd = pmd_offset(pud, addr); 427 + split_huge_pmd(vma, pmd, addr); 428 if (pmd_none_or_clear_bad(pmd)) 429 return NULL; 430 431 == Locking in hugepage aware code == 432 433 We want as much code as possible hugepage aware, as calling 434 split_huge_page() or split_huge_pmd() has a cost. 435 436 To make pagetable walks huge pmd aware, all you need to do is to call 437 pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the 438 mmap_sem in read (or write) mode to be sure an huge pmd cannot be 439 created from under you by khugepaged (khugepaged collapse_huge_page 440 takes the mmap_sem in write mode in addition to the anon_vma lock). If 441 pmd_trans_huge returns false, you just fallback in the old code 442 paths. If instead pmd_trans_huge returns true, you have to take the 443 page table lock (pmd_lock()) and re-run pmd_trans_huge. Taking the 444 page table lock will prevent the huge pmd to be converted into a 445 regular pmd from under you (split_huge_pmd can run in parallel to the 446 pagetable walk). If the second pmd_trans_huge returns false, you 447 should just drop the page table lock and fallback to the old code as 448 before. Otherwise you can proceed to process the huge pmd and the 449 hugepage natively. Once finished you can drop the page table lock. 450 451 == Refcounts and transparent huge pages == 452 453 Refcounting on THP is mostly consistent with refcounting on other compound 454 pages: 455 456 - get_page()/put_page() and GUP operate in head page's ->_refcount. 457 458 - ->_refcount in tail pages is always zero: get_page_unless_zero() never 459 succeed on tail pages. 460 461 - map/unmap of the pages with PTE entry increment/decrement ->_mapcount 462 on relevant sub-page of the compound page. 463 464 - map/unmap of the whole compound page accounted in compound_mapcount 465 (stored in first tail page). For file huge pages, we also increment 466 ->_mapcount of all sub-pages in order to have race-free detection of 467 last unmap of subpages. 468 469 PageDoubleMap() indicates that the page is *possibly* mapped with PTEs. 470 471 For anonymous pages PageDoubleMap() also indicates ->_mapcount in all 472 subpages is offset up by one. This additional reference is required to 473 get race-free detection of unmap of subpages when we have them mapped with 474 both PMDs and PTEs. 475 476 This is optimization required to lower overhead of per-subpage mapcount 477 tracking. The alternative is alter ->_mapcount in all subpages on each 478 map/unmap of the whole compound page. 479 480 For anonymous pages, we set PG_double_map when a PMD of the page got split 481 for the first time, but still have PMD mapping. The additional references 482 go away with last compound_mapcount. 483 484 File pages get PG_double_map set on first map of the page with PTE and 485 goes away when the page gets evicted from page cache. 486 487 split_huge_page internally has to distribute the refcounts in the head 488 page to the tail pages before clearing all PG_head/tail bits from the page 489 structures. It can be done easily for refcounts taken by page table 490 entries. But we don't have enough information on how to distribute any 491 additional pins (i.e. from get_user_pages). split_huge_page() fails any 492 requests to split pinned huge page: it expects page count to be equal to 493 sum of mapcount of all sub-pages plus one (split_huge_page caller must 494 have reference for head page). 495 496 split_huge_page uses migration entries to stabilize page->_refcount and 497 page->_mapcount of anonymous pages. File pages just got unmapped. 498 499 We safe against physical memory scanners too: the only legitimate way 500 scanner can get reference to a page is get_page_unless_zero(). 501 502 All tail pages have zero ->_refcount until atomic_add(). This prevents the 503 scanner from getting a reference to the tail page up to that point. After the 504 atomic_add() we don't care about the ->_refcount value. We already known how 505 many references should be uncharged from the head page. 506 507 For head page get_page_unless_zero() will succeed and we don't mind. It's 508 clear where reference should go after split: it will stay on head page. 509 510 Note that split_huge_pmd() doesn't have any limitation on refcounting: 511 pmd can be split at any point and never fails. 512 513 == Partial unmap and deferred_split_huge_page() == 514 515 Unmapping part of THP (with munmap() or other way) is not going to free 516 memory immediately. Instead, we detect that a subpage of THP is not in use 517 in page_remove_rmap() and queue the THP for splitting if memory pressure 518 comes. Splitting will free up unused subpages. 519 520 Splitting the page right away is not an option due to locking context in 521 the place where we can detect partial unmap. It's also might be 522 counterproductive since in many cases partial unmap happens during exit(2) if 523 a THP crosses a VMA boundary. 524 525 Function deferred_split_huge_page() is used to queue page for splitting. 526 The splitting itself will happen when we get memory pressure via shrinker 527 interface.