1============================ 2Transparent Hugepage Support 3============================ 4 5Objective 6========= 7 8Performance critical computing applications dealing with large memory 9working sets are already running on top of libhugetlbfs and in turn 10hugetlbfs. Transparent HugePage Support (THP) is an alternative mean of 11using huge pages for the backing of virtual memory with huge pages 12that supports the automatic promotion and demotion of page sizes and 13without the shortcomings of hugetlbfs. 14 15Currently THP only works for anonymous memory mappings and tmpfs/shmem. 16But in the future it can expand to other filesystems. 17 18.. note:: 19 in the examples below we presume that the basic page size is 4K and 20 the huge page size is 2M, although the actual numbers may vary 21 depending on the CPU architecture. 22 23The reason applications are running faster is because of two 24factors. The first factor is almost completely irrelevant and it's not 25of significant interest because it'll also have the downside of 26requiring larger clear-page copy-page in page faults which is a 27potentially negative effect. The first factor consists in taking a 28single page fault for each 2M virtual region touched by userland (so 29reducing the enter/exit kernel frequency by a 512 times factor). This 30only matters the first time the memory is accessed for the lifetime of 31a memory mapping. The second long lasting and much more important 32factor will affect all subsequent accesses to the memory for the whole 33runtime of the application. The second factor consist of two 34components: 35 361) the TLB miss will run faster (especially with virtualization using 37 nested pagetables but almost always also on bare metal without 38 virtualization) 39 402) a single TLB entry will be mapping a much larger amount of virtual 41 memory in turn reducing the number of TLB misses. With 42 virtualization and nested pagetables the TLB can be mapped of 43 larger size only if both KVM and the Linux guest are using 44 hugepages but a significant speedup already happens if only one of 45 the two is using hugepages just because of the fact the TLB miss is 46 going to run faster. 47 48Modern kernels support "multi-size THP" (mTHP), which introduces the 49ability to allocate memory in blocks that are bigger than a base page 50but smaller than traditional PMD-size (as described above), in 51increments of a power-of-2 number of pages. mTHP can back anonymous 52memory (for example 16K, 32K, 64K, etc). These THPs continue to be 53PTE-mapped, but in many cases can still provide similar benefits to 54those outlined above: Page faults are significantly reduced (by a 55factor of e.g. 4, 8, 16, etc), but latency spikes are much less 56prominent because the size of each page isn't as huge as the PMD-sized 57variant and there is less memory to clear in each page fault. Some 58architectures also employ TLB compression mechanisms to squeeze more 59entries in when a set of PTEs are virtually and physically contiguous 60and approporiately aligned. In this case, TLB misses will occur less 61often. 62 63THP can be enabled system wide or restricted to certain tasks or even 64memory ranges inside task's address space. Unless THP is completely 65disabled, there is ``khugepaged`` daemon that scans memory and 66collapses sequences of basic pages into PMD-sized huge pages. 67 68The THP behaviour is controlled via :ref:`sysfs <thp_sysfs>` 69interface and using madvise(2) and prctl(2) system calls. 70 71Transparent Hugepage Support maximizes the usefulness of free memory 72if compared to the reservation approach of hugetlbfs by allowing all 73unused memory to be used as cache or other movable (or even unmovable 74entities). It doesn't require reservation to prevent hugepage 75allocation failures to be noticeable from userland. It allows paging 76and all other advanced VM features to be available on the 77hugepages. It requires no modifications for applications to take 78advantage of it. 79 80Applications however can be further optimized to take advantage of 81this feature, like for example they've been optimized before to avoid 82a flood of mmap system calls for every malloc(4k). Optimizing userland 83is by far not mandatory and khugepaged already can take care of long 84lived page allocations even for hugepage unaware applications that 85deals with large amounts of memory. 86 87In certain cases when hugepages are enabled system wide, application 88may end up allocating more memory resources. An application may mmap a 89large region but only touch 1 byte of it, in that case a 2M page might 90be allocated instead of a 4k page for no good. This is why it's 91possible to disable hugepages system-wide and to only have them inside 92MADV_HUGEPAGE madvise regions. 93 94Embedded systems should enable hugepages only inside madvise regions 95to eliminate any risk of wasting any precious byte of memory and to 96only run faster. 97 98Applications that gets a lot of benefit from hugepages and that don't 99risk to lose memory by using hugepages, should use 100madvise(MADV_HUGEPAGE) on their critical mmapped regions. 101 102.. _thp_sysfs: 103 104sysfs 105===== 106 107Global THP controls 108------------------- 109 110Transparent Hugepage Support for anonymous memory can be entirely disabled 111(mostly for debugging purposes) or only enabled inside MADV_HUGEPAGE 112regions (to avoid the risk of consuming more memory resources) or enabled 113system wide. This can be achieved per-supported-THP-size with one of:: 114 115 echo always >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled 116 echo madvise >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled 117 echo never >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled 118 119where <size> is the hugepage size being addressed, the available sizes 120for which vary by system. 121 122For example:: 123 124 echo always >/sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled 125 126Alternatively it is possible to specify that a given hugepage size 127will inherit the top-level "enabled" value:: 128 129 echo inherit >/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/enabled 130 131For example:: 132 133 echo inherit >/sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled 134 135The top-level setting (for use with "inherit") can be set by issuing 136one of the following commands:: 137 138 echo always >/sys/kernel/mm/transparent_hugepage/enabled 139 echo madvise >/sys/kernel/mm/transparent_hugepage/enabled 140 echo never >/sys/kernel/mm/transparent_hugepage/enabled 141 142By default, PMD-sized hugepages have enabled="inherit" and all other 143hugepage sizes have enabled="never". If enabling multiple hugepage 144sizes, the kernel will select the most appropriate enabled size for a 145given allocation. 146 147It's also possible to limit defrag efforts in the VM to generate 148anonymous hugepages in case they're not immediately free to madvise 149regions or to never try to defrag memory and simply fallback to regular 150pages unless hugepages are immediately available. Clearly if we spend CPU 151time to defrag memory, we would expect to gain even more by the fact we 152use hugepages later instead of regular pages. This isn't always 153guaranteed, but it may be more likely in case the allocation is for a 154MADV_HUGEPAGE region. 155 156:: 157 158 echo always >/sys/kernel/mm/transparent_hugepage/defrag 159 echo defer >/sys/kernel/mm/transparent_hugepage/defrag 160 echo defer+madvise >/sys/kernel/mm/transparent_hugepage/defrag 161 echo madvise >/sys/kernel/mm/transparent_hugepage/defrag 162 echo never >/sys/kernel/mm/transparent_hugepage/defrag 163 164always 165 means that an application requesting THP will stall on 166 allocation failure and directly reclaim pages and compact 167 memory in an effort to allocate a THP immediately. This may be 168 desirable for virtual machines that benefit heavily from THP 169 use and are willing to delay the VM start to utilise them. 170 171defer 172 means that an application will wake kswapd in the background 173 to reclaim pages and wake kcompactd to compact memory so that 174 THP is available in the near future. It's the responsibility 175 of khugepaged to then install the THP pages later. 176 177defer+madvise 178 will enter direct reclaim and compaction like ``always``, but 179 only for regions that have used madvise(MADV_HUGEPAGE); all 180 other regions will wake kswapd in the background to reclaim 181 pages and wake kcompactd to compact memory so that THP is 182 available in the near future. 183 184madvise 185 will enter direct reclaim like ``always`` but only for regions 186 that are have used madvise(MADV_HUGEPAGE). This is the default 187 behaviour. 188 189never 190 should be self-explanatory. 191 192By default kernel tries to use huge, PMD-mappable zero page on read 193page fault to anonymous mapping. It's possible to disable huge zero 194page by writing 0 or enable it back by writing 1:: 195 196 echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page 197 echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page 198 199Some userspace (such as a test program, or an optimized memory 200allocation library) may want to know the size (in bytes) of a 201PMD-mappable transparent hugepage:: 202 203 cat /sys/kernel/mm/transparent_hugepage/hpage_pmd_size 204 205All THPs at fault and collapse time will be added to _deferred_list, 206and will therefore be split under memory presure if they are considered 207"underused". A THP is underused if the number of zero-filled pages in 208the THP is above max_ptes_none (see below). It is possible to disable 209this behaviour by writing 0 to shrink_underused, and enable it by writing 2101 to it:: 211 212 echo 0 > /sys/kernel/mm/transparent_hugepage/shrink_underused 213 echo 1 > /sys/kernel/mm/transparent_hugepage/shrink_underused 214 215khugepaged will be automatically started when PMD-sized THP is enabled 216(either of the per-size anon control or the top-level control are set 217to "always" or "madvise"), and it'll be automatically shutdown when 218PMD-sized THP is disabled (when both the per-size anon control and the 219top-level control are "never") 220 221Khugepaged controls 222------------------- 223 224.. note:: 225 khugepaged currently only searches for opportunities to collapse to 226 PMD-sized THP and no attempt is made to collapse to other THP 227 sizes. 228 229khugepaged runs usually at low frequency so while one may not want to 230invoke defrag algorithms synchronously during the page faults, it 231should be worth invoking defrag at least in khugepaged. However it's 232also possible to disable defrag in khugepaged by writing 0 or enable 233defrag in khugepaged by writing 1:: 234 235 echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag 236 echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag 237 238You can also control how many pages khugepaged should scan at each 239pass:: 240 241 /sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan 242 243and how many milliseconds to wait in khugepaged between each pass (you 244can set this to 0 to run khugepaged at 100% utilization of one core):: 245 246 /sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs 247 248and how many milliseconds to wait in khugepaged if there's an hugepage 249allocation failure to throttle the next allocation attempt:: 250 251 /sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs 252 253The khugepaged progress can be seen in the number of pages collapsed (note 254that this counter may not be an exact count of the number of pages 255collapsed, since "collapsed" could mean multiple things: (1) A PTE mapping 256being replaced by a PMD mapping, or (2) All 4K physical pages replaced by 257one 2M hugepage. Each may happen independently, or together, depending on 258the type of memory and the failures that occur. As such, this value should 259be interpreted roughly as a sign of progress, and counters in /proc/vmstat 260consulted for more accurate accounting):: 261 262 /sys/kernel/mm/transparent_hugepage/khugepaged/pages_collapsed 263 264for each pass:: 265 266 /sys/kernel/mm/transparent_hugepage/khugepaged/full_scans 267 268``max_ptes_none`` specifies how many extra small pages (that are 269not already mapped) can be allocated when collapsing a group 270of small pages into one large page:: 271 272 /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none 273 274A higher value leads to use additional memory for programs. 275A lower value leads to gain less thp performance. Value of 276max_ptes_none can waste cpu time very little, you can 277ignore it. 278 279``max_ptes_swap`` specifies how many pages can be brought in from 280swap when collapsing a group of pages into a transparent huge page:: 281 282 /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_swap 283 284A higher value can cause excessive swap IO and waste 285memory. A lower value can prevent THPs from being 286collapsed, resulting fewer pages being collapsed into 287THPs, and lower memory access performance. 288 289``max_ptes_shared`` specifies how many pages can be shared across multiple 290processes. khugepaged might treat pages of THPs as shared if any page of 291that THP is shared. Exceeding the number would block the collapse:: 292 293 /sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_shared 294 295A higher value may increase memory footprint for some workloads. 296 297Boot parameters 298=============== 299 300You can change the sysfs boot time default for the top-level "enabled" 301control by passing the parameter ``transparent_hugepage=always`` or 302``transparent_hugepage=madvise`` or ``transparent_hugepage=never`` to the 303kernel command line. 304 305Alternatively, each supported anonymous THP size can be controlled by 306passing ``thp_anon=<size>[KMG],<size>[KMG]:<state>;<size>[KMG]-<size>[KMG]:<state>``, 307where ``<size>`` is the THP size (must be a power of 2 of PAGE_SIZE and 308supported anonymous THP) and ``<state>`` is one of ``always``, ``madvise``, 309``never`` or ``inherit``. 310 311For example, the following will set 16K, 32K, 64K THP to ``always``, 312set 128K, 512K to ``inherit``, set 256K to ``madvise`` and 1M, 2M 313to ``never``:: 314 315 thp_anon=16K-64K:always;128K,512K:inherit;256K:madvise;1M-2M:never 316 317``thp_anon=`` may be specified multiple times to configure all THP sizes as 318required. If ``thp_anon=`` is specified at least once, any anon THP sizes 319not explicitly configured on the command line are implicitly set to 320``never``. 321 322``transparent_hugepage`` setting only affects the global toggle. If 323``thp_anon`` is not specified, PMD_ORDER THP will default to ``inherit``. 324However, if a valid ``thp_anon`` setting is provided by the user, the 325PMD_ORDER THP policy will be overridden. If the policy for PMD_ORDER 326is not defined within a valid ``thp_anon``, its policy will default to 327``never``. 328 329Hugepages in tmpfs/shmem 330======================== 331 332You can control hugepage allocation policy in tmpfs with mount option 333``huge=``. It can have following values: 334 335always 336 Attempt to allocate huge pages every time we need a new page; 337 338never 339 Do not allocate huge pages; 340 341within_size 342 Only allocate huge page if it will be fully within i_size. 343 Also respect fadvise()/madvise() hints; 344 345advise 346 Only allocate huge pages if requested with fadvise()/madvise(); 347 348The default policy is ``never``. 349 350``mount -o remount,huge= /mountpoint`` works fine after mount: remounting 351``huge=never`` will not attempt to break up huge pages at all, just stop more 352from being allocated. 353 354There's also sysfs knob to control hugepage allocation policy for internal 355shmem mount: /sys/kernel/mm/transparent_hugepage/shmem_enabled. The mount 356is used for SysV SHM, memfds, shared anonymous mmaps (of /dev/zero or 357MAP_ANONYMOUS), GPU drivers' DRM objects, Ashmem. 358 359In addition to policies listed above, shmem_enabled allows two further 360values: 361 362deny 363 For use in emergencies, to force the huge option off from 364 all mounts; 365force 366 Force the huge option on for all - very useful for testing; 367 368Shmem can also use "multi-size THP" (mTHP) by adding a new sysfs knob to 369control mTHP allocation: 370'/sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/shmem_enabled', 371and its value for each mTHP is essentially consistent with the global 372setting. An 'inherit' option is added to ensure compatibility with these 373global settings. Conversely, the options 'force' and 'deny' are dropped, 374which are rather testing artifacts from the old ages. 375 376always 377 Attempt to allocate <size> huge pages every time we need a new page; 378 379inherit 380 Inherit the top-level "shmem_enabled" value. By default, PMD-sized hugepages 381 have enabled="inherit" and all other hugepage sizes have enabled="never"; 382 383never 384 Do not allocate <size> huge pages; 385 386within_size 387 Only allocate <size> huge page if it will be fully within i_size. 388 Also respect fadvise()/madvise() hints; 389 390advise 391 Only allocate <size> huge pages if requested with fadvise()/madvise(); 392 393Need of application restart 394=========================== 395 396The transparent_hugepage/enabled and 397transparent_hugepage/hugepages-<size>kB/enabled values and tmpfs mount 398option only affect future behavior. So to make them effective you need 399to restart any application that could have been using hugepages. This 400also applies to the regions registered in khugepaged. 401 402Monitoring usage 403================ 404 405The number of PMD-sized anonymous transparent huge pages currently used by the 406system is available by reading the AnonHugePages field in ``/proc/meminfo``. 407To identify what applications are using PMD-sized anonymous transparent huge 408pages, it is necessary to read ``/proc/PID/smaps`` and count the AnonHugePages 409fields for each mapping. (Note that AnonHugePages only applies to traditional 410PMD-sized THP for historical reasons and should have been called 411AnonHugePmdMapped). 412 413The number of file transparent huge pages mapped to userspace is available 414by reading ShmemPmdMapped and ShmemHugePages fields in ``/proc/meminfo``. 415To identify what applications are mapping file transparent huge pages, it 416is necessary to read ``/proc/PID/smaps`` and count the FileHugeMapped fields 417for each mapping. 418 419Note that reading the smaps file is expensive and reading it 420frequently will incur overhead. 421 422There are a number of counters in ``/proc/vmstat`` that may be used to 423monitor how successfully the system is providing huge pages for use. 424 425thp_fault_alloc 426 is incremented every time a huge page is successfully 427 allocated and charged to handle a page fault. 428 429thp_collapse_alloc 430 is incremented by khugepaged when it has found 431 a range of pages to collapse into one huge page and has 432 successfully allocated a new huge page to store the data. 433 434thp_fault_fallback 435 is incremented if a page fault fails to allocate or charge 436 a huge page and instead falls back to using small pages. 437 438thp_fault_fallback_charge 439 is incremented if a page fault fails to charge a huge page and 440 instead falls back to using small pages even though the 441 allocation was successful. 442 443thp_collapse_alloc_failed 444 is incremented if khugepaged found a range 445 of pages that should be collapsed into one huge page but failed 446 the allocation. 447 448thp_file_alloc 449 is incremented every time a shmem huge page is successfully 450 allocated (Note that despite being named after "file", the counter 451 measures only shmem). 452 453thp_file_fallback 454 is incremented if a shmem huge page is attempted to be allocated 455 but fails and instead falls back to using small pages. (Note that 456 despite being named after "file", the counter measures only shmem). 457 458thp_file_fallback_charge 459 is incremented if a shmem huge page cannot be charged and instead 460 falls back to using small pages even though the allocation was 461 successful. (Note that despite being named after "file", the 462 counter measures only shmem). 463 464thp_file_mapped 465 is incremented every time a file or shmem huge page is mapped into 466 user address space. 467 468thp_split_page 469 is incremented every time a huge page is split into base 470 pages. This can happen for a variety of reasons but a common 471 reason is that a huge page is old and is being reclaimed. 472 This action implies splitting all PMD the page mapped with. 473 474thp_split_page_failed 475 is incremented if kernel fails to split huge 476 page. This can happen if the page was pinned by somebody. 477 478thp_deferred_split_page 479 is incremented when a huge page is put onto split 480 queue. This happens when a huge page is partially unmapped and 481 splitting it would free up some memory. Pages on split queue are 482 going to be split under memory pressure. 483 484thp_underused_split_page 485 is incremented when a huge page on the split queue was split 486 because it was underused. A THP is underused if the number of 487 zero pages in the THP is above a certain threshold 488 (/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none). 489 490thp_split_pmd 491 is incremented every time a PMD split into table of PTEs. 492 This can happen, for instance, when application calls mprotect() or 493 munmap() on part of huge page. It doesn't split huge page, only 494 page table entry. 495 496thp_zero_page_alloc 497 is incremented every time a huge zero page used for thp is 498 successfully allocated. Note, it doesn't count every map of 499 the huge zero page, only its allocation. 500 501thp_zero_page_alloc_failed 502 is incremented if kernel fails to allocate 503 huge zero page and falls back to using small pages. 504 505thp_swpout 506 is incremented every time a huge page is swapout in one 507 piece without splitting. 508 509thp_swpout_fallback 510 is incremented if a huge page has to be split before swapout. 511 Usually because failed to allocate some continuous swap space 512 for the huge page. 513 514In /sys/kernel/mm/transparent_hugepage/hugepages-<size>kB/stats, There are 515also individual counters for each huge page size, which can be utilized to 516monitor the system's effectiveness in providing huge pages for usage. Each 517counter has its own corresponding file. 518 519anon_fault_alloc 520 is incremented every time a huge page is successfully 521 allocated and charged to handle a page fault. 522 523anon_fault_fallback 524 is incremented if a page fault fails to allocate or charge 525 a huge page and instead falls back to using huge pages with 526 lower orders or small pages. 527 528anon_fault_fallback_charge 529 is incremented if a page fault fails to charge a huge page and 530 instead falls back to using huge pages with lower orders or 531 small pages even though the allocation was successful. 532 533swpout 534 is incremented every time a huge page is swapped out in one 535 piece without splitting. 536 537swpout_fallback 538 is incremented if a huge page has to be split before swapout. 539 Usually because failed to allocate some continuous swap space 540 for the huge page. 541 542shmem_alloc 543 is incremented every time a shmem huge page is successfully 544 allocated. 545 546shmem_fallback 547 is incremented if a shmem huge page is attempted to be allocated 548 but fails and instead falls back to using small pages. 549 550shmem_fallback_charge 551 is incremented if a shmem huge page cannot be charged and instead 552 falls back to using small pages even though the allocation was 553 successful. 554 555split 556 is incremented every time a huge page is successfully split into 557 smaller orders. This can happen for a variety of reasons but a 558 common reason is that a huge page is old and is being reclaimed. 559 560split_failed 561 is incremented if kernel fails to split huge 562 page. This can happen if the page was pinned by somebody. 563 564split_deferred 565 is incremented when a huge page is put onto split queue. 566 This happens when a huge page is partially unmapped and splitting 567 it would free up some memory. Pages on split queue are going to 568 be split under memory pressure, if splitting is possible. 569 570nr_anon 571 the number of anonymous THP we have in the whole system. These THPs 572 might be currently entirely mapped or have partially unmapped/unused 573 subpages. 574 575nr_anon_partially_mapped 576 the number of anonymous THP which are likely partially mapped, possibly 577 wasting memory, and have been queued for deferred memory reclamation. 578 Note that in corner some cases (e.g., failed migration), we might detect 579 an anonymous THP as "partially mapped" and count it here, even though it 580 is not actually partially mapped anymore. 581 582As the system ages, allocating huge pages may be expensive as the 583system uses memory compaction to copy data around memory to free a 584huge page for use. There are some counters in ``/proc/vmstat`` to help 585monitor this overhead. 586 587compact_stall 588 is incremented every time a process stalls to run 589 memory compaction so that a huge page is free for use. 590 591compact_success 592 is incremented if the system compacted memory and 593 freed a huge page for use. 594 595compact_fail 596 is incremented if the system tries to compact memory 597 but failed. 598 599It is possible to establish how long the stalls were using the function 600tracer to record how long was spent in __alloc_pages() and 601using the mm_page_alloc tracepoint to identify which allocations were 602for huge pages. 603 604Optimizing the applications 605=========================== 606 607To be guaranteed that the kernel will map a THP immediately in any 608memory region, the mmap region has to be hugepage naturally 609aligned. posix_memalign() can provide that guarantee. 610 611Hugetlbfs 612========= 613 614You can use hugetlbfs on a kernel that has transparent hugepage 615support enabled just fine as always. No difference can be noted in 616hugetlbfs other than there will be less overall fragmentation. All 617usual features belonging to hugetlbfs are preserved and 618unaffected. libhugetlbfs will also work fine as usual. 619