1= Transparent Hugepage Support =
   3== Objective ==
   5Performance critical computing applications dealing with large memory
   6working sets are already running on top of libhugetlbfs and in turn
   7hugetlbfs. Transparent Hugepage Support is an alternative means of
   8using huge pages for the backing of virtual memory with huge pages
   9that supports the automatic promotion and demotion of page sizes and
  10without the shortcomings of hugetlbfs.
  12Currently it only works for anonymous memory mappings and tmpfs/shmem.
  13But in the future it can expand to other filesystems.
  15The reason applications are running faster is because of two
  16factors. The first factor is almost completely irrelevant and it's not
  17of significant interest because it'll also have the downside of
  18requiring larger clear-page copy-page in page faults which is a
  19potentially negative effect. The first factor consists in taking a
  20single page fault for each 2M virtual region touched by userland (so
  21reducing the enter/exit kernel frequency by a 512 times factor). This
  22only matters the first time the memory is accessed for the lifetime of
  23a memory mapping. The second long lasting and much more important
  24factor will affect all subsequent accesses to the memory for the whole
  25runtime of the application. The second factor consist of two
  26components: 1) the TLB miss will run faster (especially with
  27virtualization using nested pagetables but almost always also on bare
  28metal without virtualization) and 2) a single TLB entry will be
  29mapping a much larger amount of virtual memory in turn reducing the
  30number of TLB misses. With virtualization and nested pagetables the
  31TLB can be mapped of larger size only if both KVM and the Linux guest
  32are using hugepages but a significant speedup already happens if only
  33one of the two is using hugepages just because of the fact the TLB
  34miss is going to run faster.
  36== Design ==
  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.
  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
  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)
  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)
  60Transparent Hugepage Support maximizes the usefulness of free memory
  61if compared to the reservation approach of hugetlbfs by allowing all
  62unused memory to be used as cache or other movable (or even unmovable
  63entities). It doesn't require reservation to prevent hugepage
  64allocation failures to be noticeable from userland. It allows paging
  65and all other advanced VM features to be available on the
  66hugepages. It requires no modifications for applications to take
  67advantage of it.
  69Applications however can be further optimized to take advantage of
  70this feature, like for example they've been optimized before to avoid
  71a flood of mmap system calls for every malloc(4k). Optimizing userland
  72is by far not mandatory and khugepaged already can take care of long
  73lived page allocations even for hugepage unaware applications that
  74deals with large amounts of memory.
  76In certain cases when hugepages are enabled system wide, application
  77may end up allocating more memory resources. An application may mmap a
  78large region but only touch 1 byte of it, in that case a 2M page might
  79be allocated instead of a 4k page for no good. This is why it's
  80possible to disable hugepages system-wide and to only have them inside
  81MADV_HUGEPAGE madvise regions.
  83Embedded systems should enable hugepages only inside madvise regions
  84to eliminate any risk of wasting any precious byte of memory and to
  85only run faster.
  87Applications that gets a lot of benefit from hugepages and that don't
  88risk to lose memory by using hugepages, should use
  89madvise(MADV_HUGEPAGE) on their critical mmapped regions.
  91== sysfs ==
  93Transparent Hugepage Support for anonymous memory can be entirely disabled
  94(mostly for debugging purposes) or only enabled inside MADV_HUGEPAGE
  95regions (to avoid the risk of consuming more memory resources) or enabled
  96system wide. This can be achieved with one of:
  98echo always >/sys/kernel/mm/transparent_hugepage/enabled
  99echo madvise >/sys/kernel/mm/transparent_hugepage/enabled
 100echo never >/sys/kernel/mm/transparent_hugepage/enabled
 102It's also possible to limit defrag efforts in the VM to generate
 103anonymous hugepages in case they're not immediately free to madvise
 104regions or to never try to defrag memory and simply fallback to regular
 105pages unless hugepages are immediately available. Clearly if we spend CPU
 106time to defrag memory, we would expect to gain even more by the fact we
 107use hugepages later instead of regular pages. This isn't always
 108guaranteed, but it may be more likely in case the allocation is for a
 109MADV_HUGEPAGE region.
 111echo always >/sys/kernel/mm/transparent_hugepage/defrag
 112echo defer >/sys/kernel/mm/transparent_hugepage/defrag
 113echo defer+madvise >/sys/kernel/mm/transparent_hugepage/defrag
 114echo madvise >/sys/kernel/mm/transparent_hugepage/defrag
 115echo never >/sys/kernel/mm/transparent_hugepage/defrag
 117"always" means that an application requesting THP will stall on allocation
 118failure and directly reclaim pages and compact memory in an effort to
 119allocate a THP immediately. This may be desirable for virtual machines
 120that benefit heavily from THP use and are willing to delay the VM start
 121to utilise them.
 123"defer" means that an application will wake kswapd in the background
 124to reclaim pages and wake kcompactd to compact memory so that THP is
 125available in the near future. It's the responsibility of khugepaged
 126to then install the THP pages later.
 128"defer+madvise" will enter direct reclaim and compaction like "always", but
 129only for regions that have used madvise(MADV_HUGEPAGE); all other regions
 130will wake kswapd in the background to reclaim pages and wake kcompactd to
 131compact memory so that THP is available in the near future.
 133"madvise" will enter direct reclaim like "always" but only for regions
 134that are have used madvise(MADV_HUGEPAGE). This is the default behaviour.
 136"never" should be self-explanatory.
 138By default kernel tries to use huge zero page on read page fault to
 139anonymous mapping. It's possible to disable huge zero page by writing 0
 140or enable it back by writing 1:
 142echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page
 143echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page
 145Some userspace (such as a test program, or an optimized memory allocation
 146library) may want to know the size (in bytes) of a transparent hugepage:
 148cat /sys/kernel/mm/transparent_hugepage/hpage_pmd_size
 150khugepaged will be automatically started when
 151transparent_hugepage/enabled is set to "always" or "madvise, and it'll
 152be automatically shutdown if it's set to "never".
 154khugepaged runs usually at low frequency so while one may not want to
 155invoke defrag algorithms synchronously during the page faults, it
 156should be worth invoking defrag at least in khugepaged. However it's
 157also possible to disable defrag in khugepaged by writing 0 or enable
 158defrag in khugepaged by writing 1:
 160echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag
 161echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag
 163You can also control how many pages khugepaged should scan at each
 168and how many milliseconds to wait in khugepaged between each pass (you
 169can set this to 0 to run khugepaged at 100% utilization of one core):
 173and how many milliseconds to wait in khugepaged if there's an hugepage
 174allocation failure to throttle the next allocation attempt.
 178The khugepaged progress can be seen in the number of pages collapsed:
 182for each pass:
 186max_ptes_none specifies how many extra small pages (that are
 187not already mapped) can be allocated when collapsing a group
 188of small pages into one large page.
 192A higher value leads to use additional memory for programs.
 193A lower value leads to gain less thp performance. Value of
 194max_ptes_none can waste cpu time very little, you can
 195ignore it.
 197max_ptes_swap specifies how many pages can be brought in from
 198swap when collapsing a group of pages into a transparent huge page.
 202A higher value can cause excessive swap IO and waste
 203memory. A lower value can prevent THPs from being
 204collapsed, resulting fewer pages being collapsed into
 205THPs, and lower memory access performance.
 207== Boot parameter ==
 209You can change the sysfs boot time defaults of Transparent Hugepage
 210Support by passing the parameter "transparent_hugepage=always" or
 211"transparent_hugepage=madvise" or "transparent_hugepage=never"
 212(without "") to the kernel command line.
 214== Hugepages in tmpfs/shmem ==
 216You can control hugepage allocation policy in tmpfs with mount option
 217"huge=". It can have following values:
 219  - "always":
 220    Attempt to allocate huge pages every time we need a new page;
 222  - "never":
 223    Do not allocate huge pages;
 225  - "within_size":
 226    Only allocate huge page if it will be fully within i_size.
 227    Also respect fadvise()/madvise() hints;
 229  - "advise:
 230    Only allocate huge pages if requested with fadvise()/madvise();
 232The default policy is "never".
 234"mount -o remount,huge= /mountpoint" works fine after mount: remounting
 235huge=never will not attempt to break up huge pages at all, just stop more
 236from being allocated.
 238There's also sysfs knob to control hugepage allocation policy for internal
 239shmem mount: /sys/kernel/mm/transparent_hugepage/shmem_enabled. The mount
 240is used for SysV SHM, memfds, shared anonymous mmaps (of /dev/zero or
 241MAP_ANONYMOUS), GPU drivers' DRM objects, Ashmem.
 243In addition to policies listed above, shmem_enabled allows two further
 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;
 252== Need of application restart ==
 254The transparent_hugepage/enabled values and tmpfs mount option only affect
 255future behavior. So to make them effective you need to restart any
 256application that could have been using hugepages. This also applies to the
 257regions registered in khugepaged.
 259== Monitoring usage ==
 261The number of anonymous transparent huge pages currently used by the
 262system is available by reading the AnonHugePages field in /proc/meminfo.
 263To identify what applications are using anonymous transparent huge pages,
 264it is necessary to read /proc/PID/smaps and count the AnonHugePages fields
 265for each mapping.
 267The number of file transparent huge pages mapped to userspace is available
 268by reading ShmemPmdMapped and ShmemHugePages fields in /proc/meminfo.
 269To identify what applications are mapping file transparent huge pages, it
 270is necessary to read /proc/PID/smaps and count the FileHugeMapped fields
 271for each mapping.
 273Note that reading the smaps file is expensive and reading it
 274frequently will incur overhead.
 276There are a number of counters in /proc/vmstat that may be used to
 277monitor how successfully the system is providing huge pages for use.
 279thp_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.
 283thp_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.
 287thp_fault_fallback is incremented if a page fault fails to allocate
 288        a huge page and instead falls back to using small pages.
 290thp_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.
 294thp_file_alloc is incremented every time a file huge page is successfully
 295        allocated.
 297thp_file_mapped is incremented every time a file huge page is mapped into
 298        user address space.
 300thp_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.
 305thp_split_page_failed is incremented if kernel fails to split huge
 306        page. This can happen if the page was pinned by somebody.
 308thp_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.
 313thp_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.
 318thp_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.
 323thp_zero_page_alloc_failed is incremented if kernel fails to allocate
 324        huge zero page and falls back to using small pages.
 326As the system ages, allocating huge pages may be expensive as the
 327system uses memory compaction to copy data around memory to free a
 328huge page for use. There are some counters in /proc/vmstat to help
 329monitor this overhead.
 331compact_stall is incremented every time a process stalls to run
 332        memory compaction so that a huge page is free for use.
 334compact_success is incremented if the system compacted memory and
 335        freed a huge page for use.
 337compact_fail is incremented if the system tries to compact memory
 338        but failed.
 340compact_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.
 346compact_pagemigrate_failed is incremented when the underlying mechanism
 347        for moving a page failed.
 349compact_blocks_moved is incremented each time memory compaction examines
 350        a huge page aligned range of pages.
 352It is possible to establish how long the stalls were using the function
 353tracer to record how long was spent in __alloc_pages_nodemask and
 354using the mm_page_alloc tracepoint to identify which allocations were
 355for huge pages.
 357== get_user_pages and follow_page ==
 359get_user_pages and follow_page if run on a hugepage, will return the
 360head or tail pages as usual (exactly as they would do on
 361hugetlbfs). Most gup users will only care about the actual physical
 362address of the page and its temporary pinning to release after the I/O
 363is complete, so they won't ever notice the fact the page is huge. But
 364if any driver is going to mangle over the page structure of the tail
 365page (like for checking page->mapping or other bits that are relevant
 366for the head page and not the tail page), it should be updated to jump
 367to check head page instead. Taking reference on any head/tail page would
 368prevent page from being split by anyone.
 370NOTE: these aren't new constraints to the GUP API, and they match the
 371same constrains that applies to hugetlbfs too, so any driver capable
 372of handling GUP on hugetlbfs will also work fine on transparent
 373hugepage backed mappings.
 375In case you can't handle compound pages if they're returned by
 376follow_page, the FOLL_SPLIT bit can be specified as parameter to
 377follow_page, so that it will split the hugepages before returning
 378them. Migration for example passes FOLL_SPLIT as parameter to
 379follow_page because it's not hugepage aware and in fact it can't work
 380at all on hugetlbfs (but it instead works fine on transparent
 381hugepages thanks to FOLL_SPLIT). migration simply can't deal with
 382hugepages being returned (as it's not only checking the pfn of the
 383page and pinning it during the copy but it pretends to migrate the
 384memory in regular page sizes and with regular pte/pmd mappings).
 386== Optimizing the applications ==
 388To be guaranteed that the kernel will map a 2M page immediately in any
 389memory region, the mmap region has to be hugepage naturally
 390aligned. posix_memalign() can provide that guarantee.
 392== Hugetlbfs ==
 394You can use hugetlbfs on a kernel that has transparent hugepage
 395support enabled just fine as always. No difference can be noted in
 396hugetlbfs other than there will be less overall fragmentation. All
 397usual features belonging to hugetlbfs are preserved and
 398unaffected. libhugetlbfs will also work fine as usual.
 400== Graceful fallback ==
 402Code walking pagetables but unaware about huge pmds can simply call
 403split_huge_pmd(vma, pmd, addr) where the pmd is the one returned by
 404pmd_offset. It's trivial to make the code transparent hugepage aware
 405by just grepping for "pmd_offset" and adding split_huge_pmd where
 406missing after pmd_offset returns the pmd. Thanks to the graceful
 407fallback design, with a one liner change, you can avoid to write
 408hundred if not thousand of lines of complex code to make your code
 409hugepage aware.
 411If you're not walking pagetables but you run into a physical hugepage
 412but you can't handle it natively in your code, you can split it by
 413calling split_huge_page(page). This is what the Linux VM does before
 414it tries to swapout the hugepage for example. split_huge_page() can fail
 415if the page is pinned and you must handle this correctly.
 417Example to make mremap.c transparent hugepage aware with a one liner
 420diff --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;
 426        pmd = pmd_offset(pud, addr);
 427+       split_huge_pmd(vma, pmd, addr);
 428        if (pmd_none_or_clear_bad(pmd))
 429                return NULL;
 431== Locking in hugepage aware code ==
 433We want as much code as possible hugepage aware, as calling
 434split_huge_page() or split_huge_pmd() has a cost.
 436To make pagetable walks huge pmd aware, all you need to do is to call
 437pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the
 438mmap_sem in read (or write) mode to be sure an huge pmd cannot be
 439created from under you by khugepaged (khugepaged collapse_huge_page
 440takes the mmap_sem in write mode in addition to the anon_vma lock). If
 441pmd_trans_huge returns false, you just fallback in the old code
 442paths. If instead pmd_trans_huge returns true, you have to take the
 443page table lock (pmd_lock()) and re-run pmd_trans_huge. Taking the
 444page table lock will prevent the huge pmd to be converted into a
 445regular pmd from under you (split_huge_pmd can run in parallel to the
 446pagetable walk). If the second pmd_trans_huge returns false, you
 447should just drop the page table lock and fallback to the old code as
 448before. Otherwise you can proceed to process the huge pmd and the
 449hugepage natively. Once finished you can drop the page table lock.
 451== Refcounts and transparent huge pages ==
 453Refcounting on THP is mostly consistent with refcounting on other compound
 456  - get_page()/put_page() and GUP operate in head page's ->_refcount.
 458  - ->_refcount in tail pages is always zero: get_page_unless_zero() never
 459    succeed on tail pages.
 461  - map/unmap of the pages with PTE entry increment/decrement ->_mapcount
 462    on relevant sub-page of the compound page.
 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.
 469PageDoubleMap() indicates that the page is *possibly* mapped with PTEs.
 471For anonymous pages PageDoubleMap() also indicates ->_mapcount in all
 472subpages is offset up by one. This additional reference is required to
 473get race-free detection of unmap of subpages when we have them mapped with
 474both PMDs and PTEs.
 476This is optimization required to lower overhead of per-subpage mapcount
 477tracking. The alternative is alter ->_mapcount in all subpages on each
 478map/unmap of the whole compound page.
 480For anonymous pages, we set PG_double_map when a PMD of the page got split
 481for the first time, but still have PMD mapping. The additional references
 482go away with last compound_mapcount.
 484File pages get PG_double_map set on first map of the page with PTE and
 485goes away when the page gets evicted from page cache.
 487split_huge_page internally has to distribute the refcounts in the head
 488page to the tail pages before clearing all PG_head/tail bits from the page
 489structures. It can be done easily for refcounts taken by page table
 490entries. But we don't have enough information on how to distribute any
 491additional pins (i.e. from get_user_pages). split_huge_page() fails any
 492requests to split pinned huge page: it expects page count to be equal to
 493sum of mapcount of all sub-pages plus one (split_huge_page caller must
 494have reference for head page).
 496split_huge_page uses migration entries to stabilize page->_refcount and
 497page->_mapcount of anonymous pages. File pages just got unmapped.
 499We safe against physical memory scanners too: the only legitimate way
 500scanner can get reference to a page is get_page_unless_zero().
 502All tail pages have zero ->_refcount until atomic_add(). This prevents the
 503scanner from getting a reference to the tail page up to that point. After the
 504atomic_add() we don't care about the ->_refcount value. We already known how
 505many references should be uncharged from the head page.
 507For head page get_page_unless_zero() will succeed and we don't mind. It's
 508clear where reference should go after split: it will stay on head page.
 510Note that split_huge_pmd() doesn't have any limitation on refcounting:
 511pmd can be split at any point and never fails.
 513== Partial unmap and deferred_split_huge_page() ==
 515Unmapping part of THP (with munmap() or other way) is not going to free
 516memory immediately. Instead, we detect that a subpage of THP is not in use
 517in page_remove_rmap() and queue the THP for splitting if memory pressure
 518comes. Splitting will free up unused subpages.
 520Splitting the page right away is not an option due to locking context in
 521the place where we can detect partial unmap. It's also might be
 522counterproductive since in many cases partial unmap happens during exit(2) if
 523a THP crosses a VMA boundary.
 525Function deferred_split_huge_page() is used to queue page for splitting.
 526The splitting itself will happen when we get memory pressure via shrinker
 528 kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.