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