1 .\" Copyright (C) 2001 David Gómez <davidge@jazzfree.com>
3 .\" SPDX-License-Identifier: Linux-man-pages-copyleft
5 .\" Based on comments from mm/filemap.c. Last modified on 10-06-2001
6 .\" Modified, 25 Feb 2002, Michael Kerrisk, <mtk.manpages@gmail.com>
7 .\" Added notes on MADV_DONTNEED
8 .\" 2010-06-19, mtk, Added documentation of MADV_MERGEABLE and
10 .\" 2010-06-15, Andi Kleen, Add documentation of MADV_HWPOISON.
11 .\" 2010-06-19, Andi Kleen, Add documentation of MADV_SOFT_OFFLINE.
12 .\" 2011-09-18, Doug Goldstein <cardoe@cardoe.com>
13 .\" Document MADV_HUGEPAGE and MADV_NOHUGEPAGE
15 .TH madvise 2 (date) "Linux man-pages (unreleased)"
17 madvise \- give advice about use of memory
20 .RI ( libc ", " \-lc )
23 .B #include <sys/mman.h>
25 .BI "int madvise(void " addr [. length "], size_t " length ", int " advice );
29 Feature Test Macro Requirements for glibc (see
30 .BR feature_test_macros (7)):
37 Up to and including glibc 2.19:
43 system call is used to give advice or directions to the kernel
44 about the address range beginning at address
49 only operates on whole pages, therefore
54 is rounded up to a multiple of page size.
56 the goal of such advice is to improve system or application performance.
58 Initially, the system call supported a set of "conventional"
60 values, which are also available on several other implementations.
63 is not specified in POSIX.)
64 Subsequently, a number of Linux-specific
66 values have been added.
68 .\" ======================================================================
70 .SS Conventional advice values
74 allow an application to tell the kernel how it expects to use
75 some mapped or shared memory areas, so that the kernel can choose
76 appropriate read-ahead and caching techniques.
79 values do not influence the semantics of the application
80 (except in the case of
82 but may influence its performance.
85 values listed here have analogs in the POSIX-specified
87 function, and the values have the same meanings, with the exception of
90 The advice is indicated in the
92 argument, which is one of the following:
99 Expect page references in random order.
100 (Hence, read ahead may be less useful than normally.)
103 Expect page references in sequential order.
104 (Hence, pages in the given range can be aggressively read ahead,
105 and may be freed soon after they are accessed.)
108 Expect access in the near future.
109 (Hence, it might be a good idea to read some pages ahead.)
112 Do not expect access in the near future.
113 (For the time being, the application is finished with the given range,
114 so the kernel can free resources associated with it.)
119 the semantics of memory access in the specified region are changed:
120 subsequent accesses of pages in the range will succeed, but will result
121 in either repopulating the memory contents from the
122 up-to-date contents of the underlying mapped file
123 (for shared file mappings, shared anonymous mappings,
124 and shmem-based techniques such as System V shared memory segments)
125 or zero-fill-on-demand pages for anonymous private mappings.
127 Note that, when applied to shared mappings,
129 might not lead to immediate freeing of the pages in the range.
130 The kernel is free to delay freeing the pages until an appropriate moment.
131 The resident set size (RSS) of the calling process will be immediately
135 cannot be applied to locked pages, or
138 (Pages marked with the kernel-internal
140 .\" http://lwn.net/Articles/162860/
141 flag are special memory areas that are not managed
142 by the virtual memory subsystem.
143 Such pages are typically created by device drivers that
144 map the pages into user space.)
146 Support for Huge TLB pages was added in Linux v5.18.
147 Addresses within a mapping backed by Huge TLB pages must be aligned
148 to the underlying Huge TLB page size,
149 and the range length is rounded up
150 to a multiple of the underlying Huge TLB page size.
152 .\" ======================================================================
154 .SS Linux-specific advice values
155 The following Linux-specific
157 values have no counterparts in the POSIX-specified
158 .BR posix_madvise (3),
159 and may or may not have counterparts in the
161 interface available on other implementations.
162 Note that some of these operations change the semantics of memory accesses.
164 .BR MADV_REMOVE " (since Linux 2.6.16)"
165 .\" commit f6b3ec238d12c8cc6cc71490c6e3127988460349
166 Free up a given range of pages
167 and its associated backing store.
168 This is equivalent to punching a hole in the corresponding
169 range of the backing store (see
171 Subsequent accesses in the specified address range will see
172 data with a value of zero.
173 .\" Databases want to use this feature to drop a section of their
174 .\" bufferpool (shared memory segments) - without writing back to
175 .\" disk/swap space. This feature is also useful for supporting
176 .\" hot-plug memory on UML.
178 The specified address range must be mapped shared and writable.
179 This flag cannot be applied to locked pages, or
183 In the initial implementation, only
188 .\" commit 3f31d07571eeea18a7d34db9af21d2285b807a17
189 any filesystem which supports the
191 .B FALLOC_FL_PUNCH_HOLE
194 Filesystems which do not support
199 Support for the Huge TLB filesystem was added in Linux v4.3.
201 .BR MADV_DONTFORK " (since Linux 2.6.16)"
202 .\" commit f822566165dd46ff5de9bf895cfa6c51f53bb0c4
203 .\" See http://lwn.net/Articles/171941/
204 Do not make the pages in this range available to the child after a
206 This is useful to prevent copy-on-write semantics from changing
207 the physical location of a page if the parent writes to it after a
209 (Such page relocations cause problems for hardware that
211 .\" [PATCH] madvise MADV_DONTFORK/MADV_DOFORK
212 .\" Currently, copy-on-write may change the physical address of
213 .\" a page even if the user requested that the page is pinned in
214 .\" memory (either by mlock or by get_user_pages). This happens
215 .\" if the process forks meanwhile, and the parent writes to that
216 .\" page. As a result, the page is orphaned: in case of
217 .\" get_user_pages, the application will never see any data hardware
218 .\" DMA's into this page after the COW. In case of mlock'd memory,
219 .\" the parent is not getting the realtime/security benefits of mlock.
221 .\" In particular, this affects the Infiniband modules which do DMA from
222 .\" and into user pages all the time.
224 .\" This patch adds madvise options to control whether memory range is
225 .\" inherited across fork. Useful e.g. for when hardware is doing DMA
226 .\" from/into these pages. Could also be useful to an application
227 .\" wanting to speed up its forks by cutting large areas out of
230 .\" SEE ALSO: http://lwn.net/Articles/171941/
231 .\" "Tweaks to madvise() and posix_fadvise()", 14 Feb 2006
233 .BR MADV_DOFORK " (since Linux 2.6.16)"
236 restoring the default behavior, whereby a mapping is inherited across
239 .BR MADV_HWPOISON " (since Linux 2.6.32)"
240 .\" commit 9893e49d64a4874ea67849ee2cfbf3f3d6817573
241 Poison the pages in the range specified by
245 and handle subsequent references to those pages
246 like a hardware memory corruption.
247 This operation is available only for privileged
248 .RB ( CAP_SYS_ADMIN )
250 This operation may result in the calling process receiving a
252 and the page being unmapped.
254 This feature is intended for testing of memory error-handling code;
255 it is available only if the kernel was configured with
256 .BR CONFIG_MEMORY_FAILURE .
258 .BR MADV_MERGEABLE " (since Linux 2.6.32)"
259 .\" commit f8af4da3b4c14e7267c4ffb952079af3912c51c5
260 Enable Kernel Samepage Merging (KSM) for the pages in the range specified by
264 The kernel regularly scans those areas of user memory that have
265 been marked as mergeable,
266 looking for pages with identical content.
267 These are replaced by a single write-protected page (which is automatically
268 copied if a process later wants to update the content of the page).
269 KSM merges only private anonymous pages (see
272 The KSM feature is intended for applications that generate many
273 instances of the same data (e.g., virtualization systems such as KVM).
274 It can consume a lot of processing power; use with care.
275 See the Linux kernel source file
276 .I Documentation/admin\-guide/mm/ksm.rst
283 operations are available only if the kernel was configured with
286 .BR MADV_UNMERGEABLE " (since Linux 2.6.32)"
287 Undo the effect of an earlier
289 operation on the specified address range;
290 KSM unmerges whatever pages it had merged in the address range specified by
295 .BR MADV_SOFT_OFFLINE " (since Linux 2.6.33)"
296 .\" commit afcf938ee0aac4ef95b1a23bac704c6fbeb26de6
297 Soft offline the pages in the range specified by
301 The memory of each page in the specified range is preserved
302 (i.e., when next accessed, the same content will be visible,
303 but in a new physical page frame),
304 and the original page is offlined
305 (i.e., no longer used, and taken out of normal memory management).
308 operation is invisible to (i.e., does not change the semantics of)
311 This feature is intended for testing of memory error-handling code;
312 it is available only if the kernel was configured with
313 .BR CONFIG_MEMORY_FAILURE .
315 .BR MADV_HUGEPAGE " (since Linux 2.6.38)"
316 .\" commit 0af4e98b6b095c74588af04872f83d333c958c32
317 .\" http://lwn.net/Articles/358904/
318 .\" https://lwn.net/Articles/423584/
319 Enable Transparent Huge Pages (THP) for pages in the range specified by
323 The kernel will regularly scan the areas marked as huge page candidates
324 to replace them with huge pages.
325 The kernel will also allocate huge pages directly when the region is
326 naturally aligned to the huge page size (see
327 .BR posix_memalign (2)).
329 This feature is primarily aimed at applications that use large mappings of
330 data and access large regions of that memory at a time (e.g., virtualization
331 systems such as QEMU).
332 It can very easily waste memory (e.g., a 2\ MB mapping that only ever accesses
333 1 byte will result in 2\ MB of wired memory instead of one 4\ KB page).
334 See the Linux kernel source file
335 .I Documentation/admin\-guide/mm/transhuge.rst
338 Most common kernels configurations provide
339 .BR MADV_HUGEPAGE -style
340 behavior by default, and thus
342 is normally not necessary.
343 It is mostly intended for embedded systems, where
344 .BR MADV_HUGEPAGE -style
345 behavior may not be enabled by default in the kernel.
347 this flag can be used in order to selectively enable THP.
350 is used, it should always be in regions of memory with
351 an access pattern that the developer knows in advance won't risk
352 to increase the memory footprint of the application when transparent
353 hugepages are enabled.
355 .\" commit 99cb0dbd47a15d395bf3faa78dc122bc5efe3fc0
357 automatic scan of eligible areas and replacement by huge pages works with
358 private anonymous pages (see
361 and file-backed pages.
362 For all memory types,
363 memory may only be replaced by huge pages on hugepage-aligned boundaries.
364 For file-mapped memory
365 \[em]including tmpfs (see
367 the mapping must also be naturally hugepage-aligned within the file.
371 the file must not be open for write and the mapping must be executable.
373 The VMA must not be marked
381 nor can it be stack memory or backed by a DAX-enabled device
382 (unless the DAX device is hot-plugged as System RAM).
383 The process must also not have
384 .B PR_SET_THP_DISABLE
390 .BR MADV_NOHUGEPAGE ,
393 operations are available only if the kernel was configured with
394 .B CONFIG_TRANSPARENT_HUGEPAGE
395 and file/shmem memory is only supported if the kernel was configured with
396 .BR CONFIG_READ_ONLY_THP_FOR_FS .
398 .BR MADV_NOHUGEPAGE " (since Linux 2.6.38)"
399 Ensures that memory in the address range specified by
403 will not be backed by transparent hugepages.
405 .BR MADV_COLLAPSE " (since Linux 6.1)"
406 .\" commit 7d8faaf155454f8798ec56404faca29a82689c77
407 .\" commit 34488399fa08faaf664743fa54b271eb6f9e1321
408 Perform a best-effort synchronous collapse of
409 the native pages mapped by the memory range
410 into Transparent Huge Pages (THPs).
412 operates on the current state of memory of the calling process and
413 makes no persistent changes or guarantees on how pages will be mapped,
415 or faulted in the future.
418 supports private anonymous pages (see
421 and file-backed pages.
424 for general information on memory requirements for THP.
425 If the range provided spans multiple VMAs,
426 the semantics of the collapse over each VMA is independent from the others.
427 If collapse of a given huge page-aligned/sized region fails,
428 the operation may continue to attempt collapsing
429 the remainder of the specified memory.
431 will automatically clamp the provided range to be hugepage-aligned.
433 All non-resident pages covered by the range
434 will first be swapped/faulted-in,
435 before being copied onto a freshly allocated hugepage.
436 If the native pages compose the same PTE-mapped hugepage,
437 and are suitably aligned,
438 allocation of a new hugepage may be elided and
439 collapse may happen in-place.
440 Unmapped pages will have their data directly initialized to 0
443 for every eligible hugepage-aligned/sized region to be collapsed,
444 at least one page must currently be backed by physical memory.
447 is independent of any sysfs
451 .IR /sys/kernel/mm/transparent_hugepage ,
452 both in terms of determining THP eligibility,
453 and allocation semantics.
454 See Linux kernel source file
455 .I Documentation/admin\-guide/mm/transhuge.rst
456 for more information.
460 tmpfs mount when operating on tmpfs files.
461 Allocation for the new hugepage may enter direct reclaim and/or compaction,
462 regardless of VMA flags
467 When the system has multiple NUMA nodes,
468 the hugepage will be allocated from
469 the node providing the most native pages.
471 If all hugepage-sized/aligned regions covered by the provided range were
472 either successfully collapsed,
473 or were already PMD-mapped THPs,
474 this operation will be deemed successful.
475 Note that this doesn't guarantee anything about
476 other possible mappings of the memory.
477 In the event multiple hugepage-aligned/sized areas fail to collapse,
478 only the most-recently\[en]failed code will be set in
481 .BR MADV_DONTDUMP " (since Linux 3.4)"
482 .\" commit 909af768e88867016f427264ae39d27a57b6a8ed
483 .\" commit accb61fe7bb0f5c2a4102239e4981650f9048519
484 Exclude from a core dump those pages in the range specified by
488 This is useful in applications that have large areas of memory
489 that are known not to be useful in a core dump.
492 takes precedence over the bit mask that is set via the
493 .IR /proc/ pid /coredump_filter
497 .BR MADV_DODUMP " (since Linux 3.4)"
498 Undo the effect of an earlier
501 .BR MADV_FREE " (since Linux 4.5)"
502 The application no longer requires the pages in the range specified by
506 The kernel can thus free these pages,
507 but the freeing could be delayed until memory pressure occurs.
508 For each of the pages that has been marked to be freed
509 but has not yet been freed,
510 the free operation will be canceled if the caller writes into the page.
513 operation, any stale data (i.e., dirty, unwritten pages) will be lost
514 when the kernel frees the pages.
515 However, subsequent writes to pages in the range will succeed
516 and then kernel cannot free those dirtied pages,
517 so that the caller can always see just written data.
518 If there is no subsequent write,
519 the kernel can free the pages at any time.
520 Once pages in the range have been freed, the caller will
521 see zero-fill-on-demand pages upon subsequent page references.
526 can be applied only to private anonymous pages (see
529 .\" commit 93e06c7a645343d222c9a838834a51042eebbbf7
530 when freeing pages on a swapless system,
531 the pages in the given range are freed instantly,
532 regardless of memory pressure.
534 .BR MADV_WIPEONFORK " (since Linux 4.14)"
535 .\" commit d2cd9ede6e193dd7d88b6d27399e96229a551b19
536 Present the child process with zero-filled memory in this range after a
538 This is useful in forking servers in order to ensure
539 that sensitive per-process data
540 (for example, PRNG seeds, cryptographic secrets, and so on)
541 is not handed to child processes.
545 operation can be applied only to private anonymous pages (see
548 Within the child created by
552 setting remains in place on the specified address range.
553 This setting is cleared during
556 .BR MADV_KEEPONFORK " (since Linux 4.14)"
557 .\" commit d2cd9ede6e193dd7d88b6d27399e96229a551b19
558 Undo the effect of an earlier
559 .BR MADV_WIPEONFORK .
561 .BR MADV_COLD " (since Linux 5.4)"
562 .\" commit 9c276cc65a58faf98be8e56962745ec99ab87636
563 Deactivate a given range of pages.
564 This will make the pages a more probable
565 reclaim target should there be a memory pressure.
566 This is a nondestructive operation.
567 The advice might be ignored for some pages in the range when it is not
570 .BR MADV_PAGEOUT " (since Linux 5.4)"
571 .\" commit 1a4e58cce84ee88129d5d49c064bd2852b481357
572 Reclaim a given range of pages.
573 This is done to free up memory occupied by these pages.
574 If a page is anonymous, it will be swapped out.
575 If a page is file-backed and dirty, it will be written back to the backing
577 The advice might be ignored for some pages in the range when it is not
580 .BR MADV_POPULATE_READ " (since Linux 5.14)"
581 "Populate (prefault) page tables readable,
582 faulting in all pages in the range just as if manually reading from each page;
584 avoid the actual memory access that would have been performed after handling
589 .B MADV_POPULATE_READ
590 does not hide errors,
591 can be applied to (parts of) existing mappings and will always populate
592 (prefault) page tables readable.
593 One example use case is prefaulting a file mapping,
594 reading all file content from disk;
596 pages won't be dirtied and consequently won't have to be written back to disk
597 when evicting the pages from memory.
599 Depending on the underlying mapping,
600 map the shared zeropage,
601 preallocate memory or read the underlying file;
602 files with holes might or might not preallocate blocks.
606 signal is not generated; instead, an error is returned.
609 .B MADV_POPULATE_READ
611 all page tables have been populated (prefaulted) readable once.
613 .B MADV_POPULATE_READ
615 some page tables might have been populated.
617 .B MADV_POPULATE_READ
618 cannot be applied to mappings without read permissions
619 and special mappings,
621 mappings marked with kernel-internal flags such as
625 or secret memory regions created using
626 .BR memfd_secret(2) .
629 .BR MADV_POPULATE_READ ,
630 the process can be killed at any moment when the system runs out of memory.
632 .BR MADV_POPULATE_WRITE " (since Linux 5.14)"
633 Populate (prefault) page tables writable,
634 faulting in all pages in the range just as if manually writing to each
637 avoid the actual memory access that would have been performed after handling
642 MADV_POPULATE_WRITE does not hide errors,
643 can be applied to (parts of) existing mappings and will always populate
644 (prefault) page tables writable.
645 One example use case is preallocating memory,
646 breaking any CoW (Copy on Write).
648 Depending on the underlying mapping,
649 preallocate memory or read the underlying file;
650 files with holes will preallocate blocks.
654 signal is not generated; instead, an error is returned.
657 .B MADV_POPULATE_WRITE
659 all page tables have been populated (prefaulted) writable once.
661 .B MADV_POPULATE_WRITE
663 some page tables might have been populated.
665 .B MADV_POPULATE_WRITE
666 cannot be applied to mappings without write permissions
667 and special mappings,
669 mappings marked with kernel-internal flags such as
673 or secret memory regions created using
674 .BR memfd_secret(2) .
677 .BR MADV_POPULATE_WRITE ,
678 the process can be killed at any moment when the system runs out of memory.
683 On error, it returns \-1 and
685 is set to indicate the error.
692 but the specified address range is not a shared writable mapping.
695 A kernel resource was temporarily unavailable.
698 The map exists, but the area maps something that isn't a file.
703 Could not charge hugepage to cgroup: cgroup limit exceeded.
708 .B MADV_POPULATE_READ
710 .BR MADV_POPULATE_WRITE ,
711 and populating (prefaulting) page tables failed because a
713 would have been generated on actual memory access and the reason is not a
715 (HW poisoned pages can,
719 flag described elsewhere in this page).
723 is not page-aligned or
739 and the specified address range includes locked, Huge TLB pages, or
749 and the specified address range includes locked, Huge TLB pages, or
758 .BR MADV_UNMERGEABLE ,
759 but the kernel was not configured with
768 but the specified address range includes file, Huge TLB,
777 .B MADV_POPULATE_READ
779 .BR MADV_POPULATE_WRITE ,
780 but the specified address range includes ranges with insufficient permissions
783 mappings marked with kernel-internal flags such a
787 or secret memory regions created using
788 .BR memfd_secret(2) .
793 Paging in this area would exceed the process's
794 maximum resident set size.
799 Not enough memory: paging in failed.
804 Not enough memory: could not allocate hugepage.
807 Addresses in the specified range are not currently
808 mapped, or are outside the address space of the process.
813 .B MADV_POPULATE_READ
815 .BR MADV_POPULATE_WRITE ,
816 and populating (prefaulting) page tables failed because there was not enough
823 but the caller does not have the
830 .B MADV_POPULATE_READ
832 .BR MADV_POPULATE_WRITE ,
833 and populating (prefaulting) page tables failed because a HW poisoned page
834 (HW poisoned pages can,
838 flag described elsewhere in this page)
841 Versions of this system call, implementing a wide variety of
843 values, exist on many other implementations.
844 Other implementations typically implement at least the flags listed
846 .IR "Conventional advice flags" ,
847 albeit with some variation in semantics.
849 POSIX.1-2001 describes
850 .BR posix_madvise (3)
852 .BR POSIX_MADV_NORMAL ,
853 .BR POSIX_MADV_RANDOM ,
854 .BR POSIX_MADV_SEQUENTIAL ,
855 .BR POSIX_MADV_WILLNEED ,
857 .BR POSIX_MADV_DONTNEED ,
858 and so on, with behavior close to the similarly named flags listed above.
860 The Linux implementation requires that the address
862 be page-aligned, and allows
865 If there are some parts of the specified address range
866 that are not mapped, the Linux version of
868 ignores them and applies the call to the rest (but returns
870 from the system call, as it should).
872 .I madvise(0,\ 0,\ advice)
875 is supported by the kernel and can be relied on to probe for support.
879 First appeared in 4.4BSD.
882 .\" commit d3ac21cacc24790eb45d735769f35753f5b56ceb
883 support for this system call is optional,
884 depending on the setting of the
885 .B CONFIG_ADVISE_SYSCALLS
886 configuration option.
889 .BR memfd_secret (2),
896 .BR process_madvise (2),
897 .BR posix_madvise (3),