linux/Documentation/vm/unevictable-lru.txt
<<
>>
Prefs
   1                        ==============================
   2                        UNEVICTABLE LRU INFRASTRUCTURE
   3                        ==============================
   4
   5========
   6CONTENTS
   7========
   8
   9 (*) The Unevictable LRU
  10
  11     - The unevictable page list.
  12     - Memory control group interaction.
  13     - Marking address spaces unevictable.
  14     - Detecting Unevictable Pages.
  15     - vmscan's handling of unevictable pages.
  16
  17 (*) mlock()'d pages.
  18
  19     - History.
  20     - Basic management.
  21     - mlock()/mlockall() system call handling.
  22     - Filtering special vmas.
  23     - munlock()/munlockall() system call handling.
  24     - Migrating mlocked pages.
  25     - mmap(MAP_LOCKED) system call handling.
  26     - munmap()/exit()/exec() system call handling.
  27     - try_to_unmap().
  28     - try_to_munlock() reverse map scan.
  29     - Page reclaim in shrink_*_list().
  30
  31
  32============
  33INTRODUCTION
  34============
  35
  36This document describes the Linux memory manager's "Unevictable LRU"
  37infrastructure and the use of this to manage several types of "unevictable"
  38pages.
  39
  40The document attempts to provide the overall rationale behind this mechanism
  41and the rationale for some of the design decisions that drove the
  42implementation.  The latter design rationale is discussed in the context of an
  43implementation description.  Admittedly, one can obtain the implementation
  44details - the "what does it do?" - by reading the code.  One hopes that the
  45descriptions below add value by provide the answer to "why does it do that?".
  46
  47
  48===================
  49THE UNEVICTABLE LRU
  50===================
  51
  52The Unevictable LRU facility adds an additional LRU list to track unevictable
  53pages and to hide these pages from vmscan.  This mechanism is based on a patch
  54by Larry Woodman of Red Hat to address several scalability problems with page
  55reclaim in Linux.  The problems have been observed at customer sites on large
  56memory x86_64 systems.
  57
  58To illustrate this with an example, a non-NUMA x86_64 platform with 128GB of
  59main memory will have over 32 million 4k pages in a single zone.  When a large
  60fraction of these pages are not evictable for any reason [see below], vmscan
  61will spend a lot of time scanning the LRU lists looking for the small fraction
  62of pages that are evictable.  This can result in a situation where all CPUs are
  63spending 100% of their time in vmscan for hours or days on end, with the system
  64completely unresponsive.
  65
  66The unevictable list addresses the following classes of unevictable pages:
  67
  68 (*) Those owned by ramfs.
  69
  70 (*) Those mapped into SHM_LOCK'd shared memory regions.
  71
  72 (*) Those mapped into VM_LOCKED [mlock()ed] VMAs.
  73
  74The infrastructure may also be able to handle other conditions that make pages
  75unevictable, either by definition or by circumstance, in the future.
  76
  77
  78THE UNEVICTABLE PAGE LIST
  79-------------------------
  80
  81The Unevictable LRU infrastructure consists of an additional, per-zone, LRU list
  82called the "unevictable" list and an associated page flag, PG_unevictable, to
  83indicate that the page is being managed on the unevictable list.
  84
  85The PG_unevictable flag is analogous to, and mutually exclusive with, the
  86PG_active flag in that it indicates on which LRU list a page resides when
  87PG_lru is set.
  88
  89The Unevictable LRU infrastructure maintains unevictable pages on an additional
  90LRU list for a few reasons:
  91
  92 (1) We get to "treat unevictable pages just like we treat other pages in the
  93     system - which means we get to use the same code to manipulate them, the
  94     same code to isolate them (for migrate, etc.), the same code to keep track
  95     of the statistics, etc..." [Rik van Riel]
  96
  97 (2) We want to be able to migrate unevictable pages between nodes for memory
  98     defragmentation, workload management and memory hotplug.  The linux kernel
  99     can only migrate pages that it can successfully isolate from the LRU
 100     lists.  If we were to maintain pages elsewhere than on an LRU-like list,
 101     where they can be found by isolate_lru_page(), we would prevent their
 102     migration, unless we reworked migration code to find the unevictable pages
 103     itself.
 104
 105
 106The unevictable list does not differentiate between file-backed and anonymous,
 107swap-backed pages.  This differentiation is only important while the pages are,
 108in fact, evictable.
 109
 110The unevictable list benefits from the "arrayification" of the per-zone LRU
 111lists and statistics originally proposed and posted by Christoph Lameter.
 112
 113The unevictable list does not use the LRU pagevec mechanism. Rather,
 114unevictable pages are placed directly on the page's zone's unevictable list
 115under the zone lru_lock.  This allows us to prevent the stranding of pages on
 116the unevictable list when one task has the page isolated from the LRU and other
 117tasks are changing the "evictability" state of the page.
 118
 119
 120MEMORY CONTROL GROUP INTERACTION
 121--------------------------------
 122
 123The unevictable LRU facility interacts with the memory control group [aka
 124memory controller; see Documentation/cgroups/memory.txt] by extending the
 125lru_list enum.
 126
 127The memory controller data structure automatically gets a per-zone unevictable
 128list as a result of the "arrayification" of the per-zone LRU lists (one per
 129lru_list enum element).  The memory controller tracks the movement of pages to
 130and from the unevictable list.
 131
 132When a memory control group comes under memory pressure, the controller will
 133not attempt to reclaim pages on the unevictable list.  This has a couple of
 134effects:
 135
 136 (1) Because the pages are "hidden" from reclaim on the unevictable list, the
 137     reclaim process can be more efficient, dealing only with pages that have a
 138     chance of being reclaimed.
 139
 140 (2) On the other hand, if too many of the pages charged to the control group
 141     are unevictable, the evictable portion of the working set of the tasks in
 142     the control group may not fit into the available memory.  This can cause
 143     the control group to thrash or to OOM-kill tasks.
 144
 145
 146MARKING ADDRESS SPACES UNEVICTABLE
 147----------------------------------
 148
 149For facilities such as ramfs none of the pages attached to the address space
 150may be evicted.  To prevent eviction of any such pages, the AS_UNEVICTABLE
 151address space flag is provided, and this can be manipulated by a filesystem
 152using a number of wrapper functions:
 153
 154 (*) void mapping_set_unevictable(struct address_space *mapping);
 155
 156        Mark the address space as being completely unevictable.
 157
 158 (*) void mapping_clear_unevictable(struct address_space *mapping);
 159
 160        Mark the address space as being evictable.
 161
 162 (*) int mapping_unevictable(struct address_space *mapping);
 163
 164        Query the address space, and return true if it is completely
 165        unevictable.
 166
 167These are currently used in two places in the kernel:
 168
 169 (1) By ramfs to mark the address spaces of its inodes when they are created,
 170     and this mark remains for the life of the inode.
 171
 172 (2) By SYSV SHM to mark SHM_LOCK'd address spaces until SHM_UNLOCK is called.
 173
 174     Note that SHM_LOCK is not required to page in the locked pages if they're
 175     swapped out; the application must touch the pages manually if it wants to
 176     ensure they're in memory.
 177
 178
 179DETECTING UNEVICTABLE PAGES
 180---------------------------
 181
 182The function page_evictable() in vmscan.c determines whether a page is
 183evictable or not using the query function outlined above [see section "Marking
 184address spaces unevictable"] to check the AS_UNEVICTABLE flag.
 185
 186For address spaces that are so marked after being populated (as SHM regions
 187might be), the lock action (eg: SHM_LOCK) can be lazy, and need not populate
 188the page tables for the region as does, for example, mlock(), nor need it make
 189any special effort to push any pages in the SHM_LOCK'd area to the unevictable
 190list.  Instead, vmscan will do this if and when it encounters the pages during
 191a reclamation scan.
 192
 193On an unlock action (such as SHM_UNLOCK), the unlocker (eg: shmctl()) must scan
 194the pages in the region and "rescue" them from the unevictable list if no other
 195condition is keeping them unevictable.  If an unevictable region is destroyed,
 196the pages are also "rescued" from the unevictable list in the process of
 197freeing them.
 198
 199page_evictable() also checks for mlocked pages by testing an additional page
 200flag, PG_mlocked (as wrapped by PageMlocked()).  If the page is NOT mlocked,
 201and a non-NULL VMA is supplied, page_evictable() will check whether the VMA is
 202VM_LOCKED via is_mlocked_vma().  is_mlocked_vma() will SetPageMlocked() and
 203update the appropriate statistics if the vma is VM_LOCKED.  This method allows
 204efficient "culling" of pages in the fault path that are being faulted in to
 205VM_LOCKED VMAs.
 206
 207
 208VMSCAN'S HANDLING OF UNEVICTABLE PAGES
 209--------------------------------------
 210
 211If unevictable pages are culled in the fault path, or moved to the unevictable
 212list at mlock() or mmap() time, vmscan will not encounter the pages until they
 213have become evictable again (via munlock() for example) and have been "rescued"
 214from the unevictable list.  However, there may be situations where we decide,
 215for the sake of expediency, to leave a unevictable page on one of the regular
 216active/inactive LRU lists for vmscan to deal with.  vmscan checks for such
 217pages in all of the shrink_{active|inactive|page}_list() functions and will
 218"cull" such pages that it encounters: that is, it diverts those pages to the
 219unevictable list for the zone being scanned.
 220
 221There may be situations where a page is mapped into a VM_LOCKED VMA, but the
 222page is not marked as PG_mlocked.  Such pages will make it all the way to
 223shrink_page_list() where they will be detected when vmscan walks the reverse
 224map in try_to_unmap().  If try_to_unmap() returns SWAP_MLOCK,
 225shrink_page_list() will cull the page at that point.
 226
 227To "cull" an unevictable page, vmscan simply puts the page back on the LRU list
 228using putback_lru_page() - the inverse operation to isolate_lru_page() - after
 229dropping the page lock.  Because the condition which makes the page unevictable
 230may change once the page is unlocked, putback_lru_page() will recheck the
 231unevictable state of a page that it places on the unevictable list.  If the
 232page has become unevictable, putback_lru_page() removes it from the list and
 233retries, including the page_unevictable() test.  Because such a race is a rare
 234event and movement of pages onto the unevictable list should be rare, these
 235extra evictabilty checks should not occur in the majority of calls to
 236putback_lru_page().
 237
 238
 239=============
 240MLOCKED PAGES
 241=============
 242
 243The unevictable page list is also useful for mlock(), in addition to ramfs and
 244SYSV SHM.  Note that mlock() is only available in CONFIG_MMU=y situations; in
 245NOMMU situations, all mappings are effectively mlocked.
 246
 247
 248HISTORY
 249-------
 250
 251The "Unevictable mlocked Pages" infrastructure is based on work originally
 252posted by Nick Piggin in an RFC patch entitled "mm: mlocked pages off LRU".
 253Nick posted his patch as an alternative to a patch posted by Christoph Lameter
 254to achieve the same objective: hiding mlocked pages from vmscan.
 255
 256In Nick's patch, he used one of the struct page LRU list link fields as a count
 257of VM_LOCKED VMAs that map the page.  This use of the link field for a count
 258prevented the management of the pages on an LRU list, and thus mlocked pages
 259were not migratable as isolate_lru_page() could not find them, and the LRU list
 260link field was not available to the migration subsystem.
 261
 262Nick resolved this by putting mlocked pages back on the lru list before
 263attempting to isolate them, thus abandoning the count of VM_LOCKED VMAs.  When
 264Nick's patch was integrated with the Unevictable LRU work, the count was
 265replaced by walking the reverse map to determine whether any VM_LOCKED VMAs
 266mapped the page.  More on this below.
 267
 268
 269BASIC MANAGEMENT
 270----------------
 271
 272mlocked pages - pages mapped into a VM_LOCKED VMA - are a class of unevictable
 273pages.  When such a page has been "noticed" by the memory management subsystem,
 274the page is marked with the PG_mlocked flag.  This can be manipulated using the
 275PageMlocked() functions.
 276
 277A PG_mlocked page will be placed on the unevictable list when it is added to
 278the LRU.  Such pages can be "noticed" by memory management in several places:
 279
 280 (1) in the mlock()/mlockall() system call handlers;
 281
 282 (2) in the mmap() system call handler when mmapping a region with the
 283     MAP_LOCKED flag;
 284
 285 (3) mmapping a region in a task that has called mlockall() with the MCL_FUTURE
 286     flag
 287
 288 (4) in the fault path, if mlocked pages are "culled" in the fault path,
 289     and when a VM_LOCKED stack segment is expanded; or
 290
 291 (5) as mentioned above, in vmscan:shrink_page_list() when attempting to
 292     reclaim a page in a VM_LOCKED VMA via try_to_unmap()
 293
 294all of which result in the VM_LOCKED flag being set for the VMA if it doesn't
 295already have it set.
 296
 297mlocked pages become unlocked and rescued from the unevictable list when:
 298
 299 (1) mapped in a range unlocked via the munlock()/munlockall() system calls;
 300
 301 (2) munmap()'d out of the last VM_LOCKED VMA that maps the page, including
 302     unmapping at task exit;
 303
 304 (3) when the page is truncated from the last VM_LOCKED VMA of an mmapped file;
 305     or
 306
 307 (4) before a page is COW'd in a VM_LOCKED VMA.
 308
 309
 310mlock()/mlockall() SYSTEM CALL HANDLING
 311---------------------------------------
 312
 313Both [do_]mlock() and [do_]mlockall() system call handlers call mlock_fixup()
 314for each VMA in the range specified by the call.  In the case of mlockall(),
 315this is the entire active address space of the task.  Note that mlock_fixup()
 316is used for both mlocking and munlocking a range of memory.  A call to mlock()
 317an already VM_LOCKED VMA, or to munlock() a VMA that is not VM_LOCKED is
 318treated as a no-op, and mlock_fixup() simply returns.
 319
 320If the VMA passes some filtering as described in "Filtering Special Vmas"
 321below, mlock_fixup() will attempt to merge the VMA with its neighbors or split
 322off a subset of the VMA if the range does not cover the entire VMA.  Once the
 323VMA has been merged or split or neither, mlock_fixup() will call
 324__mlock_vma_pages_range() to fault in the pages via get_user_pages() and to
 325mark the pages as mlocked via mlock_vma_page().
 326
 327Note that the VMA being mlocked might be mapped with PROT_NONE.  In this case,
 328get_user_pages() will be unable to fault in the pages.  That's okay.  If pages
 329do end up getting faulted into this VM_LOCKED VMA, we'll handle them in the
 330fault path or in vmscan.
 331
 332Also note that a page returned by get_user_pages() could be truncated or
 333migrated out from under us, while we're trying to mlock it.  To detect this,
 334__mlock_vma_pages_range() checks page_mapping() after acquiring the page lock.
 335If the page is still associated with its mapping, we'll go ahead and call
 336mlock_vma_page().  If the mapping is gone, we just unlock the page and move on.
 337In the worst case, this will result in a page mapped in a VM_LOCKED VMA
 338remaining on a normal LRU list without being PageMlocked().  Again, vmscan will
 339detect and cull such pages.
 340
 341mlock_vma_page() will call TestSetPageMlocked() for each page returned by
 342get_user_pages().  We use TestSetPageMlocked() because the page might already
 343be mlocked by another task/VMA and we don't want to do extra work.  We
 344especially do not want to count an mlocked page more than once in the
 345statistics.  If the page was already mlocked, mlock_vma_page() need do nothing
 346more.
 347
 348If the page was NOT already mlocked, mlock_vma_page() attempts to isolate the
 349page from the LRU, as it is likely on the appropriate active or inactive list
 350at that time.  If the isolate_lru_page() succeeds, mlock_vma_page() will put
 351back the page - by calling putback_lru_page() - which will notice that the page
 352is now mlocked and divert the page to the zone's unevictable list.  If
 353mlock_vma_page() is unable to isolate the page from the LRU, vmscan will handle
 354it later if and when it attempts to reclaim the page.
 355
 356
 357FILTERING SPECIAL VMAS
 358----------------------
 359
 360mlock_fixup() filters several classes of "special" VMAs:
 361
 3621) VMAs with VM_IO or VM_PFNMAP set are skipped entirely.  The pages behind
 363   these mappings are inherently pinned, so we don't need to mark them as
 364   mlocked.  In any case, most of the pages have no struct page in which to so
 365   mark the page.  Because of this, get_user_pages() will fail for these VMAs,
 366   so there is no sense in attempting to visit them.
 367
 3682) VMAs mapping hugetlbfs page are already effectively pinned into memory.  We
 369   neither need nor want to mlock() these pages.  However, to preserve the
 370   prior behavior of mlock() - before the unevictable/mlock changes -
 371   mlock_fixup() will call make_pages_present() in the hugetlbfs VMA range to
 372   allocate the huge pages and populate the ptes.
 373
 3743) VMAs with VM_DONTEXPAND or VM_RESERVED are generally userspace mappings of
 375   kernel pages, such as the VDSO page, relay channel pages, etc.  These pages
 376   are inherently unevictable and are not managed on the LRU lists.
 377   mlock_fixup() treats these VMAs the same as hugetlbfs VMAs.  It calls
 378   make_pages_present() to populate the ptes.
 379
 380Note that for all of these special VMAs, mlock_fixup() does not set the
 381VM_LOCKED flag.  Therefore, we won't have to deal with them later during
 382munlock(), munmap() or task exit.  Neither does mlock_fixup() account these
 383VMAs against the task's "locked_vm".
 384
 385
 386munlock()/munlockall() SYSTEM CALL HANDLING
 387-------------------------------------------
 388
 389The munlock() and munlockall() system calls are handled by the same functions -
 390do_mlock[all]() - as the mlock() and mlockall() system calls with the unlock vs
 391lock operation indicated by an argument.  So, these system calls are also
 392handled by mlock_fixup().  Again, if called for an already munlocked VMA,
 393mlock_fixup() simply returns.  Because of the VMA filtering discussed above,
 394VM_LOCKED will not be set in any "special" VMAs.  So, these VMAs will be
 395ignored for munlock.
 396
 397If the VMA is VM_LOCKED, mlock_fixup() again attempts to merge or split off the
 398specified range.  The range is then munlocked via the function
 399__mlock_vma_pages_range() - the same function used to mlock a VMA range -
 400passing a flag to indicate that munlock() is being performed.
 401
 402Because the VMA access protections could have been changed to PROT_NONE after
 403faulting in and mlocking pages, get_user_pages() was unreliable for visiting
 404these pages for munlocking.  Because we don't want to leave pages mlocked,
 405get_user_pages() was enhanced to accept a flag to ignore the permissions when
 406fetching the pages - all of which should be resident as a result of previous
 407mlocking.
 408
 409For munlock(), __mlock_vma_pages_range() unlocks individual pages by calling
 410munlock_vma_page().  munlock_vma_page() unconditionally clears the PG_mlocked
 411flag using TestClearPageMlocked().  As with mlock_vma_page(),
 412munlock_vma_page() use the Test*PageMlocked() function to handle the case where
 413the page might have already been unlocked by another task.  If the page was
 414mlocked, munlock_vma_page() updates that zone statistics for the number of
 415mlocked pages.  Note, however, that at this point we haven't checked whether
 416the page is mapped by other VM_LOCKED VMAs.
 417
 418We can't call try_to_munlock(), the function that walks the reverse map to
 419check for other VM_LOCKED VMAs, without first isolating the page from the LRU.
 420try_to_munlock() is a variant of try_to_unmap() and thus requires that the page
 421not be on an LRU list [more on these below].  However, the call to
 422isolate_lru_page() could fail, in which case we couldn't try_to_munlock().  So,
 423we go ahead and clear PG_mlocked up front, as this might be the only chance we
 424have.  If we can successfully isolate the page, we go ahead and
 425try_to_munlock(), which will restore the PG_mlocked flag and update the zone
 426page statistics if it finds another VMA holding the page mlocked.  If we fail
 427to isolate the page, we'll have left a potentially mlocked page on the LRU.
 428This is fine, because we'll catch it later if and if vmscan tries to reclaim
 429the page.  This should be relatively rare.
 430
 431
 432MIGRATING MLOCKED PAGES
 433-----------------------
 434
 435A page that is being migrated has been isolated from the LRU lists and is held
 436locked across unmapping of the page, updating the page's address space entry
 437and copying the contents and state, until the page table entry has been
 438replaced with an entry that refers to the new page.  Linux supports migration
 439of mlocked pages and other unevictable pages.  This involves simply moving the
 440PG_mlocked and PG_unevictable states from the old page to the new page.
 441
 442Note that page migration can race with mlocking or munlocking of the same page.
 443This has been discussed from the mlock/munlock perspective in the respective
 444sections above.  Both processes (migration and m[un]locking) hold the page
 445locked.  This provides the first level of synchronization.  Page migration
 446zeros out the page_mapping of the old page before unlocking it, so m[un]lock
 447can skip these pages by testing the page mapping under page lock.
 448
 449To complete page migration, we place the new and old pages back onto the LRU
 450after dropping the page lock.  The "unneeded" page - old page on success, new
 451page on failure - will be freed when the reference count held by the migration
 452process is released.  To ensure that we don't strand pages on the unevictable
 453list because of a race between munlock and migration, page migration uses the
 454putback_lru_page() function to add migrated pages back to the LRU.
 455
 456
 457mmap(MAP_LOCKED) SYSTEM CALL HANDLING
 458-------------------------------------
 459
 460In addition the the mlock()/mlockall() system calls, an application can request
 461that a region of memory be mlocked supplying the MAP_LOCKED flag to the mmap()
 462call.  Furthermore, any mmap() call or brk() call that expands the heap by a
 463task that has previously called mlockall() with the MCL_FUTURE flag will result
 464in the newly mapped memory being mlocked.  Before the unevictable/mlock
 465changes, the kernel simply called make_pages_present() to allocate pages and
 466populate the page table.
 467
 468To mlock a range of memory under the unevictable/mlock infrastructure, the
 469mmap() handler and task address space expansion functions call
 470mlock_vma_pages_range() specifying the vma and the address range to mlock.
 471mlock_vma_pages_range() filters VMAs like mlock_fixup(), as described above in
 472"Filtering Special VMAs".  It will clear the VM_LOCKED flag, which will have
 473already been set by the caller, in filtered VMAs.  Thus these VMA's need not be
 474visited for munlock when the region is unmapped.
 475
 476For "normal" VMAs, mlock_vma_pages_range() calls __mlock_vma_pages_range() to
 477fault/allocate the pages and mlock them.  Again, like mlock_fixup(),
 478mlock_vma_pages_range() downgrades the mmap semaphore to read mode before
 479attempting to fault/allocate and mlock the pages and "upgrades" the semaphore
 480back to write mode before returning.
 481
 482The callers of mlock_vma_pages_range() will have already added the memory range
 483to be mlocked to the task's "locked_vm".  To account for filtered VMAs,
 484mlock_vma_pages_range() returns the number of pages NOT mlocked.  All of the
 485callers then subtract a non-negative return value from the task's locked_vm.  A
 486negative return value represent an error - for example, from get_user_pages()
 487attempting to fault in a VMA with PROT_NONE access.  In this case, we leave the
 488memory range accounted as locked_vm, as the protections could be changed later
 489and pages allocated into that region.
 490
 491
 492munmap()/exit()/exec() SYSTEM CALL HANDLING
 493-------------------------------------------
 494
 495When unmapping an mlocked region of memory, whether by an explicit call to
 496munmap() or via an internal unmap from exit() or exec() processing, we must
 497munlock the pages if we're removing the last VM_LOCKED VMA that maps the pages.
 498Before the unevictable/mlock changes, mlocking did not mark the pages in any
 499way, so unmapping them required no processing.
 500
 501To munlock a range of memory under the unevictable/mlock infrastructure, the
 502munmap() handler and task address space call tear down function
 503munlock_vma_pages_all().  The name reflects the observation that one always
 504specifies the entire VMA range when munlock()ing during unmap of a region.
 505Because of the VMA filtering when mlocking() regions, only "normal" VMAs that
 506actually contain mlocked pages will be passed to munlock_vma_pages_all().
 507
 508munlock_vma_pages_all() clears the VM_LOCKED VMA flag and, like mlock_fixup()
 509for the munlock case, calls __munlock_vma_pages_range() to walk the page table
 510for the VMA's memory range and munlock_vma_page() each resident page mapped by
 511the VMA.  This effectively munlocks the page, only if this is the last
 512VM_LOCKED VMA that maps the page.
 513
 514
 515try_to_unmap()
 516--------------
 517
 518Pages can, of course, be mapped into multiple VMAs.  Some of these VMAs may
 519have VM_LOCKED flag set.  It is possible for a page mapped into one or more
 520VM_LOCKED VMAs not to have the PG_mlocked flag set and therefore reside on one
 521of the active or inactive LRU lists.  This could happen if, for example, a task
 522in the process of munlocking the page could not isolate the page from the LRU.
 523As a result, vmscan/shrink_page_list() might encounter such a page as described
 524in section "vmscan's handling of unevictable pages".  To handle this situation,
 525try_to_unmap() checks for VM_LOCKED VMAs while it is walking a page's reverse
 526map.
 527
 528try_to_unmap() is always called, by either vmscan for reclaim or for page
 529migration, with the argument page locked and isolated from the LRU.  Separate
 530functions handle anonymous and mapped file pages, as these types of pages have
 531different reverse map mechanisms.
 532
 533 (*) try_to_unmap_anon()
 534
 535     To unmap anonymous pages, each VMA in the list anchored in the anon_vma
 536     must be visited - at least until a VM_LOCKED VMA is encountered.  If the
 537     page is being unmapped for migration, VM_LOCKED VMAs do not stop the
 538     process because mlocked pages are migratable.  However, for reclaim, if
 539     the page is mapped into a VM_LOCKED VMA, the scan stops.
 540
 541     try_to_unmap_anon() attempts to acquire in read mode the mmap semaphore of
 542     the mm_struct to which the VMA belongs.  If this is successful, it will
 543     mlock the page via mlock_vma_page() - we wouldn't have gotten to
 544     try_to_unmap_anon() if the page were already mlocked - and will return
 545     SWAP_MLOCK, indicating that the page is unevictable.
 546
 547     If the mmap semaphore cannot be acquired, we are not sure whether the page
 548     is really unevictable or not.  In this case, try_to_unmap_anon() will
 549     return SWAP_AGAIN.
 550
 551 (*) try_to_unmap_file() - linear mappings
 552
 553     Unmapping of a mapped file page works the same as for anonymous mappings,
 554     except that the scan visits all VMAs that map the page's index/page offset
 555     in the page's mapping's reverse map priority search tree.  It also visits
 556     each VMA in the page's mapping's non-linear list, if the list is
 557     non-empty.
 558
 559     As for anonymous pages, on encountering a VM_LOCKED VMA for a mapped file
 560     page, try_to_unmap_file() will attempt to acquire the associated
 561     mm_struct's mmap semaphore to mlock the page, returning SWAP_MLOCK if this
 562     is successful, and SWAP_AGAIN, if not.
 563
 564 (*) try_to_unmap_file() - non-linear mappings
 565
 566     If a page's mapping contains a non-empty non-linear mapping VMA list, then
 567     try_to_un{map|lock}() must also visit each VMA in that list to determine
 568     whether the page is mapped in a VM_LOCKED VMA.  Again, the scan must visit
 569     all VMAs in the non-linear list to ensure that the pages is not/should not
 570     be mlocked.
 571
 572     If a VM_LOCKED VMA is found in the list, the scan could terminate.
 573     However, there is no easy way to determine whether the page is actually
 574     mapped in a given VMA - either for unmapping or testing whether the
 575     VM_LOCKED VMA actually pins the page.
 576
 577     try_to_unmap_file() handles non-linear mappings by scanning a certain
 578     number of pages - a "cluster" - in each non-linear VMA associated with the
 579     page's mapping, for each file mapped page that vmscan tries to unmap.  If
 580     this happens to unmap the page we're trying to unmap, try_to_unmap() will
 581     notice this on return (page_mapcount(page) will be 0) and return
 582     SWAP_SUCCESS.  Otherwise, it will return SWAP_AGAIN, causing vmscan to
 583     recirculate this page.  We take advantage of the cluster scan in
 584     try_to_unmap_cluster() as follows:
 585
 586        For each non-linear VMA, try_to_unmap_cluster() attempts to acquire the
 587        mmap semaphore of the associated mm_struct for read without blocking.
 588
 589        If this attempt is successful and the VMA is VM_LOCKED,
 590        try_to_unmap_cluster() will retain the mmap semaphore for the scan;
 591        otherwise it drops it here.
 592
 593        Then, for each page in the cluster, if we're holding the mmap semaphore
 594        for a locked VMA, try_to_unmap_cluster() calls mlock_vma_page() to
 595        mlock the page.  This call is a no-op if the page is already locked,
 596        but will mlock any pages in the non-linear mapping that happen to be
 597        unlocked.
 598
 599        If one of the pages so mlocked is the page passed in to try_to_unmap(),
 600        try_to_unmap_cluster() will return SWAP_MLOCK, rather than the default
 601        SWAP_AGAIN.  This will allow vmscan to cull the page, rather than
 602        recirculating it on the inactive list.
 603
 604        Again, if try_to_unmap_cluster() cannot acquire the VMA's mmap sem, it
 605        returns SWAP_AGAIN, indicating that the page is mapped by a VM_LOCKED
 606        VMA, but couldn't be mlocked.
 607
 608
 609try_to_munlock() REVERSE MAP SCAN
 610---------------------------------
 611
 612 [!] TODO/FIXME: a better name might be page_mlocked() - analogous to the
 613     page_referenced() reverse map walker.
 614
 615When munlock_vma_page() [see section "munlock()/munlockall() System Call
 616Handling" above] tries to munlock a page, it needs to determine whether or not
 617the page is mapped by any VM_LOCKED VMA without actually attempting to unmap
 618all PTEs from the page.  For this purpose, the unevictable/mlock infrastructure
 619introduced a variant of try_to_unmap() called try_to_munlock().
 620
 621try_to_munlock() calls the same functions as try_to_unmap() for anonymous and
 622mapped file pages with an additional argument specifying unlock versus unmap
 623processing.  Again, these functions walk the respective reverse maps looking
 624for VM_LOCKED VMAs.  When such a VMA is found for anonymous pages and file
 625pages mapped in linear VMAs, as in the try_to_unmap() case, the functions
 626attempt to acquire the associated mmap semaphore, mlock the page via
 627mlock_vma_page() and return SWAP_MLOCK.  This effectively undoes the
 628pre-clearing of the page's PG_mlocked done by munlock_vma_page.
 629
 630If try_to_unmap() is unable to acquire a VM_LOCKED VMA's associated mmap
 631semaphore, it will return SWAP_AGAIN.  This will allow shrink_page_list() to
 632recycle the page on the inactive list and hope that it has better luck with the
 633page next time.
 634
 635For file pages mapped into non-linear VMAs, the try_to_munlock() logic works
 636slightly differently.  On encountering a VM_LOCKED non-linear VMA that might
 637map the page, try_to_munlock() returns SWAP_AGAIN without actually mlocking the
 638page.  munlock_vma_page() will just leave the page unlocked and let vmscan deal
 639with it - the usual fallback position.
 640
 641Note that try_to_munlock()'s reverse map walk must visit every VMA in a page's
 642reverse map to determine that a page is NOT mapped into any VM_LOCKED VMA.
 643However, the scan can terminate when it encounters a VM_LOCKED VMA and can
 644successfully acquire the VMA's mmap semaphore for read and mlock the page.
 645Although try_to_munlock() might be called a great many times when munlocking a
 646large region or tearing down a large address space that has been mlocked via
 647mlockall(), overall this is a fairly rare event.
 648
 649
 650PAGE RECLAIM IN shrink_*_list()
 651-------------------------------
 652
 653shrink_active_list() culls any obviously unevictable pages - i.e.
 654!page_evictable(page, NULL) - diverting these to the unevictable list.
 655However, shrink_active_list() only sees unevictable pages that made it onto the
 656active/inactive lru lists.  Note that these pages do not have PageUnevictable
 657set - otherwise they would be on the unevictable list and shrink_active_list
 658would never see them.
 659
 660Some examples of these unevictable pages on the LRU lists are:
 661
 662 (1) ramfs pages that have been placed on the LRU lists when first allocated.
 663
 664 (2) SHM_LOCK'd shared memory pages.  shmctl(SHM_LOCK) does not attempt to
 665     allocate or fault in the pages in the shared memory region.  This happens
 666     when an application accesses the page the first time after SHM_LOCK'ing
 667     the segment.
 668
 669 (3) mlocked pages that could not be isolated from the LRU and moved to the
 670     unevictable list in mlock_vma_page().
 671
 672 (4) Pages mapped into multiple VM_LOCKED VMAs, but try_to_munlock() couldn't
 673     acquire the VMA's mmap semaphore to test the flags and set PageMlocked.
 674     munlock_vma_page() was forced to let the page back on to the normal LRU
 675     list for vmscan to handle.
 676
 677shrink_inactive_list() also diverts any unevictable pages that it finds on the
 678inactive lists to the appropriate zone's unevictable list.
 679
 680shrink_inactive_list() should only see SHM_LOCK'd pages that became SHM_LOCK'd
 681after shrink_active_list() had moved them to the inactive list, or pages mapped
 682into VM_LOCKED VMAs that munlock_vma_page() couldn't isolate from the LRU to
 683recheck via try_to_munlock().  shrink_inactive_list() won't notice the latter,
 684but will pass on to shrink_page_list().
 685
 686shrink_page_list() again culls obviously unevictable pages that it could
 687encounter for similar reason to shrink_inactive_list().  Pages mapped into
 688VM_LOCKED VMAs but without PG_mlocked set will make it all the way to
 689try_to_unmap().  shrink_page_list() will divert them to the unevictable list
 690when try_to_unmap() returns SWAP_MLOCK, as discussed above.
 691
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.