linux/mm/mlock.c
<<
>>
Prefs
   1/*
   2 *      linux/mm/mlock.c
   3 *
   4 *  (C) Copyright 1995 Linus Torvalds
   5 *  (C) Copyright 2002 Christoph Hellwig
   6 */
   7
   8#include <linux/capability.h>
   9#include <linux/mman.h>
  10#include <linux/mm.h>
  11#include <linux/swap.h>
  12#include <linux/swapops.h>
  13#include <linux/pagemap.h>
  14#include <linux/mempolicy.h>
  15#include <linux/syscalls.h>
  16#include <linux/sched.h>
  17#include <linux/module.h>
  18#include <linux/rmap.h>
  19#include <linux/mmzone.h>
  20#include <linux/hugetlb.h>
  21
  22#include "internal.h"
  23
  24int can_do_mlock(void)
  25{
  26        if (capable(CAP_IPC_LOCK))
  27                return 1;
  28        if (current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur != 0)
  29                return 1;
  30        return 0;
  31}
  32EXPORT_SYMBOL(can_do_mlock);
  33
  34#ifdef CONFIG_UNEVICTABLE_LRU
  35/*
  36 * Mlocked pages are marked with PageMlocked() flag for efficient testing
  37 * in vmscan and, possibly, the fault path; and to support semi-accurate
  38 * statistics.
  39 *
  40 * An mlocked page [PageMlocked(page)] is unevictable.  As such, it will
  41 * be placed on the LRU "unevictable" list, rather than the [in]active lists.
  42 * The unevictable list is an LRU sibling list to the [in]active lists.
  43 * PageUnevictable is set to indicate the unevictable state.
  44 *
  45 * When lazy mlocking via vmscan, it is important to ensure that the
  46 * vma's VM_LOCKED status is not concurrently being modified, otherwise we
  47 * may have mlocked a page that is being munlocked. So lazy mlock must take
  48 * the mmap_sem for read, and verify that the vma really is locked
  49 * (see mm/rmap.c).
  50 */
  51
  52/*
  53 *  LRU accounting for clear_page_mlock()
  54 */
  55void __clear_page_mlock(struct page *page)
  56{
  57        VM_BUG_ON(!PageLocked(page));
  58
  59        if (!page->mapping) {   /* truncated ? */
  60                return;
  61        }
  62
  63        dec_zone_page_state(page, NR_MLOCK);
  64        count_vm_event(UNEVICTABLE_PGCLEARED);
  65        if (!isolate_lru_page(page)) {
  66                putback_lru_page(page);
  67        } else {
  68                /*
  69                 * We lost the race. the page already moved to evictable list.
  70                 */
  71                if (PageUnevictable(page))
  72                        count_vm_event(UNEVICTABLE_PGSTRANDED);
  73        }
  74}
  75
  76/*
  77 * Mark page as mlocked if not already.
  78 * If page on LRU, isolate and putback to move to unevictable list.
  79 */
  80void mlock_vma_page(struct page *page)
  81{
  82        BUG_ON(!PageLocked(page));
  83
  84        if (!TestSetPageMlocked(page)) {
  85                inc_zone_page_state(page, NR_MLOCK);
  86                count_vm_event(UNEVICTABLE_PGMLOCKED);
  87                if (!isolate_lru_page(page))
  88                        putback_lru_page(page);
  89        }
  90}
  91
  92/*
  93 * called from munlock()/munmap() path with page supposedly on the LRU.
  94 *
  95 * Note:  unlike mlock_vma_page(), we can't just clear the PageMlocked
  96 * [in try_to_munlock()] and then attempt to isolate the page.  We must
  97 * isolate the page to keep others from messing with its unevictable
  98 * and mlocked state while trying to munlock.  However, we pre-clear the
  99 * mlocked state anyway as we might lose the isolation race and we might
 100 * not get another chance to clear PageMlocked.  If we successfully
 101 * isolate the page and try_to_munlock() detects other VM_LOCKED vmas
 102 * mapping the page, it will restore the PageMlocked state, unless the page
 103 * is mapped in a non-linear vma.  So, we go ahead and SetPageMlocked(),
 104 * perhaps redundantly.
 105 * If we lose the isolation race, and the page is mapped by other VM_LOCKED
 106 * vmas, we'll detect this in vmscan--via try_to_munlock() or try_to_unmap()
 107 * either of which will restore the PageMlocked state by calling
 108 * mlock_vma_page() above, if it can grab the vma's mmap sem.
 109 */
 110static void munlock_vma_page(struct page *page)
 111{
 112        BUG_ON(!PageLocked(page));
 113
 114        if (TestClearPageMlocked(page)) {
 115                dec_zone_page_state(page, NR_MLOCK);
 116                if (!isolate_lru_page(page)) {
 117                        int ret = try_to_munlock(page);
 118                        /*
 119                         * did try_to_unlock() succeed or punt?
 120                         */
 121                        if (ret == SWAP_SUCCESS || ret == SWAP_AGAIN)
 122                                count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 123
 124                        putback_lru_page(page);
 125                } else {
 126                        /*
 127                         * We lost the race.  let try_to_unmap() deal
 128                         * with it.  At least we get the page state and
 129                         * mlock stats right.  However, page is still on
 130                         * the noreclaim list.  We'll fix that up when
 131                         * the page is eventually freed or we scan the
 132                         * noreclaim list.
 133                         */
 134                        if (PageUnevictable(page))
 135                                count_vm_event(UNEVICTABLE_PGSTRANDED);
 136                        else
 137                                count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 138                }
 139        }
 140}
 141
 142/**
 143 * __mlock_vma_pages_range() -  mlock/munlock a range of pages in the vma.
 144 * @vma:   target vma
 145 * @start: start address
 146 * @end:   end address
 147 * @mlock: 0 indicate munlock, otherwise mlock.
 148 *
 149 * If @mlock == 0, unlock an mlocked range;
 150 * else mlock the range of pages.  This takes care of making the pages present ,
 151 * too.
 152 *
 153 * return 0 on success, negative error code on error.
 154 *
 155 * vma->vm_mm->mmap_sem must be held for at least read.
 156 */
 157static long __mlock_vma_pages_range(struct vm_area_struct *vma,
 158                                   unsigned long start, unsigned long end,
 159                                   int mlock)
 160{
 161        struct mm_struct *mm = vma->vm_mm;
 162        unsigned long addr = start;
 163        struct page *pages[16]; /* 16 gives a reasonable batch */
 164        int nr_pages = (end - start) / PAGE_SIZE;
 165        int ret = 0;
 166        int gup_flags = 0;
 167
 168        VM_BUG_ON(start & ~PAGE_MASK);
 169        VM_BUG_ON(end   & ~PAGE_MASK);
 170        VM_BUG_ON(start < vma->vm_start);
 171        VM_BUG_ON(end   > vma->vm_end);
 172        VM_BUG_ON((!rwsem_is_locked(&mm->mmap_sem)) &&
 173                  (atomic_read(&mm->mm_users) != 0));
 174
 175        /*
 176         * mlock:   don't page populate if page has PROT_NONE permission.
 177         * munlock: the pages always do munlock althrough
 178         *          its has PROT_NONE permission.
 179         */
 180        if (!mlock)
 181                gup_flags |= GUP_FLAGS_IGNORE_VMA_PERMISSIONS;
 182
 183        if (vma->vm_flags & VM_WRITE)
 184                gup_flags |= GUP_FLAGS_WRITE;
 185
 186        while (nr_pages > 0) {
 187                int i;
 188
 189                cond_resched();
 190
 191                /*
 192                 * get_user_pages makes pages present if we are
 193                 * setting mlock. and this extra reference count will
 194                 * disable migration of this page.  However, page may
 195                 * still be truncated out from under us.
 196                 */
 197                ret = __get_user_pages(current, mm, addr,
 198                                min_t(int, nr_pages, ARRAY_SIZE(pages)),
 199                                gup_flags, pages, NULL);
 200                /*
 201                 * This can happen for, e.g., VM_NONLINEAR regions before
 202                 * a page has been allocated and mapped at a given offset,
 203                 * or for addresses that map beyond end of a file.
 204                 * We'll mlock the the pages if/when they get faulted in.
 205                 */
 206                if (ret < 0)
 207                        break;
 208                if (ret == 0) {
 209                        /*
 210                         * We know the vma is there, so the only time
 211                         * we cannot get a single page should be an
 212                         * error (ret < 0) case.
 213                         */
 214                        WARN_ON(1);
 215                        break;
 216                }
 217
 218                lru_add_drain();        /* push cached pages to LRU */
 219
 220                for (i = 0; i < ret; i++) {
 221                        struct page *page = pages[i];
 222
 223                        lock_page(page);
 224                        /*
 225                         * Because we lock page here and migration is blocked
 226                         * by the elevated reference, we need only check for
 227                         * page truncation (file-cache only).
 228                         */
 229                        if (page->mapping) {
 230                                if (mlock)
 231                                        mlock_vma_page(page);
 232                                else
 233                                        munlock_vma_page(page);
 234                        }
 235                        unlock_page(page);
 236                        put_page(page);         /* ref from get_user_pages() */
 237
 238                        /*
 239                         * here we assume that get_user_pages() has given us
 240                         * a list of virtually contiguous pages.
 241                         */
 242                        addr += PAGE_SIZE;      /* for next get_user_pages() */
 243                        nr_pages--;
 244                }
 245                ret = 0;
 246        }
 247
 248        return ret;     /* count entire vma as locked_vm */
 249}
 250
 251/*
 252 * convert get_user_pages() return value to posix mlock() error
 253 */
 254static int __mlock_posix_error_return(long retval)
 255{
 256        if (retval == -EFAULT)
 257                retval = -ENOMEM;
 258        else if (retval == -ENOMEM)
 259                retval = -EAGAIN;
 260        return retval;
 261}
 262
 263#else /* CONFIG_UNEVICTABLE_LRU */
 264
 265/*
 266 * Just make pages present if VM_LOCKED.  No-op if unlocking.
 267 */
 268static long __mlock_vma_pages_range(struct vm_area_struct *vma,
 269                                   unsigned long start, unsigned long end,
 270                                   int mlock)
 271{
 272        if (mlock && (vma->vm_flags & VM_LOCKED))
 273                return make_pages_present(start, end);
 274        return 0;
 275}
 276
 277static inline int __mlock_posix_error_return(long retval)
 278{
 279        return 0;
 280}
 281
 282#endif /* CONFIG_UNEVICTABLE_LRU */
 283
 284/**
 285 * mlock_vma_pages_range() - mlock pages in specified vma range.
 286 * @vma - the vma containing the specfied address range
 287 * @start - starting address in @vma to mlock
 288 * @end   - end address [+1] in @vma to mlock
 289 *
 290 * For mmap()/mremap()/expansion of mlocked vma.
 291 *
 292 * return 0 on success for "normal" vmas.
 293 *
 294 * return number of pages [> 0] to be removed from locked_vm on success
 295 * of "special" vmas.
 296 *
 297 * return negative error if vma spanning @start-@range disappears while
 298 * mmap semaphore is dropped.  Unlikely?
 299 */
 300long mlock_vma_pages_range(struct vm_area_struct *vma,
 301                        unsigned long start, unsigned long end)
 302{
 303        struct mm_struct *mm = vma->vm_mm;
 304        int nr_pages = (end - start) / PAGE_SIZE;
 305        BUG_ON(!(vma->vm_flags & VM_LOCKED));
 306
 307        /*
 308         * filter unlockable vmas
 309         */
 310        if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 311                goto no_mlock;
 312
 313        if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 314                        is_vm_hugetlb_page(vma) ||
 315                        vma == get_gate_vma(current))) {
 316                long error;
 317                downgrade_write(&mm->mmap_sem);
 318
 319                error = __mlock_vma_pages_range(vma, start, end, 1);
 320
 321                up_read(&mm->mmap_sem);
 322                /* vma can change or disappear */
 323                down_write(&mm->mmap_sem);
 324                vma = find_vma(mm, start);
 325                /* non-NULL vma must contain @start, but need to check @end */
 326                if (!vma ||  end > vma->vm_end)
 327                        return -ENOMEM;
 328
 329                return 0;       /* hide other errors from mmap(), et al */
 330        }
 331
 332        /*
 333         * User mapped kernel pages or huge pages:
 334         * make these pages present to populate the ptes, but
 335         * fall thru' to reset VM_LOCKED--no need to unlock, and
 336         * return nr_pages so these don't get counted against task's
 337         * locked limit.  huge pages are already counted against
 338         * locked vm limit.
 339         */
 340        make_pages_present(start, end);
 341
 342no_mlock:
 343        vma->vm_flags &= ~VM_LOCKED;    /* and don't come back! */
 344        return nr_pages;                /* error or pages NOT mlocked */
 345}
 346
 347
 348/*
 349 * munlock_vma_pages_range() - munlock all pages in the vma range.'
 350 * @vma - vma containing range to be munlock()ed.
 351 * @start - start address in @vma of the range
 352 * @end - end of range in @vma.
 353 *
 354 *  For mremap(), munmap() and exit().
 355 *
 356 * Called with @vma VM_LOCKED.
 357 *
 358 * Returns with VM_LOCKED cleared.  Callers must be prepared to
 359 * deal with this.
 360 *
 361 * We don't save and restore VM_LOCKED here because pages are
 362 * still on lru.  In unmap path, pages might be scanned by reclaim
 363 * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 364 * free them.  This will result in freeing mlocked pages.
 365 */
 366void munlock_vma_pages_range(struct vm_area_struct *vma,
 367                           unsigned long start, unsigned long end)
 368{
 369        vma->vm_flags &= ~VM_LOCKED;
 370        __mlock_vma_pages_range(vma, start, end, 0);
 371}
 372
 373/*
 374 * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 375 *
 376 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 377 * munlock is a no-op.  However, for some special vmas, we go ahead and
 378 * populate the ptes via make_pages_present().
 379 *
 380 * For vmas that pass the filters, merge/split as appropriate.
 381 */
 382static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 383        unsigned long start, unsigned long end, unsigned int newflags)
 384{
 385        struct mm_struct *mm = vma->vm_mm;
 386        pgoff_t pgoff;
 387        int nr_pages;
 388        int ret = 0;
 389        int lock = newflags & VM_LOCKED;
 390
 391        if (newflags == vma->vm_flags ||
 392                        (vma->vm_flags & (VM_IO | VM_PFNMAP)))
 393                goto out;       /* don't set VM_LOCKED,  don't count */
 394
 395        if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 396                        is_vm_hugetlb_page(vma) ||
 397                        vma == get_gate_vma(current)) {
 398                if (lock)
 399                        make_pages_present(start, end);
 400                goto out;       /* don't set VM_LOCKED,  don't count */
 401        }
 402
 403        pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 404        *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 405                          vma->vm_file, pgoff, vma_policy(vma));
 406        if (*prev) {
 407                vma = *prev;
 408                goto success;
 409        }
 410
 411        if (start != vma->vm_start) {
 412                ret = split_vma(mm, vma, start, 1);
 413                if (ret)
 414                        goto out;
 415        }
 416
 417        if (end != vma->vm_end) {
 418                ret = split_vma(mm, vma, end, 0);
 419                if (ret)
 420                        goto out;
 421        }
 422
 423success:
 424        /*
 425         * Keep track of amount of locked VM.
 426         */
 427        nr_pages = (end - start) >> PAGE_SHIFT;
 428        if (!lock)
 429                nr_pages = -nr_pages;
 430        mm->locked_vm += nr_pages;
 431
 432        /*
 433         * vm_flags is protected by the mmap_sem held in write mode.
 434         * It's okay if try_to_unmap_one unmaps a page just after we
 435         * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 436         */
 437        vma->vm_flags = newflags;
 438
 439        if (lock) {
 440                /*
 441                 * mmap_sem is currently held for write.  Downgrade the write
 442                 * lock to a read lock so that other faults, mmap scans, ...
 443                 * while we fault in all pages.
 444                 */
 445                downgrade_write(&mm->mmap_sem);
 446
 447                ret = __mlock_vma_pages_range(vma, start, end, 1);
 448
 449                /*
 450                 * Need to reacquire mmap sem in write mode, as our callers
 451                 * expect this.  We have no support for atomically upgrading
 452                 * a sem to write, so we need to check for ranges while sem
 453                 * is unlocked.
 454                 */
 455                up_read(&mm->mmap_sem);
 456                /* vma can change or disappear */
 457                down_write(&mm->mmap_sem);
 458                *prev = find_vma(mm, start);
 459                /* non-NULL *prev must contain @start, but need to check @end */
 460                if (!(*prev) || end > (*prev)->vm_end)
 461                        ret = -ENOMEM;
 462                else if (ret > 0) {
 463                        mm->locked_vm -= ret;
 464                        ret = 0;
 465                } else
 466                        ret = __mlock_posix_error_return(ret); /* translate if needed */
 467        } else {
 468                /*
 469                 * TODO:  for unlocking, pages will already be resident, so
 470                 * we don't need to wait for allocations/reclaim/pagein, ...
 471                 * However, unlocking a very large region can still take a
 472                 * while.  Should we downgrade the semaphore for both lock
 473                 * AND unlock ?
 474                 */
 475                __mlock_vma_pages_range(vma, start, end, 0);
 476        }
 477
 478out:
 479        *prev = vma;
 480        return ret;
 481}
 482
 483static int do_mlock(unsigned long start, size_t len, int on)
 484{
 485        unsigned long nstart, end, tmp;
 486        struct vm_area_struct * vma, * prev;
 487        int error;
 488
 489        len = PAGE_ALIGN(len);
 490        end = start + len;
 491        if (end < start)
 492                return -EINVAL;
 493        if (end == start)
 494                return 0;
 495        vma = find_vma_prev(current->mm, start, &prev);
 496        if (!vma || vma->vm_start > start)
 497                return -ENOMEM;
 498
 499        if (start > vma->vm_start)
 500                prev = vma;
 501
 502        for (nstart = start ; ; ) {
 503                unsigned int newflags;
 504
 505                /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 506
 507                newflags = vma->vm_flags | VM_LOCKED;
 508                if (!on)
 509                        newflags &= ~VM_LOCKED;
 510
 511                tmp = vma->vm_end;
 512                if (tmp > end)
 513                        tmp = end;
 514                error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 515                if (error)
 516                        break;
 517                nstart = tmp;
 518                if (nstart < prev->vm_end)
 519                        nstart = prev->vm_end;
 520                if (nstart >= end)
 521                        break;
 522
 523                vma = prev->vm_next;
 524                if (!vma || vma->vm_start != nstart) {
 525                        error = -ENOMEM;
 526                        break;
 527                }
 528        }
 529        return error;
 530}
 531
 532SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 533{
 534        unsigned long locked;
 535        unsigned long lock_limit;
 536        int error = -ENOMEM;
 537
 538        if (!can_do_mlock())
 539                return -EPERM;
 540
 541        lru_add_drain_all();    /* flush pagevec */
 542
 543        down_write(&current->mm->mmap_sem);
 544        len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 545        start &= PAGE_MASK;
 546
 547        locked = len >> PAGE_SHIFT;
 548        locked += current->mm->locked_vm;
 549
 550        lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 551        lock_limit >>= PAGE_SHIFT;
 552
 553        /* check against resource limits */
 554        if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 555                error = do_mlock(start, len, 1);
 556        up_write(&current->mm->mmap_sem);
 557        return error;
 558}
 559
 560SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 561{
 562        int ret;
 563
 564        down_write(&current->mm->mmap_sem);
 565        len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 566        start &= PAGE_MASK;
 567        ret = do_mlock(start, len, 0);
 568        up_write(&current->mm->mmap_sem);
 569        return ret;
 570}
 571
 572static int do_mlockall(int flags)
 573{
 574        struct vm_area_struct * vma, * prev = NULL;
 575        unsigned int def_flags = 0;
 576
 577        if (flags & MCL_FUTURE)
 578                def_flags = VM_LOCKED;
 579        current->mm->def_flags = def_flags;
 580        if (flags == MCL_FUTURE)
 581                goto out;
 582
 583        for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 584                unsigned int newflags;
 585
 586                newflags = vma->vm_flags | VM_LOCKED;
 587                if (!(flags & MCL_CURRENT))
 588                        newflags &= ~VM_LOCKED;
 589
 590                /* Ignore errors */
 591                mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 592        }
 593out:
 594        return 0;
 595}
 596
 597SYSCALL_DEFINE1(mlockall, int, flags)
 598{
 599        unsigned long lock_limit;
 600        int ret = -EINVAL;
 601
 602        if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 603                goto out;
 604
 605        ret = -EPERM;
 606        if (!can_do_mlock())
 607                goto out;
 608
 609        lru_add_drain_all();    /* flush pagevec */
 610
 611        down_write(&current->mm->mmap_sem);
 612
 613        lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 614        lock_limit >>= PAGE_SHIFT;
 615
 616        ret = -ENOMEM;
 617        if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 618            capable(CAP_IPC_LOCK))
 619                ret = do_mlockall(flags);
 620        up_write(&current->mm->mmap_sem);
 621out:
 622        return ret;
 623}
 624
 625SYSCALL_DEFINE0(munlockall)
 626{
 627        int ret;
 628
 629        down_write(&current->mm->mmap_sem);
 630        ret = do_mlockall(0);
 631        up_write(&current->mm->mmap_sem);
 632        return ret;
 633}
 634
 635/*
 636 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 637 * shm segments) get accounted against the user_struct instead.
 638 */
 639static DEFINE_SPINLOCK(shmlock_user_lock);
 640
 641int user_shm_lock(size_t size, struct user_struct *user)
 642{
 643        unsigned long lock_limit, locked;
 644        int allowed = 0;
 645
 646        locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 647        lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 648        if (lock_limit == RLIM_INFINITY)
 649                allowed = 1;
 650        lock_limit >>= PAGE_SHIFT;
 651        spin_lock(&shmlock_user_lock);
 652        if (!allowed &&
 653            locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 654                goto out;
 655        get_uid(user);
 656        user->locked_shm += locked;
 657        allowed = 1;
 658out:
 659        spin_unlock(&shmlock_user_lock);
 660        return allowed;
 661}
 662
 663void user_shm_unlock(size_t size, struct user_struct *user)
 664{
 665        spin_lock(&shmlock_user_lock);
 666        user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 667        spin_unlock(&shmlock_user_lock);
 668        free_uid(user);
 669}
 670