mm/mlock.c

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