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  * called from munlock()/munmap() path with page supposedly on the LRU.
  93  *
  94  * Note:  unlike mlock_vma_page(), we can't just clear the PageMlocked
  95  * [in try_to_munlock()] and then attempt to isolate the page.  We must
  96  * isolate the page to keep others from messing with its unevictable
  97  * and mlocked state while trying to munlock.  However, we pre-clear the
  98  * mlocked state anyway as we might lose the isolation race and we might
  99  * not get another chance to clear PageMlocked.  If we successfully
 100  * isolate the page and try_to_munlock() detects other VM_LOCKED vmas
 101  * mapping the page, it will restore the PageMlocked state, unless the page
 102  * is mapped in a non-linear vma.  So, we go ahead and SetPageMlocked(),
 103  * perhaps redundantly.
 104  * If we lose the isolation race, and the page is mapped by other VM_LOCKED
 105  * vmas, we'll detect this in vmscan--via try_to_munlock() or try_to_unmap()
 106  * either of which will restore the PageMlocked state by calling
 107  * mlock_vma_page() above, if it can grab the vma's mmap sem.
 108  */
 109 static void munlock_vma_page(struct page *page)
 110 {
 111         BUG_ON(!PageLocked(page));
 112
 113         if (TestClearPageMlocked(page)) {
 114                 dec_zone_page_state(page, NR_MLOCK);
 115                 if (!isolate_lru_page(page)) {
 116                         int ret = try_to_munlock(page);
 117                         /*
 118                          * did try_to_unlock() succeed or punt?
 119                          */
 120                         if (ret == SWAP_SUCCESS || ret == SWAP_AGAIN)
 121                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 122
 123                         putback_lru_page(page);
 124                 } else {
 125                         /*
 126                          * We lost the race.  let try_to_unmap() deal
 127                          * with it.  At least we get the page state and
 128                          * mlock stats right.  However, page is still on
 129                          * the noreclaim list.  We'll fix that up when
 130                          * the page is eventually freed or we scan the
 131                          * noreclaim list.
 132                          */
 133                         if (PageUnevictable(page))
 134                                 count_vm_event(UNEVICTABLE_PGSTRANDED);
 135                         else
 136                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 137                 }
 138         }
 139 }
 140
 141 /**
 142  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
 143  * @vma:   target vma
 144  * @start: start address
 145  * @end:   end address
 146  *
 147  * This takes care of making the pages present too.
 148  *
 149  * return 0 on success, negative error code on error.
 150  *
 151  * vma->vm_mm->mmap_sem must be held for at least read.
 152  */
 153 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
 154                                     unsigned long start, unsigned long end)
 155 {
 156         struct mm_struct *mm = vma->vm_mm;
 157         unsigned long addr = start;
 158         struct page *pages[16]; /* 16 gives a reasonable batch */
 159         int nr_pages = (end - start) / PAGE_SIZE;
 160         int ret = 0;
 161         int gup_flags;
 162
 163         VM_BUG_ON(start & ~PAGE_MASK);
 164         VM_BUG_ON(end   & ~PAGE_MASK);
 165         VM_BUG_ON(start < vma->vm_start);
 166         VM_BUG_ON(end   > vma->vm_end);
 167         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 168
 169         gup_flags = 0;
 170         if (vma->vm_flags & VM_WRITE)
 171                 gup_flags = GUP_FLAGS_WRITE;
 172
 173         while (nr_pages > 0) {
 174                 int i;
 175
 176                 cond_resched();
 177
 178                 /*
 179                  * get_user_pages makes pages present if we are
 180                  * setting mlock. and this extra reference count will
 181                  * disable migration of this page.  However, page may
 182                  * still be truncated out from under us.
 183                  */
 184                 ret = __get_user_pages(current, mm, addr,
 185                                 min_t(int, nr_pages, ARRAY_SIZE(pages)),
 186                                 gup_flags, pages, NULL);
 187                 /*
 188                  * This can happen for, e.g., VM_NONLINEAR regions before
 189                  * a page has been allocated and mapped at a given offset,
 190                  * or for addresses that map beyond end of a file.
 191                  * We'll mlock the pages if/when they get faulted in.
 192                  */
 193                 if (ret < 0)
 194                         break;
 195
 196                 lru_add_drain();        /* push cached pages to LRU */
 197
 198                 for (i = 0; i < ret; i++) {
 199                         struct page *page = pages[i];
 200
 201                         lock_page(page);
 202                         /*
 203                          * Because we lock page here and migration is blocked
 204                          * by the elevated reference, we need only check for
 205                          * file-cache page truncation.  This page->mapping
 206                          * check also neatly skips over the ZERO_PAGE(),
 207                          * though if that's common we'd prefer not to lock it.
 208                          */
 209                         if (page->mapping)
 210                                 mlock_vma_page(page);
 211                         unlock_page(page);
 212                         put_page(page); /* ref from get_user_pages() */
 213                 }
 214
 215                 addr += ret * PAGE_SIZE;
 216                 nr_pages -= ret;
 217                 ret = 0;
 218         }
 219
 220         return ret;     /* 0 or negative error code */
 221 }
 222
 223 /*
 224  * convert get_user_pages() return value to posix mlock() error
 225  */
 226 static int __mlock_posix_error_return(long retval)
 227 {
 228         if (retval == -EFAULT)
 229                 retval = -ENOMEM;
 230         else if (retval == -ENOMEM)
 231                 retval = -EAGAIN;
 232         return retval;
 233 }
 234
 235 /**
 236  * mlock_vma_pages_range() - mlock pages in specified vma range.
 237  * @vma - the vma containing the specfied address range
 238  * @start - starting address in @vma to mlock
 239  * @end   - end address [+1] in @vma to mlock
 240  *
 241  * For mmap()/mremap()/expansion of mlocked vma.
 242  *
 243  * return 0 on success for "normal" vmas.
 244  *
 245  * return number of pages [> 0] to be removed from locked_vm on success
 246  * of "special" vmas.
 247  */
 248 long mlock_vma_pages_range(struct vm_area_struct *vma,
 249                         unsigned long start, unsigned long end)
 250 {
 251         int nr_pages = (end - start) / PAGE_SIZE;
 252         BUG_ON(!(vma->vm_flags & VM_LOCKED));
 253
 254         /*
 255          * filter unlockable vmas
 256          */
 257         if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 258                 goto no_mlock;
 259
 260         if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 261                         is_vm_hugetlb_page(vma) ||
 262                         vma == get_gate_vma(current))) {
 263
 264                 __mlock_vma_pages_range(vma, start, end);
 265
 266                 /* Hide errors from mmap() and other callers */
 267                 return 0;
 268         }
 269
 270         /*
 271          * User mapped kernel pages or huge pages:
 272          * make these pages present to populate the ptes, but
 273          * fall thru' to reset VM_LOCKED--no need to unlock, and
 274          * return nr_pages so these don't get counted against task's
 275          * locked limit.  huge pages are already counted against
 276          * locked vm limit.
 277          */
 278         make_pages_present(start, end);
 279
 280 no_mlock:
 281         vma->vm_flags &= ~VM_LOCKED;    /* and don't come back! */
 282         return nr_pages;                /* error or pages NOT mlocked */
 283 }
 284
 285 /*
 286  * munlock_vma_pages_range() - munlock all pages in the vma range.'
 287  * @vma - vma containing range to be munlock()ed.
 288  * @start - start address in @vma of the range
 289  * @end - end of range in @vma.
 290  *
 291  *  For mremap(), munmap() and exit().
 292  *
 293  * Called with @vma VM_LOCKED.
 294  *
 295  * Returns with VM_LOCKED cleared.  Callers must be prepared to
 296  * deal with this.
 297  *
 298  * We don't save and restore VM_LOCKED here because pages are
 299  * still on lru.  In unmap path, pages might be scanned by reclaim
 300  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 301  * free them.  This will result in freeing mlocked pages.
 302  */
 303 void munlock_vma_pages_range(struct vm_area_struct *vma,
 304                              unsigned long start, unsigned long end)
 305 {
 306         unsigned long addr;
 307
 308         lru_add_drain();
 309         vma->vm_flags &= ~VM_LOCKED;
 310
 311         for (addr = start; addr < end; addr += PAGE_SIZE) {
 312                 struct page *page = follow_page(vma, addr, FOLL_GET);
 313                 if (page) {
 314                         lock_page(page);
 315                         if (page->mapping)
 316                                 munlock_vma_page(page);
 317                         unlock_page(page);
 318                         put_page(page);
 319                 }
 320                 cond_resched();
 321         }
 322 }
 323
 324 /*
 325  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 326  *
 327  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 328  * munlock is a no-op.  However, for some special vmas, we go ahead and
 329  * populate the ptes via make_pages_present().
 330  *
 331  * For vmas that pass the filters, merge/split as appropriate.
 332  */
 333 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 334         unsigned long start, unsigned long end, unsigned int newflags)
 335 {
 336         struct mm_struct *mm = vma->vm_mm;
 337         pgoff_t pgoff;
 338         int nr_pages;
 339         int ret = 0;
 340         int lock = newflags & VM_LOCKED;
 341
 342         if (newflags == vma->vm_flags ||
 343                         (vma->vm_flags & (VM_IO | VM_PFNMAP)))
 344                 goto out;       /* don't set VM_LOCKED,  don't count */
 345
 346         if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 347                         is_vm_hugetlb_page(vma) ||
 348                         vma == get_gate_vma(current)) {
 349                 if (lock)
 350                         make_pages_present(start, end);
 351                 goto out;       /* don't set VM_LOCKED,  don't count */
 352         }
 353
 354         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 355         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 356                           vma->vm_file, pgoff, vma_policy(vma));
 357         if (*prev) {
 358                 vma = *prev;
 359                 goto success;
 360         }
 361
 362         if (start != vma->vm_start) {
 363                 ret = split_vma(mm, vma, start, 1);
 364                 if (ret)
 365                         goto out;
 366         }
 367
 368         if (end != vma->vm_end) {
 369                 ret = split_vma(mm, vma, end, 0);
 370                 if (ret)
 371                         goto out;
 372         }
 373
 374 success:
 375         /*
 376          * Keep track of amount of locked VM.
 377          */
 378         nr_pages = (end - start) >> PAGE_SHIFT;
 379         if (!lock)
 380                 nr_pages = -nr_pages;
 381         mm->locked_vm += nr_pages;
 382
 383         /*
 384          * vm_flags is protected by the mmap_sem held in write mode.
 385          * It's okay if try_to_unmap_one unmaps a page just after we
 386          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 387          */
 388
 389         if (lock) {
 390                 vma->vm_flags = newflags;
 391                 ret = __mlock_vma_pages_range(vma, start, end);
 392                 if (ret < 0)
 393                         ret = __mlock_posix_error_return(ret);
 394         } else {
 395                 munlock_vma_pages_range(vma, start, end);
 396         }
 397
 398 out:
 399         *prev = vma;
 400         return ret;
 401 }
 402
 403 static int do_mlock(unsigned long start, size_t len, int on)
 404 {
 405         unsigned long nstart, end, tmp;
 406         struct vm_area_struct * vma, * prev;
 407         int error;
 408
 409         len = PAGE_ALIGN(len);
 410         end = start + len;
 411         if (end < start)
 412                 return -EINVAL;
 413         if (end == start)
 414                 return 0;
 415         vma = find_vma_prev(current->mm, start, &prev);
 416         if (!vma || vma->vm_start > start)
 417                 return -ENOMEM;
 418
 419         if (start > vma->vm_start)
 420                 prev = vma;
 421
 422         for (nstart = start ; ; ) {
 423                 unsigned int newflags;
 424
 425                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 426
 427                 newflags = vma->vm_flags | VM_LOCKED;
 428                 if (!on)
 429                         newflags &= ~VM_LOCKED;
 430
 431                 tmp = vma->vm_end;
 432                 if (tmp > end)
 433                         tmp = end;
 434                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 435                 if (error)
 436                         break;
 437                 nstart = tmp;
 438                 if (nstart < prev->vm_end)
 439                         nstart = prev->vm_end;
 440                 if (nstart >= end)
 441                         break;
 442
 443                 vma = prev->vm_next;
 444                 if (!vma || vma->vm_start != nstart) {
 445                         error = -ENOMEM;
 446                         break;
 447                 }
 448         }
 449         return error;
 450 }
 451
 452 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 453 {
 454         unsigned long locked;
 455         unsigned long lock_limit;
 456         int error = -ENOMEM;
 457
 458         if (!can_do_mlock())
 459                 return -EPERM;
 460
 461         lru_add_drain_all();    /* flush pagevec */
 462
 463         down_write(&current->mm->mmap_sem);
 464         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 465         start &= PAGE_MASK;
 466
 467         locked = len >> PAGE_SHIFT;
 468         locked += current->mm->locked_vm;
 469
 470         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 471         lock_limit >>= PAGE_SHIFT;
 472
 473         /* check against resource limits */
 474         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 475                 error = do_mlock(start, len, 1);
 476         up_write(&current->mm->mmap_sem);
 477         return error;
 478 }
 479
 480 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 481 {
 482         int ret;
 483
 484         down_write(&current->mm->mmap_sem);
 485         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 486         start &= PAGE_MASK;
 487         ret = do_mlock(start, len, 0);
 488         up_write(&current->mm->mmap_sem);
 489         return ret;
 490 }
 491
 492 static int do_mlockall(int flags)
 493 {
 494         struct vm_area_struct * vma, * prev = NULL;
 495         unsigned int def_flags = 0;
 496
 497         if (flags & MCL_FUTURE)
 498                 def_flags = VM_LOCKED;
 499         current->mm->def_flags = def_flags;
 500         if (flags == MCL_FUTURE)
 501                 goto out;
 502
 503         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 504                 unsigned int newflags;
 505
 506                 newflags = vma->vm_flags | VM_LOCKED;
 507                 if (!(flags & MCL_CURRENT))
 508                         newflags &= ~VM_LOCKED;
 509
 510                 /* Ignore errors */
 511                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 512         }
 513 out:
 514         return 0;
 515 }
 516
 517 SYSCALL_DEFINE1(mlockall, int, flags)
 518 {
 519         unsigned long lock_limit;
 520         int ret = -EINVAL;
 521
 522         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 523                 goto out;
 524
 525         ret = -EPERM;
 526         if (!can_do_mlock())
 527                 goto out;
 528
 529         lru_add_drain_all();    /* flush pagevec */
 530
 531         down_write(&current->mm->mmap_sem);
 532
 533         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 534         lock_limit >>= PAGE_SHIFT;
 535
 536         ret = -ENOMEM;
 537         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 538             capable(CAP_IPC_LOCK))
 539                 ret = do_mlockall(flags);
 540         up_write(&current->mm->mmap_sem);
 541 out:
 542         return ret;
 543 }
 544
 545 SYSCALL_DEFINE0(munlockall)
 546 {
 547         int ret;
 548
 549         down_write(&current->mm->mmap_sem);
 550         ret = do_mlockall(0);
 551         up_write(&current->mm->mmap_sem);
 552         return ret;
 553 }
 554
 555 /*
 556  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 557  * shm segments) get accounted against the user_struct instead.
 558  */
 559 static DEFINE_SPINLOCK(shmlock_user_lock);
 560
 561 int user_shm_lock(size_t size, struct user_struct *user)
 562 {
 563         unsigned long lock_limit, locked;
 564         int allowed = 0;
 565
 566         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 567         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 568         if (lock_limit == RLIM_INFINITY)
 569                 allowed = 1;
 570         lock_limit >>= PAGE_SHIFT;
 571         spin_lock(&shmlock_user_lock);
 572         if (!allowed &&
 573             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 574                 goto out;
 575         get_uid(user);
 576         user->locked_shm += locked;
 577         allowed = 1;
 578 out:
 579         spin_unlock(&shmlock_user_lock);
 580         return allowed;
 581 }
 582
 583 void user_shm_unlock(size_t size, struct user_struct *user)
 584 {
 585         spin_lock(&shmlock_user_lock);
 586         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 587         spin_unlock(&shmlock_user_lock);
 588         free_uid(user);
 589 }
 590
 591 int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim,
 592                           size_t size)
 593 {
 594         unsigned long lim, vm, pgsz;
 595         int error = -ENOMEM;
 596
 597         pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
 598
 599         down_write(&mm->mmap_sem);
 600
 601         lim = rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
 602         vm   = mm->total_vm + pgsz;
 603         if (lim < vm)
 604                 goto out;
 605
 606         lim = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
 607         vm   = mm->locked_vm + pgsz;
 608         if (lim < vm)
 609                 goto out;
 610
 611         mm->total_vm  += pgsz;
 612         mm->locked_vm += pgsz;
 613
 614         error = 0;
 615  out:
 616         up_write(&mm->mmap_sem);
 617         return error;
 618 }
 619
 620 void refund_locked_memory(struct mm_struct *mm, size_t size)
 621 {
 622         unsigned long pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
 623
 624         down_write(&mm->mmap_sem);
 625
 626         mm->total_vm  -= pgsz;
 627         mm->locked_vm -= pgsz;
 628
 629         up_write(&mm->mmap_sem);
 630 }