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/export.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 (rlimit(RLIMIT_MEMLOCK) != 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         if (!TestClearPageMlocked(page))
  57                 return;
  58
  59         mod_zone_page_state(page_zone(page), NR_MLOCK,
  60                             -hpage_nr_pages(page));
  61         count_vm_event(UNEVICTABLE_PGCLEARED);
  62         if (!isolate_lru_page(page)) {
  63                 putback_lru_page(page);
  64         } else {
  65                 /*
  66                  * We lost the race. the page already moved to evictable list.
  67                  */
  68                 if (PageUnevictable(page))
  69                         count_vm_event(UNEVICTABLE_PGSTRANDED);
  70         }
  71 }
  72
  73 /*
  74  * Mark page as mlocked if not already.
  75  * If page on LRU, isolate and putback to move to unevictable list.
  76  */
  77 void mlock_vma_page(struct page *page)
  78 {
  79         BUG_ON(!PageLocked(page));
  80
  81         if (!TestSetPageMlocked(page)) {
  82                 mod_zone_page_state(page_zone(page), NR_MLOCK,
  83                                     hpage_nr_pages(page));
  84                 count_vm_event(UNEVICTABLE_PGMLOCKED);
  85                 if (!isolate_lru_page(page))
  86                         putback_lru_page(page);
  87         }
  88 }
  89
  90 /**
  91  * munlock_vma_page - munlock a vma page
  92  * @page - page to be unlocked
  93  *
  94  * called from munlock()/munmap() path with page supposedly on the LRU.
  95  * When we munlock a page, because the vma where we found the page is being
  96  * munlock()ed or munmap()ed, we want to check whether other vmas hold the
  97  * page locked so that we can leave it on the unevictable lru list and not
  98  * bother vmscan with it.  However, to walk the page's rmap list in
  99  * try_to_munlock() we must isolate the page from the LRU.  If some other
 100  * task has removed the page from the LRU, we won't be able to do that.
 101  * So we clear the PageMlocked as we might not get another chance.  If we
 102  * can't isolate the page, we leave it for putback_lru_page() and vmscan
 103  * [page_referenced()/try_to_unmap()] to deal with.
 104  */
 105 unsigned int munlock_vma_page(struct page *page)
 106 {
 107         unsigned int page_mask = 0;
 108
 109         BUG_ON(!PageLocked(page));
 110
 111         if (TestClearPageMlocked(page)) {
 112                 unsigned int nr_pages = hpage_nr_pages(page);
 113                 mod_zone_page_state(page_zone(page), NR_MLOCK, -nr_pages);
 114                 page_mask = nr_pages - 1;
 115                 if (!isolate_lru_page(page)) {
 116                         int ret = SWAP_AGAIN;
 117
 118                         /*
 119                          * Optimization: if the page was mapped just once,
 120                          * that's our mapping and we don't need to check all the
 121                          * other vmas.
 122                          */
 123                         if (page_mapcount(page) > 1)
 124                                 ret = try_to_munlock(page);
 125                         /*
 126                          * did try_to_unlock() succeed or punt?
 127                          */
 128                         if (ret != SWAP_MLOCK)
 129                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 130
 131                         putback_lru_page(page);
 132                 } else {
 133                         /*
 134                          * Some other task has removed the page from the LRU.
 135                          * putback_lru_page() will take care of removing the
 136                          * page from the unevictable list, if necessary.
 137                          * vmscan [page_referenced()] will move the page back
 138                          * to the unevictable list if some other vma has it
 139                          * mlocked.
 140                          */
 141                         if (PageUnevictable(page))
 142                                 count_vm_event(UNEVICTABLE_PGSTRANDED);
 143                         else
 144                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 145                 }
 146         }
 147
 148         return page_mask;
 149 }
 150
 151 /**
 152  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
 153  * @vma:   target vma
 154  * @start: start address
 155  * @end:   end address
 156  *
 157  * This takes care of making the pages present too.
 158  *
 159  * return 0 on success, negative error code on error.
 160  *
 161  * vma->vm_mm->mmap_sem must be held for at least read.
 162  */
 163 long __mlock_vma_pages_range(struct vm_area_struct *vma,
 164                 unsigned long start, unsigned long end, int *nonblocking)
 165 {
 166         struct mm_struct *mm = vma->vm_mm;
 167         unsigned long nr_pages = (end - start) / PAGE_SIZE;
 168         int gup_flags;
 169
 170         VM_BUG_ON(start & ~PAGE_MASK);
 171         VM_BUG_ON(end   & ~PAGE_MASK);
 172         VM_BUG_ON(start < vma->vm_start);
 173         VM_BUG_ON(end   > vma->vm_end);
 174         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 175
 176         gup_flags = FOLL_TOUCH | FOLL_MLOCK;
 177         /*
 178          * We want to touch writable mappings with a write fault in order
 179          * to break COW, except for shared mappings because these don't COW
 180          * and we would not want to dirty them for nothing.
 181          */
 182         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
 183                 gup_flags |= FOLL_WRITE;
 184
 185         /*
 186          * We want mlock to succeed for regions that have any permissions
 187          * other than PROT_NONE.
 188          */
 189         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
 190                 gup_flags |= FOLL_FORCE;
 191
 192         /*
 193          * We made sure addr is within a VMA, so the following will
 194          * not result in a stack expansion that recurses back here.
 195          */
 196         return __get_user_pages(current, mm, start, nr_pages, gup_flags,
 197                                 NULL, NULL, nonblocking);
 198 }
 199
 200 /*
 201  * convert get_user_pages() return value to posix mlock() error
 202  */
 203 static int __mlock_posix_error_return(long retval)
 204 {
 205         if (retval == -EFAULT)
 206                 retval = -ENOMEM;
 207         else if (retval == -ENOMEM)
 208                 retval = -EAGAIN;
 209         return retval;
 210 }
 211
 212 /*
 213  * munlock_vma_pages_range() - munlock all pages in the vma range.'
 214  * @vma - vma containing range to be munlock()ed.
 215  * @start - start address in @vma of the range
 216  * @end - end of range in @vma.
 217  *
 218  *  For mremap(), munmap() and exit().
 219  *
 220  * Called with @vma VM_LOCKED.
 221  *
 222  * Returns with VM_LOCKED cleared.  Callers must be prepared to
 223  * deal with this.
 224  *
 225  * We don't save and restore VM_LOCKED here because pages are
 226  * still on lru.  In unmap path, pages might be scanned by reclaim
 227  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 228  * free them.  This will result in freeing mlocked pages.
 229  */
 230 void munlock_vma_pages_range(struct vm_area_struct *vma,
 231                              unsigned long start, unsigned long end)
 232 {
 233         vma->vm_flags &= ~VM_LOCKED;
 234
 235         while (start < end) {
 236                 struct page *page;
 237                 unsigned int page_mask, page_increm;
 238
 239                 /*
 240                  * Although FOLL_DUMP is intended for get_dump_page(),
 241                  * it just so happens that its special treatment of the
 242                  * ZERO_PAGE (returning an error instead of doing get_page)
 243                  * suits munlock very well (and if somehow an abnormal page
 244                  * has sneaked into the range, we won't oops here: great).
 245                  */
 246                 page = follow_page_mask(vma, start, FOLL_GET | FOLL_DUMP,
 247                                         &page_mask);
 248                 if (page && !IS_ERR(page)) {
 249                         lock_page(page);
 250                         lru_add_drain();
 251                         /*
 252                          * Any THP page found by follow_page_mask() may have
 253                          * gotten split before reaching munlock_vma_page(),
 254                          * so we need to recompute the page_mask here.
 255                          */
 256                         page_mask = munlock_vma_page(page);
 257                         unlock_page(page);
 258                         put_page(page);
 259                 }
 260                 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
 261                 start += page_increm * PAGE_SIZE;
 262                 cond_resched();
 263         }
 264 }
 265
 266 /*
 267  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 268  *
 269  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 270  * munlock is a no-op.  However, for some special vmas, we go ahead and
 271  * populate the ptes.
 272  *
 273  * For vmas that pass the filters, merge/split as appropriate.
 274  */
 275 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 276         unsigned long start, unsigned long end, vm_flags_t newflags)
 277 {
 278         struct mm_struct *mm = vma->vm_mm;
 279         pgoff_t pgoff;
 280         int nr_pages;
 281         int ret = 0;
 282         int lock = !!(newflags & VM_LOCKED);
 283
 284         if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
 285             is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
 286                 goto out;       /* don't set VM_LOCKED,  don't count */
 287
 288         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 289         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 290                           vma->vm_file, pgoff, vma_policy(vma));
 291         if (*prev) {
 292                 vma = *prev;
 293                 goto success;
 294         }
 295
 296         if (start != vma->vm_start) {
 297                 ret = split_vma(mm, vma, start, 1);
 298                 if (ret)
 299                         goto out;
 300         }
 301
 302         if (end != vma->vm_end) {
 303                 ret = split_vma(mm, vma, end, 0);
 304                 if (ret)
 305                         goto out;
 306         }
 307
 308 success:
 309         /*
 310          * Keep track of amount of locked VM.
 311          */
 312         nr_pages = (end - start) >> PAGE_SHIFT;
 313         if (!lock)
 314                 nr_pages = -nr_pages;
 315         mm->locked_vm += nr_pages;
 316
 317         /*
 318          * vm_flags is protected by the mmap_sem held in write mode.
 319          * It's okay if try_to_unmap_one unmaps a page just after we
 320          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 321          */
 322
 323         if (lock)
 324                 vma->vm_flags = newflags;
 325         else
 326                 munlock_vma_pages_range(vma, start, end);
 327
 328 out:
 329         *prev = vma;
 330         return ret;
 331 }
 332
 333 static int do_mlock(unsigned long start, size_t len, int on)
 334 {
 335         unsigned long nstart, end, tmp;
 336         struct vm_area_struct * vma, * prev;
 337         int error;
 338
 339         VM_BUG_ON(start & ~PAGE_MASK);
 340         VM_BUG_ON(len != PAGE_ALIGN(len));
 341         end = start + len;
 342         if (end < start)
 343                 return -EINVAL;
 344         if (end == start)
 345                 return 0;
 346         vma = find_vma(current->mm, start);
 347         if (!vma || vma->vm_start > start)
 348                 return -ENOMEM;
 349
 350         prev = vma->vm_prev;
 351         if (start > vma->vm_start)
 352                 prev = vma;
 353
 354         for (nstart = start ; ; ) {
 355                 vm_flags_t newflags;
 356
 357                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 358
 359                 newflags = vma->vm_flags & ~VM_LOCKED;
 360                 if (on)
 361                         newflags |= VM_LOCKED | VM_POPULATE;
 362
 363                 tmp = vma->vm_end;
 364                 if (tmp > end)
 365                         tmp = end;
 366                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 367                 if (error)
 368                         break;
 369                 nstart = tmp;
 370                 if (nstart < prev->vm_end)
 371                         nstart = prev->vm_end;
 372                 if (nstart >= end)
 373                         break;
 374
 375                 vma = prev->vm_next;
 376                 if (!vma || vma->vm_start != nstart) {
 377                         error = -ENOMEM;
 378                         break;
 379                 }
 380         }
 381         return error;
 382 }
 383
 384 /*
 385  * __mm_populate - populate and/or mlock pages within a range of address space.
 386  *
 387  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
 388  * flags. VMAs must be already marked with the desired vm_flags, and
 389  * mmap_sem must not be held.
 390  */
 391 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
 392 {
 393         struct mm_struct *mm = current->mm;
 394         unsigned long end, nstart, nend;
 395         struct vm_area_struct *vma = NULL;
 396         int locked = 0;
 397         long ret = 0;
 398
 399         VM_BUG_ON(start & ~PAGE_MASK);
 400         VM_BUG_ON(len != PAGE_ALIGN(len));
 401         end = start + len;
 402
 403         for (nstart = start; nstart < end; nstart = nend) {
 404                 /*
 405                  * We want to fault in pages for [nstart; end) address range.
 406                  * Find first corresponding VMA.
 407                  */
 408                 if (!locked) {
 409                         locked = 1;
 410                         down_read(&mm->mmap_sem);
 411                         vma = find_vma(mm, nstart);
 412                 } else if (nstart >= vma->vm_end)
 413                         vma = vma->vm_next;
 414                 if (!vma || vma->vm_start >= end)
 415                         break;
 416                 /*
 417                  * Set [nstart; nend) to intersection of desired address
 418                  * range with the first VMA. Also, skip undesirable VMA types.
 419                  */
 420                 nend = min(end, vma->vm_end);
 421                 if ((vma->vm_flags & (VM_IO | VM_PFNMAP | VM_POPULATE)) !=
 422                     VM_POPULATE)
 423                         continue;
 424                 if (nstart < vma->vm_start)
 425                         nstart = vma->vm_start;
 426                 /*
 427                  * Now fault in a range of pages. __mlock_vma_pages_range()
 428                  * double checks the vma flags, so that it won't mlock pages
 429                  * if the vma was already munlocked.
 430                  */
 431                 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
 432                 if (ret < 0) {
 433                         if (ignore_errors) {
 434                                 ret = 0;
 435                                 continue;       /* continue at next VMA */
 436                         }
 437                         ret = __mlock_posix_error_return(ret);
 438                         break;
 439                 }
 440                 nend = nstart + ret * PAGE_SIZE;
 441                 ret = 0;
 442         }
 443         if (locked)
 444                 up_read(&mm->mmap_sem);
 445         return ret;     /* 0 or negative error code */
 446 }
 447
 448 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 449 {
 450         unsigned long locked;
 451         unsigned long lock_limit;
 452         int error = -ENOMEM;
 453
 454         if (!can_do_mlock())
 455                 return -EPERM;
 456
 457         lru_add_drain_all();    /* flush pagevec */
 458
 459         down_write(&current->mm->mmap_sem);
 460         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 461         start &= PAGE_MASK;
 462
 463         locked = len >> PAGE_SHIFT;
 464         locked += current->mm->locked_vm;
 465
 466         lock_limit = rlimit(RLIMIT_MEMLOCK);
 467         lock_limit >>= PAGE_SHIFT;
 468
 469         /* check against resource limits */
 470         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 471                 error = do_mlock(start, len, 1);
 472         up_write(&current->mm->mmap_sem);
 473         if (!error)
 474                 error = __mm_populate(start, len, 0);
 475         return error;
 476 }
 477
 478 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 479 {
 480         int ret;
 481
 482         down_write(&current->mm->mmap_sem);
 483         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 484         start &= PAGE_MASK;
 485         ret = do_mlock(start, len, 0);
 486         up_write(&current->mm->mmap_sem);
 487         return ret;
 488 }
 489
 490 static int do_mlockall(int flags)
 491 {
 492         struct vm_area_struct * vma, * prev = NULL;
 493
 494         if (flags & MCL_FUTURE)
 495                 current->mm->def_flags |= VM_LOCKED | VM_POPULATE;
 496         else
 497                 current->mm->def_flags &= ~(VM_LOCKED | VM_POPULATE);
 498         if (flags == MCL_FUTURE)
 499                 goto out;
 500
 501         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 502                 vm_flags_t newflags;
 503
 504                 newflags = vma->vm_flags & ~VM_LOCKED;
 505                 if (flags & MCL_CURRENT)
 506                         newflags |= VM_LOCKED | VM_POPULATE;
 507
 508                 /* Ignore errors */
 509                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 510         }
 511 out:
 512         return 0;
 513 }
 514
 515 SYSCALL_DEFINE1(mlockall, int, flags)
 516 {
 517         unsigned long lock_limit;
 518         int ret = -EINVAL;
 519
 520         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 521                 goto out;
 522
 523         ret = -EPERM;
 524         if (!can_do_mlock())
 525                 goto out;
 526
 527         if (flags & MCL_CURRENT)
 528                 lru_add_drain_all();    /* flush pagevec */
 529
 530         down_write(&current->mm->mmap_sem);
 531
 532         lock_limit = rlimit(RLIMIT_MEMLOCK);
 533         lock_limit >>= PAGE_SHIFT;
 534
 535         ret = -ENOMEM;
 536         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 537             capable(CAP_IPC_LOCK))
 538                 ret = do_mlockall(flags);
 539         up_write(&current->mm->mmap_sem);
 540         if (!ret && (flags & MCL_CURRENT))
 541                 mm_populate(0, TASK_SIZE);
 542 out:
 543         return ret;
 544 }
 545
 546 SYSCALL_DEFINE0(munlockall)
 547 {
 548         int ret;
 549
 550         down_write(&current->mm->mmap_sem);
 551         ret = do_mlockall(0);
 552         up_write(&current->mm->mmap_sem);
 553         return ret;
 554 }
 555
 556 /*
 557  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 558  * shm segments) get accounted against the user_struct instead.
 559  */
 560 static DEFINE_SPINLOCK(shmlock_user_lock);
 561
 562 int user_shm_lock(size_t size, struct user_struct *user)
 563 {
 564         unsigned long lock_limit, locked;
 565         int allowed = 0;
 566
 567         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 568         lock_limit = rlimit(RLIMIT_MEMLOCK);
 569         if (lock_limit == RLIM_INFINITY)
 570                 allowed = 1;
 571         lock_limit >>= PAGE_SHIFT;
 572         spin_lock(&shmlock_user_lock);
 573         if (!allowed &&
 574             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 575                 goto out;
 576         get_uid(user);
 577         user->locked_shm += locked;
 578         allowed = 1;
 579 out:
 580         spin_unlock(&shmlock_user_lock);
 581         return allowed;
 582 }
 583
 584 void user_shm_unlock(size_t size, struct user_struct *user)
 585 {
 586         spin_lock(&shmlock_user_lock);
 587         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 588         spin_unlock(&shmlock_user_lock);
 589         free_uid(user);
 590 }