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