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 (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         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                                     int *nonblocking)
 153 {
 154         struct mm_struct *mm = vma->vm_mm;
 155         unsigned long addr = start;
 156         int nr_pages = (end - start) / PAGE_SIZE;
 157         int gup_flags;
 158
 159         VM_BUG_ON(start & ~PAGE_MASK);
 160         VM_BUG_ON(end   & ~PAGE_MASK);
 161         VM_BUG_ON(start < vma->vm_start);
 162         VM_BUG_ON(end   > vma->vm_end);
 163         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 164
 165         gup_flags = FOLL_TOUCH | FOLL_MLOCK;
 166         /*
 167          * We want to touch writable mappings with a write fault in order
 168          * to break COW, except for shared mappings because these don't COW
 169          * and we would not want to dirty them for nothing.
 170          */
 171         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
 172                 gup_flags |= FOLL_WRITE;
 173
 174         /*
 175          * We want mlock to succeed for regions that have any permissions
 176          * other than PROT_NONE.
 177          */
 178         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
 179                 gup_flags |= FOLL_FORCE;
 180
 181         return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
 182                                 NULL, NULL, nonblocking);
 183 }
 184
 185 /*
 186  * convert get_user_pages() return value to posix mlock() error
 187  */
 188 static int __mlock_posix_error_return(long retval)
 189 {
 190         if (retval == -EFAULT)
 191                 retval = -ENOMEM;
 192         else if (retval == -ENOMEM)
 193                 retval = -EAGAIN;
 194         return retval;
 195 }
 196
 197 /**
 198  * mlock_vma_pages_range() - mlock pages in specified vma range.
 199  * @vma - the vma containing the specfied address range
 200  * @start - starting address in @vma to mlock
 201  * @end   - end address [+1] in @vma to mlock
 202  *
 203  * For mmap()/mremap()/expansion of mlocked vma.
 204  *
 205  * return 0 on success for "normal" vmas.
 206  *
 207  * return number of pages [> 0] to be removed from locked_vm on success
 208  * of "special" vmas.
 209  */
 210 long mlock_vma_pages_range(struct vm_area_struct *vma,
 211                         unsigned long start, unsigned long end)
 212 {
 213         int nr_pages = (end - start) / PAGE_SIZE;
 214         BUG_ON(!(vma->vm_flags & VM_LOCKED));
 215
 216         /*
 217          * filter unlockable vmas
 218          */
 219         if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 220                 goto no_mlock;
 221
 222         if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 223                         is_vm_hugetlb_page(vma) ||
 224                         vma == get_gate_vma(current))) {
 225
 226                 __mlock_vma_pages_range(vma, start, end, NULL);
 227
 228                 /* Hide errors from mmap() and other callers */
 229                 return 0;
 230         }
 231
 232         /*
 233          * User mapped kernel pages or huge pages:
 234          * make these pages present to populate the ptes, but
 235          * fall thru' to reset VM_LOCKED--no need to unlock, and
 236          * return nr_pages so these don't get counted against task's
 237          * locked limit.  huge pages are already counted against
 238          * locked vm limit.
 239          */
 240         make_pages_present(start, end);
 241
 242 no_mlock:
 243         vma->vm_flags &= ~VM_LOCKED;    /* and don't come back! */
 244         return nr_pages;                /* error or pages NOT mlocked */
 245 }
 246
 247 /*
 248  * munlock_vma_pages_range() - munlock all pages in the vma range.'
 249  * @vma - vma containing range to be munlock()ed.
 250  * @start - start address in @vma of the range
 251  * @end - end of range in @vma.
 252  *
 253  *  For mremap(), munmap() and exit().
 254  *
 255  * Called with @vma VM_LOCKED.
 256  *
 257  * Returns with VM_LOCKED cleared.  Callers must be prepared to
 258  * deal with this.
 259  *
 260  * We don't save and restore VM_LOCKED here because pages are
 261  * still on lru.  In unmap path, pages might be scanned by reclaim
 262  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 263  * free them.  This will result in freeing mlocked pages.
 264  */
 265 void munlock_vma_pages_range(struct vm_area_struct *vma,
 266                              unsigned long start, unsigned long end)
 267 {
 268         unsigned long addr;
 269
 270         lru_add_drain();
 271         vma->vm_flags &= ~VM_LOCKED;
 272
 273         for (addr = start; addr < end; addr += PAGE_SIZE) {
 274                 struct page *page;
 275                 /*
 276                  * Although FOLL_DUMP is intended for get_dump_page(),
 277                  * it just so happens that its special treatment of the
 278                  * ZERO_PAGE (returning an error instead of doing get_page)
 279                  * suits munlock very well (and if somehow an abnormal page
 280                  * has sneaked into the range, we won't oops here: great).
 281                  */
 282                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
 283                 if (page && !IS_ERR(page)) {
 284                         lock_page(page);
 285                         /*
 286                          * Like in __mlock_vma_pages_range(),
 287                          * because we lock page here and migration is
 288                          * blocked by the elevated reference, we need
 289                          * only check for file-cache page truncation.
 290                          */
 291                         if (page->mapping)
 292                                 munlock_vma_page(page);
 293                         unlock_page(page);
 294                         put_page(page);
 295                 }
 296                 cond_resched();
 297         }
 298 }
 299
 300 /*
 301  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 302  *
 303  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 304  * munlock is a no-op.  However, for some special vmas, we go ahead and
 305  * populate the ptes via make_pages_present().
 306  *
 307  * For vmas that pass the filters, merge/split as appropriate.
 308  */
 309 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 310         unsigned long start, unsigned long end, unsigned int newflags)
 311 {
 312         struct mm_struct *mm = vma->vm_mm;
 313         pgoff_t pgoff;
 314         int nr_pages;
 315         int ret = 0;
 316         int lock = newflags & VM_LOCKED;
 317
 318         if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
 319             is_vm_hugetlb_page(vma) || vma == get_gate_vma(current))
 320                 goto out;       /* don't set VM_LOCKED,  don't count */
 321
 322         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 323         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 324                           vma->vm_file, pgoff, vma_policy(vma));
 325         if (*prev) {
 326                 vma = *prev;
 327                 goto success;
 328         }
 329
 330         if (start != vma->vm_start) {
 331                 ret = split_vma(mm, vma, start, 1);
 332                 if (ret)
 333                         goto out;
 334         }
 335
 336         if (end != vma->vm_end) {
 337                 ret = split_vma(mm, vma, end, 0);
 338                 if (ret)
 339                         goto out;
 340         }
 341
 342 success:
 343         /*
 344          * Keep track of amount of locked VM.
 345          */
 346         nr_pages = (end - start) >> PAGE_SHIFT;
 347         if (!lock)
 348                 nr_pages = -nr_pages;
 349         mm->locked_vm += nr_pages;
 350
 351         /*
 352          * vm_flags is protected by the mmap_sem held in write mode.
 353          * It's okay if try_to_unmap_one unmaps a page just after we
 354          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 355          */
 356
 357         if (lock)
 358                 vma->vm_flags = newflags;
 359         else
 360                 munlock_vma_pages_range(vma, start, end);
 361
 362 out:
 363         *prev = vma;
 364         return ret;
 365 }
 366
 367 static int do_mlock(unsigned long start, size_t len, int on)
 368 {
 369         unsigned long nstart, end, tmp;
 370         struct vm_area_struct * vma, * prev;
 371         int error;
 372
 373         VM_BUG_ON(start & ~PAGE_MASK);
 374         VM_BUG_ON(len != PAGE_ALIGN(len));
 375         end = start + len;
 376         if (end < start)
 377                 return -EINVAL;
 378         if (end == start)
 379                 return 0;
 380         vma = find_vma_prev(current->mm, start, &prev);
 381         if (!vma || vma->vm_start > start)
 382                 return -ENOMEM;
 383
 384         if (start > vma->vm_start)
 385                 prev = vma;
 386
 387         for (nstart = start ; ; ) {
 388                 unsigned int newflags;
 389
 390                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 391
 392                 newflags = vma->vm_flags | VM_LOCKED;
 393                 if (!on)
 394                         newflags &= ~VM_LOCKED;
 395
 396                 tmp = vma->vm_end;
 397                 if (tmp > end)
 398                         tmp = end;
 399                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 400                 if (error)
 401                         break;
 402                 nstart = tmp;
 403                 if (nstart < prev->vm_end)
 404                         nstart = prev->vm_end;
 405                 if (nstart >= end)
 406                         break;
 407
 408                 vma = prev->vm_next;
 409                 if (!vma || vma->vm_start != nstart) {
 410                         error = -ENOMEM;
 411                         break;
 412                 }
 413         }
 414         return error;
 415 }
 416
 417 static int do_mlock_pages(unsigned long start, size_t len, int ignore_errors)
 418 {
 419         struct mm_struct *mm = current->mm;
 420         unsigned long end, nstart, nend;
 421         struct vm_area_struct *vma = NULL;
 422         int locked = 0;
 423         int ret = 0;
 424
 425         VM_BUG_ON(start & ~PAGE_MASK);
 426         VM_BUG_ON(len != PAGE_ALIGN(len));
 427         end = start + len;
 428
 429         for (nstart = start; nstart < end; nstart = nend) {
 430                 /*
 431                  * We want to fault in pages for [nstart; end) address range.
 432                  * Find first corresponding VMA.
 433                  */
 434                 if (!locked) {
 435                         locked = 1;
 436                         down_read(&mm->mmap_sem);
 437                         vma = find_vma(mm, nstart);
 438                 } else if (nstart >= vma->vm_end)
 439                         vma = vma->vm_next;
 440                 if (!vma || vma->vm_start >= end)
 441                         break;
 442                 /*
 443                  * Set [nstart; nend) to intersection of desired address
 444                  * range with the first VMA. Also, skip undesirable VMA types.
 445                  */
 446                 nend = min(end, vma->vm_end);
 447                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 448                         continue;
 449                 if (nstart < vma->vm_start)
 450                         nstart = vma->vm_start;
 451                 /*
 452                  * Now fault in a range of pages. __mlock_vma_pages_range()
 453                  * double checks the vma flags, so that it won't mlock pages
 454                  * if the vma was already munlocked.
 455                  */
 456                 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
 457                 if (ret < 0) {
 458                         if (ignore_errors) {
 459                                 ret = 0;
 460                                 continue;       /* continue at next VMA */
 461                         }
 462                         ret = __mlock_posix_error_return(ret);
 463                         break;
 464                 }
 465                 nend = nstart + ret * PAGE_SIZE;
 466                 ret = 0;
 467         }
 468         if (locked)
 469                 up_read(&mm->mmap_sem);
 470         return ret;     /* 0 or negative error code */
 471 }
 472
 473 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 474 {
 475         unsigned long locked;
 476         unsigned long lock_limit;
 477         int error = -ENOMEM;
 478
 479         if (!can_do_mlock())
 480                 return -EPERM;
 481
 482         lru_add_drain_all();    /* flush pagevec */
 483
 484         down_write(&current->mm->mmap_sem);
 485         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 486         start &= PAGE_MASK;
 487
 488         locked = len >> PAGE_SHIFT;
 489         locked += current->mm->locked_vm;
 490
 491         lock_limit = rlimit(RLIMIT_MEMLOCK);
 492         lock_limit >>= PAGE_SHIFT;
 493
 494         /* check against resource limits */
 495         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 496                 error = do_mlock(start, len, 1);
 497         up_write(&current->mm->mmap_sem);
 498         if (!error)
 499                 error = do_mlock_pages(start, len, 0);
 500         return error;
 501 }
 502
 503 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 504 {
 505         int ret;
 506
 507         down_write(&current->mm->mmap_sem);
 508         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 509         start &= PAGE_MASK;
 510         ret = do_mlock(start, len, 0);
 511         up_write(&current->mm->mmap_sem);
 512         return ret;
 513 }
 514
 515 static int do_mlockall(int flags)
 516 {
 517         struct vm_area_struct * vma, * prev = NULL;
 518         unsigned int def_flags = 0;
 519
 520         if (flags & MCL_FUTURE)
 521                 def_flags = VM_LOCKED;
 522         current->mm->def_flags = def_flags;
 523         if (flags == MCL_FUTURE)
 524                 goto out;
 525
 526         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 527                 unsigned int newflags;
 528
 529                 newflags = vma->vm_flags | VM_LOCKED;
 530                 if (!(flags & MCL_CURRENT))
 531                         newflags &= ~VM_LOCKED;
 532
 533                 /* Ignore errors */
 534                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 535         }
 536 out:
 537         return 0;
 538 }
 539
 540 SYSCALL_DEFINE1(mlockall, int, flags)
 541 {
 542         unsigned long lock_limit;
 543         int ret = -EINVAL;
 544
 545         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 546                 goto out;
 547
 548         ret = -EPERM;
 549         if (!can_do_mlock())
 550                 goto out;
 551
 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 }