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         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 = SWAP_AGAIN;
 114
 115                         /*
 116                          * Optimization: if the page was mapped just once,
 117                          * that's our mapping and we don't need to check all the
 118                          * other vmas.
 119                          */
 120                         if (page_mapcount(page) > 1)
 121                                 ret = try_to_munlock(page);
 122                         /*
 123                          * did try_to_unlock() succeed or punt?
 124                          */
 125                         if (ret != SWAP_MLOCK)
 126                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 127
 128                         putback_lru_page(page);
 129                 } else {
 130                         /*
 131                          * Some other task has removed the page from the LRU.
 132                          * putback_lru_page() will take care of removing the
 133                          * page from the unevictable list, if necessary.
 134                          * vmscan [page_referenced()] will move the page back
 135                          * to the unevictable list if some other vma has it
 136                          * mlocked.
 137                          */
 138                         if (PageUnevictable(page))
 139                                 count_vm_event(UNEVICTABLE_PGSTRANDED);
 140                         else
 141                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 142                 }
 143         }
 144 }
 145
 146 /**
 147  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
 148  * @vma:   target vma
 149  * @start: start address
 150  * @end:   end address
 151  *
 152  * This takes care of making the pages present too.
 153  *
 154  * return 0 on success, negative error code on error.
 155  *
 156  * vma->vm_mm->mmap_sem must be held for at least read.
 157  */
 158 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
 159                                     unsigned long start, unsigned long end,
 160                                     int *nonblocking)
 161 {
 162         struct mm_struct *mm = vma->vm_mm;
 163         unsigned long addr = start;
 164         int nr_pages = (end - start) / PAGE_SIZE;
 165         int gup_flags;
 166
 167         VM_BUG_ON(start & ~PAGE_MASK);
 168         VM_BUG_ON(end   & ~PAGE_MASK);
 169         VM_BUG_ON(start < vma->vm_start);
 170         VM_BUG_ON(end   > vma->vm_end);
 171         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 172
 173         gup_flags = FOLL_TOUCH | FOLL_MLOCK;
 174         /*
 175          * We want to touch writable mappings with a write fault in order
 176          * to break COW, except for shared mappings because these don't COW
 177          * and we would not want to dirty them for nothing.
 178          */
 179         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
 180                 gup_flags |= FOLL_WRITE;
 181
 182         /*
 183          * We want mlock to succeed for regions that have any permissions
 184          * other than PROT_NONE.
 185          */
 186         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
 187                 gup_flags |= FOLL_FORCE;
 188
 189         return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
 190                                 NULL, NULL, nonblocking);
 191 }
 192
 193 /*
 194  * convert get_user_pages() return value to posix mlock() error
 195  */
 196 static int __mlock_posix_error_return(long retval)
 197 {
 198         if (retval == -EFAULT)
 199                 retval = -ENOMEM;
 200         else if (retval == -ENOMEM)
 201                 retval = -EAGAIN;
 202         return retval;
 203 }
 204
 205 /**
 206  * mlock_vma_pages_range() - mlock pages in specified vma range.
 207  * @vma - the vma containing the specfied address range
 208  * @start - starting address in @vma to mlock
 209  * @end   - end address [+1] in @vma to mlock
 210  *
 211  * For mmap()/mremap()/expansion of mlocked vma.
 212  *
 213  * return 0 on success for "normal" vmas.
 214  *
 215  * return number of pages [> 0] to be removed from locked_vm on success
 216  * of "special" vmas.
 217  */
 218 long mlock_vma_pages_range(struct vm_area_struct *vma,
 219                         unsigned long start, unsigned long end)
 220 {
 221         int nr_pages = (end - start) / PAGE_SIZE;
 222         BUG_ON(!(vma->vm_flags & VM_LOCKED));
 223
 224         /*
 225          * filter unlockable vmas
 226          */
 227         if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 228                 goto no_mlock;
 229
 230         if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 231                         is_vm_hugetlb_page(vma) ||
 232                         vma == get_gate_vma(current->mm))) {
 233
 234                 __mlock_vma_pages_range(vma, start, end, NULL);
 235
 236                 /* Hide errors from mmap() and other callers */
 237                 return 0;
 238         }
 239
 240         /*
 241          * User mapped kernel pages or huge pages:
 242          * make these pages present to populate the ptes, but
 243          * fall thru' to reset VM_LOCKED--no need to unlock, and
 244          * return nr_pages so these don't get counted against task's
 245          * locked limit.  huge pages are already counted against
 246          * locked vm limit.
 247          */
 248         make_pages_present(start, end);
 249
 250 no_mlock:
 251         vma->vm_flags &= ~VM_LOCKED;    /* and don't come back! */
 252         return nr_pages;                /* error or pages NOT mlocked */
 253 }
 254
 255 /*
 256  * munlock_vma_pages_range() - munlock all pages in the vma range.'
 257  * @vma - vma containing range to be munlock()ed.
 258  * @start - start address in @vma of the range
 259  * @end - end of range in @vma.
 260  *
 261  *  For mremap(), munmap() and exit().
 262  *
 263  * Called with @vma VM_LOCKED.
 264  *
 265  * Returns with VM_LOCKED cleared.  Callers must be prepared to
 266  * deal with this.
 267  *
 268  * We don't save and restore VM_LOCKED here because pages are
 269  * still on lru.  In unmap path, pages might be scanned by reclaim
 270  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 271  * free them.  This will result in freeing mlocked pages.
 272  */
 273 void munlock_vma_pages_range(struct vm_area_struct *vma,
 274                              unsigned long start, unsigned long end)
 275 {
 276         unsigned long addr;
 277
 278         lru_add_drain();
 279         vma->vm_flags &= ~VM_LOCKED;
 280
 281         for (addr = start; addr < end; addr += PAGE_SIZE) {
 282                 struct page *page;
 283                 /*
 284                  * Although FOLL_DUMP is intended for get_dump_page(),
 285                  * it just so happens that its special treatment of the
 286                  * ZERO_PAGE (returning an error instead of doing get_page)
 287                  * suits munlock very well (and if somehow an abnormal page
 288                  * has sneaked into the range, we won't oops here: great).
 289                  */
 290                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
 291                 if (page && !IS_ERR(page)) {
 292                         lock_page(page);
 293                         /*
 294                          * Like in __mlock_vma_pages_range(),
 295                          * because we lock page here and migration is
 296                          * blocked by the elevated reference, we need
 297                          * only check for file-cache page truncation.
 298                          */
 299                         if (page->mapping)
 300                                 munlock_vma_page(page);
 301                         unlock_page(page);
 302                         put_page(page);
 303                 }
 304                 cond_resched();
 305         }
 306 }
 307
 308 /*
 309  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 310  *
 311  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 312  * munlock is a no-op.  However, for some special vmas, we go ahead and
 313  * populate the ptes via make_pages_present().
 314  *
 315  * For vmas that pass the filters, merge/split as appropriate.
 316  */
 317 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 318         unsigned long start, unsigned long end, vm_flags_t newflags)
 319 {
 320         struct mm_struct *mm = vma->vm_mm;
 321         pgoff_t pgoff;
 322         int nr_pages;
 323         int ret = 0;
 324         int lock = !!(newflags & VM_LOCKED);
 325
 326         if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
 327             is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
 328                 goto out;       /* don't set VM_LOCKED,  don't count */
 329
 330         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 331         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 332                           vma->vm_file, pgoff, vma_policy(vma));
 333         if (*prev) {
 334                 vma = *prev;
 335                 goto success;
 336         }
 337
 338         if (start != vma->vm_start) {
 339                 ret = split_vma(mm, vma, start, 1);
 340                 if (ret)
 341                         goto out;
 342         }
 343
 344         if (end != vma->vm_end) {
 345                 ret = split_vma(mm, vma, end, 0);
 346                 if (ret)
 347                         goto out;
 348         }
 349
 350 success:
 351         /*
 352          * Keep track of amount of locked VM.
 353          */
 354         nr_pages = (end - start) >> PAGE_SHIFT;
 355         if (!lock)
 356                 nr_pages = -nr_pages;
 357         mm->locked_vm += nr_pages;
 358
 359         /*
 360          * vm_flags is protected by the mmap_sem held in write mode.
 361          * It's okay if try_to_unmap_one unmaps a page just after we
 362          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 363          */
 364
 365         if (lock)
 366                 vma->vm_flags = newflags;
 367         else
 368                 munlock_vma_pages_range(vma, start, end);
 369
 370 out:
 371         *prev = vma;
 372         return ret;
 373 }
 374
 375 static int do_mlock(unsigned long start, size_t len, int on)
 376 {
 377         unsigned long nstart, end, tmp;
 378         struct vm_area_struct * vma, * prev;
 379         int error;
 380
 381         VM_BUG_ON(start & ~PAGE_MASK);
 382         VM_BUG_ON(len != PAGE_ALIGN(len));
 383         end = start + len;
 384         if (end < start)
 385                 return -EINVAL;
 386         if (end == start)
 387                 return 0;
 388         vma = find_vma_prev(current->mm, start, &prev);
 389         if (!vma || vma->vm_start > start)
 390                 return -ENOMEM;
 391
 392         if (start > vma->vm_start)
 393                 prev = vma;
 394
 395         for (nstart = start ; ; ) {
 396                 vm_flags_t newflags;
 397
 398                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 399
 400                 newflags = vma->vm_flags | VM_LOCKED;
 401                 if (!on)
 402                         newflags &= ~VM_LOCKED;
 403
 404                 tmp = vma->vm_end;
 405                 if (tmp > end)
 406                         tmp = end;
 407                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 408                 if (error)
 409                         break;
 410                 nstart = tmp;
 411                 if (nstart < prev->vm_end)
 412                         nstart = prev->vm_end;
 413                 if (nstart >= end)
 414                         break;
 415
 416                 vma = prev->vm_next;
 417                 if (!vma || vma->vm_start != nstart) {
 418                         error = -ENOMEM;
 419                         break;
 420                 }
 421         }
 422         return error;
 423 }
 424
 425 static int do_mlock_pages(unsigned long start, size_t len, int ignore_errors)
 426 {
 427         struct mm_struct *mm = current->mm;
 428         unsigned long end, nstart, nend;
 429         struct vm_area_struct *vma = NULL;
 430         int locked = 0;
 431         int ret = 0;
 432
 433         VM_BUG_ON(start & ~PAGE_MASK);
 434         VM_BUG_ON(len != PAGE_ALIGN(len));
 435         end = start + len;
 436
 437         for (nstart = start; nstart < end; nstart = nend) {
 438                 /*
 439                  * We want to fault in pages for [nstart; end) address range.
 440                  * Find first corresponding VMA.
 441                  */
 442                 if (!locked) {
 443                         locked = 1;
 444                         down_read(&mm->mmap_sem);
 445                         vma = find_vma(mm, nstart);
 446                 } else if (nstart >= vma->vm_end)
 447                         vma = vma->vm_next;
 448                 if (!vma || vma->vm_start >= end)
 449                         break;
 450                 /*
 451                  * Set [nstart; nend) to intersection of desired address
 452                  * range with the first VMA. Also, skip undesirable VMA types.
 453                  */
 454                 nend = min(end, vma->vm_end);
 455                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 456                         continue;
 457                 if (nstart < vma->vm_start)
 458                         nstart = vma->vm_start;
 459                 /*
 460                  * Now fault in a range of pages. __mlock_vma_pages_range()
 461                  * double checks the vma flags, so that it won't mlock pages
 462                  * if the vma was already munlocked.
 463                  */
 464                 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
 465                 if (ret < 0) {
 466                         if (ignore_errors) {
 467                                 ret = 0;
 468                                 continue;       /* continue at next VMA */
 469                         }
 470                         ret = __mlock_posix_error_return(ret);
 471                         break;
 472                 }
 473                 nend = nstart + ret * PAGE_SIZE;
 474                 ret = 0;
 475         }
 476         if (locked)
 477                 up_read(&mm->mmap_sem);
 478         return ret;     /* 0 or negative error code */
 479 }
 480
 481 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 482 {
 483         unsigned long locked;
 484         unsigned long lock_limit;
 485         int error = -ENOMEM;
 486
 487         if (!can_do_mlock())
 488                 return -EPERM;
 489
 490         lru_add_drain_all();    /* flush pagevec */
 491
 492         down_write(&current->mm->mmap_sem);
 493         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 494         start &= PAGE_MASK;
 495
 496         locked = len >> PAGE_SHIFT;
 497         locked += current->mm->locked_vm;
 498
 499         lock_limit = rlimit(RLIMIT_MEMLOCK);
 500         lock_limit >>= PAGE_SHIFT;
 501
 502         /* check against resource limits */
 503         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 504                 error = do_mlock(start, len, 1);
 505         up_write(&current->mm->mmap_sem);
 506         if (!error)
 507                 error = do_mlock_pages(start, len, 0);
 508         return error;
 509 }
 510
 511 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 512 {
 513         int ret;
 514
 515         down_write(&current->mm->mmap_sem);
 516         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 517         start &= PAGE_MASK;
 518         ret = do_mlock(start, len, 0);
 519         up_write(&current->mm->mmap_sem);
 520         return ret;
 521 }
 522
 523 static int do_mlockall(int flags)
 524 {
 525         struct vm_area_struct * vma, * prev = NULL;
 526         unsigned int def_flags = 0;
 527
 528         if (flags & MCL_FUTURE)
 529                 def_flags = VM_LOCKED;
 530         current->mm->def_flags = def_flags;
 531         if (flags == MCL_FUTURE)
 532                 goto out;
 533
 534         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 535                 vm_flags_t newflags;
 536
 537                 newflags = vma->vm_flags | VM_LOCKED;
 538                 if (!(flags & MCL_CURRENT))
 539                         newflags &= ~VM_LOCKED;
 540
 541                 /* Ignore errors */
 542                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 543         }
 544 out:
 545         return 0;
 546 }
 547
 548 SYSCALL_DEFINE1(mlockall, int, flags)
 549 {
 550         unsigned long lock_limit;
 551         int ret = -EINVAL;
 552
 553         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 554                 goto out;
 555
 556         ret = -EPERM;
 557         if (!can_do_mlock())
 558                 goto out;
 559
 560         if (flags & MCL_CURRENT)
 561                 lru_add_drain_all();    /* flush pagevec */
 562
 563         down_write(&current->mm->mmap_sem);
 564
 565         lock_limit = rlimit(RLIMIT_MEMLOCK);
 566         lock_limit >>= PAGE_SHIFT;
 567
 568         ret = -ENOMEM;
 569         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 570             capable(CAP_IPC_LOCK))
 571                 ret = do_mlockall(flags);
 572         up_write(&current->mm->mmap_sem);
 573         if (!ret && (flags & MCL_CURRENT)) {
 574                 /* Ignore errors */
 575                 do_mlock_pages(0, TASK_SIZE, 1);
 576         }
 577 out:
 578         return ret;
 579 }
 580
 581 SYSCALL_DEFINE0(munlockall)
 582 {
 583         int ret;
 584
 585         down_write(&current->mm->mmap_sem);
 586         ret = do_mlockall(0);
 587         up_write(&current->mm->mmap_sem);
 588         return ret;
 589 }
 590
 591 /*
 592  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 593  * shm segments) get accounted against the user_struct instead.
 594  */
 595 static DEFINE_SPINLOCK(shmlock_user_lock);
 596
 597 int user_shm_lock(size_t size, struct user_struct *user)
 598 {
 599         unsigned long lock_limit, locked;
 600         int allowed = 0;
 601
 602         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 603         lock_limit = rlimit(RLIMIT_MEMLOCK);
 604         if (lock_limit == RLIM_INFINITY)
 605                 allowed = 1;
 606         lock_limit >>= PAGE_SHIFT;
 607         spin_lock(&shmlock_user_lock);
 608         if (!allowed &&
 609             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 610                 goto out;
 611         get_uid(user);
 612         user->locked_shm += locked;
 613         allowed = 1;
 614 out:
 615         spin_unlock(&shmlock_user_lock);
 616         return allowed;
 617 }
 618
 619 void user_shm_unlock(size_t size, struct user_struct *user)
 620 {
 621         spin_lock(&shmlock_user_lock);
 622         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 623         spin_unlock(&shmlock_user_lock);
 624         free_uid(user);
 625 }