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 void munlock_vma_page(struct page *page)
 106 {
 107         BUG_ON(!PageLocked(page));
 108
 109         if (TestClearPageMlocked(page)) {
 110                 mod_zone_page_state(page_zone(page), NR_MLOCK,
 111                                     -hpage_nr_pages(page));
 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) ||
 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                         munlock_vma_page(page);
 294                         unlock_page(page);
 295                         put_page(page);
 296                 }
 297                 cond_resched();
 298         }
 299 }
 300
 301 /*
 302  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 303  *
 304  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 305  * munlock is a no-op.  However, for some special vmas, we go ahead and
 306  * populate the ptes via make_pages_present().
 307  *
 308  * For vmas that pass the filters, merge/split as appropriate.
 309  */
 310 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 311         unsigned long start, unsigned long end, vm_flags_t newflags)
 312 {
 313         struct mm_struct *mm = vma->vm_mm;
 314         pgoff_t pgoff;
 315         int nr_pages;
 316         int ret = 0;
 317         int lock = !!(newflags & VM_LOCKED);
 318
 319         if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
 320             is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
 321                 goto out;       /* don't set VM_LOCKED,  don't count */
 322
 323         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 324         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 325                           vma->vm_file, pgoff, vma_policy(vma));
 326         if (*prev) {
 327                 vma = *prev;
 328                 goto success;
 329         }
 330
 331         if (start != vma->vm_start) {
 332                 ret = split_vma(mm, vma, start, 1);
 333                 if (ret)
 334                         goto out;
 335         }
 336
 337         if (end != vma->vm_end) {
 338                 ret = split_vma(mm, vma, end, 0);
 339                 if (ret)
 340                         goto out;
 341         }
 342
 343 success:
 344         /*
 345          * Keep track of amount of locked VM.
 346          */
 347         nr_pages = (end - start) >> PAGE_SHIFT;
 348         if (!lock)
 349                 nr_pages = -nr_pages;
 350         mm->locked_vm += nr_pages;
 351
 352         /*
 353          * vm_flags is protected by the mmap_sem held in write mode.
 354          * It's okay if try_to_unmap_one unmaps a page just after we
 355          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 356          */
 357
 358         if (lock)
 359                 vma->vm_flags = newflags;
 360         else
 361                 munlock_vma_pages_range(vma, start, end);
 362
 363 out:
 364         *prev = vma;
 365         return ret;
 366 }
 367
 368 static int do_mlock(unsigned long start, size_t len, int on)
 369 {
 370         unsigned long nstart, end, tmp;
 371         struct vm_area_struct * vma, * prev;
 372         int error;
 373
 374         VM_BUG_ON(start & ~PAGE_MASK);
 375         VM_BUG_ON(len != PAGE_ALIGN(len));
 376         end = start + len;
 377         if (end < start)
 378                 return -EINVAL;
 379         if (end == start)
 380                 return 0;
 381         vma = find_vma(current->mm, start);
 382         if (!vma || vma->vm_start > start)
 383                 return -ENOMEM;
 384
 385         prev = vma->vm_prev;
 386         if (start > vma->vm_start)
 387                 prev = vma;
 388
 389         for (nstart = start ; ; ) {
 390                 vm_flags_t newflags;
 391
 392                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 393
 394                 newflags = vma->vm_flags | VM_LOCKED;
 395                 if (!on)
 396                         newflags &= ~VM_LOCKED;
 397
 398                 tmp = vma->vm_end;
 399                 if (tmp > end)
 400                         tmp = end;
 401                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 402                 if (error)
 403                         break;
 404                 nstart = tmp;
 405                 if (nstart < prev->vm_end)
 406                         nstart = prev->vm_end;
 407                 if (nstart >= end)
 408                         break;
 409
 410                 vma = prev->vm_next;
 411                 if (!vma || vma->vm_start != nstart) {
 412                         error = -ENOMEM;
 413                         break;
 414                 }
 415         }
 416         return error;
 417 }
 418
 419 static int do_mlock_pages(unsigned long start, size_t len, int ignore_errors)
 420 {
 421         struct mm_struct *mm = current->mm;
 422         unsigned long end, nstart, nend;
 423         struct vm_area_struct *vma = NULL;
 424         int locked = 0;
 425         int ret = 0;
 426
 427         VM_BUG_ON(start & ~PAGE_MASK);
 428         VM_BUG_ON(len != PAGE_ALIGN(len));
 429         end = start + len;
 430
 431         for (nstart = start; nstart < end; nstart = nend) {
 432                 /*
 433                  * We want to fault in pages for [nstart; end) address range.
 434                  * Find first corresponding VMA.
 435                  */
 436                 if (!locked) {
 437                         locked = 1;
 438                         down_read(&mm->mmap_sem);
 439                         vma = find_vma(mm, nstart);
 440                 } else if (nstart >= vma->vm_end)
 441                         vma = vma->vm_next;
 442                 if (!vma || vma->vm_start >= end)
 443                         break;
 444                 /*
 445                  * Set [nstart; nend) to intersection of desired address
 446                  * range with the first VMA. Also, skip undesirable VMA types.
 447                  */
 448                 nend = min(end, vma->vm_end);
 449                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 450                         continue;
 451                 if (nstart < vma->vm_start)
 452                         nstart = vma->vm_start;
 453                 /*
 454                  * Now fault in a range of pages. __mlock_vma_pages_range()
 455                  * double checks the vma flags, so that it won't mlock pages
 456                  * if the vma was already munlocked.
 457                  */
 458                 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
 459                 if (ret < 0) {
 460                         if (ignore_errors) {
 461                                 ret = 0;
 462                                 continue;       /* continue at next VMA */
 463                         }
 464                         ret = __mlock_posix_error_return(ret);
 465                         break;
 466                 }
 467                 nend = nstart + ret * PAGE_SIZE;
 468                 ret = 0;
 469         }
 470         if (locked)
 471                 up_read(&mm->mmap_sem);
 472         return ret;     /* 0 or negative error code */
 473 }
 474
 475 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 476 {
 477         unsigned long locked;
 478         unsigned long lock_limit;
 479         int error = -ENOMEM;
 480
 481         if (!can_do_mlock())
 482                 return -EPERM;
 483
 484         lru_add_drain_all();    /* flush pagevec */
 485
 486         down_write(&current->mm->mmap_sem);
 487         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 488         start &= PAGE_MASK;
 489
 490         locked = len >> PAGE_SHIFT;
 491         locked += current->mm->locked_vm;
 492
 493         lock_limit = rlimit(RLIMIT_MEMLOCK);
 494         lock_limit >>= PAGE_SHIFT;
 495
 496         /* check against resource limits */
 497         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 498                 error = do_mlock(start, len, 1);
 499         up_write(&current->mm->mmap_sem);
 500         if (!error)
 501                 error = do_mlock_pages(start, len, 0);
 502         return error;
 503 }
 504
 505 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 506 {
 507         int ret;
 508
 509         down_write(&current->mm->mmap_sem);
 510         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 511         start &= PAGE_MASK;
 512         ret = do_mlock(start, len, 0);
 513         up_write(&current->mm->mmap_sem);
 514         return ret;
 515 }
 516
 517 static int do_mlockall(int flags)
 518 {
 519         struct vm_area_struct * vma, * prev = NULL;
 520         unsigned int def_flags = 0;
 521
 522         if (flags & MCL_FUTURE)
 523                 def_flags = VM_LOCKED;
 524         current->mm->def_flags = def_flags;
 525         if (flags == MCL_FUTURE)
 526                 goto out;
 527
 528         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 529                 vm_flags_t newflags;
 530
 531                 newflags = vma->vm_flags | VM_LOCKED;
 532                 if (!(flags & MCL_CURRENT))
 533                         newflags &= ~VM_LOCKED;
 534
 535                 /* Ignore errors */
 536                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 537         }
 538 out:
 539         return 0;
 540 }
 541
 542 SYSCALL_DEFINE1(mlockall, int, flags)
 543 {
 544         unsigned long lock_limit;
 545         int ret = -EINVAL;
 546
 547         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 548                 goto out;
 549
 550         ret = -EPERM;
 551         if (!can_do_mlock())
 552                 goto out;
 553
 554         if (flags & MCL_CURRENT)
 555                 lru_add_drain_all();    /* flush pagevec */
 556
 557         down_write(&current->mm->mmap_sem);
 558
 559         lock_limit = rlimit(RLIMIT_MEMLOCK);
 560         lock_limit >>= PAGE_SHIFT;
 561
 562         ret = -ENOMEM;
 563         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 564             capable(CAP_IPC_LOCK))
 565                 ret = do_mlockall(flags);
 566         up_write(&current->mm->mmap_sem);
 567         if (!ret && (flags & MCL_CURRENT)) {
 568                 /* Ignore errors */
 569                 do_mlock_pages(0, TASK_SIZE, 1);
 570         }
 571 out:
 572         return ret;
 573 }
 574
 575 SYSCALL_DEFINE0(munlockall)
 576 {
 577         int ret;
 578
 579         down_write(&current->mm->mmap_sem);
 580         ret = do_mlockall(0);
 581         up_write(&current->mm->mmap_sem);
 582         return ret;
 583 }
 584
 585 /*
 586  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 587  * shm segments) get accounted against the user_struct instead.
 588  */
 589 static DEFINE_SPINLOCK(shmlock_user_lock);
 590
 591 int user_shm_lock(size_t size, struct user_struct *user)
 592 {
 593         unsigned long lock_limit, locked;
 594         int allowed = 0;
 595
 596         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 597         lock_limit = rlimit(RLIMIT_MEMLOCK);
 598         if (lock_limit == RLIM_INFINITY)
 599                 allowed = 1;
 600         lock_limit >>= PAGE_SHIFT;
 601         spin_lock(&shmlock_user_lock);
 602         if (!allowed &&
 603             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 604                 goto out;
 605         get_uid(user);
 606         user->locked_shm += locked;
 607         allowed = 1;
 608 out:
 609         spin_unlock(&shmlock_user_lock);
 610         return allowed;
 611 }
 612
 613 void user_shm_unlock(size_t size, struct user_struct *user)
 614 {
 615         spin_lock(&shmlock_user_lock);
 616         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 617         spin_unlock(&shmlock_user_lock);
 618         free_uid(user);
 619 }