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 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
 139 {
 140         return (vma->vm_flags & VM_GROWSDOWN) &&
 141                 (vma->vm_start == addr) &&
 142                 !vma_stack_continue(vma->vm_prev, addr);
 143 }
 144
 145 /**
 146  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
 147  * @vma:   target vma
 148  * @start: start address
 149  * @end:   end address
 150  *
 151  * This takes care of making the pages present too.
 152  *
 153  * return 0 on success, negative error code on error.
 154  *
 155  * vma->vm_mm->mmap_sem must be held for at least read.
 156  */
 157 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
 158                                     unsigned long start, unsigned long end,
 159                                     int *nonblocking)
 160 {
 161         struct mm_struct *mm = vma->vm_mm;
 162         unsigned long addr = start;
 163         int nr_pages = (end - start) / PAGE_SIZE;
 164         int gup_flags;
 165
 166         VM_BUG_ON(start & ~PAGE_MASK);
 167         VM_BUG_ON(end   & ~PAGE_MASK);
 168         VM_BUG_ON(start < vma->vm_start);
 169         VM_BUG_ON(end   > vma->vm_end);
 170         VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem));
 171
 172         gup_flags = FOLL_TOUCH;
 173         /*
 174          * We want to touch writable mappings with a write fault in order
 175          * to break COW, except for shared mappings because these don't COW
 176          * and we would not want to dirty them for nothing.
 177          */
 178         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
 179                 gup_flags |= FOLL_WRITE;
 180
 181         /*
 182          * We want mlock to succeed for regions that have any permissions
 183          * other than PROT_NONE.
 184          */
 185         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
 186                 gup_flags |= FOLL_FORCE;
 187
 188         if (vma->vm_flags & VM_LOCKED)
 189                 gup_flags |= FOLL_MLOCK;
 190
 191         /* We don't try to access the guard page of a stack vma */
 192         if (stack_guard_page(vma, start)) {
 193                 addr += PAGE_SIZE;
 194                 nr_pages--;
 195         }
 196
 197         return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
 198                                 NULL, NULL, nonblocking);
 199 }
 200
 201 /*
 202  * convert get_user_pages() return value to posix mlock() error
 203  */
 204 static int __mlock_posix_error_return(long retval)
 205 {
 206         if (retval == -EFAULT)
 207                 retval = -ENOMEM;
 208         else if (retval == -ENOMEM)
 209                 retval = -EAGAIN;
 210         return retval;
 211 }
 212
 213 /**
 214  * mlock_vma_pages_range() - mlock pages in specified vma range.
 215  * @vma - the vma containing the specfied address range
 216  * @start - starting address in @vma to mlock
 217  * @end   - end address [+1] in @vma to mlock
 218  *
 219  * For mmap()/mremap()/expansion of mlocked vma.
 220  *
 221  * return 0 on success for "normal" vmas.
 222  *
 223  * return number of pages [> 0] to be removed from locked_vm on success
 224  * of "special" vmas.
 225  */
 226 long mlock_vma_pages_range(struct vm_area_struct *vma,
 227                         unsigned long start, unsigned long end)
 228 {
 229         int nr_pages = (end - start) / PAGE_SIZE;
 230         BUG_ON(!(vma->vm_flags & VM_LOCKED));
 231
 232         /*
 233          * filter unlockable vmas
 234          */
 235         if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 236                 goto no_mlock;
 237
 238         if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 239                         is_vm_hugetlb_page(vma) ||
 240                         vma == get_gate_vma(current->mm))) {
 241
 242                 __mlock_vma_pages_range(vma, start, end, NULL);
 243
 244                 /* Hide errors from mmap() and other callers */
 245                 return 0;
 246         }
 247
 248         /*
 249          * User mapped kernel pages or huge pages:
 250          * make these pages present to populate the ptes, but
 251          * fall thru' to reset VM_LOCKED--no need to unlock, and
 252          * return nr_pages so these don't get counted against task's
 253          * locked limit.  huge pages are already counted against
 254          * locked vm limit.
 255          */
 256         make_pages_present(start, end);
 257
 258 no_mlock:
 259         vma->vm_flags &= ~VM_LOCKED;    /* and don't come back! */
 260         return nr_pages;                /* error or pages NOT mlocked */
 261 }
 262
 263 /*
 264  * munlock_vma_pages_range() - munlock all pages in the vma range.'
 265  * @vma - vma containing range to be munlock()ed.
 266  * @start - start address in @vma of the range
 267  * @end - end of range in @vma.
 268  *
 269  *  For mremap(), munmap() and exit().
 270  *
 271  * Called with @vma VM_LOCKED.
 272  *
 273  * Returns with VM_LOCKED cleared.  Callers must be prepared to
 274  * deal with this.
 275  *
 276  * We don't save and restore VM_LOCKED here because pages are
 277  * still on lru.  In unmap path, pages might be scanned by reclaim
 278  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 279  * free them.  This will result in freeing mlocked pages.
 280  */
 281 void munlock_vma_pages_range(struct vm_area_struct *vma,
 282                              unsigned long start, unsigned long end)
 283 {
 284         unsigned long addr;
 285
 286         lru_add_drain();
 287         vma->vm_flags &= ~VM_LOCKED;
 288
 289         for (addr = start; addr < end; addr += PAGE_SIZE) {
 290                 struct page *page;
 291                 /*
 292                  * Although FOLL_DUMP is intended for get_dump_page(),
 293                  * it just so happens that its special treatment of the
 294                  * ZERO_PAGE (returning an error instead of doing get_page)
 295                  * suits munlock very well (and if somehow an abnormal page
 296                  * has sneaked into the range, we won't oops here: great).
 297                  */
 298                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
 299                 if (page && !IS_ERR(page)) {
 300                         lock_page(page);
 301                         /*
 302                          * Like in __mlock_vma_pages_range(),
 303                          * because we lock page here and migration is
 304                          * blocked by the elevated reference, we need
 305                          * only check for file-cache page truncation.
 306                          */
 307                         if (page->mapping)
 308                                 munlock_vma_page(page);
 309                         unlock_page(page);
 310                         put_page(page);
 311                 }
 312                 cond_resched();
 313         }
 314 }
 315
 316 /*
 317  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 318  *
 319  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 320  * munlock is a no-op.  However, for some special vmas, we go ahead and
 321  * populate the ptes via make_pages_present().
 322  *
 323  * For vmas that pass the filters, merge/split as appropriate.
 324  */
 325 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 326         unsigned long start, unsigned long end, unsigned int newflags)
 327 {
 328         struct mm_struct *mm = vma->vm_mm;
 329         pgoff_t pgoff;
 330         int nr_pages;
 331         int ret = 0;
 332         int lock = newflags & VM_LOCKED;
 333
 334         if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
 335             is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
 336                 goto out;       /* don't set VM_LOCKED,  don't count */
 337
 338         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 339         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 340                           vma->vm_file, pgoff, vma_policy(vma));
 341         if (*prev) {
 342                 vma = *prev;
 343                 goto success;
 344         }
 345
 346         if (start != vma->vm_start) {
 347                 ret = split_vma(mm, vma, start, 1);
 348                 if (ret)
 349                         goto out;
 350         }
 351
 352         if (end != vma->vm_end) {
 353                 ret = split_vma(mm, vma, end, 0);
 354                 if (ret)
 355                         goto out;
 356         }
 357
 358 success:
 359         /*
 360          * Keep track of amount of locked VM.
 361          */
 362         nr_pages = (end - start) >> PAGE_SHIFT;
 363         if (!lock)
 364                 nr_pages = -nr_pages;
 365         mm->locked_vm += nr_pages;
 366
 367         /*
 368          * vm_flags is protected by the mmap_sem held in write mode.
 369          * It's okay if try_to_unmap_one unmaps a page just after we
 370          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 371          */
 372
 373         if (lock)
 374                 vma->vm_flags = newflags;
 375         else
 376                 munlock_vma_pages_range(vma, start, end);
 377
 378 out:
 379         *prev = vma;
 380         return ret;
 381 }
 382
 383 static int do_mlock(unsigned long start, size_t len, int on)
 384 {
 385         unsigned long nstart, end, tmp;
 386         struct vm_area_struct * vma, * prev;
 387         int error;
 388
 389         VM_BUG_ON(start & ~PAGE_MASK);
 390         VM_BUG_ON(len != PAGE_ALIGN(len));
 391         end = start + len;
 392         if (end < start)
 393                 return -EINVAL;
 394         if (end == start)
 395                 return 0;
 396         vma = find_vma_prev(current->mm, start, &prev);
 397         if (!vma || vma->vm_start > start)
 398                 return -ENOMEM;
 399
 400         if (start > vma->vm_start)
 401                 prev = vma;
 402
 403         for (nstart = start ; ; ) {
 404                 unsigned int newflags;
 405
 406                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 407
 408                 newflags = vma->vm_flags | VM_LOCKED;
 409                 if (!on)
 410                         newflags &= ~VM_LOCKED;
 411
 412                 tmp = vma->vm_end;
 413                 if (tmp > end)
 414                         tmp = end;
 415                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 416                 if (error)
 417                         break;
 418                 nstart = tmp;
 419                 if (nstart < prev->vm_end)
 420                         nstart = prev->vm_end;
 421                 if (nstart >= end)
 422                         break;
 423
 424                 vma = prev->vm_next;
 425                 if (!vma || vma->vm_start != nstart) {
 426                         error = -ENOMEM;
 427                         break;
 428                 }
 429         }
 430         return error;
 431 }
 432
 433 static int do_mlock_pages(unsigned long start, size_t len, int ignore_errors)
 434 {
 435         struct mm_struct *mm = current->mm;
 436         unsigned long end, nstart, nend;
 437         struct vm_area_struct *vma = NULL;
 438         int locked = 0;
 439         int ret = 0;
 440
 441         VM_BUG_ON(start & ~PAGE_MASK);
 442         VM_BUG_ON(len != PAGE_ALIGN(len));
 443         end = start + len;
 444
 445         for (nstart = start; nstart < end; nstart = nend) {
 446                 /*
 447                  * We want to fault in pages for [nstart; end) address range.
 448                  * Find first corresponding VMA.
 449                  */
 450                 if (!locked) {
 451                         locked = 1;
 452                         down_read(&mm->mmap_sem);
 453                         vma = find_vma(mm, nstart);
 454                 } else if (nstart >= vma->vm_end)
 455                         vma = vma->vm_next;
 456                 if (!vma || vma->vm_start >= end)
 457                         break;
 458                 /*
 459                  * Set [nstart; nend) to intersection of desired address
 460                  * range with the first VMA. Also, skip undesirable VMA types.
 461                  */
 462                 nend = min(end, vma->vm_end);
 463                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 464                         continue;
 465                 if (nstart < vma->vm_start)
 466                         nstart = vma->vm_start;
 467                 /*
 468                  * Now fault in a range of pages. __mlock_vma_pages_range()
 469                  * double checks the vma flags, so that it won't mlock pages
 470                  * if the vma was already munlocked.
 471                  */
 472                 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
 473                 if (ret < 0) {
 474                         if (ignore_errors) {
 475                                 ret = 0;
 476                                 continue;       /* continue at next VMA */
 477                         }
 478                         ret = __mlock_posix_error_return(ret);
 479                         break;
 480                 }
 481                 nend = nstart + ret * PAGE_SIZE;
 482                 ret = 0;
 483         }
 484         if (locked)
 485                 up_read(&mm->mmap_sem);
 486         return ret;     /* 0 or negative error code */
 487 }
 488
 489 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 490 {
 491         unsigned long locked;
 492         unsigned long lock_limit;
 493         int error = -ENOMEM;
 494
 495         if (!can_do_mlock())
 496                 return -EPERM;
 497
 498         lru_add_drain_all();    /* flush pagevec */
 499
 500         down_write(&current->mm->mmap_sem);
 501         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 502         start &= PAGE_MASK;
 503
 504         locked = len >> PAGE_SHIFT;
 505         locked += current->mm->locked_vm;
 506
 507         lock_limit = rlimit(RLIMIT_MEMLOCK);
 508         lock_limit >>= PAGE_SHIFT;
 509
 510         /* check against resource limits */
 511         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 512                 error = do_mlock(start, len, 1);
 513         up_write(&current->mm->mmap_sem);
 514         if (!error)
 515                 error = do_mlock_pages(start, len, 0);
 516         return error;
 517 }
 518
 519 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 520 {
 521         int ret;
 522
 523         down_write(&current->mm->mmap_sem);
 524         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 525         start &= PAGE_MASK;
 526         ret = do_mlock(start, len, 0);
 527         up_write(&current->mm->mmap_sem);
 528         return ret;
 529 }
 530
 531 static int do_mlockall(int flags)
 532 {
 533         struct vm_area_struct * vma, * prev = NULL;
 534         unsigned int def_flags = 0;
 535
 536         if (flags & MCL_FUTURE)
 537                 def_flags = VM_LOCKED;
 538         current->mm->def_flags = def_flags;
 539         if (flags == MCL_FUTURE)
 540                 goto out;
 541
 542         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 543                 unsigned int newflags;
 544
 545                 newflags = vma->vm_flags | VM_LOCKED;
 546                 if (!(flags & MCL_CURRENT))
 547                         newflags &= ~VM_LOCKED;
 548
 549                 /* Ignore errors */
 550                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 551         }
 552 out:
 553         return 0;
 554 }
 555
 556 SYSCALL_DEFINE1(mlockall, int, flags)
 557 {
 558         unsigned long lock_limit;
 559         int ret = -EINVAL;
 560
 561         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 562                 goto out;
 563
 564         ret = -EPERM;
 565         if (!can_do_mlock())
 566                 goto out;
 567
 568         lru_add_drain_all();    /* flush pagevec */
 569
 570         down_write(&current->mm->mmap_sem);
 571
 572         lock_limit = rlimit(RLIMIT_MEMLOCK);
 573         lock_limit >>= PAGE_SHIFT;
 574
 575         ret = -ENOMEM;
 576         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 577             capable(CAP_IPC_LOCK))
 578                 ret = do_mlockall(flags);
 579         up_write(&current->mm->mmap_sem);
 580         if (!ret && (flags & MCL_CURRENT)) {
 581                 /* Ignore errors */
 582                 do_mlock_pages(0, TASK_SIZE, 1);
 583         }
 584 out:
 585         return ret;
 586 }
 587
 588 SYSCALL_DEFINE0(munlockall)
 589 {
 590         int ret;
 591
 592         down_write(&current->mm->mmap_sem);
 593         ret = do_mlockall(0);
 594         up_write(&current->mm->mmap_sem);
 595         return ret;
 596 }
 597
 598 /*
 599  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 600  * shm segments) get accounted against the user_struct instead.
 601  */
 602 static DEFINE_SPINLOCK(shmlock_user_lock);
 603
 604 int user_shm_lock(size_t size, struct user_struct *user)
 605 {
 606         unsigned long lock_limit, locked;
 607         int allowed = 0;
 608
 609         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 610         lock_limit = rlimit(RLIMIT_MEMLOCK);
 611         if (lock_limit == RLIM_INFINITY)
 612                 allowed = 1;
 613         lock_limit >>= PAGE_SHIFT;
 614         spin_lock(&shmlock_user_lock);
 615         if (!allowed &&
 616             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 617                 goto out;
 618         get_uid(user);
 619         user->locked_shm += locked;
 620         allowed = 1;
 621 out:
 622         spin_unlock(&shmlock_user_lock);
 623         return allowed;
 624 }
 625
 626 void user_shm_unlock(size_t size, struct user_struct *user)
 627 {
 628         spin_lock(&shmlock_user_lock);
 629         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 630         spin_unlock(&shmlock_user_lock);
 631         free_uid(user);
 632 }