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