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 (current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur != 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  * called from munlock()/munmap() path with page supposedly on the LRU.
  93  *
  94  * Note:  unlike mlock_vma_page(), we can't just clear the PageMlocked
  95  * [in try_to_munlock()] and then attempt to isolate the page.  We must
  96  * isolate the page to keep others from messing with its unevictable
  97  * and mlocked state while trying to munlock.  However, we pre-clear the
  98  * mlocked state anyway as we might lose the isolation race and we might
  99  * not get another chance to clear PageMlocked.  If we successfully
 100  * isolate the page and try_to_munlock() detects other VM_LOCKED vmas
 101  * mapping the page, it will restore the PageMlocked state, unless the page
 102  * is mapped in a non-linear vma.  So, we go ahead and SetPageMlocked(),
 103  * perhaps redundantly.
 104  * If we lose the isolation race, and the page is mapped by other VM_LOCKED
 105  * vmas, we'll detect this in vmscan--via try_to_munlock() or try_to_unmap()
 106  * either of which will restore the PageMlocked state by calling
 107  * mlock_vma_page() above, if it can grab the vma's mmap sem.
 108  */
 109 static void munlock_vma_page(struct page *page)
 110 {
 111         BUG_ON(!PageLocked(page));
 112
 113         if (TestClearPageMlocked(page)) {
 114                 dec_zone_page_state(page, NR_MLOCK);
 115                 if (!isolate_lru_page(page)) {
 116                         int ret = try_to_munlock(page);
 117                         /*
 118                          * did try_to_unlock() succeed or punt?
 119                          */
 120                         if (ret == SWAP_SUCCESS || ret == SWAP_AGAIN)
 121                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 122
 123                         putback_lru_page(page);
 124                 } else {
 125                         /*
 126                          * We lost the race.  let try_to_unmap() deal
 127                          * with it.  At least we get the page state and
 128                          * mlock stats right.  However, page is still on
 129                          * the noreclaim list.  We'll fix that up when
 130                          * the page is eventually freed or we scan the
 131                          * noreclaim list.
 132                          */
 133                         if (PageUnevictable(page))
 134                                 count_vm_event(UNEVICTABLE_PGSTRANDED);
 135                         else
 136                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 137                 }
 138         }
 139 }
 140
 141 /**
 142  * __mlock_vma_pages_range() -  mlock/munlock a range of pages in the vma.
 143  * @vma:   target vma
 144  * @start: start address
 145  * @end:   end address
 146  * @mlock: 0 indicate munlock, otherwise mlock.
 147  *
 148  * If @mlock == 0, unlock an mlocked range;
 149  * else mlock the range of pages.  This takes care of making the pages present ,
 150  * too.
 151  *
 152  * return 0 on success, negative error code on error.
 153  *
 154  * vma->vm_mm->mmap_sem must be held for at least read.
 155  */
 156 static long __mlock_vma_pages_range(struct vm_area_struct *vma,
 157                                    unsigned long start, unsigned long end,
 158                                    int mlock)
 159 {
 160         struct mm_struct *mm = vma->vm_mm;
 161         unsigned long addr = start;
 162         struct page *pages[16]; /* 16 gives a reasonable batch */
 163         int nr_pages = (end - start) / PAGE_SIZE;
 164         int ret = 0;
 165         int gup_flags = 0;
 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                   (atomic_read(&mm->mm_users) != 0));
 173
 174         /*
 175          * mlock:   don't page populate if vma has PROT_NONE permission.
 176          * munlock: always do munlock although the vma has PROT_NONE
 177          *          permission, or SIGKILL is pending.
 178          */
 179         if (!mlock)
 180                 gup_flags |= GUP_FLAGS_IGNORE_VMA_PERMISSIONS |
 181                              GUP_FLAGS_IGNORE_SIGKILL;
 182
 183         if (vma->vm_flags & VM_WRITE)
 184                 gup_flags |= GUP_FLAGS_WRITE;
 185
 186         while (nr_pages > 0) {
 187                 int i;
 188
 189                 cond_resched();
 190
 191                 /*
 192                  * get_user_pages makes pages present if we are
 193                  * setting mlock. and this extra reference count will
 194                  * disable migration of this page.  However, page may
 195                  * still be truncated out from under us.
 196                  */
 197                 ret = __get_user_pages(current, mm, addr,
 198                                 min_t(int, nr_pages, ARRAY_SIZE(pages)),
 199                                 gup_flags, pages, NULL);
 200                 /*
 201                  * This can happen for, e.g., VM_NONLINEAR regions before
 202                  * a page has been allocated and mapped at a given offset,
 203                  * or for addresses that map beyond end of a file.
 204                  * We'll mlock the the pages if/when they get faulted in.
 205                  */
 206                 if (ret < 0)
 207                         break;
 208                 if (ret == 0) {
 209                         /*
 210                          * We know the vma is there, so the only time
 211                          * we cannot get a single page should be an
 212                          * error (ret < 0) case.
 213                          */
 214                         WARN_ON(1);
 215                         break;
 216                 }
 217
 218                 lru_add_drain();        /* push cached pages to LRU */
 219
 220                 for (i = 0; i < ret; i++) {
 221                         struct page *page = pages[i];
 222
 223                         lock_page(page);
 224                         /*
 225                          * Because we lock page here and migration is blocked
 226                          * by the elevated reference, we need only check for
 227                          * page truncation (file-cache only).
 228                          */
 229                         if (page->mapping) {
 230                                 if (mlock)
 231                                         mlock_vma_page(page);
 232                                 else
 233                                         munlock_vma_page(page);
 234                         }
 235                         unlock_page(page);
 236                         put_page(page);         /* ref from get_user_pages() */
 237
 238                         /*
 239                          * here we assume that get_user_pages() has given us
 240                          * a list of virtually contiguous pages.
 241                          */
 242                         addr += PAGE_SIZE;      /* for next get_user_pages() */
 243                         nr_pages--;
 244                 }
 245                 ret = 0;
 246         }
 247
 248         return ret;     /* count entire vma as locked_vm */
 249 }
 250
 251 /*
 252  * convert get_user_pages() return value to posix mlock() error
 253  */
 254 static int __mlock_posix_error_return(long retval)
 255 {
 256         if (retval == -EFAULT)
 257                 retval = -ENOMEM;
 258         else if (retval == -ENOMEM)
 259                 retval = -EAGAIN;
 260         return retval;
 261 }
 262
 263 /**
 264  * mlock_vma_pages_range() - mlock pages in specified vma range.
 265  * @vma - the vma containing the specfied address range
 266  * @start - starting address in @vma to mlock
 267  * @end   - end address [+1] in @vma to mlock
 268  *
 269  * For mmap()/mremap()/expansion of mlocked vma.
 270  *
 271  * return 0 on success for "normal" vmas.
 272  *
 273  * return number of pages [> 0] to be removed from locked_vm on success
 274  * of "special" vmas.
 275  */
 276 long mlock_vma_pages_range(struct vm_area_struct *vma,
 277                         unsigned long start, unsigned long end)
 278 {
 279         int nr_pages = (end - start) / PAGE_SIZE;
 280         BUG_ON(!(vma->vm_flags & VM_LOCKED));
 281
 282         /*
 283          * filter unlockable vmas
 284          */
 285         if (vma->vm_flags & (VM_IO | VM_PFNMAP))
 286                 goto no_mlock;
 287
 288         if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 289                         is_vm_hugetlb_page(vma) ||
 290                         vma == get_gate_vma(current))) {
 291
 292                 __mlock_vma_pages_range(vma, start, end, 1);
 293
 294                 /* Hide errors from mmap() and other callers */
 295                 return 0;
 296         }
 297
 298         /*
 299          * User mapped kernel pages or huge pages:
 300          * make these pages present to populate the ptes, but
 301          * fall thru' to reset VM_LOCKED--no need to unlock, and
 302          * return nr_pages so these don't get counted against task's
 303          * locked limit.  huge pages are already counted against
 304          * locked vm limit.
 305          */
 306         make_pages_present(start, end);
 307
 308 no_mlock:
 309         vma->vm_flags &= ~VM_LOCKED;    /* and don't come back! */
 310         return nr_pages;                /* error or pages NOT mlocked */
 311 }
 312
 313
 314 /*
 315  * munlock_vma_pages_range() - munlock all pages in the vma range.'
 316  * @vma - vma containing range to be munlock()ed.
 317  * @start - start address in @vma of the range
 318  * @end - end of range in @vma.
 319  *
 320  *  For mremap(), munmap() and exit().
 321  *
 322  * Called with @vma VM_LOCKED.
 323  *
 324  * Returns with VM_LOCKED cleared.  Callers must be prepared to
 325  * deal with this.
 326  *
 327  * We don't save and restore VM_LOCKED here because pages are
 328  * still on lru.  In unmap path, pages might be scanned by reclaim
 329  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 330  * free them.  This will result in freeing mlocked pages.
 331  */
 332 void munlock_vma_pages_range(struct vm_area_struct *vma,
 333                            unsigned long start, unsigned long end)
 334 {
 335         vma->vm_flags &= ~VM_LOCKED;
 336         __mlock_vma_pages_range(vma, start, end, 0);
 337 }
 338
 339 /*
 340  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 341  *
 342  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 343  * munlock is a no-op.  However, for some special vmas, we go ahead and
 344  * populate the ptes via make_pages_present().
 345  *
 346  * For vmas that pass the filters, merge/split as appropriate.
 347  */
 348 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 349         unsigned long start, unsigned long end, unsigned int newflags)
 350 {
 351         struct mm_struct *mm = vma->vm_mm;
 352         pgoff_t pgoff;
 353         int nr_pages;
 354         int ret = 0;
 355         int lock = newflags & VM_LOCKED;
 356
 357         if (newflags == vma->vm_flags ||
 358                         (vma->vm_flags & (VM_IO | VM_PFNMAP)))
 359                 goto out;       /* don't set VM_LOCKED,  don't count */
 360
 361         if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 362                         is_vm_hugetlb_page(vma) ||
 363                         vma == get_gate_vma(current)) {
 364                 if (lock)
 365                         make_pages_present(start, end);
 366                 goto out;       /* don't set VM_LOCKED,  don't count */
 367         }
 368
 369         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 370         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 371                           vma->vm_file, pgoff, vma_policy(vma));
 372         if (*prev) {
 373                 vma = *prev;
 374                 goto success;
 375         }
 376
 377         if (start != vma->vm_start) {
 378                 ret = split_vma(mm, vma, start, 1);
 379                 if (ret)
 380                         goto out;
 381         }
 382
 383         if (end != vma->vm_end) {
 384                 ret = split_vma(mm, vma, end, 0);
 385                 if (ret)
 386                         goto out;
 387         }
 388
 389 success:
 390         /*
 391          * Keep track of amount of locked VM.
 392          */
 393         nr_pages = (end - start) >> PAGE_SHIFT;
 394         if (!lock)
 395                 nr_pages = -nr_pages;
 396         mm->locked_vm += nr_pages;
 397
 398         /*
 399          * vm_flags is protected by the mmap_sem held in write mode.
 400          * It's okay if try_to_unmap_one unmaps a page just after we
 401          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 402          */
 403         vma->vm_flags = newflags;
 404
 405         if (lock) {
 406                 ret = __mlock_vma_pages_range(vma, start, end, 1);
 407
 408                 if (ret > 0) {
 409                         mm->locked_vm -= ret;
 410                         ret = 0;
 411                 } else
 412                         ret = __mlock_posix_error_return(ret); /* translate if needed */
 413         } else {
 414                 __mlock_vma_pages_range(vma, start, end, 0);
 415         }
 416
 417 out:
 418         *prev = vma;
 419         return ret;
 420 }
 421
 422 static int do_mlock(unsigned long start, size_t len, int on)
 423 {
 424         unsigned long nstart, end, tmp;
 425         struct vm_area_struct * vma, * prev;
 426         int error;
 427
 428         len = PAGE_ALIGN(len);
 429         end = start + len;
 430         if (end < start)
 431                 return -EINVAL;
 432         if (end == start)
 433                 return 0;
 434         vma = find_vma_prev(current->mm, start, &prev);
 435         if (!vma || vma->vm_start > start)
 436                 return -ENOMEM;
 437
 438         if (start > vma->vm_start)
 439                 prev = vma;
 440
 441         for (nstart = start ; ; ) {
 442                 unsigned int newflags;
 443
 444                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 445
 446                 newflags = vma->vm_flags | VM_LOCKED;
 447                 if (!on)
 448                         newflags &= ~VM_LOCKED;
 449
 450                 tmp = vma->vm_end;
 451                 if (tmp > end)
 452                         tmp = end;
 453                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 454                 if (error)
 455                         break;
 456                 nstart = tmp;
 457                 if (nstart < prev->vm_end)
 458                         nstart = prev->vm_end;
 459                 if (nstart >= end)
 460                         break;
 461
 462                 vma = prev->vm_next;
 463                 if (!vma || vma->vm_start != nstart) {
 464                         error = -ENOMEM;
 465                         break;
 466                 }
 467         }
 468         return error;
 469 }
 470
 471 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 472 {
 473         unsigned long locked;
 474         unsigned long lock_limit;
 475         int error = -ENOMEM;
 476
 477         if (!can_do_mlock())
 478                 return -EPERM;
 479
 480         lru_add_drain_all();    /* flush pagevec */
 481
 482         down_write(&current->mm->mmap_sem);
 483         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 484         start &= PAGE_MASK;
 485
 486         locked = len >> PAGE_SHIFT;
 487         locked += current->mm->locked_vm;
 488
 489         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 490         lock_limit >>= PAGE_SHIFT;
 491
 492         /* check against resource limits */
 493         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 494                 error = do_mlock(start, len, 1);
 495         up_write(&current->mm->mmap_sem);
 496         return error;
 497 }
 498
 499 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 500 {
 501         int ret;
 502
 503         down_write(&current->mm->mmap_sem);
 504         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 505         start &= PAGE_MASK;
 506         ret = do_mlock(start, len, 0);
 507         up_write(&current->mm->mmap_sem);
 508         return ret;
 509 }
 510
 511 static int do_mlockall(int flags)
 512 {
 513         struct vm_area_struct * vma, * prev = NULL;
 514         unsigned int def_flags = 0;
 515
 516         if (flags & MCL_FUTURE)
 517                 def_flags = VM_LOCKED;
 518         current->mm->def_flags = def_flags;
 519         if (flags == MCL_FUTURE)
 520                 goto out;
 521
 522         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 523                 unsigned int newflags;
 524
 525                 newflags = vma->vm_flags | VM_LOCKED;
 526                 if (!(flags & MCL_CURRENT))
 527                         newflags &= ~VM_LOCKED;
 528
 529                 /* Ignore errors */
 530                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 531         }
 532 out:
 533         return 0;
 534 }
 535
 536 SYSCALL_DEFINE1(mlockall, int, flags)
 537 {
 538         unsigned long lock_limit;
 539         int ret = -EINVAL;
 540
 541         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 542                 goto out;
 543
 544         ret = -EPERM;
 545         if (!can_do_mlock())
 546                 goto out;
 547
 548         lru_add_drain_all();    /* flush pagevec */
 549
 550         down_write(&current->mm->mmap_sem);
 551
 552         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 553         lock_limit >>= PAGE_SHIFT;
 554
 555         ret = -ENOMEM;
 556         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 557             capable(CAP_IPC_LOCK))
 558                 ret = do_mlockall(flags);
 559         up_write(&current->mm->mmap_sem);
 560 out:
 561         return ret;
 562 }
 563
 564 SYSCALL_DEFINE0(munlockall)
 565 {
 566         int ret;
 567
 568         down_write(&current->mm->mmap_sem);
 569         ret = do_mlockall(0);
 570         up_write(&current->mm->mmap_sem);
 571         return ret;
 572 }
 573
 574 /*
 575  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 576  * shm segments) get accounted against the user_struct instead.
 577  */
 578 static DEFINE_SPINLOCK(shmlock_user_lock);
 579
 580 int user_shm_lock(size_t size, struct user_struct *user)
 581 {
 582         unsigned long lock_limit, locked;
 583         int allowed = 0;
 584
 585         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 586         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 587         if (lock_limit == RLIM_INFINITY)
 588                 allowed = 1;
 589         lock_limit >>= PAGE_SHIFT;
 590         spin_lock(&shmlock_user_lock);
 591         if (!allowed &&
 592             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 593                 goto out;
 594         get_uid(user);
 595         user->locked_shm += locked;
 596         allowed = 1;
 597 out:
 598         spin_unlock(&shmlock_user_lock);
 599         return allowed;
 600 }
 601
 602 void user_shm_unlock(size_t size, struct user_struct *user)
 603 {
 604         spin_lock(&shmlock_user_lock);
 605         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 606         spin_unlock(&shmlock_user_lock);
 607         free_uid(user);
 608 }
 609
 610 int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim,
 611                           size_t size)
 612 {
 613         unsigned long lim, vm, pgsz;
 614         int error = -ENOMEM;
 615
 616         pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
 617
 618         down_write(&mm->mmap_sem);
 619
 620         lim = rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
 621         vm   = mm->total_vm + pgsz;
 622         if (lim < vm)
 623                 goto out;
 624
 625         lim = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
 626         vm   = mm->locked_vm + pgsz;
 627         if (lim < vm)
 628                 goto out;
 629
 630         mm->total_vm  += pgsz;
 631         mm->locked_vm += pgsz;
 632
 633         error = 0;
 634  out:
 635         up_write(&mm->mmap_sem);
 636         return error;
 637 }
 638
 639 void refund_locked_memory(struct mm_struct *mm, size_t size)
 640 {
 641         unsigned long pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT;
 642
 643         down_write(&mm->mmap_sem);
 644
 645         mm->total_vm  -= pgsz;
 646         mm->locked_vm -= pgsz;
 647
 648         up_write(&mm->mmap_sem);
 649 }