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