mm/mlock.c

   1 /*
   2  *      linux/mm/mlock.c
   3  *
   4  *  (C) Copyright 1995 Linus Torvalds
   5  *  (C) Copyright 2002 Christoph Hellwig
   6  */
   7
   8 #include <linux/capability.h>
   9 #include <linux/mman.h>
  10 #include <linux/mm.h>
  11 #include <linux/swap.h>
  12 #include <linux/swapops.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/mempolicy.h>
  15 #include <linux/syscalls.h>
  16 #include <linux/sched.h>
  17 #include <linux/export.h>
  18 #include <linux/rmap.h>
  19 #include <linux/mmzone.h>
  20 #include <linux/hugetlb.h>
  21
  22 #include "internal.h"
  23
  24 int can_do_mlock(void)
  25 {
  26         if (capable(CAP_IPC_LOCK))
  27                 return 1;
  28         if (rlimit(RLIMIT_MEMLOCK) != 0)
  29                 return 1;
  30         return 0;
  31 }
  32 EXPORT_SYMBOL(can_do_mlock);
  33
  34 /*
  35  * Mlocked pages are marked with PageMlocked() flag for efficient testing
  36  * in vmscan and, possibly, the fault path; and to support semi-accurate
  37  * statistics.
  38  *
  39  * An mlocked page [PageMlocked(page)] is unevictable.  As such, it will
  40  * be placed on the LRU "unevictable" list, rather than the [in]active lists.
  41  * The unevictable list is an LRU sibling list to the [in]active lists.
  42  * PageUnevictable is set to indicate the unevictable state.
  43  *
  44  * When lazy mlocking via vmscan, it is important to ensure that the
  45  * vma's VM_LOCKED status is not concurrently being modified, otherwise we
  46  * may have mlocked a page that is being munlocked. So lazy mlock must take
  47  * the mmap_sem for read, and verify that the vma really is locked
  48  * (see mm/rmap.c).
  49  */
  50
  51 /*
  52  *  LRU accounting for clear_page_mlock()
  53  */
  54 void clear_page_mlock(struct page *page)
  55 {
  56         if (!TestClearPageMlocked(page))
  57                 return;
  58
  59         mod_zone_page_state(page_zone(page), NR_MLOCK,
  60                             -hpage_nr_pages(page));
  61         count_vm_event(UNEVICTABLE_PGCLEARED);
  62         if (!isolate_lru_page(page)) {
  63                 putback_lru_page(page);
  64         } else {
  65                 /*
  66                  * We lost the race. the page already moved to evictable list.
  67                  */
  68                 if (PageUnevictable(page))
  69                         count_vm_event(UNEVICTABLE_PGSTRANDED);
  70         }
  71 }
  72
  73 /*
  74  * Mark page as mlocked if not already.
  75  * If page on LRU, isolate and putback to move to unevictable list.
  76  */
  77 void mlock_vma_page(struct page *page)
  78 {
  79         BUG_ON(!PageLocked(page));
  80
  81         if (!TestSetPageMlocked(page)) {
  82                 mod_zone_page_state(page_zone(page), NR_MLOCK,
  83                                     hpage_nr_pages(page));
  84                 count_vm_event(UNEVICTABLE_PGMLOCKED);
  85                 if (!isolate_lru_page(page))
  86                         putback_lru_page(page);
  87         }
  88 }
  89
  90 /**
  91  * munlock_vma_page - munlock a vma page
  92  * @page - page to be unlocked
  93  *
  94  * called from munlock()/munmap() path with page supposedly on the LRU.
  95  * When we munlock a page, because the vma where we found the page is being
  96  * munlock()ed or munmap()ed, we want to check whether other vmas hold the
  97  * page locked so that we can leave it on the unevictable lru list and not
  98  * bother vmscan with it.  However, to walk the page's rmap list in
  99  * try_to_munlock() we must isolate the page from the LRU.  If some other
 100  * task has removed the page from the LRU, we won't be able to do that.
 101  * So we clear the PageMlocked as we might not get another chance.  If we
 102  * can't isolate the page, we leave it for putback_lru_page() and vmscan
 103  * [page_referenced()/try_to_unmap()] to deal with.
 104  */
 105 void munlock_vma_page(struct page *page)
 106 {
 107         BUG_ON(!PageLocked(page));
 108
 109         if (TestClearPageMlocked(page)) {
 110                 mod_zone_page_state(page_zone(page), NR_MLOCK,
 111                                     -hpage_nr_pages(page));
 112                 if (!isolate_lru_page(page)) {
 113                         int ret = SWAP_AGAIN;
 114
 115                         /*
 116                          * Optimization: if the page was mapped just once,
 117                          * that's our mapping and we don't need to check all the
 118                          * other vmas.
 119                          */
 120                         if (page_mapcount(page) > 1)
 121                                 ret = try_to_munlock(page);
 122                         /*
 123                          * did try_to_unlock() succeed or punt?
 124                          */
 125                         if (ret != SWAP_MLOCK)
 126                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 127
 128                         putback_lru_page(page);
 129                 } else {
 130                         /*
 131                          * Some other task has removed the page from the LRU.
 132                          * putback_lru_page() will take care of removing the
 133                          * page from the unevictable list, if necessary.
 134                          * vmscan [page_referenced()] will move the page back
 135                          * to the unevictable list if some other vma has it
 136                          * mlocked.
 137                          */
 138                         if (PageUnevictable(page))
 139                                 count_vm_event(UNEVICTABLE_PGSTRANDED);
 140                         else
 141                                 count_vm_event(UNEVICTABLE_PGMUNLOCKED);
 142                 }
 143         }
 144 }
 145
 146 /**
 147  * __mlock_vma_pages_range() -  mlock a range of pages in the vma.
 148  * @vma:   target vma
 149  * @start: start address
 150  * @end:   end address
 151  *
 152  * This takes care of making the pages present too.
 153  *
 154  * return 0 on success, negative error code on error.
 155  *
 156  * vma->vm_mm->mmap_sem must be held for at least read.
 157  */
 158 long __mlock_vma_pages_range(struct vm_area_struct *vma,
 159                 unsigned long start, unsigned long end, int *nonblocking)
 160 {
 161         struct mm_struct *mm = vma->vm_mm;
 162         unsigned long addr = start;
 163         unsigned long 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 | FOLL_MLOCK;
 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         /*
 189          * We made sure addr is within a VMA, so the following will
 190          * not result in a stack expansion that recurses back here.
 191          */
 192         return __get_user_pages(current, mm, addr, nr_pages, gup_flags,
 193                                 NULL, NULL, nonblocking);
 194 }
 195
 196 /*
 197  * convert get_user_pages() return value to posix mlock() error
 198  */
 199 static int __mlock_posix_error_return(long retval)
 200 {
 201         if (retval == -EFAULT)
 202                 retval = -ENOMEM;
 203         else if (retval == -ENOMEM)
 204                 retval = -EAGAIN;
 205         return retval;
 206 }
 207
 208 /*
 209  * munlock_vma_pages_range() - munlock all pages in the vma range.'
 210  * @vma - vma containing range to be munlock()ed.
 211  * @start - start address in @vma of the range
 212  * @end - end of range in @vma.
 213  *
 214  *  For mremap(), munmap() and exit().
 215  *
 216  * Called with @vma VM_LOCKED.
 217  *
 218  * Returns with VM_LOCKED cleared.  Callers must be prepared to
 219  * deal with this.
 220  *
 221  * We don't save and restore VM_LOCKED here because pages are
 222  * still on lru.  In unmap path, pages might be scanned by reclaim
 223  * and re-mlocked by try_to_{munlock|unmap} before we unmap and
 224  * free them.  This will result in freeing mlocked pages.
 225  */
 226 void munlock_vma_pages_range(struct vm_area_struct *vma,
 227                              unsigned long start, unsigned long end)
 228 {
 229         unsigned long addr;
 230
 231         lru_add_drain();
 232         vma->vm_flags &= ~VM_LOCKED;
 233
 234         for (addr = start; addr < end; addr += PAGE_SIZE) {
 235                 struct page *page;
 236                 /*
 237                  * Although FOLL_DUMP is intended for get_dump_page(),
 238                  * it just so happens that its special treatment of the
 239                  * ZERO_PAGE (returning an error instead of doing get_page)
 240                  * suits munlock very well (and if somehow an abnormal page
 241                  * has sneaked into the range, we won't oops here: great).
 242                  */
 243                 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
 244                 if (page && !IS_ERR(page)) {
 245                         lock_page(page);
 246                         munlock_vma_page(page);
 247                         unlock_page(page);
 248                         put_page(page);
 249                 }
 250                 cond_resched();
 251         }
 252 }
 253
 254 /*
 255  * mlock_fixup  - handle mlock[all]/munlock[all] requests.
 256  *
 257  * Filters out "special" vmas -- VM_LOCKED never gets set for these, and
 258  * munlock is a no-op.  However, for some special vmas, we go ahead and
 259  * populate the ptes.
 260  *
 261  * For vmas that pass the filters, merge/split as appropriate.
 262  */
 263 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 264         unsigned long start, unsigned long end, vm_flags_t newflags)
 265 {
 266         struct mm_struct *mm = vma->vm_mm;
 267         pgoff_t pgoff;
 268         int nr_pages;
 269         int ret = 0;
 270         int lock = !!(newflags & VM_LOCKED);
 271
 272         if (newflags == vma->vm_flags || (vma->vm_flags & VM_SPECIAL) ||
 273             is_vm_hugetlb_page(vma) || vma == get_gate_vma(current->mm))
 274                 goto out;       /* don't set VM_LOCKED,  don't count */
 275
 276         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 277         *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma,
 278                           vma->vm_file, pgoff, vma_policy(vma));
 279         if (*prev) {
 280                 vma = *prev;
 281                 goto success;
 282         }
 283
 284         if (start != vma->vm_start) {
 285                 ret = split_vma(mm, vma, start, 1);
 286                 if (ret)
 287                         goto out;
 288         }
 289
 290         if (end != vma->vm_end) {
 291                 ret = split_vma(mm, vma, end, 0);
 292                 if (ret)
 293                         goto out;
 294         }
 295
 296 success:
 297         /*
 298          * Keep track of amount of locked VM.
 299          */
 300         nr_pages = (end - start) >> PAGE_SHIFT;
 301         if (!lock)
 302                 nr_pages = -nr_pages;
 303         mm->locked_vm += nr_pages;
 304
 305         /*
 306          * vm_flags is protected by the mmap_sem held in write mode.
 307          * It's okay if try_to_unmap_one unmaps a page just after we
 308          * set VM_LOCKED, __mlock_vma_pages_range will bring it back.
 309          */
 310
 311         if (lock)
 312                 vma->vm_flags = newflags;
 313         else
 314                 munlock_vma_pages_range(vma, start, end);
 315
 316 out:
 317         *prev = vma;
 318         return ret;
 319 }
 320
 321 static int do_mlock(unsigned long start, size_t len, int on)
 322 {
 323         unsigned long nstart, end, tmp;
 324         struct vm_area_struct * vma, * prev;
 325         int error;
 326
 327         VM_BUG_ON(start & ~PAGE_MASK);
 328         VM_BUG_ON(len != PAGE_ALIGN(len));
 329         end = start + len;
 330         if (end < start)
 331                 return -EINVAL;
 332         if (end == start)
 333                 return 0;
 334         vma = find_vma(current->mm, start);
 335         if (!vma || vma->vm_start > start)
 336                 return -ENOMEM;
 337
 338         prev = vma->vm_prev;
 339         if (start > vma->vm_start)
 340                 prev = vma;
 341
 342         for (nstart = start ; ; ) {
 343                 vm_flags_t newflags;
 344
 345                 /* Here we know that  vma->vm_start <= nstart < vma->vm_end. */
 346
 347                 newflags = vma->vm_flags & ~VM_LOCKED;
 348                 if (on)
 349                         newflags |= VM_LOCKED | VM_POPULATE;
 350
 351                 tmp = vma->vm_end;
 352                 if (tmp > end)
 353                         tmp = end;
 354                 error = mlock_fixup(vma, &prev, nstart, tmp, newflags);
 355                 if (error)
 356                         break;
 357                 nstart = tmp;
 358                 if (nstart < prev->vm_end)
 359                         nstart = prev->vm_end;
 360                 if (nstart >= end)
 361                         break;
 362
 363                 vma = prev->vm_next;
 364                 if (!vma || vma->vm_start != nstart) {
 365                         error = -ENOMEM;
 366                         break;
 367                 }
 368         }
 369         return error;
 370 }
 371
 372 /*
 373  * __mm_populate - populate and/or mlock pages within a range of address space.
 374  *
 375  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
 376  * flags. VMAs must be already marked with the desired vm_flags, and
 377  * mmap_sem must not be held.
 378  */
 379 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
 380 {
 381         struct mm_struct *mm = current->mm;
 382         unsigned long end, nstart, nend;
 383         struct vm_area_struct *vma = NULL;
 384         int locked = 0;
 385         long ret = 0;
 386
 387         VM_BUG_ON(start & ~PAGE_MASK);
 388         VM_BUG_ON(len != PAGE_ALIGN(len));
 389         end = start + len;
 390
 391         for (nstart = start; nstart < end; nstart = nend) {
 392                 /*
 393                  * We want to fault in pages for [nstart; end) address range.
 394                  * Find first corresponding VMA.
 395                  */
 396                 if (!locked) {
 397                         locked = 1;
 398                         down_read(&mm->mmap_sem);
 399                         vma = find_vma(mm, nstart);
 400                 } else if (nstart >= vma->vm_end)
 401                         vma = vma->vm_next;
 402                 if (!vma || vma->vm_start >= end)
 403                         break;
 404                 /*
 405                  * Set [nstart; nend) to intersection of desired address
 406                  * range with the first VMA. Also, skip undesirable VMA types.
 407                  */
 408                 nend = min(end, vma->vm_end);
 409                 if ((vma->vm_flags & (VM_IO | VM_PFNMAP | VM_POPULATE)) !=
 410                     VM_POPULATE)
 411                         continue;
 412                 if (nstart < vma->vm_start)
 413                         nstart = vma->vm_start;
 414                 /*
 415                  * Now fault in a range of pages. __mlock_vma_pages_range()
 416                  * double checks the vma flags, so that it won't mlock pages
 417                  * if the vma was already munlocked.
 418                  */
 419                 ret = __mlock_vma_pages_range(vma, nstart, nend, &locked);
 420                 if (ret < 0) {
 421                         if (ignore_errors) {
 422                                 ret = 0;
 423                                 continue;       /* continue at next VMA */
 424                         }
 425                         ret = __mlock_posix_error_return(ret);
 426                         break;
 427                 }
 428                 nend = nstart + ret * PAGE_SIZE;
 429                 ret = 0;
 430         }
 431         if (locked)
 432                 up_read(&mm->mmap_sem);
 433         return ret;     /* 0 or negative error code */
 434 }
 435
 436 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len)
 437 {
 438         unsigned long locked;
 439         unsigned long lock_limit;
 440         int error = -ENOMEM;
 441
 442         if (!can_do_mlock())
 443                 return -EPERM;
 444
 445         lru_add_drain_all();    /* flush pagevec */
 446
 447         down_write(&current->mm->mmap_sem);
 448         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 449         start &= PAGE_MASK;
 450
 451         locked = len >> PAGE_SHIFT;
 452         locked += current->mm->locked_vm;
 453
 454         lock_limit = rlimit(RLIMIT_MEMLOCK);
 455         lock_limit >>= PAGE_SHIFT;
 456
 457         /* check against resource limits */
 458         if ((locked <= lock_limit) || capable(CAP_IPC_LOCK))
 459                 error = do_mlock(start, len, 1);
 460         up_write(&current->mm->mmap_sem);
 461         if (!error)
 462                 error = __mm_populate(start, len, 0);
 463         return error;
 464 }
 465
 466 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len)
 467 {
 468         int ret;
 469
 470         down_write(&current->mm->mmap_sem);
 471         len = PAGE_ALIGN(len + (start & ~PAGE_MASK));
 472         start &= PAGE_MASK;
 473         ret = do_mlock(start, len, 0);
 474         up_write(&current->mm->mmap_sem);
 475         return ret;
 476 }
 477
 478 static int do_mlockall(int flags)
 479 {
 480         struct vm_area_struct * vma, * prev = NULL;
 481
 482         if (flags & MCL_FUTURE)
 483                 current->mm->def_flags |= VM_LOCKED | VM_POPULATE;
 484         else
 485                 current->mm->def_flags &= ~(VM_LOCKED | VM_POPULATE);
 486         if (flags == MCL_FUTURE)
 487                 goto out;
 488
 489         for (vma = current->mm->mmap; vma ; vma = prev->vm_next) {
 490                 vm_flags_t newflags;
 491
 492                 newflags = vma->vm_flags & ~VM_LOCKED;
 493                 if (flags & MCL_CURRENT)
 494                         newflags |= VM_LOCKED | VM_POPULATE;
 495
 496                 /* Ignore errors */
 497                 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags);
 498         }
 499 out:
 500         return 0;
 501 }
 502
 503 SYSCALL_DEFINE1(mlockall, int, flags)
 504 {
 505         unsigned long lock_limit;
 506         int ret = -EINVAL;
 507
 508         if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
 509                 goto out;
 510
 511         ret = -EPERM;
 512         if (!can_do_mlock())
 513                 goto out;
 514
 515         if (flags & MCL_CURRENT)
 516                 lru_add_drain_all();    /* flush pagevec */
 517
 518         down_write(&current->mm->mmap_sem);
 519
 520         lock_limit = rlimit(RLIMIT_MEMLOCK);
 521         lock_limit >>= PAGE_SHIFT;
 522
 523         ret = -ENOMEM;
 524         if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) ||
 525             capable(CAP_IPC_LOCK))
 526                 ret = do_mlockall(flags);
 527         up_write(&current->mm->mmap_sem);
 528         if (!ret && (flags & MCL_CURRENT))
 529                 mm_populate(0, TASK_SIZE);
 530 out:
 531         return ret;
 532 }
 533
 534 SYSCALL_DEFINE0(munlockall)
 535 {
 536         int ret;
 537
 538         down_write(&current->mm->mmap_sem);
 539         ret = do_mlockall(0);
 540         up_write(&current->mm->mmap_sem);
 541         return ret;
 542 }
 543
 544 /*
 545  * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB
 546  * shm segments) get accounted against the user_struct instead.
 547  */
 548 static DEFINE_SPINLOCK(shmlock_user_lock);
 549
 550 int user_shm_lock(size_t size, struct user_struct *user)
 551 {
 552         unsigned long lock_limit, locked;
 553         int allowed = 0;
 554
 555         locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 556         lock_limit = rlimit(RLIMIT_MEMLOCK);
 557         if (lock_limit == RLIM_INFINITY)
 558                 allowed = 1;
 559         lock_limit >>= PAGE_SHIFT;
 560         spin_lock(&shmlock_user_lock);
 561         if (!allowed &&
 562             locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK))
 563                 goto out;
 564         get_uid(user);
 565         user->locked_shm += locked;
 566         allowed = 1;
 567 out:
 568         spin_unlock(&shmlock_user_lock);
 569         return allowed;
 570 }
 571
 572 void user_shm_unlock(size_t size, struct user_struct *user)
 573 {
 574         spin_lock(&shmlock_user_lock);
 575         user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 576         spin_unlock(&shmlock_user_lock);
 577         free_uid(user);
 578 }