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