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 /* Is the vma a continuation of the stack vma above it? */ 139 static inline int vma_stack_continue(struct vm_area_struct *vma, unsigned long addr) 140 { 141 return vma && (vma->vm_end == addr) && (vma->vm_flags & VM_GROWSDOWN); 142 } 143 144 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr) 145 { 146 return (vma->vm_flags & VM_GROWSDOWN) && 147 (vma->vm_start == addr) && 148 !vma_stack_continue(vma->vm_prev, addr); 149 } 150 151 /** 152 * __mlock_vma_pages_range() - mlock a range of pages in the vma. 153 * @vma: target vma 154 * @start: start address 155 * @end: end address 156 * 157 * This takes care of making the pages present too. 158 * 159 * return 0 on success, negative error code on error. 160 * 161 * vma->vm_mm->mmap_sem must be held for at least read. 162 */ 163 static long __mlock_vma_pages_range(struct vm_area_struct *vma, 164 unsigned long start, unsigned long end) 165 { 166 struct mm_struct *mm = vma->vm_mm; 167 unsigned long addr = start; 168 struct page *pages[16]; /* 16 gives a reasonable batch */ 169 int nr_pages = (end - start) / PAGE_SIZE; 170 int ret = 0; 171 int gup_flags; 172 173 VM_BUG_ON(start & ~PAGE_MASK); 174 VM_BUG_ON(end & ~PAGE_MASK); 175 VM_BUG_ON(start < vma->vm_start); 176 VM_BUG_ON(end > vma->vm_end); 177 VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); 178 179 gup_flags = FOLL_TOUCH | FOLL_GET; 180 if (vma->vm_flags & VM_WRITE) 181 gup_flags |= FOLL_WRITE; 182 183 /* We don't try to access the guard page of a stack vma */ 184 if (stack_guard_page(vma, start)) { 185 addr += PAGE_SIZE; 186 nr_pages--; 187 } 188 189 while (nr_pages > 0) { 190 int i; 191 192 cond_resched(); 193 194 /* 195 * get_user_pages makes pages present if we are 196 * setting mlock. and this extra reference count will 197 * disable migration of this page. However, page may 198 * still be truncated out from under us. 199 */ 200 ret = __get_user_pages(current, mm, addr, 201 min_t(int, nr_pages, ARRAY_SIZE(pages)), 202 gup_flags, pages, NULL); 203 /* 204 * This can happen for, e.g., VM_NONLINEAR regions before 205 * a page has been allocated and mapped at a given offset, 206 * or for addresses that map beyond end of a file. 207 * We'll mlock the pages if/when they get faulted in. 208 */ 209 if (ret < 0) 210 break; 211 212 lru_add_drain(); /* push cached pages to LRU */ 213 214 for (i = 0; i < ret; i++) { 215 struct page *page = pages[i]; 216 217 if (page->mapping) { 218 /* 219 * That preliminary check is mainly to avoid 220 * the pointless overhead of lock_page on the 221 * ZERO_PAGE: which might bounce very badly if 222 * there is contention. However, we're still 223 * dirtying its cacheline with get/put_page: 224 * we'll add another __get_user_pages flag to 225 * avoid it if that case turns out to matter. 226 */ 227 lock_page(page); 228 /* 229 * Because we lock page here and migration is 230 * blocked by the elevated reference, we need 231 * only check for file-cache page truncation. 232 */ 233 if (page->mapping) 234 mlock_vma_page(page); 235 unlock_page(page); 236 } 237 put_page(page); /* ref from get_user_pages() */ 238 } 239 240 addr += ret * PAGE_SIZE; 241 nr_pages -= ret; 242 ret = 0; 243 } 244 245 return ret; /* 0 or negative error code */ 246 } 247 248 /* 249 * convert get_user_pages() return value to posix mlock() error 250 */ 251 static int __mlock_posix_error_return(long retval) 252 { 253 if (retval == -EFAULT) 254 retval = -ENOMEM; 255 else if (retval == -ENOMEM) 256 retval = -EAGAIN; 257 return retval; 258 } 259 260 /** 261 * mlock_vma_pages_range() - mlock pages in specified vma range. 262 * @vma - the vma containing the specfied address range 263 * @start - starting address in @vma to mlock 264 * @end - end address [+1] in @vma to mlock 265 * 266 * For mmap()/mremap()/expansion of mlocked vma. 267 * 268 * return 0 on success for "normal" vmas. 269 * 270 * return number of pages [> 0] to be removed from locked_vm on success 271 * of "special" vmas. 272 */ 273 long mlock_vma_pages_range(struct vm_area_struct *vma, 274 unsigned long start, unsigned long end) 275 { 276 int nr_pages = (end - start) / PAGE_SIZE; 277 BUG_ON(!(vma->vm_flags & VM_LOCKED)); 278 279 /* 280 * filter unlockable vmas 281 */ 282 if (vma->vm_flags & (VM_IO | VM_PFNMAP)) 283 goto no_mlock; 284 285 if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) || 286 is_vm_hugetlb_page(vma) || 287 vma == get_gate_vma(current))) { 288 289 __mlock_vma_pages_range(vma, start, end); 290 291 /* Hide errors from mmap() and other callers */ 292 return 0; 293 } 294 295 /* 296 * User mapped kernel pages or huge pages: 297 * make these pages present to populate the ptes, but 298 * fall thru' to reset VM_LOCKED--no need to unlock, and 299 * return nr_pages so these don't get counted against task's 300 * locked limit. huge pages are already counted against 301 * locked vm limit. 302 */ 303 make_pages_present(start, end); 304 305 no_mlock: 306 vma->vm_flags &= ~VM_LOCKED; /* and don't come back! */ 307 return nr_pages; /* error or pages NOT mlocked */ 308 } 309 310 /* 311 * munlock_vma_pages_range() - munlock all pages in the vma range.' 312 * @vma - vma containing range to be munlock()ed. 313 * @start - start address in @vma of the range 314 * @end - end of range in @vma. 315 * 316 * For mremap(), munmap() and exit(). 317 * 318 * Called with @vma VM_LOCKED. 319 * 320 * Returns with VM_LOCKED cleared. Callers must be prepared to 321 * deal with this. 322 * 323 * We don't save and restore VM_LOCKED here because pages are 324 * still on lru. In unmap path, pages might be scanned by reclaim 325 * and re-mlocked by try_to_{munlock|unmap} before we unmap and 326 * free them. This will result in freeing mlocked pages. 327 */ 328 void munlock_vma_pages_range(struct vm_area_struct *vma, 329 unsigned long start, unsigned long end) 330 { 331 unsigned long addr; 332 333 lru_add_drain(); 334 vma->vm_flags &= ~VM_LOCKED; 335 336 for (addr = start; addr < end; addr += PAGE_SIZE) { 337 struct page *page; 338 /* 339 * Although FOLL_DUMP is intended for get_dump_page(), 340 * it just so happens that its special treatment of the 341 * ZERO_PAGE (returning an error instead of doing get_page) 342 * suits munlock very well (and if somehow an abnormal page 343 * has sneaked into the range, we won't oops here: great). 344 */ 345 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP); 346 if (page && !IS_ERR(page)) { 347 lock_page(page); 348 /* 349 * Like in __mlock_vma_pages_range(), 350 * because we lock page here and migration is 351 * blocked by the elevated reference, we need 352 * only check for file-cache page truncation. 353 */ 354 if (page->mapping) 355 munlock_vma_page(page); 356 unlock_page(page); 357 put_page(page); 358 } 359 cond_resched(); 360 } 361 } 362 363 /* 364 * mlock_fixup - handle mlock[all]/munlock[all] requests. 365 * 366 * Filters out "special" vmas -- VM_LOCKED never gets set for these, and 367 * munlock is a no-op. However, for some special vmas, we go ahead and 368 * populate the ptes via make_pages_present(). 369 * 370 * For vmas that pass the filters, merge/split as appropriate. 371 */ 372 static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev, 373 unsigned long start, unsigned long end, unsigned int newflags) 374 { 375 struct mm_struct *mm = vma->vm_mm; 376 pgoff_t pgoff; 377 int nr_pages; 378 int ret = 0; 379 int lock = newflags & VM_LOCKED; 380 381 if (newflags == vma->vm_flags || 382 (vma->vm_flags & (VM_IO | VM_PFNMAP))) 383 goto out; /* don't set VM_LOCKED, don't count */ 384 385 if ((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) || 386 is_vm_hugetlb_page(vma) || 387 vma == get_gate_vma(current)) { 388 if (lock) 389 make_pages_present(start, end); 390 goto out; /* don't set VM_LOCKED, don't count */ 391 } 392 393 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); 394 *prev = vma_merge(mm, *prev, start, end, newflags, vma->anon_vma, 395 vma->vm_file, pgoff, vma_policy(vma)); 396 if (*prev) { 397 vma = *prev; 398 goto success; 399 } 400 401 if (start != vma->vm_start) { 402 ret = split_vma(mm, vma, start, 1); 403 if (ret) 404 goto out; 405 } 406 407 if (end != vma->vm_end) { 408 ret = split_vma(mm, vma, end, 0); 409 if (ret) 410 goto out; 411 } 412 413 success: 414 /* 415 * Keep track of amount of locked VM. 416 */ 417 nr_pages = (end - start) >> PAGE_SHIFT; 418 if (!lock) 419 nr_pages = -nr_pages; 420 mm->locked_vm += nr_pages; 421 422 /* 423 * vm_flags is protected by the mmap_sem held in write mode. 424 * It's okay if try_to_unmap_one unmaps a page just after we 425 * set VM_LOCKED, __mlock_vma_pages_range will bring it back. 426 */ 427 428 if (lock) { 429 vma->vm_flags = newflags; 430 ret = __mlock_vma_pages_range(vma, start, end); 431 if (ret < 0) 432 ret = __mlock_posix_error_return(ret); 433 } else { 434 munlock_vma_pages_range(vma, start, end); 435 } 436 437 out: 438 *prev = vma; 439 return ret; 440 } 441 442 static int do_mlock(unsigned long start, size_t len, int on) 443 { 444 unsigned long nstart, end, tmp; 445 struct vm_area_struct * vma, * prev; 446 int error; 447 448 len = PAGE_ALIGN(len); 449 end = start + len; 450 if (end < start) 451 return -EINVAL; 452 if (end == start) 453 return 0; 454 vma = find_vma_prev(current->mm, start, &prev); 455 if (!vma || vma->vm_start > start) 456 return -ENOMEM; 457 458 if (start > vma->vm_start) 459 prev = vma; 460 461 for (nstart = start ; ; ) { 462 unsigned int newflags; 463 464 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */ 465 466 newflags = vma->vm_flags | VM_LOCKED; 467 if (!on) 468 newflags &= ~VM_LOCKED; 469 470 tmp = vma->vm_end; 471 if (tmp > end) 472 tmp = end; 473 error = mlock_fixup(vma, &prev, nstart, tmp, newflags); 474 if (error) 475 break; 476 nstart = tmp; 477 if (nstart < prev->vm_end) 478 nstart = prev->vm_end; 479 if (nstart >= end) 480 break; 481 482 vma = prev->vm_next; 483 if (!vma || vma->vm_start != nstart) { 484 error = -ENOMEM; 485 break; 486 } 487 } 488 return error; 489 } 490 491 SYSCALL_DEFINE2(mlock, unsigned long, start, size_t, len) 492 { 493 unsigned long locked; 494 unsigned long lock_limit; 495 int error = -ENOMEM; 496 497 if (!can_do_mlock()) 498 return -EPERM; 499 500 lru_add_drain_all(); /* flush pagevec */ 501 502 down_write(¤t->mm->mmap_sem); 503 len = PAGE_ALIGN(len + (start & ~PAGE_MASK)); 504 start &= PAGE_MASK; 505 506 locked = len >> PAGE_SHIFT; 507 locked += current->mm->locked_vm; 508 509 lock_limit = rlimit(RLIMIT_MEMLOCK); 510 lock_limit >>= PAGE_SHIFT; 511 512 /* check against resource limits */ 513 if ((locked <= lock_limit) || capable(CAP_IPC_LOCK)) 514 error = do_mlock(start, len, 1); 515 up_write(¤t->mm->mmap_sem); 516 return error; 517 } 518 519 SYSCALL_DEFINE2(munlock, unsigned long, start, size_t, len) 520 { 521 int ret; 522 523 down_write(¤t->mm->mmap_sem); 524 len = PAGE_ALIGN(len + (start & ~PAGE_MASK)); 525 start &= PAGE_MASK; 526 ret = do_mlock(start, len, 0); 527 up_write(¤t->mm->mmap_sem); 528 return ret; 529 } 530 531 static int do_mlockall(int flags) 532 { 533 struct vm_area_struct * vma, * prev = NULL; 534 unsigned int def_flags = 0; 535 536 if (flags & MCL_FUTURE) 537 def_flags = VM_LOCKED; 538 current->mm->def_flags = def_flags; 539 if (flags == MCL_FUTURE) 540 goto out; 541 542 for (vma = current->mm->mmap; vma ; vma = prev->vm_next) { 543 unsigned int newflags; 544 545 newflags = vma->vm_flags | VM_LOCKED; 546 if (!(flags & MCL_CURRENT)) 547 newflags &= ~VM_LOCKED; 548 549 /* Ignore errors */ 550 mlock_fixup(vma, &prev, vma->vm_start, vma->vm_end, newflags); 551 } 552 out: 553 return 0; 554 } 555 556 SYSCALL_DEFINE1(mlockall, int, flags) 557 { 558 unsigned long lock_limit; 559 int ret = -EINVAL; 560 561 if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE))) 562 goto out; 563 564 ret = -EPERM; 565 if (!can_do_mlock()) 566 goto out; 567 568 lru_add_drain_all(); /* flush pagevec */ 569 570 down_write(¤t->mm->mmap_sem); 571 572 lock_limit = rlimit(RLIMIT_MEMLOCK); 573 lock_limit >>= PAGE_SHIFT; 574 575 ret = -ENOMEM; 576 if (!(flags & MCL_CURRENT) || (current->mm->total_vm <= lock_limit) || 577 capable(CAP_IPC_LOCK)) 578 ret = do_mlockall(flags); 579 up_write(¤t->mm->mmap_sem); 580 out: 581 return ret; 582 } 583 584 SYSCALL_DEFINE0(munlockall) 585 { 586 int ret; 587 588 down_write(¤t->mm->mmap_sem); 589 ret = do_mlockall(0); 590 up_write(¤t->mm->mmap_sem); 591 return ret; 592 } 593 594 /* 595 * Objects with different lifetime than processes (SHM_LOCK and SHM_HUGETLB 596 * shm segments) get accounted against the user_struct instead. 597 */ 598 static DEFINE_SPINLOCK(shmlock_user_lock); 599 600 int user_shm_lock(size_t size, struct user_struct *user) 601 { 602 unsigned long lock_limit, locked; 603 int allowed = 0; 604 605 locked = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 606 lock_limit = rlimit(RLIMIT_MEMLOCK); 607 if (lock_limit == RLIM_INFINITY) 608 allowed = 1; 609 lock_limit >>= PAGE_SHIFT; 610 spin_lock(&shmlock_user_lock); 611 if (!allowed && 612 locked + user->locked_shm > lock_limit && !capable(CAP_IPC_LOCK)) 613 goto out; 614 get_uid(user); 615 user->locked_shm += locked; 616 allowed = 1; 617 out: 618 spin_unlock(&shmlock_user_lock); 619 return allowed; 620 } 621 622 void user_shm_unlock(size_t size, struct user_struct *user) 623 { 624 spin_lock(&shmlock_user_lock); 625 user->locked_shm -= (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 626 spin_unlock(&shmlock_user_lock); 627 free_uid(user); 628 } 629