1 // SPDX-License-Identifier: GPL-2.0-only 2 #include <linux/kernel.h> 3 #include <linux/errno.h> 4 #include <linux/err.h> 5 #include <linux/spinlock.h> 6 7 #include <linux/mm.h> 8 #include <linux/memremap.h> 9 #include <linux/pagemap.h> 10 #include <linux/rmap.h> 11 #include <linux/swap.h> 12 #include <linux/swapops.h> 13 #include <linux/secretmem.h> 14 15 #include <linux/sched/signal.h> 16 #include <linux/rwsem.h> 17 #include <linux/hugetlb.h> 18 #include <linux/migrate.h> 19 #include <linux/mm_inline.h> 20 #include <linux/sched/mm.h> 21 22 #include <asm/mmu_context.h> 23 #include <asm/tlbflush.h> 24 25 #include "internal.h" 26 27 struct follow_page_context { 28 struct dev_pagemap *pgmap; 29 unsigned int page_mask; 30 }; 31 32 static inline void sanity_check_pinned_pages(struct page **pages, 33 unsigned long npages) 34 { 35 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 36 return; 37 38 /* 39 * We only pin anonymous pages if they are exclusive. Once pinned, we 40 * can no longer turn them possibly shared and PageAnonExclusive() will 41 * stick around until the page is freed. 42 * 43 * We'd like to verify that our pinned anonymous pages are still mapped 44 * exclusively. The issue with anon THP is that we don't know how 45 * they are/were mapped when pinning them. However, for anon 46 * THP we can assume that either the given page (PTE-mapped THP) or 47 * the head page (PMD-mapped THP) should be PageAnonExclusive(). If 48 * neither is the case, there is certainly something wrong. 49 */ 50 for (; npages; npages--, pages++) { 51 struct page *page = *pages; 52 struct folio *folio = page_folio(page); 53 54 if (!folio_test_anon(folio)) 55 continue; 56 if (!folio_test_large(folio) || folio_test_hugetlb(folio)) 57 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page), page); 58 else 59 /* Either a PTE-mapped or a PMD-mapped THP. */ 60 VM_BUG_ON_PAGE(!PageAnonExclusive(&folio->page) && 61 !PageAnonExclusive(page), page); 62 } 63 } 64 65 /* 66 * Return the folio with ref appropriately incremented, 67 * or NULL if that failed. 68 */ 69 static inline struct folio *try_get_folio(struct page *page, int refs) 70 { 71 struct folio *folio; 72 73 retry: 74 folio = page_folio(page); 75 if (WARN_ON_ONCE(folio_ref_count(folio) < 0)) 76 return NULL; 77 if (unlikely(!folio_ref_try_add_rcu(folio, refs))) 78 return NULL; 79 80 /* 81 * At this point we have a stable reference to the folio; but it 82 * could be that between calling page_folio() and the refcount 83 * increment, the folio was split, in which case we'd end up 84 * holding a reference on a folio that has nothing to do with the page 85 * we were given anymore. 86 * So now that the folio is stable, recheck that the page still 87 * belongs to this folio. 88 */ 89 if (unlikely(page_folio(page) != folio)) { 90 if (!put_devmap_managed_page_refs(&folio->page, refs)) 91 folio_put_refs(folio, refs); 92 goto retry; 93 } 94 95 return folio; 96 } 97 98 /** 99 * try_grab_folio() - Attempt to get or pin a folio. 100 * @page: pointer to page to be grabbed 101 * @refs: the value to (effectively) add to the folio's refcount 102 * @flags: gup flags: these are the FOLL_* flag values. 103 * 104 * "grab" names in this file mean, "look at flags to decide whether to use 105 * FOLL_PIN or FOLL_GET behavior, when incrementing the folio's refcount. 106 * 107 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the 108 * same time. (That's true throughout the get_user_pages*() and 109 * pin_user_pages*() APIs.) Cases: 110 * 111 * FOLL_GET: folio's refcount will be incremented by @refs. 112 * 113 * FOLL_PIN on large folios: folio's refcount will be incremented by 114 * @refs, and its compound_pincount will be incremented by @refs. 115 * 116 * FOLL_PIN on single-page folios: folio's refcount will be incremented by 117 * @refs * GUP_PIN_COUNTING_BIAS. 118 * 119 * Return: The folio containing @page (with refcount appropriately 120 * incremented) for success, or NULL upon failure. If neither FOLL_GET 121 * nor FOLL_PIN was set, that's considered failure, and furthermore, 122 * a likely bug in the caller, so a warning is also emitted. 123 */ 124 struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags) 125 { 126 if (flags & FOLL_GET) 127 return try_get_folio(page, refs); 128 else if (flags & FOLL_PIN) { 129 struct folio *folio; 130 131 /* 132 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a 133 * right zone, so fail and let the caller fall back to the slow 134 * path. 135 */ 136 if (unlikely((flags & FOLL_LONGTERM) && 137 !is_longterm_pinnable_page(page))) 138 return NULL; 139 140 /* 141 * CAUTION: Don't use compound_head() on the page before this 142 * point, the result won't be stable. 143 */ 144 folio = try_get_folio(page, refs); 145 if (!folio) 146 return NULL; 147 148 /* 149 * When pinning a large folio, use an exact count to track it. 150 * 151 * However, be sure to *also* increment the normal folio 152 * refcount field at least once, so that the folio really 153 * is pinned. That's why the refcount from the earlier 154 * try_get_folio() is left intact. 155 */ 156 if (folio_test_large(folio)) 157 atomic_add(refs, folio_pincount_ptr(folio)); 158 else 159 folio_ref_add(folio, 160 refs * (GUP_PIN_COUNTING_BIAS - 1)); 161 /* 162 * Adjust the pincount before re-checking the PTE for changes. 163 * This is essentially a smp_mb() and is paired with a memory 164 * barrier in page_try_share_anon_rmap(). 165 */ 166 smp_mb__after_atomic(); 167 168 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, refs); 169 170 return folio; 171 } 172 173 WARN_ON_ONCE(1); 174 return NULL; 175 } 176 177 static void gup_put_folio(struct folio *folio, int refs, unsigned int flags) 178 { 179 if (flags & FOLL_PIN) { 180 node_stat_mod_folio(folio, NR_FOLL_PIN_RELEASED, refs); 181 if (folio_test_large(folio)) 182 atomic_sub(refs, folio_pincount_ptr(folio)); 183 else 184 refs *= GUP_PIN_COUNTING_BIAS; 185 } 186 187 if (!put_devmap_managed_page_refs(&folio->page, refs)) 188 folio_put_refs(folio, refs); 189 } 190 191 /** 192 * try_grab_page() - elevate a page's refcount by a flag-dependent amount 193 * @page: pointer to page to be grabbed 194 * @flags: gup flags: these are the FOLL_* flag values. 195 * 196 * This might not do anything at all, depending on the flags argument. 197 * 198 * "grab" names in this file mean, "look at flags to decide whether to use 199 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount. 200 * 201 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same 202 * time. Cases: please see the try_grab_folio() documentation, with 203 * "refs=1". 204 * 205 * Return: true for success, or if no action was required (if neither FOLL_PIN 206 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or 207 * FOLL_PIN was set, but the page could not be grabbed. 208 */ 209 bool __must_check try_grab_page(struct page *page, unsigned int flags) 210 { 211 struct folio *folio = page_folio(page); 212 213 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN)); 214 if (WARN_ON_ONCE(folio_ref_count(folio) <= 0)) 215 return false; 216 217 if (flags & FOLL_GET) 218 folio_ref_inc(folio); 219 else if (flags & FOLL_PIN) { 220 /* 221 * Similar to try_grab_folio(): be sure to *also* 222 * increment the normal page refcount field at least once, 223 * so that the page really is pinned. 224 */ 225 if (folio_test_large(folio)) { 226 folio_ref_add(folio, 1); 227 atomic_add(1, folio_pincount_ptr(folio)); 228 } else { 229 folio_ref_add(folio, GUP_PIN_COUNTING_BIAS); 230 } 231 232 node_stat_mod_folio(folio, NR_FOLL_PIN_ACQUIRED, 1); 233 } 234 235 return true; 236 } 237 238 /** 239 * unpin_user_page() - release a dma-pinned page 240 * @page: pointer to page to be released 241 * 242 * Pages that were pinned via pin_user_pages*() must be released via either 243 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so 244 * that such pages can be separately tracked and uniquely handled. In 245 * particular, interactions with RDMA and filesystems need special handling. 246 */ 247 void unpin_user_page(struct page *page) 248 { 249 sanity_check_pinned_pages(&page, 1); 250 gup_put_folio(page_folio(page), 1, FOLL_PIN); 251 } 252 EXPORT_SYMBOL(unpin_user_page); 253 254 static inline struct folio *gup_folio_range_next(struct page *start, 255 unsigned long npages, unsigned long i, unsigned int *ntails) 256 { 257 struct page *next = nth_page(start, i); 258 struct folio *folio = page_folio(next); 259 unsigned int nr = 1; 260 261 if (folio_test_large(folio)) 262 nr = min_t(unsigned int, npages - i, 263 folio_nr_pages(folio) - folio_page_idx(folio, next)); 264 265 *ntails = nr; 266 return folio; 267 } 268 269 static inline struct folio *gup_folio_next(struct page **list, 270 unsigned long npages, unsigned long i, unsigned int *ntails) 271 { 272 struct folio *folio = page_folio(list[i]); 273 unsigned int nr; 274 275 for (nr = i + 1; nr < npages; nr++) { 276 if (page_folio(list[nr]) != folio) 277 break; 278 } 279 280 *ntails = nr - i; 281 return folio; 282 } 283 284 /** 285 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages 286 * @pages: array of pages to be maybe marked dirty, and definitely released. 287 * @npages: number of pages in the @pages array. 288 * @make_dirty: whether to mark the pages dirty 289 * 290 * "gup-pinned page" refers to a page that has had one of the get_user_pages() 291 * variants called on that page. 292 * 293 * For each page in the @pages array, make that page (or its head page, if a 294 * compound page) dirty, if @make_dirty is true, and if the page was previously 295 * listed as clean. In any case, releases all pages using unpin_user_page(), 296 * possibly via unpin_user_pages(), for the non-dirty case. 297 * 298 * Please see the unpin_user_page() documentation for details. 299 * 300 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is 301 * required, then the caller should a) verify that this is really correct, 302 * because _lock() is usually required, and b) hand code it: 303 * set_page_dirty_lock(), unpin_user_page(). 304 * 305 */ 306 void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages, 307 bool make_dirty) 308 { 309 unsigned long i; 310 struct folio *folio; 311 unsigned int nr; 312 313 if (!make_dirty) { 314 unpin_user_pages(pages, npages); 315 return; 316 } 317 318 sanity_check_pinned_pages(pages, npages); 319 for (i = 0; i < npages; i += nr) { 320 folio = gup_folio_next(pages, npages, i, &nr); 321 /* 322 * Checking PageDirty at this point may race with 323 * clear_page_dirty_for_io(), but that's OK. Two key 324 * cases: 325 * 326 * 1) This code sees the page as already dirty, so it 327 * skips the call to set_page_dirty(). That could happen 328 * because clear_page_dirty_for_io() called 329 * page_mkclean(), followed by set_page_dirty(). 330 * However, now the page is going to get written back, 331 * which meets the original intention of setting it 332 * dirty, so all is well: clear_page_dirty_for_io() goes 333 * on to call TestClearPageDirty(), and write the page 334 * back. 335 * 336 * 2) This code sees the page as clean, so it calls 337 * set_page_dirty(). The page stays dirty, despite being 338 * written back, so it gets written back again in the 339 * next writeback cycle. This is harmless. 340 */ 341 if (!folio_test_dirty(folio)) { 342 folio_lock(folio); 343 folio_mark_dirty(folio); 344 folio_unlock(folio); 345 } 346 gup_put_folio(folio, nr, FOLL_PIN); 347 } 348 } 349 EXPORT_SYMBOL(unpin_user_pages_dirty_lock); 350 351 /** 352 * unpin_user_page_range_dirty_lock() - release and optionally dirty 353 * gup-pinned page range 354 * 355 * @page: the starting page of a range maybe marked dirty, and definitely released. 356 * @npages: number of consecutive pages to release. 357 * @make_dirty: whether to mark the pages dirty 358 * 359 * "gup-pinned page range" refers to a range of pages that has had one of the 360 * pin_user_pages() variants called on that page. 361 * 362 * For the page ranges defined by [page .. page+npages], make that range (or 363 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the 364 * page range was previously listed as clean. 365 * 366 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is 367 * required, then the caller should a) verify that this is really correct, 368 * because _lock() is usually required, and b) hand code it: 369 * set_page_dirty_lock(), unpin_user_page(). 370 * 371 */ 372 void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages, 373 bool make_dirty) 374 { 375 unsigned long i; 376 struct folio *folio; 377 unsigned int nr; 378 379 for (i = 0; i < npages; i += nr) { 380 folio = gup_folio_range_next(page, npages, i, &nr); 381 if (make_dirty && !folio_test_dirty(folio)) { 382 folio_lock(folio); 383 folio_mark_dirty(folio); 384 folio_unlock(folio); 385 } 386 gup_put_folio(folio, nr, FOLL_PIN); 387 } 388 } 389 EXPORT_SYMBOL(unpin_user_page_range_dirty_lock); 390 391 static void unpin_user_pages_lockless(struct page **pages, unsigned long npages) 392 { 393 unsigned long i; 394 struct folio *folio; 395 unsigned int nr; 396 397 /* 398 * Don't perform any sanity checks because we might have raced with 399 * fork() and some anonymous pages might now actually be shared -- 400 * which is why we're unpinning after all. 401 */ 402 for (i = 0; i < npages; i += nr) { 403 folio = gup_folio_next(pages, npages, i, &nr); 404 gup_put_folio(folio, nr, FOLL_PIN); 405 } 406 } 407 408 /** 409 * unpin_user_pages() - release an array of gup-pinned pages. 410 * @pages: array of pages to be marked dirty and released. 411 * @npages: number of pages in the @pages array. 412 * 413 * For each page in the @pages array, release the page using unpin_user_page(). 414 * 415 * Please see the unpin_user_page() documentation for details. 416 */ 417 void unpin_user_pages(struct page **pages, unsigned long npages) 418 { 419 unsigned long i; 420 struct folio *folio; 421 unsigned int nr; 422 423 /* 424 * If this WARN_ON() fires, then the system *might* be leaking pages (by 425 * leaving them pinned), but probably not. More likely, gup/pup returned 426 * a hard -ERRNO error to the caller, who erroneously passed it here. 427 */ 428 if (WARN_ON(IS_ERR_VALUE(npages))) 429 return; 430 431 sanity_check_pinned_pages(pages, npages); 432 for (i = 0; i < npages; i += nr) { 433 folio = gup_folio_next(pages, npages, i, &nr); 434 gup_put_folio(folio, nr, FOLL_PIN); 435 } 436 } 437 EXPORT_SYMBOL(unpin_user_pages); 438 439 /* 440 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's 441 * lifecycle. Avoid setting the bit unless necessary, or it might cause write 442 * cache bouncing on large SMP machines for concurrent pinned gups. 443 */ 444 static inline void mm_set_has_pinned_flag(unsigned long *mm_flags) 445 { 446 if (!test_bit(MMF_HAS_PINNED, mm_flags)) 447 set_bit(MMF_HAS_PINNED, mm_flags); 448 } 449 450 #ifdef CONFIG_MMU 451 static struct page *no_page_table(struct vm_area_struct *vma, 452 unsigned int flags) 453 { 454 /* 455 * When core dumping an enormous anonymous area that nobody 456 * has touched so far, we don't want to allocate unnecessary pages or 457 * page tables. Return error instead of NULL to skip handle_mm_fault, 458 * then get_dump_page() will return NULL to leave a hole in the dump. 459 * But we can only make this optimization where a hole would surely 460 * be zero-filled if handle_mm_fault() actually did handle it. 461 */ 462 if ((flags & FOLL_DUMP) && 463 (vma_is_anonymous(vma) || !vma->vm_ops->fault)) 464 return ERR_PTR(-EFAULT); 465 return NULL; 466 } 467 468 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address, 469 pte_t *pte, unsigned int flags) 470 { 471 if (flags & FOLL_TOUCH) { 472 pte_t entry = *pte; 473 474 if (flags & FOLL_WRITE) 475 entry = pte_mkdirty(entry); 476 entry = pte_mkyoung(entry); 477 478 if (!pte_same(*pte, entry)) { 479 set_pte_at(vma->vm_mm, address, pte, entry); 480 update_mmu_cache(vma, address, pte); 481 } 482 } 483 484 /* Proper page table entry exists, but no corresponding struct page */ 485 return -EEXIST; 486 } 487 488 /* FOLL_FORCE can write to even unwritable PTEs in COW mappings. */ 489 static inline bool can_follow_write_pte(pte_t pte, struct page *page, 490 struct vm_area_struct *vma, 491 unsigned int flags) 492 { 493 /* If the pte is writable, we can write to the page. */ 494 if (pte_write(pte)) 495 return true; 496 497 /* Maybe FOLL_FORCE is set to override it? */ 498 if (!(flags & FOLL_FORCE)) 499 return false; 500 501 /* But FOLL_FORCE has no effect on shared mappings */ 502 if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED)) 503 return false; 504 505 /* ... or read-only private ones */ 506 if (!(vma->vm_flags & VM_MAYWRITE)) 507 return false; 508 509 /* ... or already writable ones that just need to take a write fault */ 510 if (vma->vm_flags & VM_WRITE) 511 return false; 512 513 /* 514 * See can_change_pte_writable(): we broke COW and could map the page 515 * writable if we have an exclusive anonymous page ... 516 */ 517 if (!page || !PageAnon(page) || !PageAnonExclusive(page)) 518 return false; 519 520 /* ... and a write-fault isn't required for other reasons. */ 521 if (vma_soft_dirty_enabled(vma) && !pte_soft_dirty(pte)) 522 return false; 523 return !userfaultfd_pte_wp(vma, pte); 524 } 525 526 static struct page *follow_page_pte(struct vm_area_struct *vma, 527 unsigned long address, pmd_t *pmd, unsigned int flags, 528 struct dev_pagemap **pgmap) 529 { 530 struct mm_struct *mm = vma->vm_mm; 531 struct page *page; 532 spinlock_t *ptl; 533 pte_t *ptep, pte; 534 int ret; 535 536 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 537 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == 538 (FOLL_PIN | FOLL_GET))) 539 return ERR_PTR(-EINVAL); 540 retry: 541 if (unlikely(pmd_bad(*pmd))) 542 return no_page_table(vma, flags); 543 544 ptep = pte_offset_map_lock(mm, pmd, address, &ptl); 545 pte = *ptep; 546 if (!pte_present(pte)) { 547 swp_entry_t entry; 548 /* 549 * KSM's break_ksm() relies upon recognizing a ksm page 550 * even while it is being migrated, so for that case we 551 * need migration_entry_wait(). 552 */ 553 if (likely(!(flags & FOLL_MIGRATION))) 554 goto no_page; 555 if (pte_none(pte)) 556 goto no_page; 557 entry = pte_to_swp_entry(pte); 558 if (!is_migration_entry(entry)) 559 goto no_page; 560 pte_unmap_unlock(ptep, ptl); 561 migration_entry_wait(mm, pmd, address); 562 goto retry; 563 } 564 if ((flags & FOLL_NUMA) && pte_protnone(pte)) 565 goto no_page; 566 567 page = vm_normal_page(vma, address, pte); 568 569 /* 570 * We only care about anon pages in can_follow_write_pte() and don't 571 * have to worry about pte_devmap() because they are never anon. 572 */ 573 if ((flags & FOLL_WRITE) && 574 !can_follow_write_pte(pte, page, vma, flags)) { 575 page = NULL; 576 goto out; 577 } 578 579 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) { 580 /* 581 * Only return device mapping pages in the FOLL_GET or FOLL_PIN 582 * case since they are only valid while holding the pgmap 583 * reference. 584 */ 585 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap); 586 if (*pgmap) 587 page = pte_page(pte); 588 else 589 goto no_page; 590 } else if (unlikely(!page)) { 591 if (flags & FOLL_DUMP) { 592 /* Avoid special (like zero) pages in core dumps */ 593 page = ERR_PTR(-EFAULT); 594 goto out; 595 } 596 597 if (is_zero_pfn(pte_pfn(pte))) { 598 page = pte_page(pte); 599 } else { 600 ret = follow_pfn_pte(vma, address, ptep, flags); 601 page = ERR_PTR(ret); 602 goto out; 603 } 604 } 605 606 if (!pte_write(pte) && gup_must_unshare(flags, page)) { 607 page = ERR_PTR(-EMLINK); 608 goto out; 609 } 610 611 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && 612 !PageAnonExclusive(page), page); 613 614 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */ 615 if (unlikely(!try_grab_page(page, flags))) { 616 page = ERR_PTR(-ENOMEM); 617 goto out; 618 } 619 /* 620 * We need to make the page accessible if and only if we are going 621 * to access its content (the FOLL_PIN case). Please see 622 * Documentation/core-api/pin_user_pages.rst for details. 623 */ 624 if (flags & FOLL_PIN) { 625 ret = arch_make_page_accessible(page); 626 if (ret) { 627 unpin_user_page(page); 628 page = ERR_PTR(ret); 629 goto out; 630 } 631 } 632 if (flags & FOLL_TOUCH) { 633 if ((flags & FOLL_WRITE) && 634 !pte_dirty(pte) && !PageDirty(page)) 635 set_page_dirty(page); 636 /* 637 * pte_mkyoung() would be more correct here, but atomic care 638 * is needed to avoid losing the dirty bit: it is easier to use 639 * mark_page_accessed(). 640 */ 641 mark_page_accessed(page); 642 } 643 out: 644 pte_unmap_unlock(ptep, ptl); 645 return page; 646 no_page: 647 pte_unmap_unlock(ptep, ptl); 648 if (!pte_none(pte)) 649 return NULL; 650 return no_page_table(vma, flags); 651 } 652 653 static struct page *follow_pmd_mask(struct vm_area_struct *vma, 654 unsigned long address, pud_t *pudp, 655 unsigned int flags, 656 struct follow_page_context *ctx) 657 { 658 pmd_t *pmd, pmdval; 659 spinlock_t *ptl; 660 struct page *page; 661 struct mm_struct *mm = vma->vm_mm; 662 663 pmd = pmd_offset(pudp, address); 664 /* 665 * The READ_ONCE() will stabilize the pmdval in a register or 666 * on the stack so that it will stop changing under the code. 667 */ 668 pmdval = READ_ONCE(*pmd); 669 if (pmd_none(pmdval)) 670 return no_page_table(vma, flags); 671 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) { 672 page = follow_huge_pmd(mm, address, pmd, flags); 673 if (page) 674 return page; 675 return no_page_table(vma, flags); 676 } 677 if (is_hugepd(__hugepd(pmd_val(pmdval)))) { 678 page = follow_huge_pd(vma, address, 679 __hugepd(pmd_val(pmdval)), flags, 680 PMD_SHIFT); 681 if (page) 682 return page; 683 return no_page_table(vma, flags); 684 } 685 retry: 686 if (!pmd_present(pmdval)) { 687 /* 688 * Should never reach here, if thp migration is not supported; 689 * Otherwise, it must be a thp migration entry. 690 */ 691 VM_BUG_ON(!thp_migration_supported() || 692 !is_pmd_migration_entry(pmdval)); 693 694 if (likely(!(flags & FOLL_MIGRATION))) 695 return no_page_table(vma, flags); 696 697 pmd_migration_entry_wait(mm, pmd); 698 pmdval = READ_ONCE(*pmd); 699 /* 700 * MADV_DONTNEED may convert the pmd to null because 701 * mmap_lock is held in read mode 702 */ 703 if (pmd_none(pmdval)) 704 return no_page_table(vma, flags); 705 goto retry; 706 } 707 if (pmd_devmap(pmdval)) { 708 ptl = pmd_lock(mm, pmd); 709 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap); 710 spin_unlock(ptl); 711 if (page) 712 return page; 713 } 714 if (likely(!pmd_trans_huge(pmdval))) 715 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); 716 717 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval)) 718 return no_page_table(vma, flags); 719 720 retry_locked: 721 ptl = pmd_lock(mm, pmd); 722 if (unlikely(pmd_none(*pmd))) { 723 spin_unlock(ptl); 724 return no_page_table(vma, flags); 725 } 726 if (unlikely(!pmd_present(*pmd))) { 727 spin_unlock(ptl); 728 if (likely(!(flags & FOLL_MIGRATION))) 729 return no_page_table(vma, flags); 730 pmd_migration_entry_wait(mm, pmd); 731 goto retry_locked; 732 } 733 if (unlikely(!pmd_trans_huge(*pmd))) { 734 spin_unlock(ptl); 735 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); 736 } 737 if (flags & FOLL_SPLIT_PMD) { 738 int ret; 739 page = pmd_page(*pmd); 740 if (is_huge_zero_page(page)) { 741 spin_unlock(ptl); 742 ret = 0; 743 split_huge_pmd(vma, pmd, address); 744 if (pmd_trans_unstable(pmd)) 745 ret = -EBUSY; 746 } else { 747 spin_unlock(ptl); 748 split_huge_pmd(vma, pmd, address); 749 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0; 750 } 751 752 return ret ? ERR_PTR(ret) : 753 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap); 754 } 755 page = follow_trans_huge_pmd(vma, address, pmd, flags); 756 spin_unlock(ptl); 757 ctx->page_mask = HPAGE_PMD_NR - 1; 758 return page; 759 } 760 761 static struct page *follow_pud_mask(struct vm_area_struct *vma, 762 unsigned long address, p4d_t *p4dp, 763 unsigned int flags, 764 struct follow_page_context *ctx) 765 { 766 pud_t *pud; 767 spinlock_t *ptl; 768 struct page *page; 769 struct mm_struct *mm = vma->vm_mm; 770 771 pud = pud_offset(p4dp, address); 772 if (pud_none(*pud)) 773 return no_page_table(vma, flags); 774 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) { 775 page = follow_huge_pud(mm, address, pud, flags); 776 if (page) 777 return page; 778 return no_page_table(vma, flags); 779 } 780 if (is_hugepd(__hugepd(pud_val(*pud)))) { 781 page = follow_huge_pd(vma, address, 782 __hugepd(pud_val(*pud)), flags, 783 PUD_SHIFT); 784 if (page) 785 return page; 786 return no_page_table(vma, flags); 787 } 788 if (pud_devmap(*pud)) { 789 ptl = pud_lock(mm, pud); 790 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap); 791 spin_unlock(ptl); 792 if (page) 793 return page; 794 } 795 if (unlikely(pud_bad(*pud))) 796 return no_page_table(vma, flags); 797 798 return follow_pmd_mask(vma, address, pud, flags, ctx); 799 } 800 801 static struct page *follow_p4d_mask(struct vm_area_struct *vma, 802 unsigned long address, pgd_t *pgdp, 803 unsigned int flags, 804 struct follow_page_context *ctx) 805 { 806 p4d_t *p4d; 807 struct page *page; 808 809 p4d = p4d_offset(pgdp, address); 810 if (p4d_none(*p4d)) 811 return no_page_table(vma, flags); 812 BUILD_BUG_ON(p4d_huge(*p4d)); 813 if (unlikely(p4d_bad(*p4d))) 814 return no_page_table(vma, flags); 815 816 if (is_hugepd(__hugepd(p4d_val(*p4d)))) { 817 page = follow_huge_pd(vma, address, 818 __hugepd(p4d_val(*p4d)), flags, 819 P4D_SHIFT); 820 if (page) 821 return page; 822 return no_page_table(vma, flags); 823 } 824 return follow_pud_mask(vma, address, p4d, flags, ctx); 825 } 826 827 /** 828 * follow_page_mask - look up a page descriptor from a user-virtual address 829 * @vma: vm_area_struct mapping @address 830 * @address: virtual address to look up 831 * @flags: flags modifying lookup behaviour 832 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a 833 * pointer to output page_mask 834 * 835 * @flags can have FOLL_ flags set, defined in <linux/mm.h> 836 * 837 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches 838 * the device's dev_pagemap metadata to avoid repeating expensive lookups. 839 * 840 * When getting an anonymous page and the caller has to trigger unsharing 841 * of a shared anonymous page first, -EMLINK is returned. The caller should 842 * trigger a fault with FAULT_FLAG_UNSHARE set. Note that unsharing is only 843 * relevant with FOLL_PIN and !FOLL_WRITE. 844 * 845 * On output, the @ctx->page_mask is set according to the size of the page. 846 * 847 * Return: the mapped (struct page *), %NULL if no mapping exists, or 848 * an error pointer if there is a mapping to something not represented 849 * by a page descriptor (see also vm_normal_page()). 850 */ 851 static struct page *follow_page_mask(struct vm_area_struct *vma, 852 unsigned long address, unsigned int flags, 853 struct follow_page_context *ctx) 854 { 855 pgd_t *pgd; 856 struct page *page; 857 struct mm_struct *mm = vma->vm_mm; 858 859 ctx->page_mask = 0; 860 861 /* make this handle hugepd */ 862 page = follow_huge_addr(mm, address, flags & FOLL_WRITE); 863 if (!IS_ERR(page)) { 864 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN)); 865 return page; 866 } 867 868 pgd = pgd_offset(mm, address); 869 870 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) 871 return no_page_table(vma, flags); 872 873 if (pgd_huge(*pgd)) { 874 page = follow_huge_pgd(mm, address, pgd, flags); 875 if (page) 876 return page; 877 return no_page_table(vma, flags); 878 } 879 if (is_hugepd(__hugepd(pgd_val(*pgd)))) { 880 page = follow_huge_pd(vma, address, 881 __hugepd(pgd_val(*pgd)), flags, 882 PGDIR_SHIFT); 883 if (page) 884 return page; 885 return no_page_table(vma, flags); 886 } 887 888 return follow_p4d_mask(vma, address, pgd, flags, ctx); 889 } 890 891 struct page *follow_page(struct vm_area_struct *vma, unsigned long address, 892 unsigned int foll_flags) 893 { 894 struct follow_page_context ctx = { NULL }; 895 struct page *page; 896 897 if (vma_is_secretmem(vma)) 898 return NULL; 899 900 if (foll_flags & FOLL_PIN) 901 return NULL; 902 903 page = follow_page_mask(vma, address, foll_flags, &ctx); 904 if (ctx.pgmap) 905 put_dev_pagemap(ctx.pgmap); 906 return page; 907 } 908 909 static int get_gate_page(struct mm_struct *mm, unsigned long address, 910 unsigned int gup_flags, struct vm_area_struct **vma, 911 struct page **page) 912 { 913 pgd_t *pgd; 914 p4d_t *p4d; 915 pud_t *pud; 916 pmd_t *pmd; 917 pte_t *pte; 918 int ret = -EFAULT; 919 920 /* user gate pages are read-only */ 921 if (gup_flags & FOLL_WRITE) 922 return -EFAULT; 923 if (address > TASK_SIZE) 924 pgd = pgd_offset_k(address); 925 else 926 pgd = pgd_offset_gate(mm, address); 927 if (pgd_none(*pgd)) 928 return -EFAULT; 929 p4d = p4d_offset(pgd, address); 930 if (p4d_none(*p4d)) 931 return -EFAULT; 932 pud = pud_offset(p4d, address); 933 if (pud_none(*pud)) 934 return -EFAULT; 935 pmd = pmd_offset(pud, address); 936 if (!pmd_present(*pmd)) 937 return -EFAULT; 938 VM_BUG_ON(pmd_trans_huge(*pmd)); 939 pte = pte_offset_map(pmd, address); 940 if (pte_none(*pte)) 941 goto unmap; 942 *vma = get_gate_vma(mm); 943 if (!page) 944 goto out; 945 *page = vm_normal_page(*vma, address, *pte); 946 if (!*page) { 947 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte))) 948 goto unmap; 949 *page = pte_page(*pte); 950 } 951 if (unlikely(!try_grab_page(*page, gup_flags))) { 952 ret = -ENOMEM; 953 goto unmap; 954 } 955 out: 956 ret = 0; 957 unmap: 958 pte_unmap(pte); 959 return ret; 960 } 961 962 /* 963 * mmap_lock must be held on entry. If @locked != NULL and *@flags 964 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it 965 * is, *@locked will be set to 0 and -EBUSY returned. 966 */ 967 static int faultin_page(struct vm_area_struct *vma, 968 unsigned long address, unsigned int *flags, bool unshare, 969 int *locked) 970 { 971 unsigned int fault_flags = 0; 972 vm_fault_t ret; 973 974 if (*flags & FOLL_NOFAULT) 975 return -EFAULT; 976 if (*flags & FOLL_WRITE) 977 fault_flags |= FAULT_FLAG_WRITE; 978 if (*flags & FOLL_REMOTE) 979 fault_flags |= FAULT_FLAG_REMOTE; 980 if (locked) 981 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; 982 if (*flags & FOLL_NOWAIT) 983 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT; 984 if (*flags & FOLL_TRIED) { 985 /* 986 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED 987 * can co-exist 988 */ 989 fault_flags |= FAULT_FLAG_TRIED; 990 } 991 if (unshare) { 992 fault_flags |= FAULT_FLAG_UNSHARE; 993 /* FAULT_FLAG_WRITE and FAULT_FLAG_UNSHARE are incompatible */ 994 VM_BUG_ON(fault_flags & FAULT_FLAG_WRITE); 995 } 996 997 ret = handle_mm_fault(vma, address, fault_flags, NULL); 998 999 if (ret & VM_FAULT_COMPLETED) { 1000 /* 1001 * With FAULT_FLAG_RETRY_NOWAIT we'll never release the 1002 * mmap lock in the page fault handler. Sanity check this. 1003 */ 1004 WARN_ON_ONCE(fault_flags & FAULT_FLAG_RETRY_NOWAIT); 1005 if (locked) 1006 *locked = 0; 1007 /* 1008 * We should do the same as VM_FAULT_RETRY, but let's not 1009 * return -EBUSY since that's not reflecting the reality of 1010 * what has happened - we've just fully completed a page 1011 * fault, with the mmap lock released. Use -EAGAIN to show 1012 * that we want to take the mmap lock _again_. 1013 */ 1014 return -EAGAIN; 1015 } 1016 1017 if (ret & VM_FAULT_ERROR) { 1018 int err = vm_fault_to_errno(ret, *flags); 1019 1020 if (err) 1021 return err; 1022 BUG(); 1023 } 1024 1025 if (ret & VM_FAULT_RETRY) { 1026 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) 1027 *locked = 0; 1028 return -EBUSY; 1029 } 1030 1031 return 0; 1032 } 1033 1034 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags) 1035 { 1036 vm_flags_t vm_flags = vma->vm_flags; 1037 int write = (gup_flags & FOLL_WRITE); 1038 int foreign = (gup_flags & FOLL_REMOTE); 1039 1040 if (vm_flags & (VM_IO | VM_PFNMAP)) 1041 return -EFAULT; 1042 1043 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma)) 1044 return -EFAULT; 1045 1046 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma)) 1047 return -EOPNOTSUPP; 1048 1049 if (vma_is_secretmem(vma)) 1050 return -EFAULT; 1051 1052 if (write) { 1053 if (!(vm_flags & VM_WRITE)) { 1054 if (!(gup_flags & FOLL_FORCE)) 1055 return -EFAULT; 1056 /* 1057 * We used to let the write,force case do COW in a 1058 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could 1059 * set a breakpoint in a read-only mapping of an 1060 * executable, without corrupting the file (yet only 1061 * when that file had been opened for writing!). 1062 * Anon pages in shared mappings are surprising: now 1063 * just reject it. 1064 */ 1065 if (!is_cow_mapping(vm_flags)) 1066 return -EFAULT; 1067 } 1068 } else if (!(vm_flags & VM_READ)) { 1069 if (!(gup_flags & FOLL_FORCE)) 1070 return -EFAULT; 1071 /* 1072 * Is there actually any vma we can reach here which does not 1073 * have VM_MAYREAD set? 1074 */ 1075 if (!(vm_flags & VM_MAYREAD)) 1076 return -EFAULT; 1077 } 1078 /* 1079 * gups are always data accesses, not instruction 1080 * fetches, so execute=false here 1081 */ 1082 if (!arch_vma_access_permitted(vma, write, false, foreign)) 1083 return -EFAULT; 1084 return 0; 1085 } 1086 1087 /** 1088 * __get_user_pages() - pin user pages in memory 1089 * @mm: mm_struct of target mm 1090 * @start: starting user address 1091 * @nr_pages: number of pages from start to pin 1092 * @gup_flags: flags modifying pin behaviour 1093 * @pages: array that receives pointers to the pages pinned. 1094 * Should be at least nr_pages long. Or NULL, if caller 1095 * only intends to ensure the pages are faulted in. 1096 * @vmas: array of pointers to vmas corresponding to each page. 1097 * Or NULL if the caller does not require them. 1098 * @locked: whether we're still with the mmap_lock held 1099 * 1100 * Returns either number of pages pinned (which may be less than the 1101 * number requested), or an error. Details about the return value: 1102 * 1103 * -- If nr_pages is 0, returns 0. 1104 * -- If nr_pages is >0, but no pages were pinned, returns -errno. 1105 * -- If nr_pages is >0, and some pages were pinned, returns the number of 1106 * pages pinned. Again, this may be less than nr_pages. 1107 * -- 0 return value is possible when the fault would need to be retried. 1108 * 1109 * The caller is responsible for releasing returned @pages, via put_page(). 1110 * 1111 * @vmas are valid only as long as mmap_lock is held. 1112 * 1113 * Must be called with mmap_lock held. It may be released. See below. 1114 * 1115 * __get_user_pages walks a process's page tables and takes a reference to 1116 * each struct page that each user address corresponds to at a given 1117 * instant. That is, it takes the page that would be accessed if a user 1118 * thread accesses the given user virtual address at that instant. 1119 * 1120 * This does not guarantee that the page exists in the user mappings when 1121 * __get_user_pages returns, and there may even be a completely different 1122 * page there in some cases (eg. if mmapped pagecache has been invalidated 1123 * and subsequently re faulted). However it does guarantee that the page 1124 * won't be freed completely. And mostly callers simply care that the page 1125 * contains data that was valid *at some point in time*. Typically, an IO 1126 * or similar operation cannot guarantee anything stronger anyway because 1127 * locks can't be held over the syscall boundary. 1128 * 1129 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If 1130 * the page is written to, set_page_dirty (or set_page_dirty_lock, as 1131 * appropriate) must be called after the page is finished with, and 1132 * before put_page is called. 1133 * 1134 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is 1135 * released by an up_read(). That can happen if @gup_flags does not 1136 * have FOLL_NOWAIT. 1137 * 1138 * A caller using such a combination of @locked and @gup_flags 1139 * must therefore hold the mmap_lock for reading only, and recognize 1140 * when it's been released. Otherwise, it must be held for either 1141 * reading or writing and will not be released. 1142 * 1143 * In most cases, get_user_pages or get_user_pages_fast should be used 1144 * instead of __get_user_pages. __get_user_pages should be used only if 1145 * you need some special @gup_flags. 1146 */ 1147 static long __get_user_pages(struct mm_struct *mm, 1148 unsigned long start, unsigned long nr_pages, 1149 unsigned int gup_flags, struct page **pages, 1150 struct vm_area_struct **vmas, int *locked) 1151 { 1152 long ret = 0, i = 0; 1153 struct vm_area_struct *vma = NULL; 1154 struct follow_page_context ctx = { NULL }; 1155 1156 if (!nr_pages) 1157 return 0; 1158 1159 start = untagged_addr(start); 1160 1161 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN))); 1162 1163 /* 1164 * If FOLL_FORCE is set then do not force a full fault as the hinting 1165 * fault information is unrelated to the reference behaviour of a task 1166 * using the address space 1167 */ 1168 if (!(gup_flags & FOLL_FORCE)) 1169 gup_flags |= FOLL_NUMA; 1170 1171 do { 1172 struct page *page; 1173 unsigned int foll_flags = gup_flags; 1174 unsigned int page_increm; 1175 1176 /* first iteration or cross vma bound */ 1177 if (!vma || start >= vma->vm_end) { 1178 vma = find_extend_vma(mm, start); 1179 if (!vma && in_gate_area(mm, start)) { 1180 ret = get_gate_page(mm, start & PAGE_MASK, 1181 gup_flags, &vma, 1182 pages ? &pages[i] : NULL); 1183 if (ret) 1184 goto out; 1185 ctx.page_mask = 0; 1186 goto next_page; 1187 } 1188 1189 if (!vma) { 1190 ret = -EFAULT; 1191 goto out; 1192 } 1193 ret = check_vma_flags(vma, gup_flags); 1194 if (ret) 1195 goto out; 1196 1197 if (is_vm_hugetlb_page(vma)) { 1198 i = follow_hugetlb_page(mm, vma, pages, vmas, 1199 &start, &nr_pages, i, 1200 gup_flags, locked); 1201 if (locked && *locked == 0) { 1202 /* 1203 * We've got a VM_FAULT_RETRY 1204 * and we've lost mmap_lock. 1205 * We must stop here. 1206 */ 1207 BUG_ON(gup_flags & FOLL_NOWAIT); 1208 goto out; 1209 } 1210 continue; 1211 } 1212 } 1213 retry: 1214 /* 1215 * If we have a pending SIGKILL, don't keep faulting pages and 1216 * potentially allocating memory. 1217 */ 1218 if (fatal_signal_pending(current)) { 1219 ret = -EINTR; 1220 goto out; 1221 } 1222 cond_resched(); 1223 1224 page = follow_page_mask(vma, start, foll_flags, &ctx); 1225 if (!page || PTR_ERR(page) == -EMLINK) { 1226 ret = faultin_page(vma, start, &foll_flags, 1227 PTR_ERR(page) == -EMLINK, locked); 1228 switch (ret) { 1229 case 0: 1230 goto retry; 1231 case -EBUSY: 1232 case -EAGAIN: 1233 ret = 0; 1234 fallthrough; 1235 case -EFAULT: 1236 case -ENOMEM: 1237 case -EHWPOISON: 1238 goto out; 1239 } 1240 BUG(); 1241 } else if (PTR_ERR(page) == -EEXIST) { 1242 /* 1243 * Proper page table entry exists, but no corresponding 1244 * struct page. If the caller expects **pages to be 1245 * filled in, bail out now, because that can't be done 1246 * for this page. 1247 */ 1248 if (pages) { 1249 ret = PTR_ERR(page); 1250 goto out; 1251 } 1252 1253 goto next_page; 1254 } else if (IS_ERR(page)) { 1255 ret = PTR_ERR(page); 1256 goto out; 1257 } 1258 if (pages) { 1259 pages[i] = page; 1260 flush_anon_page(vma, page, start); 1261 flush_dcache_page(page); 1262 ctx.page_mask = 0; 1263 } 1264 next_page: 1265 if (vmas) { 1266 vmas[i] = vma; 1267 ctx.page_mask = 0; 1268 } 1269 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask); 1270 if (page_increm > nr_pages) 1271 page_increm = nr_pages; 1272 i += page_increm; 1273 start += page_increm * PAGE_SIZE; 1274 nr_pages -= page_increm; 1275 } while (nr_pages); 1276 out: 1277 if (ctx.pgmap) 1278 put_dev_pagemap(ctx.pgmap); 1279 return i ? i : ret; 1280 } 1281 1282 static bool vma_permits_fault(struct vm_area_struct *vma, 1283 unsigned int fault_flags) 1284 { 1285 bool write = !!(fault_flags & FAULT_FLAG_WRITE); 1286 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE); 1287 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ; 1288 1289 if (!(vm_flags & vma->vm_flags)) 1290 return false; 1291 1292 /* 1293 * The architecture might have a hardware protection 1294 * mechanism other than read/write that can deny access. 1295 * 1296 * gup always represents data access, not instruction 1297 * fetches, so execute=false here: 1298 */ 1299 if (!arch_vma_access_permitted(vma, write, false, foreign)) 1300 return false; 1301 1302 return true; 1303 } 1304 1305 /** 1306 * fixup_user_fault() - manually resolve a user page fault 1307 * @mm: mm_struct of target mm 1308 * @address: user address 1309 * @fault_flags:flags to pass down to handle_mm_fault() 1310 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller 1311 * does not allow retry. If NULL, the caller must guarantee 1312 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY. 1313 * 1314 * This is meant to be called in the specific scenario where for locking reasons 1315 * we try to access user memory in atomic context (within a pagefault_disable() 1316 * section), this returns -EFAULT, and we want to resolve the user fault before 1317 * trying again. 1318 * 1319 * Typically this is meant to be used by the futex code. 1320 * 1321 * The main difference with get_user_pages() is that this function will 1322 * unconditionally call handle_mm_fault() which will in turn perform all the 1323 * necessary SW fixup of the dirty and young bits in the PTE, while 1324 * get_user_pages() only guarantees to update these in the struct page. 1325 * 1326 * This is important for some architectures where those bits also gate the 1327 * access permission to the page because they are maintained in software. On 1328 * such architectures, gup() will not be enough to make a subsequent access 1329 * succeed. 1330 * 1331 * This function will not return with an unlocked mmap_lock. So it has not the 1332 * same semantics wrt the @mm->mmap_lock as does filemap_fault(). 1333 */ 1334 int fixup_user_fault(struct mm_struct *mm, 1335 unsigned long address, unsigned int fault_flags, 1336 bool *unlocked) 1337 { 1338 struct vm_area_struct *vma; 1339 vm_fault_t ret; 1340 1341 address = untagged_addr(address); 1342 1343 if (unlocked) 1344 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; 1345 1346 retry: 1347 vma = find_extend_vma(mm, address); 1348 if (!vma || address < vma->vm_start) 1349 return -EFAULT; 1350 1351 if (!vma_permits_fault(vma, fault_flags)) 1352 return -EFAULT; 1353 1354 if ((fault_flags & FAULT_FLAG_KILLABLE) && 1355 fatal_signal_pending(current)) 1356 return -EINTR; 1357 1358 ret = handle_mm_fault(vma, address, fault_flags, NULL); 1359 1360 if (ret & VM_FAULT_COMPLETED) { 1361 /* 1362 * NOTE: it's a pity that we need to retake the lock here 1363 * to pair with the unlock() in the callers. Ideally we 1364 * could tell the callers so they do not need to unlock. 1365 */ 1366 mmap_read_lock(mm); 1367 *unlocked = true; 1368 return 0; 1369 } 1370 1371 if (ret & VM_FAULT_ERROR) { 1372 int err = vm_fault_to_errno(ret, 0); 1373 1374 if (err) 1375 return err; 1376 BUG(); 1377 } 1378 1379 if (ret & VM_FAULT_RETRY) { 1380 mmap_read_lock(mm); 1381 *unlocked = true; 1382 fault_flags |= FAULT_FLAG_TRIED; 1383 goto retry; 1384 } 1385 1386 return 0; 1387 } 1388 EXPORT_SYMBOL_GPL(fixup_user_fault); 1389 1390 /* 1391 * Please note that this function, unlike __get_user_pages will not 1392 * return 0 for nr_pages > 0 without FOLL_NOWAIT 1393 */ 1394 static __always_inline long __get_user_pages_locked(struct mm_struct *mm, 1395 unsigned long start, 1396 unsigned long nr_pages, 1397 struct page **pages, 1398 struct vm_area_struct **vmas, 1399 int *locked, 1400 unsigned int flags) 1401 { 1402 long ret, pages_done; 1403 bool lock_dropped; 1404 1405 if (locked) { 1406 /* if VM_FAULT_RETRY can be returned, vmas become invalid */ 1407 BUG_ON(vmas); 1408 /* check caller initialized locked */ 1409 BUG_ON(*locked != 1); 1410 } 1411 1412 if (flags & FOLL_PIN) 1413 mm_set_has_pinned_flag(&mm->flags); 1414 1415 /* 1416 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior 1417 * is to set FOLL_GET if the caller wants pages[] filled in (but has 1418 * carelessly failed to specify FOLL_GET), so keep doing that, but only 1419 * for FOLL_GET, not for the newer FOLL_PIN. 1420 * 1421 * FOLL_PIN always expects pages to be non-null, but no need to assert 1422 * that here, as any failures will be obvious enough. 1423 */ 1424 if (pages && !(flags & FOLL_PIN)) 1425 flags |= FOLL_GET; 1426 1427 pages_done = 0; 1428 lock_dropped = false; 1429 for (;;) { 1430 ret = __get_user_pages(mm, start, nr_pages, flags, pages, 1431 vmas, locked); 1432 if (!locked) 1433 /* VM_FAULT_RETRY couldn't trigger, bypass */ 1434 return ret; 1435 1436 /* VM_FAULT_RETRY or VM_FAULT_COMPLETED cannot return errors */ 1437 if (!*locked) { 1438 BUG_ON(ret < 0); 1439 BUG_ON(ret >= nr_pages); 1440 } 1441 1442 if (ret > 0) { 1443 nr_pages -= ret; 1444 pages_done += ret; 1445 if (!nr_pages) 1446 break; 1447 } 1448 if (*locked) { 1449 /* 1450 * VM_FAULT_RETRY didn't trigger or it was a 1451 * FOLL_NOWAIT. 1452 */ 1453 if (!pages_done) 1454 pages_done = ret; 1455 break; 1456 } 1457 /* 1458 * VM_FAULT_RETRY triggered, so seek to the faulting offset. 1459 * For the prefault case (!pages) we only update counts. 1460 */ 1461 if (likely(pages)) 1462 pages += ret; 1463 start += ret << PAGE_SHIFT; 1464 lock_dropped = true; 1465 1466 retry: 1467 /* 1468 * Repeat on the address that fired VM_FAULT_RETRY 1469 * with both FAULT_FLAG_ALLOW_RETRY and 1470 * FAULT_FLAG_TRIED. Note that GUP can be interrupted 1471 * by fatal signals, so we need to check it before we 1472 * start trying again otherwise it can loop forever. 1473 */ 1474 1475 if (fatal_signal_pending(current)) { 1476 if (!pages_done) 1477 pages_done = -EINTR; 1478 break; 1479 } 1480 1481 ret = mmap_read_lock_killable(mm); 1482 if (ret) { 1483 BUG_ON(ret > 0); 1484 if (!pages_done) 1485 pages_done = ret; 1486 break; 1487 } 1488 1489 *locked = 1; 1490 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED, 1491 pages, NULL, locked); 1492 if (!*locked) { 1493 /* Continue to retry until we succeeded */ 1494 BUG_ON(ret != 0); 1495 goto retry; 1496 } 1497 if (ret != 1) { 1498 BUG_ON(ret > 1); 1499 if (!pages_done) 1500 pages_done = ret; 1501 break; 1502 } 1503 nr_pages--; 1504 pages_done++; 1505 if (!nr_pages) 1506 break; 1507 if (likely(pages)) 1508 pages++; 1509 start += PAGE_SIZE; 1510 } 1511 if (lock_dropped && *locked) { 1512 /* 1513 * We must let the caller know we temporarily dropped the lock 1514 * and so the critical section protected by it was lost. 1515 */ 1516 mmap_read_unlock(mm); 1517 *locked = 0; 1518 } 1519 return pages_done; 1520 } 1521 1522 /** 1523 * populate_vma_page_range() - populate a range of pages in the vma. 1524 * @vma: target vma 1525 * @start: start address 1526 * @end: end address 1527 * @locked: whether the mmap_lock is still held 1528 * 1529 * This takes care of mlocking the pages too if VM_LOCKED is set. 1530 * 1531 * Return either number of pages pinned in the vma, or a negative error 1532 * code on error. 1533 * 1534 * vma->vm_mm->mmap_lock must be held. 1535 * 1536 * If @locked is NULL, it may be held for read or write and will 1537 * be unperturbed. 1538 * 1539 * If @locked is non-NULL, it must held for read only and may be 1540 * released. If it's released, *@locked will be set to 0. 1541 */ 1542 long populate_vma_page_range(struct vm_area_struct *vma, 1543 unsigned long start, unsigned long end, int *locked) 1544 { 1545 struct mm_struct *mm = vma->vm_mm; 1546 unsigned long nr_pages = (end - start) / PAGE_SIZE; 1547 int gup_flags; 1548 long ret; 1549 1550 VM_BUG_ON(!PAGE_ALIGNED(start)); 1551 VM_BUG_ON(!PAGE_ALIGNED(end)); 1552 VM_BUG_ON_VMA(start < vma->vm_start, vma); 1553 VM_BUG_ON_VMA(end > vma->vm_end, vma); 1554 mmap_assert_locked(mm); 1555 1556 /* 1557 * Rightly or wrongly, the VM_LOCKONFAULT case has never used 1558 * faultin_page() to break COW, so it has no work to do here. 1559 */ 1560 if (vma->vm_flags & VM_LOCKONFAULT) 1561 return nr_pages; 1562 1563 gup_flags = FOLL_TOUCH; 1564 /* 1565 * We want to touch writable mappings with a write fault in order 1566 * to break COW, except for shared mappings because these don't COW 1567 * and we would not want to dirty them for nothing. 1568 */ 1569 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE) 1570 gup_flags |= FOLL_WRITE; 1571 1572 /* 1573 * We want mlock to succeed for regions that have any permissions 1574 * other than PROT_NONE. 1575 */ 1576 if (vma_is_accessible(vma)) 1577 gup_flags |= FOLL_FORCE; 1578 1579 /* 1580 * We made sure addr is within a VMA, so the following will 1581 * not result in a stack expansion that recurses back here. 1582 */ 1583 ret = __get_user_pages(mm, start, nr_pages, gup_flags, 1584 NULL, NULL, locked); 1585 lru_add_drain(); 1586 return ret; 1587 } 1588 1589 /* 1590 * faultin_vma_page_range() - populate (prefault) page tables inside the 1591 * given VMA range readable/writable 1592 * 1593 * This takes care of mlocking the pages, too, if VM_LOCKED is set. 1594 * 1595 * @vma: target vma 1596 * @start: start address 1597 * @end: end address 1598 * @write: whether to prefault readable or writable 1599 * @locked: whether the mmap_lock is still held 1600 * 1601 * Returns either number of processed pages in the vma, or a negative error 1602 * code on error (see __get_user_pages()). 1603 * 1604 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and 1605 * covered by the VMA. 1606 * 1607 * If @locked is NULL, it may be held for read or write and will be unperturbed. 1608 * 1609 * If @locked is non-NULL, it must held for read only and may be released. If 1610 * it's released, *@locked will be set to 0. 1611 */ 1612 long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start, 1613 unsigned long end, bool write, int *locked) 1614 { 1615 struct mm_struct *mm = vma->vm_mm; 1616 unsigned long nr_pages = (end - start) / PAGE_SIZE; 1617 int gup_flags; 1618 long ret; 1619 1620 VM_BUG_ON(!PAGE_ALIGNED(start)); 1621 VM_BUG_ON(!PAGE_ALIGNED(end)); 1622 VM_BUG_ON_VMA(start < vma->vm_start, vma); 1623 VM_BUG_ON_VMA(end > vma->vm_end, vma); 1624 mmap_assert_locked(mm); 1625 1626 /* 1627 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark 1628 * the page dirty with FOLL_WRITE -- which doesn't make a 1629 * difference with !FOLL_FORCE, because the page is writable 1630 * in the page table. 1631 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit 1632 * a poisoned page. 1633 * !FOLL_FORCE: Require proper access permissions. 1634 */ 1635 gup_flags = FOLL_TOUCH | FOLL_HWPOISON; 1636 if (write) 1637 gup_flags |= FOLL_WRITE; 1638 1639 /* 1640 * We want to report -EINVAL instead of -EFAULT for any permission 1641 * problems or incompatible mappings. 1642 */ 1643 if (check_vma_flags(vma, gup_flags)) 1644 return -EINVAL; 1645 1646 ret = __get_user_pages(mm, start, nr_pages, gup_flags, 1647 NULL, NULL, locked); 1648 lru_add_drain(); 1649 return ret; 1650 } 1651 1652 /* 1653 * __mm_populate - populate and/or mlock pages within a range of address space. 1654 * 1655 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap 1656 * flags. VMAs must be already marked with the desired vm_flags, and 1657 * mmap_lock must not be held. 1658 */ 1659 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors) 1660 { 1661 struct mm_struct *mm = current->mm; 1662 unsigned long end, nstart, nend; 1663 struct vm_area_struct *vma = NULL; 1664 int locked = 0; 1665 long ret = 0; 1666 1667 end = start + len; 1668 1669 for (nstart = start; nstart < end; nstart = nend) { 1670 /* 1671 * We want to fault in pages for [nstart; end) address range. 1672 * Find first corresponding VMA. 1673 */ 1674 if (!locked) { 1675 locked = 1; 1676 mmap_read_lock(mm); 1677 vma = find_vma(mm, nstart); 1678 } else if (nstart >= vma->vm_end) 1679 vma = vma->vm_next; 1680 if (!vma || vma->vm_start >= end) 1681 break; 1682 /* 1683 * Set [nstart; nend) to intersection of desired address 1684 * range with the first VMA. Also, skip undesirable VMA types. 1685 */ 1686 nend = min(end, vma->vm_end); 1687 if (vma->vm_flags & (VM_IO | VM_PFNMAP)) 1688 continue; 1689 if (nstart < vma->vm_start) 1690 nstart = vma->vm_start; 1691 /* 1692 * Now fault in a range of pages. populate_vma_page_range() 1693 * double checks the vma flags, so that it won't mlock pages 1694 * if the vma was already munlocked. 1695 */ 1696 ret = populate_vma_page_range(vma, nstart, nend, &locked); 1697 if (ret < 0) { 1698 if (ignore_errors) { 1699 ret = 0; 1700 continue; /* continue at next VMA */ 1701 } 1702 break; 1703 } 1704 nend = nstart + ret * PAGE_SIZE; 1705 ret = 0; 1706 } 1707 if (locked) 1708 mmap_read_unlock(mm); 1709 return ret; /* 0 or negative error code */ 1710 } 1711 #else /* CONFIG_MMU */ 1712 static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start, 1713 unsigned long nr_pages, struct page **pages, 1714 struct vm_area_struct **vmas, int *locked, 1715 unsigned int foll_flags) 1716 { 1717 struct vm_area_struct *vma; 1718 unsigned long vm_flags; 1719 long i; 1720 1721 /* calculate required read or write permissions. 1722 * If FOLL_FORCE is set, we only require the "MAY" flags. 1723 */ 1724 vm_flags = (foll_flags & FOLL_WRITE) ? 1725 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD); 1726 vm_flags &= (foll_flags & FOLL_FORCE) ? 1727 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE); 1728 1729 for (i = 0; i < nr_pages; i++) { 1730 vma = find_vma(mm, start); 1731 if (!vma) 1732 goto finish_or_fault; 1733 1734 /* protect what we can, including chardevs */ 1735 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) || 1736 !(vm_flags & vma->vm_flags)) 1737 goto finish_or_fault; 1738 1739 if (pages) { 1740 pages[i] = virt_to_page((void *)start); 1741 if (pages[i]) 1742 get_page(pages[i]); 1743 } 1744 if (vmas) 1745 vmas[i] = vma; 1746 start = (start + PAGE_SIZE) & PAGE_MASK; 1747 } 1748 1749 return i; 1750 1751 finish_or_fault: 1752 return i ? : -EFAULT; 1753 } 1754 #endif /* !CONFIG_MMU */ 1755 1756 /** 1757 * fault_in_writeable - fault in userspace address range for writing 1758 * @uaddr: start of address range 1759 * @size: size of address range 1760 * 1761 * Returns the number of bytes not faulted in (like copy_to_user() and 1762 * copy_from_user()). 1763 */ 1764 size_t fault_in_writeable(char __user *uaddr, size_t size) 1765 { 1766 char __user *start = uaddr, *end; 1767 1768 if (unlikely(size == 0)) 1769 return 0; 1770 if (!user_write_access_begin(uaddr, size)) 1771 return size; 1772 if (!PAGE_ALIGNED(uaddr)) { 1773 unsafe_put_user(0, uaddr, out); 1774 uaddr = (char __user *)PAGE_ALIGN((unsigned long)uaddr); 1775 } 1776 end = (char __user *)PAGE_ALIGN((unsigned long)start + size); 1777 if (unlikely(end < start)) 1778 end = NULL; 1779 while (uaddr != end) { 1780 unsafe_put_user(0, uaddr, out); 1781 uaddr += PAGE_SIZE; 1782 } 1783 1784 out: 1785 user_write_access_end(); 1786 if (size > uaddr - start) 1787 return size - (uaddr - start); 1788 return 0; 1789 } 1790 EXPORT_SYMBOL(fault_in_writeable); 1791 1792 /** 1793 * fault_in_subpage_writeable - fault in an address range for writing 1794 * @uaddr: start of address range 1795 * @size: size of address range 1796 * 1797 * Fault in a user address range for writing while checking for permissions at 1798 * sub-page granularity (e.g. arm64 MTE). This function should be used when 1799 * the caller cannot guarantee forward progress of a copy_to_user() loop. 1800 * 1801 * Returns the number of bytes not faulted in (like copy_to_user() and 1802 * copy_from_user()). 1803 */ 1804 size_t fault_in_subpage_writeable(char __user *uaddr, size_t size) 1805 { 1806 size_t faulted_in; 1807 1808 /* 1809 * Attempt faulting in at page granularity first for page table 1810 * permission checking. The arch-specific probe_subpage_writeable() 1811 * functions may not check for this. 1812 */ 1813 faulted_in = size - fault_in_writeable(uaddr, size); 1814 if (faulted_in) 1815 faulted_in -= probe_subpage_writeable(uaddr, faulted_in); 1816 1817 return size - faulted_in; 1818 } 1819 EXPORT_SYMBOL(fault_in_subpage_writeable); 1820 1821 /* 1822 * fault_in_safe_writeable - fault in an address range for writing 1823 * @uaddr: start of address range 1824 * @size: length of address range 1825 * 1826 * Faults in an address range for writing. This is primarily useful when we 1827 * already know that some or all of the pages in the address range aren't in 1828 * memory. 1829 * 1830 * Unlike fault_in_writeable(), this function is non-destructive. 1831 * 1832 * Note that we don't pin or otherwise hold the pages referenced that we fault 1833 * in. There's no guarantee that they'll stay in memory for any duration of 1834 * time. 1835 * 1836 * Returns the number of bytes not faulted in, like copy_to_user() and 1837 * copy_from_user(). 1838 */ 1839 size_t fault_in_safe_writeable(const char __user *uaddr, size_t size) 1840 { 1841 unsigned long start = (unsigned long)uaddr, end; 1842 struct mm_struct *mm = current->mm; 1843 bool unlocked = false; 1844 1845 if (unlikely(size == 0)) 1846 return 0; 1847 end = PAGE_ALIGN(start + size); 1848 if (end < start) 1849 end = 0; 1850 1851 mmap_read_lock(mm); 1852 do { 1853 if (fixup_user_fault(mm, start, FAULT_FLAG_WRITE, &unlocked)) 1854 break; 1855 start = (start + PAGE_SIZE) & PAGE_MASK; 1856 } while (start != end); 1857 mmap_read_unlock(mm); 1858 1859 if (size > (unsigned long)uaddr - start) 1860 return size - ((unsigned long)uaddr - start); 1861 return 0; 1862 } 1863 EXPORT_SYMBOL(fault_in_safe_writeable); 1864 1865 /** 1866 * fault_in_readable - fault in userspace address range for reading 1867 * @uaddr: start of user address range 1868 * @size: size of user address range 1869 * 1870 * Returns the number of bytes not faulted in (like copy_to_user() and 1871 * copy_from_user()). 1872 */ 1873 size_t fault_in_readable(const char __user *uaddr, size_t size) 1874 { 1875 const char __user *start = uaddr, *end; 1876 volatile char c; 1877 1878 if (unlikely(size == 0)) 1879 return 0; 1880 if (!user_read_access_begin(uaddr, size)) 1881 return size; 1882 if (!PAGE_ALIGNED(uaddr)) { 1883 unsafe_get_user(c, uaddr, out); 1884 uaddr = (const char __user *)PAGE_ALIGN((unsigned long)uaddr); 1885 } 1886 end = (const char __user *)PAGE_ALIGN((unsigned long)start + size); 1887 if (unlikely(end < start)) 1888 end = NULL; 1889 while (uaddr != end) { 1890 unsafe_get_user(c, uaddr, out); 1891 uaddr += PAGE_SIZE; 1892 } 1893 1894 out: 1895 user_read_access_end(); 1896 (void)c; 1897 if (size > uaddr - start) 1898 return size - (uaddr - start); 1899 return 0; 1900 } 1901 EXPORT_SYMBOL(fault_in_readable); 1902 1903 /** 1904 * get_dump_page() - pin user page in memory while writing it to core dump 1905 * @addr: user address 1906 * 1907 * Returns struct page pointer of user page pinned for dump, 1908 * to be freed afterwards by put_page(). 1909 * 1910 * Returns NULL on any kind of failure - a hole must then be inserted into 1911 * the corefile, to preserve alignment with its headers; and also returns 1912 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found - 1913 * allowing a hole to be left in the corefile to save disk space. 1914 * 1915 * Called without mmap_lock (takes and releases the mmap_lock by itself). 1916 */ 1917 #ifdef CONFIG_ELF_CORE 1918 struct page *get_dump_page(unsigned long addr) 1919 { 1920 struct mm_struct *mm = current->mm; 1921 struct page *page; 1922 int locked = 1; 1923 int ret; 1924 1925 if (mmap_read_lock_killable(mm)) 1926 return NULL; 1927 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked, 1928 FOLL_FORCE | FOLL_DUMP | FOLL_GET); 1929 if (locked) 1930 mmap_read_unlock(mm); 1931 return (ret == 1) ? page : NULL; 1932 } 1933 #endif /* CONFIG_ELF_CORE */ 1934 1935 #ifdef CONFIG_MIGRATION 1936 /* 1937 * Returns the number of collected pages. Return value is always >= 0. 1938 */ 1939 static unsigned long collect_longterm_unpinnable_pages( 1940 struct list_head *movable_page_list, 1941 unsigned long nr_pages, 1942 struct page **pages) 1943 { 1944 unsigned long i, collected = 0; 1945 struct folio *prev_folio = NULL; 1946 bool drain_allow = true; 1947 1948 for (i = 0; i < nr_pages; i++) { 1949 struct folio *folio = page_folio(pages[i]); 1950 1951 if (folio == prev_folio) 1952 continue; 1953 prev_folio = folio; 1954 1955 if (folio_is_longterm_pinnable(folio)) 1956 continue; 1957 1958 collected++; 1959 1960 if (folio_is_device_coherent(folio)) 1961 continue; 1962 1963 if (folio_test_hugetlb(folio)) { 1964 isolate_hugetlb(&folio->page, movable_page_list); 1965 continue; 1966 } 1967 1968 if (!folio_test_lru(folio) && drain_allow) { 1969 lru_add_drain_all(); 1970 drain_allow = false; 1971 } 1972 1973 if (!folio_isolate_lru(folio)) 1974 continue; 1975 1976 list_add_tail(&folio->lru, movable_page_list); 1977 node_stat_mod_folio(folio, 1978 NR_ISOLATED_ANON + folio_is_file_lru(folio), 1979 folio_nr_pages(folio)); 1980 } 1981 1982 return collected; 1983 } 1984 1985 /* 1986 * Unpins all pages and migrates device coherent pages and movable_page_list. 1987 * Returns -EAGAIN if all pages were successfully migrated or -errno for failure 1988 * (or partial success). 1989 */ 1990 static int migrate_longterm_unpinnable_pages( 1991 struct list_head *movable_page_list, 1992 unsigned long nr_pages, 1993 struct page **pages) 1994 { 1995 int ret; 1996 unsigned long i; 1997 1998 for (i = 0; i < nr_pages; i++) { 1999 struct folio *folio = page_folio(pages[i]); 2000 2001 if (folio_is_device_coherent(folio)) { 2002 /* 2003 * Migration will fail if the page is pinned, so convert 2004 * the pin on the source page to a normal reference. 2005 */ 2006 pages[i] = NULL; 2007 folio_get(folio); 2008 gup_put_folio(folio, 1, FOLL_PIN); 2009 2010 if (migrate_device_coherent_page(&folio->page)) { 2011 ret = -EBUSY; 2012 goto err; 2013 } 2014 2015 continue; 2016 } 2017 2018 /* 2019 * We can't migrate pages with unexpected references, so drop 2020 * the reference obtained by __get_user_pages_locked(). 2021 * Migrating pages have been added to movable_page_list after 2022 * calling folio_isolate_lru() which takes a reference so the 2023 * page won't be freed if it's migrating. 2024 */ 2025 unpin_user_page(pages[i]); 2026 pages[i] = NULL; 2027 } 2028 2029 if (!list_empty(movable_page_list)) { 2030 struct migration_target_control mtc = { 2031 .nid = NUMA_NO_NODE, 2032 .gfp_mask = GFP_USER | __GFP_NOWARN, 2033 }; 2034 2035 if (migrate_pages(movable_page_list, alloc_migration_target, 2036 NULL, (unsigned long)&mtc, MIGRATE_SYNC, 2037 MR_LONGTERM_PIN, NULL)) { 2038 ret = -ENOMEM; 2039 goto err; 2040 } 2041 } 2042 2043 putback_movable_pages(movable_page_list); 2044 2045 return -EAGAIN; 2046 2047 err: 2048 for (i = 0; i < nr_pages; i++) 2049 if (pages[i]) 2050 unpin_user_page(pages[i]); 2051 putback_movable_pages(movable_page_list); 2052 2053 return ret; 2054 } 2055 2056 /* 2057 * Check whether all pages are *allowed* to be pinned. Rather confusingly, all 2058 * pages in the range are required to be pinned via FOLL_PIN, before calling 2059 * this routine. 2060 * 2061 * If any pages in the range are not allowed to be pinned, then this routine 2062 * will migrate those pages away, unpin all the pages in the range and return 2063 * -EAGAIN. The caller should re-pin the entire range with FOLL_PIN and then 2064 * call this routine again. 2065 * 2066 * If an error other than -EAGAIN occurs, this indicates a migration failure. 2067 * The caller should give up, and propagate the error back up the call stack. 2068 * 2069 * If everything is OK and all pages in the range are allowed to be pinned, then 2070 * this routine leaves all pages pinned and returns zero for success. 2071 */ 2072 static long check_and_migrate_movable_pages(unsigned long nr_pages, 2073 struct page **pages) 2074 { 2075 unsigned long collected; 2076 LIST_HEAD(movable_page_list); 2077 2078 collected = collect_longterm_unpinnable_pages(&movable_page_list, 2079 nr_pages, pages); 2080 if (!collected) 2081 return 0; 2082 2083 return migrate_longterm_unpinnable_pages(&movable_page_list, nr_pages, 2084 pages); 2085 } 2086 #else 2087 static long check_and_migrate_movable_pages(unsigned long nr_pages, 2088 struct page **pages) 2089 { 2090 return 0; 2091 } 2092 #endif /* CONFIG_MIGRATION */ 2093 2094 /* 2095 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which 2096 * allows us to process the FOLL_LONGTERM flag. 2097 */ 2098 static long __gup_longterm_locked(struct mm_struct *mm, 2099 unsigned long start, 2100 unsigned long nr_pages, 2101 struct page **pages, 2102 struct vm_area_struct **vmas, 2103 unsigned int gup_flags) 2104 { 2105 unsigned int flags; 2106 long rc, nr_pinned_pages; 2107 2108 if (!(gup_flags & FOLL_LONGTERM)) 2109 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas, 2110 NULL, gup_flags); 2111 2112 /* 2113 * If we get to this point then FOLL_LONGTERM is set, and FOLL_LONGTERM 2114 * implies FOLL_PIN (although the reverse is not true). Therefore it is 2115 * correct to unconditionally call check_and_migrate_movable_pages() 2116 * which assumes pages have been pinned via FOLL_PIN. 2117 * 2118 * Enforce the above reasoning by asserting that FOLL_PIN is set. 2119 */ 2120 if (WARN_ON(!(gup_flags & FOLL_PIN))) 2121 return -EINVAL; 2122 flags = memalloc_pin_save(); 2123 do { 2124 nr_pinned_pages = __get_user_pages_locked(mm, start, nr_pages, 2125 pages, vmas, NULL, 2126 gup_flags); 2127 if (nr_pinned_pages <= 0) { 2128 rc = nr_pinned_pages; 2129 break; 2130 } 2131 rc = check_and_migrate_movable_pages(nr_pinned_pages, pages); 2132 } while (rc == -EAGAIN); 2133 memalloc_pin_restore(flags); 2134 2135 return rc ? rc : nr_pinned_pages; 2136 } 2137 2138 static bool is_valid_gup_flags(unsigned int gup_flags) 2139 { 2140 /* 2141 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs, 2142 * never directly by the caller, so enforce that with an assertion: 2143 */ 2144 if (WARN_ON_ONCE(gup_flags & FOLL_PIN)) 2145 return false; 2146 /* 2147 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying 2148 * that is, FOLL_LONGTERM is a specific case, more restrictive case of 2149 * FOLL_PIN. 2150 */ 2151 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM)) 2152 return false; 2153 2154 return true; 2155 } 2156 2157 #ifdef CONFIG_MMU 2158 static long __get_user_pages_remote(struct mm_struct *mm, 2159 unsigned long start, unsigned long nr_pages, 2160 unsigned int gup_flags, struct page **pages, 2161 struct vm_area_struct **vmas, int *locked) 2162 { 2163 /* 2164 * Parts of FOLL_LONGTERM behavior are incompatible with 2165 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on 2166 * vmas. However, this only comes up if locked is set, and there are 2167 * callers that do request FOLL_LONGTERM, but do not set locked. So, 2168 * allow what we can. 2169 */ 2170 if (gup_flags & FOLL_LONGTERM) { 2171 if (WARN_ON_ONCE(locked)) 2172 return -EINVAL; 2173 /* 2174 * This will check the vmas (even if our vmas arg is NULL) 2175 * and return -ENOTSUPP if DAX isn't allowed in this case: 2176 */ 2177 return __gup_longterm_locked(mm, start, nr_pages, pages, 2178 vmas, gup_flags | FOLL_TOUCH | 2179 FOLL_REMOTE); 2180 } 2181 2182 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas, 2183 locked, 2184 gup_flags | FOLL_TOUCH | FOLL_REMOTE); 2185 } 2186 2187 /** 2188 * get_user_pages_remote() - pin user pages in memory 2189 * @mm: mm_struct of target mm 2190 * @start: starting user address 2191 * @nr_pages: number of pages from start to pin 2192 * @gup_flags: flags modifying lookup behaviour 2193 * @pages: array that receives pointers to the pages pinned. 2194 * Should be at least nr_pages long. Or NULL, if caller 2195 * only intends to ensure the pages are faulted in. 2196 * @vmas: array of pointers to vmas corresponding to each page. 2197 * Or NULL if the caller does not require them. 2198 * @locked: pointer to lock flag indicating whether lock is held and 2199 * subsequently whether VM_FAULT_RETRY functionality can be 2200 * utilised. Lock must initially be held. 2201 * 2202 * Returns either number of pages pinned (which may be less than the 2203 * number requested), or an error. Details about the return value: 2204 * 2205 * -- If nr_pages is 0, returns 0. 2206 * -- If nr_pages is >0, but no pages were pinned, returns -errno. 2207 * -- If nr_pages is >0, and some pages were pinned, returns the number of 2208 * pages pinned. Again, this may be less than nr_pages. 2209 * 2210 * The caller is responsible for releasing returned @pages, via put_page(). 2211 * 2212 * @vmas are valid only as long as mmap_lock is held. 2213 * 2214 * Must be called with mmap_lock held for read or write. 2215 * 2216 * get_user_pages_remote walks a process's page tables and takes a reference 2217 * to each struct page that each user address corresponds to at a given 2218 * instant. That is, it takes the page that would be accessed if a user 2219 * thread accesses the given user virtual address at that instant. 2220 * 2221 * This does not guarantee that the page exists in the user mappings when 2222 * get_user_pages_remote returns, and there may even be a completely different 2223 * page there in some cases (eg. if mmapped pagecache has been invalidated 2224 * and subsequently re faulted). However it does guarantee that the page 2225 * won't be freed completely. And mostly callers simply care that the page 2226 * contains data that was valid *at some point in time*. Typically, an IO 2227 * or similar operation cannot guarantee anything stronger anyway because 2228 * locks can't be held over the syscall boundary. 2229 * 2230 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page 2231 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must 2232 * be called after the page is finished with, and before put_page is called. 2233 * 2234 * get_user_pages_remote is typically used for fewer-copy IO operations, 2235 * to get a handle on the memory by some means other than accesses 2236 * via the user virtual addresses. The pages may be submitted for 2237 * DMA to devices or accessed via their kernel linear mapping (via the 2238 * kmap APIs). Care should be taken to use the correct cache flushing APIs. 2239 * 2240 * See also get_user_pages_fast, for performance critical applications. 2241 * 2242 * get_user_pages_remote should be phased out in favor of 2243 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing 2244 * should use get_user_pages_remote because it cannot pass 2245 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault. 2246 */ 2247 long get_user_pages_remote(struct mm_struct *mm, 2248 unsigned long start, unsigned long nr_pages, 2249 unsigned int gup_flags, struct page **pages, 2250 struct vm_area_struct **vmas, int *locked) 2251 { 2252 if (!is_valid_gup_flags(gup_flags)) 2253 return -EINVAL; 2254 2255 return __get_user_pages_remote(mm, start, nr_pages, gup_flags, 2256 pages, vmas, locked); 2257 } 2258 EXPORT_SYMBOL(get_user_pages_remote); 2259 2260 #else /* CONFIG_MMU */ 2261 long get_user_pages_remote(struct mm_struct *mm, 2262 unsigned long start, unsigned long nr_pages, 2263 unsigned int gup_flags, struct page **pages, 2264 struct vm_area_struct **vmas, int *locked) 2265 { 2266 return 0; 2267 } 2268 2269 static long __get_user_pages_remote(struct mm_struct *mm, 2270 unsigned long start, unsigned long nr_pages, 2271 unsigned int gup_flags, struct page **pages, 2272 struct vm_area_struct **vmas, int *locked) 2273 { 2274 return 0; 2275 } 2276 #endif /* !CONFIG_MMU */ 2277 2278 /** 2279 * get_user_pages() - pin user pages in memory 2280 * @start: starting user address 2281 * @nr_pages: number of pages from start to pin 2282 * @gup_flags: flags modifying lookup behaviour 2283 * @pages: array that receives pointers to the pages pinned. 2284 * Should be at least nr_pages long. Or NULL, if caller 2285 * only intends to ensure the pages are faulted in. 2286 * @vmas: array of pointers to vmas corresponding to each page. 2287 * Or NULL if the caller does not require them. 2288 * 2289 * This is the same as get_user_pages_remote(), just with a less-flexible 2290 * calling convention where we assume that the mm being operated on belongs to 2291 * the current task, and doesn't allow passing of a locked parameter. We also 2292 * obviously don't pass FOLL_REMOTE in here. 2293 */ 2294 long get_user_pages(unsigned long start, unsigned long nr_pages, 2295 unsigned int gup_flags, struct page **pages, 2296 struct vm_area_struct **vmas) 2297 { 2298 if (!is_valid_gup_flags(gup_flags)) 2299 return -EINVAL; 2300 2301 return __gup_longterm_locked(current->mm, start, nr_pages, 2302 pages, vmas, gup_flags | FOLL_TOUCH); 2303 } 2304 EXPORT_SYMBOL(get_user_pages); 2305 2306 /* 2307 * get_user_pages_unlocked() is suitable to replace the form: 2308 * 2309 * mmap_read_lock(mm); 2310 * get_user_pages(mm, ..., pages, NULL); 2311 * mmap_read_unlock(mm); 2312 * 2313 * with: 2314 * 2315 * get_user_pages_unlocked(mm, ..., pages); 2316 * 2317 * It is functionally equivalent to get_user_pages_fast so 2318 * get_user_pages_fast should be used instead if specific gup_flags 2319 * (e.g. FOLL_FORCE) are not required. 2320 */ 2321 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages, 2322 struct page **pages, unsigned int gup_flags) 2323 { 2324 struct mm_struct *mm = current->mm; 2325 int locked = 1; 2326 long ret; 2327 2328 /* 2329 * FIXME: Current FOLL_LONGTERM behavior is incompatible with 2330 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on 2331 * vmas. As there are no users of this flag in this call we simply 2332 * disallow this option for now. 2333 */ 2334 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM)) 2335 return -EINVAL; 2336 2337 mmap_read_lock(mm); 2338 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL, 2339 &locked, gup_flags | FOLL_TOUCH); 2340 if (locked) 2341 mmap_read_unlock(mm); 2342 return ret; 2343 } 2344 EXPORT_SYMBOL(get_user_pages_unlocked); 2345 2346 /* 2347 * Fast GUP 2348 * 2349 * get_user_pages_fast attempts to pin user pages by walking the page 2350 * tables directly and avoids taking locks. Thus the walker needs to be 2351 * protected from page table pages being freed from under it, and should 2352 * block any THP splits. 2353 * 2354 * One way to achieve this is to have the walker disable interrupts, and 2355 * rely on IPIs from the TLB flushing code blocking before the page table 2356 * pages are freed. This is unsuitable for architectures that do not need 2357 * to broadcast an IPI when invalidating TLBs. 2358 * 2359 * Another way to achieve this is to batch up page table containing pages 2360 * belonging to more than one mm_user, then rcu_sched a callback to free those 2361 * pages. Disabling interrupts will allow the fast_gup walker to both block 2362 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs 2363 * (which is a relatively rare event). The code below adopts this strategy. 2364 * 2365 * Before activating this code, please be aware that the following assumptions 2366 * are currently made: 2367 * 2368 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to 2369 * free pages containing page tables or TLB flushing requires IPI broadcast. 2370 * 2371 * *) ptes can be read atomically by the architecture. 2372 * 2373 * *) access_ok is sufficient to validate userspace address ranges. 2374 * 2375 * The last two assumptions can be relaxed by the addition of helper functions. 2376 * 2377 * This code is based heavily on the PowerPC implementation by Nick Piggin. 2378 */ 2379 #ifdef CONFIG_HAVE_FAST_GUP 2380 2381 static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start, 2382 unsigned int flags, 2383 struct page **pages) 2384 { 2385 while ((*nr) - nr_start) { 2386 struct page *page = pages[--(*nr)]; 2387 2388 ClearPageReferenced(page); 2389 if (flags & FOLL_PIN) 2390 unpin_user_page(page); 2391 else 2392 put_page(page); 2393 } 2394 } 2395 2396 #ifdef CONFIG_ARCH_HAS_PTE_SPECIAL 2397 /* 2398 * Fast-gup relies on pte change detection to avoid concurrent pgtable 2399 * operations. 2400 * 2401 * To pin the page, fast-gup needs to do below in order: 2402 * (1) pin the page (by prefetching pte), then (2) check pte not changed. 2403 * 2404 * For the rest of pgtable operations where pgtable updates can be racy 2405 * with fast-gup, we need to do (1) clear pte, then (2) check whether page 2406 * is pinned. 2407 * 2408 * Above will work for all pte-level operations, including THP split. 2409 * 2410 * For THP collapse, it's a bit more complicated because fast-gup may be 2411 * walking a pgtable page that is being freed (pte is still valid but pmd 2412 * can be cleared already). To avoid race in such condition, we need to 2413 * also check pmd here to make sure pmd doesn't change (corresponds to 2414 * pmdp_collapse_flush() in the THP collapse code path). 2415 */ 2416 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr, 2417 unsigned long end, unsigned int flags, 2418 struct page **pages, int *nr) 2419 { 2420 struct dev_pagemap *pgmap = NULL; 2421 int nr_start = *nr, ret = 0; 2422 pte_t *ptep, *ptem; 2423 2424 ptem = ptep = pte_offset_map(&pmd, addr); 2425 do { 2426 pte_t pte = ptep_get_lockless(ptep); 2427 struct page *page; 2428 struct folio *folio; 2429 2430 /* 2431 * Similar to the PMD case below, NUMA hinting must take slow 2432 * path using the pte_protnone check. 2433 */ 2434 if (pte_protnone(pte)) 2435 goto pte_unmap; 2436 2437 if (!pte_access_permitted(pte, flags & FOLL_WRITE)) 2438 goto pte_unmap; 2439 2440 if (pte_devmap(pte)) { 2441 if (unlikely(flags & FOLL_LONGTERM)) 2442 goto pte_unmap; 2443 2444 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap); 2445 if (unlikely(!pgmap)) { 2446 undo_dev_pagemap(nr, nr_start, flags, pages); 2447 goto pte_unmap; 2448 } 2449 } else if (pte_special(pte)) 2450 goto pte_unmap; 2451 2452 VM_BUG_ON(!pfn_valid(pte_pfn(pte))); 2453 page = pte_page(pte); 2454 2455 folio = try_grab_folio(page, 1, flags); 2456 if (!folio) 2457 goto pte_unmap; 2458 2459 if (unlikely(page_is_secretmem(page))) { 2460 gup_put_folio(folio, 1, flags); 2461 goto pte_unmap; 2462 } 2463 2464 if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) || 2465 unlikely(pte_val(pte) != pte_val(*ptep))) { 2466 gup_put_folio(folio, 1, flags); 2467 goto pte_unmap; 2468 } 2469 2470 if (!pte_write(pte) && gup_must_unshare(flags, page)) { 2471 gup_put_folio(folio, 1, flags); 2472 goto pte_unmap; 2473 } 2474 2475 /* 2476 * We need to make the page accessible if and only if we are 2477 * going to access its content (the FOLL_PIN case). Please 2478 * see Documentation/core-api/pin_user_pages.rst for 2479 * details. 2480 */ 2481 if (flags & FOLL_PIN) { 2482 ret = arch_make_page_accessible(page); 2483 if (ret) { 2484 gup_put_folio(folio, 1, flags); 2485 goto pte_unmap; 2486 } 2487 } 2488 folio_set_referenced(folio); 2489 pages[*nr] = page; 2490 (*nr)++; 2491 } while (ptep++, addr += PAGE_SIZE, addr != end); 2492 2493 ret = 1; 2494 2495 pte_unmap: 2496 if (pgmap) 2497 put_dev_pagemap(pgmap); 2498 pte_unmap(ptem); 2499 return ret; 2500 } 2501 #else 2502 2503 /* 2504 * If we can't determine whether or not a pte is special, then fail immediately 2505 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not 2506 * to be special. 2507 * 2508 * For a futex to be placed on a THP tail page, get_futex_key requires a 2509 * get_user_pages_fast_only implementation that can pin pages. Thus it's still 2510 * useful to have gup_huge_pmd even if we can't operate on ptes. 2511 */ 2512 static int gup_pte_range(pmd_t pmd, pmd_t *pmdp, unsigned long addr, 2513 unsigned long end, unsigned int flags, 2514 struct page **pages, int *nr) 2515 { 2516 return 0; 2517 } 2518 #endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */ 2519 2520 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE) 2521 static int __gup_device_huge(unsigned long pfn, unsigned long addr, 2522 unsigned long end, unsigned int flags, 2523 struct page **pages, int *nr) 2524 { 2525 int nr_start = *nr; 2526 struct dev_pagemap *pgmap = NULL; 2527 2528 do { 2529 struct page *page = pfn_to_page(pfn); 2530 2531 pgmap = get_dev_pagemap(pfn, pgmap); 2532 if (unlikely(!pgmap)) { 2533 undo_dev_pagemap(nr, nr_start, flags, pages); 2534 break; 2535 } 2536 SetPageReferenced(page); 2537 pages[*nr] = page; 2538 if (unlikely(!try_grab_page(page, flags))) { 2539 undo_dev_pagemap(nr, nr_start, flags, pages); 2540 break; 2541 } 2542 (*nr)++; 2543 pfn++; 2544 } while (addr += PAGE_SIZE, addr != end); 2545 2546 put_dev_pagemap(pgmap); 2547 return addr == end; 2548 } 2549 2550 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr, 2551 unsigned long end, unsigned int flags, 2552 struct page **pages, int *nr) 2553 { 2554 unsigned long fault_pfn; 2555 int nr_start = *nr; 2556 2557 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT); 2558 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr)) 2559 return 0; 2560 2561 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { 2562 undo_dev_pagemap(nr, nr_start, flags, pages); 2563 return 0; 2564 } 2565 return 1; 2566 } 2567 2568 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr, 2569 unsigned long end, unsigned int flags, 2570 struct page **pages, int *nr) 2571 { 2572 unsigned long fault_pfn; 2573 int nr_start = *nr; 2574 2575 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); 2576 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr)) 2577 return 0; 2578 2579 if (unlikely(pud_val(orig) != pud_val(*pudp))) { 2580 undo_dev_pagemap(nr, nr_start, flags, pages); 2581 return 0; 2582 } 2583 return 1; 2584 } 2585 #else 2586 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr, 2587 unsigned long end, unsigned int flags, 2588 struct page **pages, int *nr) 2589 { 2590 BUILD_BUG(); 2591 return 0; 2592 } 2593 2594 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr, 2595 unsigned long end, unsigned int flags, 2596 struct page **pages, int *nr) 2597 { 2598 BUILD_BUG(); 2599 return 0; 2600 } 2601 #endif 2602 2603 static int record_subpages(struct page *page, unsigned long addr, 2604 unsigned long end, struct page **pages) 2605 { 2606 int nr; 2607 2608 for (nr = 0; addr != end; nr++, addr += PAGE_SIZE) 2609 pages[nr] = nth_page(page, nr); 2610 2611 return nr; 2612 } 2613 2614 #ifdef CONFIG_ARCH_HAS_HUGEPD 2615 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, 2616 unsigned long sz) 2617 { 2618 unsigned long __boundary = (addr + sz) & ~(sz-1); 2619 return (__boundary - 1 < end - 1) ? __boundary : end; 2620 } 2621 2622 static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, 2623 unsigned long end, unsigned int flags, 2624 struct page **pages, int *nr) 2625 { 2626 unsigned long pte_end; 2627 struct page *page; 2628 struct folio *folio; 2629 pte_t pte; 2630 int refs; 2631 2632 pte_end = (addr + sz) & ~(sz-1); 2633 if (pte_end < end) 2634 end = pte_end; 2635 2636 pte = huge_ptep_get(ptep); 2637 2638 if (!pte_access_permitted(pte, flags & FOLL_WRITE)) 2639 return 0; 2640 2641 /* hugepages are never "special" */ 2642 VM_BUG_ON(!pfn_valid(pte_pfn(pte))); 2643 2644 page = nth_page(pte_page(pte), (addr & (sz - 1)) >> PAGE_SHIFT); 2645 refs = record_subpages(page, addr, end, pages + *nr); 2646 2647 folio = try_grab_folio(page, refs, flags); 2648 if (!folio) 2649 return 0; 2650 2651 if (unlikely(pte_val(pte) != pte_val(*ptep))) { 2652 gup_put_folio(folio, refs, flags); 2653 return 0; 2654 } 2655 2656 if (!pte_write(pte) && gup_must_unshare(flags, &folio->page)) { 2657 gup_put_folio(folio, refs, flags); 2658 return 0; 2659 } 2660 2661 *nr += refs; 2662 folio_set_referenced(folio); 2663 return 1; 2664 } 2665 2666 static int gup_huge_pd(hugepd_t hugepd, unsigned long addr, 2667 unsigned int pdshift, unsigned long end, unsigned int flags, 2668 struct page **pages, int *nr) 2669 { 2670 pte_t *ptep; 2671 unsigned long sz = 1UL << hugepd_shift(hugepd); 2672 unsigned long next; 2673 2674 ptep = hugepte_offset(hugepd, addr, pdshift); 2675 do { 2676 next = hugepte_addr_end(addr, end, sz); 2677 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr)) 2678 return 0; 2679 } while (ptep++, addr = next, addr != end); 2680 2681 return 1; 2682 } 2683 #else 2684 static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr, 2685 unsigned int pdshift, unsigned long end, unsigned int flags, 2686 struct page **pages, int *nr) 2687 { 2688 return 0; 2689 } 2690 #endif /* CONFIG_ARCH_HAS_HUGEPD */ 2691 2692 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr, 2693 unsigned long end, unsigned int flags, 2694 struct page **pages, int *nr) 2695 { 2696 struct page *page; 2697 struct folio *folio; 2698 int refs; 2699 2700 if (!pmd_access_permitted(orig, flags & FOLL_WRITE)) 2701 return 0; 2702 2703 if (pmd_devmap(orig)) { 2704 if (unlikely(flags & FOLL_LONGTERM)) 2705 return 0; 2706 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags, 2707 pages, nr); 2708 } 2709 2710 page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT); 2711 refs = record_subpages(page, addr, end, pages + *nr); 2712 2713 folio = try_grab_folio(page, refs, flags); 2714 if (!folio) 2715 return 0; 2716 2717 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) { 2718 gup_put_folio(folio, refs, flags); 2719 return 0; 2720 } 2721 2722 if (!pmd_write(orig) && gup_must_unshare(flags, &folio->page)) { 2723 gup_put_folio(folio, refs, flags); 2724 return 0; 2725 } 2726 2727 *nr += refs; 2728 folio_set_referenced(folio); 2729 return 1; 2730 } 2731 2732 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr, 2733 unsigned long end, unsigned int flags, 2734 struct page **pages, int *nr) 2735 { 2736 struct page *page; 2737 struct folio *folio; 2738 int refs; 2739 2740 if (!pud_access_permitted(orig, flags & FOLL_WRITE)) 2741 return 0; 2742 2743 if (pud_devmap(orig)) { 2744 if (unlikely(flags & FOLL_LONGTERM)) 2745 return 0; 2746 return __gup_device_huge_pud(orig, pudp, addr, end, flags, 2747 pages, nr); 2748 } 2749 2750 page = nth_page(pud_page(orig), (addr & ~PUD_MASK) >> PAGE_SHIFT); 2751 refs = record_subpages(page, addr, end, pages + *nr); 2752 2753 folio = try_grab_folio(page, refs, flags); 2754 if (!folio) 2755 return 0; 2756 2757 if (unlikely(pud_val(orig) != pud_val(*pudp))) { 2758 gup_put_folio(folio, refs, flags); 2759 return 0; 2760 } 2761 2762 if (!pud_write(orig) && gup_must_unshare(flags, &folio->page)) { 2763 gup_put_folio(folio, refs, flags); 2764 return 0; 2765 } 2766 2767 *nr += refs; 2768 folio_set_referenced(folio); 2769 return 1; 2770 } 2771 2772 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr, 2773 unsigned long end, unsigned int flags, 2774 struct page **pages, int *nr) 2775 { 2776 int refs; 2777 struct page *page; 2778 struct folio *folio; 2779 2780 if (!pgd_access_permitted(orig, flags & FOLL_WRITE)) 2781 return 0; 2782 2783 BUILD_BUG_ON(pgd_devmap(orig)); 2784 2785 page = nth_page(pgd_page(orig), (addr & ~PGDIR_MASK) >> PAGE_SHIFT); 2786 refs = record_subpages(page, addr, end, pages + *nr); 2787 2788 folio = try_grab_folio(page, refs, flags); 2789 if (!folio) 2790 return 0; 2791 2792 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) { 2793 gup_put_folio(folio, refs, flags); 2794 return 0; 2795 } 2796 2797 *nr += refs; 2798 folio_set_referenced(folio); 2799 return 1; 2800 } 2801 2802 static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end, 2803 unsigned int flags, struct page **pages, int *nr) 2804 { 2805 unsigned long next; 2806 pmd_t *pmdp; 2807 2808 pmdp = pmd_offset_lockless(pudp, pud, addr); 2809 do { 2810 pmd_t pmd = READ_ONCE(*pmdp); 2811 2812 next = pmd_addr_end(addr, end); 2813 if (!pmd_present(pmd)) 2814 return 0; 2815 2816 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) || 2817 pmd_devmap(pmd))) { 2818 /* 2819 * NUMA hinting faults need to be handled in the GUP 2820 * slowpath for accounting purposes and so that they 2821 * can be serialised against THP migration. 2822 */ 2823 if (pmd_protnone(pmd)) 2824 return 0; 2825 2826 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags, 2827 pages, nr)) 2828 return 0; 2829 2830 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) { 2831 /* 2832 * architecture have different format for hugetlbfs 2833 * pmd format and THP pmd format 2834 */ 2835 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr, 2836 PMD_SHIFT, next, flags, pages, nr)) 2837 return 0; 2838 } else if (!gup_pte_range(pmd, pmdp, addr, next, flags, pages, nr)) 2839 return 0; 2840 } while (pmdp++, addr = next, addr != end); 2841 2842 return 1; 2843 } 2844 2845 static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end, 2846 unsigned int flags, struct page **pages, int *nr) 2847 { 2848 unsigned long next; 2849 pud_t *pudp; 2850 2851 pudp = pud_offset_lockless(p4dp, p4d, addr); 2852 do { 2853 pud_t pud = READ_ONCE(*pudp); 2854 2855 next = pud_addr_end(addr, end); 2856 if (unlikely(!pud_present(pud))) 2857 return 0; 2858 if (unlikely(pud_huge(pud))) { 2859 if (!gup_huge_pud(pud, pudp, addr, next, flags, 2860 pages, nr)) 2861 return 0; 2862 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) { 2863 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr, 2864 PUD_SHIFT, next, flags, pages, nr)) 2865 return 0; 2866 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr)) 2867 return 0; 2868 } while (pudp++, addr = next, addr != end); 2869 2870 return 1; 2871 } 2872 2873 static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end, 2874 unsigned int flags, struct page **pages, int *nr) 2875 { 2876 unsigned long next; 2877 p4d_t *p4dp; 2878 2879 p4dp = p4d_offset_lockless(pgdp, pgd, addr); 2880 do { 2881 p4d_t p4d = READ_ONCE(*p4dp); 2882 2883 next = p4d_addr_end(addr, end); 2884 if (p4d_none(p4d)) 2885 return 0; 2886 BUILD_BUG_ON(p4d_huge(p4d)); 2887 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) { 2888 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr, 2889 P4D_SHIFT, next, flags, pages, nr)) 2890 return 0; 2891 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr)) 2892 return 0; 2893 } while (p4dp++, addr = next, addr != end); 2894 2895 return 1; 2896 } 2897 2898 static void gup_pgd_range(unsigned long addr, unsigned long end, 2899 unsigned int flags, struct page **pages, int *nr) 2900 { 2901 unsigned long next; 2902 pgd_t *pgdp; 2903 2904 pgdp = pgd_offset(current->mm, addr); 2905 do { 2906 pgd_t pgd = READ_ONCE(*pgdp); 2907 2908 next = pgd_addr_end(addr, end); 2909 if (pgd_none(pgd)) 2910 return; 2911 if (unlikely(pgd_huge(pgd))) { 2912 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags, 2913 pages, nr)) 2914 return; 2915 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) { 2916 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr, 2917 PGDIR_SHIFT, next, flags, pages, nr)) 2918 return; 2919 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr)) 2920 return; 2921 } while (pgdp++, addr = next, addr != end); 2922 } 2923 #else 2924 static inline void gup_pgd_range(unsigned long addr, unsigned long end, 2925 unsigned int flags, struct page **pages, int *nr) 2926 { 2927 } 2928 #endif /* CONFIG_HAVE_FAST_GUP */ 2929 2930 #ifndef gup_fast_permitted 2931 /* 2932 * Check if it's allowed to use get_user_pages_fast_only() for the range, or 2933 * we need to fall back to the slow version: 2934 */ 2935 static bool gup_fast_permitted(unsigned long start, unsigned long end) 2936 { 2937 return true; 2938 } 2939 #endif 2940 2941 static int __gup_longterm_unlocked(unsigned long start, int nr_pages, 2942 unsigned int gup_flags, struct page **pages) 2943 { 2944 int ret; 2945 2946 /* 2947 * FIXME: FOLL_LONGTERM does not work with 2948 * get_user_pages_unlocked() (see comments in that function) 2949 */ 2950 if (gup_flags & FOLL_LONGTERM) { 2951 mmap_read_lock(current->mm); 2952 ret = __gup_longterm_locked(current->mm, 2953 start, nr_pages, 2954 pages, NULL, gup_flags); 2955 mmap_read_unlock(current->mm); 2956 } else { 2957 ret = get_user_pages_unlocked(start, nr_pages, 2958 pages, gup_flags); 2959 } 2960 2961 return ret; 2962 } 2963 2964 static unsigned long lockless_pages_from_mm(unsigned long start, 2965 unsigned long end, 2966 unsigned int gup_flags, 2967 struct page **pages) 2968 { 2969 unsigned long flags; 2970 int nr_pinned = 0; 2971 unsigned seq; 2972 2973 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) || 2974 !gup_fast_permitted(start, end)) 2975 return 0; 2976 2977 if (gup_flags & FOLL_PIN) { 2978 seq = raw_read_seqcount(¤t->mm->write_protect_seq); 2979 if (seq & 1) 2980 return 0; 2981 } 2982 2983 /* 2984 * Disable interrupts. The nested form is used, in order to allow full, 2985 * general purpose use of this routine. 2986 * 2987 * With interrupts disabled, we block page table pages from being freed 2988 * from under us. See struct mmu_table_batch comments in 2989 * include/asm-generic/tlb.h for more details. 2990 * 2991 * We do not adopt an rcu_read_lock() here as we also want to block IPIs 2992 * that come from THPs splitting. 2993 */ 2994 local_irq_save(flags); 2995 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned); 2996 local_irq_restore(flags); 2997 2998 /* 2999 * When pinning pages for DMA there could be a concurrent write protect 3000 * from fork() via copy_page_range(), in this case always fail fast GUP. 3001 */ 3002 if (gup_flags & FOLL_PIN) { 3003 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) { 3004 unpin_user_pages_lockless(pages, nr_pinned); 3005 return 0; 3006 } else { 3007 sanity_check_pinned_pages(pages, nr_pinned); 3008 } 3009 } 3010 return nr_pinned; 3011 } 3012 3013 static int internal_get_user_pages_fast(unsigned long start, 3014 unsigned long nr_pages, 3015 unsigned int gup_flags, 3016 struct page **pages) 3017 { 3018 unsigned long len, end; 3019 unsigned long nr_pinned; 3020 int ret; 3021 3022 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM | 3023 FOLL_FORCE | FOLL_PIN | FOLL_GET | 3024 FOLL_FAST_ONLY | FOLL_NOFAULT))) 3025 return -EINVAL; 3026 3027 if (gup_flags & FOLL_PIN) 3028 mm_set_has_pinned_flag(¤t->mm->flags); 3029 3030 if (!(gup_flags & FOLL_FAST_ONLY)) 3031 might_lock_read(¤t->mm->mmap_lock); 3032 3033 start = untagged_addr(start) & PAGE_MASK; 3034 len = nr_pages << PAGE_SHIFT; 3035 if (check_add_overflow(start, len, &end)) 3036 return 0; 3037 if (unlikely(!access_ok((void __user *)start, len))) 3038 return -EFAULT; 3039 3040 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages); 3041 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY) 3042 return nr_pinned; 3043 3044 /* Slow path: try to get the remaining pages with get_user_pages */ 3045 start += nr_pinned << PAGE_SHIFT; 3046 pages += nr_pinned; 3047 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, gup_flags, 3048 pages); 3049 if (ret < 0) { 3050 /* 3051 * The caller has to unpin the pages we already pinned so 3052 * returning -errno is not an option 3053 */ 3054 if (nr_pinned) 3055 return nr_pinned; 3056 return ret; 3057 } 3058 return ret + nr_pinned; 3059 } 3060 3061 /** 3062 * get_user_pages_fast_only() - pin user pages in memory 3063 * @start: starting user address 3064 * @nr_pages: number of pages from start to pin 3065 * @gup_flags: flags modifying pin behaviour 3066 * @pages: array that receives pointers to the pages pinned. 3067 * Should be at least nr_pages long. 3068 * 3069 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to 3070 * the regular GUP. 3071 * Note a difference with get_user_pages_fast: this always returns the 3072 * number of pages pinned, 0 if no pages were pinned. 3073 * 3074 * If the architecture does not support this function, simply return with no 3075 * pages pinned. 3076 * 3077 * Careful, careful! COW breaking can go either way, so a non-write 3078 * access can get ambiguous page results. If you call this function without 3079 * 'write' set, you'd better be sure that you're ok with that ambiguity. 3080 */ 3081 int get_user_pages_fast_only(unsigned long start, int nr_pages, 3082 unsigned int gup_flags, struct page **pages) 3083 { 3084 int nr_pinned; 3085 /* 3086 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET, 3087 * because gup fast is always a "pin with a +1 page refcount" request. 3088 * 3089 * FOLL_FAST_ONLY is required in order to match the API description of 3090 * this routine: no fall back to regular ("slow") GUP. 3091 */ 3092 gup_flags |= FOLL_GET | FOLL_FAST_ONLY; 3093 3094 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags, 3095 pages); 3096 3097 /* 3098 * As specified in the API description above, this routine is not 3099 * allowed to return negative values. However, the common core 3100 * routine internal_get_user_pages_fast() *can* return -errno. 3101 * Therefore, correct for that here: 3102 */ 3103 if (nr_pinned < 0) 3104 nr_pinned = 0; 3105 3106 return nr_pinned; 3107 } 3108 EXPORT_SYMBOL_GPL(get_user_pages_fast_only); 3109 3110 /** 3111 * get_user_pages_fast() - pin user pages in memory 3112 * @start: starting user address 3113 * @nr_pages: number of pages from start to pin 3114 * @gup_flags: flags modifying pin behaviour 3115 * @pages: array that receives pointers to the pages pinned. 3116 * Should be at least nr_pages long. 3117 * 3118 * Attempt to pin user pages in memory without taking mm->mmap_lock. 3119 * If not successful, it will fall back to taking the lock and 3120 * calling get_user_pages(). 3121 * 3122 * Returns number of pages pinned. This may be fewer than the number requested. 3123 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns 3124 * -errno. 3125 */ 3126 int get_user_pages_fast(unsigned long start, int nr_pages, 3127 unsigned int gup_flags, struct page **pages) 3128 { 3129 if (!is_valid_gup_flags(gup_flags)) 3130 return -EINVAL; 3131 3132 /* 3133 * The caller may or may not have explicitly set FOLL_GET; either way is 3134 * OK. However, internally (within mm/gup.c), gup fast variants must set 3135 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount" 3136 * request. 3137 */ 3138 gup_flags |= FOLL_GET; 3139 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages); 3140 } 3141 EXPORT_SYMBOL_GPL(get_user_pages_fast); 3142 3143 /** 3144 * pin_user_pages_fast() - pin user pages in memory without taking locks 3145 * 3146 * @start: starting user address 3147 * @nr_pages: number of pages from start to pin 3148 * @gup_flags: flags modifying pin behaviour 3149 * @pages: array that receives pointers to the pages pinned. 3150 * Should be at least nr_pages long. 3151 * 3152 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See 3153 * get_user_pages_fast() for documentation on the function arguments, because 3154 * the arguments here are identical. 3155 * 3156 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please 3157 * see Documentation/core-api/pin_user_pages.rst for further details. 3158 */ 3159 int pin_user_pages_fast(unsigned long start, int nr_pages, 3160 unsigned int gup_flags, struct page **pages) 3161 { 3162 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 3163 if (WARN_ON_ONCE(gup_flags & FOLL_GET)) 3164 return -EINVAL; 3165 3166 if (WARN_ON_ONCE(!pages)) 3167 return -EINVAL; 3168 3169 gup_flags |= FOLL_PIN; 3170 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages); 3171 } 3172 EXPORT_SYMBOL_GPL(pin_user_pages_fast); 3173 3174 /* 3175 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior 3176 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET. 3177 * 3178 * The API rules are the same, too: no negative values may be returned. 3179 */ 3180 int pin_user_pages_fast_only(unsigned long start, int nr_pages, 3181 unsigned int gup_flags, struct page **pages) 3182 { 3183 int nr_pinned; 3184 3185 /* 3186 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API 3187 * rules require returning 0, rather than -errno: 3188 */ 3189 if (WARN_ON_ONCE(gup_flags & FOLL_GET)) 3190 return 0; 3191 3192 if (WARN_ON_ONCE(!pages)) 3193 return 0; 3194 /* 3195 * FOLL_FAST_ONLY is required in order to match the API description of 3196 * this routine: no fall back to regular ("slow") GUP. 3197 */ 3198 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY); 3199 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags, 3200 pages); 3201 /* 3202 * This routine is not allowed to return negative values. However, 3203 * internal_get_user_pages_fast() *can* return -errno. Therefore, 3204 * correct for that here: 3205 */ 3206 if (nr_pinned < 0) 3207 nr_pinned = 0; 3208 3209 return nr_pinned; 3210 } 3211 EXPORT_SYMBOL_GPL(pin_user_pages_fast_only); 3212 3213 /** 3214 * pin_user_pages_remote() - pin pages of a remote process 3215 * 3216 * @mm: mm_struct of target mm 3217 * @start: starting user address 3218 * @nr_pages: number of pages from start to pin 3219 * @gup_flags: flags modifying lookup behaviour 3220 * @pages: array that receives pointers to the pages pinned. 3221 * Should be at least nr_pages long. 3222 * @vmas: array of pointers to vmas corresponding to each page. 3223 * Or NULL if the caller does not require them. 3224 * @locked: pointer to lock flag indicating whether lock is held and 3225 * subsequently whether VM_FAULT_RETRY functionality can be 3226 * utilised. Lock must initially be held. 3227 * 3228 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See 3229 * get_user_pages_remote() for documentation on the function arguments, because 3230 * the arguments here are identical. 3231 * 3232 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please 3233 * see Documentation/core-api/pin_user_pages.rst for details. 3234 */ 3235 long pin_user_pages_remote(struct mm_struct *mm, 3236 unsigned long start, unsigned long nr_pages, 3237 unsigned int gup_flags, struct page **pages, 3238 struct vm_area_struct **vmas, int *locked) 3239 { 3240 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 3241 if (WARN_ON_ONCE(gup_flags & FOLL_GET)) 3242 return -EINVAL; 3243 3244 if (WARN_ON_ONCE(!pages)) 3245 return -EINVAL; 3246 3247 gup_flags |= FOLL_PIN; 3248 return __get_user_pages_remote(mm, start, nr_pages, gup_flags, 3249 pages, vmas, locked); 3250 } 3251 EXPORT_SYMBOL(pin_user_pages_remote); 3252 3253 /** 3254 * pin_user_pages() - pin user pages in memory for use by other devices 3255 * 3256 * @start: starting user address 3257 * @nr_pages: number of pages from start to pin 3258 * @gup_flags: flags modifying lookup behaviour 3259 * @pages: array that receives pointers to the pages pinned. 3260 * Should be at least nr_pages long. 3261 * @vmas: array of pointers to vmas corresponding to each page. 3262 * Or NULL if the caller does not require them. 3263 * 3264 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and 3265 * FOLL_PIN is set. 3266 * 3267 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please 3268 * see Documentation/core-api/pin_user_pages.rst for details. 3269 */ 3270 long pin_user_pages(unsigned long start, unsigned long nr_pages, 3271 unsigned int gup_flags, struct page **pages, 3272 struct vm_area_struct **vmas) 3273 { 3274 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 3275 if (WARN_ON_ONCE(gup_flags & FOLL_GET)) 3276 return -EINVAL; 3277 3278 if (WARN_ON_ONCE(!pages)) 3279 return -EINVAL; 3280 3281 gup_flags |= FOLL_PIN; 3282 return __gup_longterm_locked(current->mm, start, nr_pages, 3283 pages, vmas, gup_flags); 3284 } 3285 EXPORT_SYMBOL(pin_user_pages); 3286 3287 /* 3288 * pin_user_pages_unlocked() is the FOLL_PIN variant of 3289 * get_user_pages_unlocked(). Behavior is the same, except that this one sets 3290 * FOLL_PIN and rejects FOLL_GET. 3291 */ 3292 long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages, 3293 struct page **pages, unsigned int gup_flags) 3294 { 3295 /* FOLL_GET and FOLL_PIN are mutually exclusive. */ 3296 if (WARN_ON_ONCE(gup_flags & FOLL_GET)) 3297 return -EINVAL; 3298 3299 if (WARN_ON_ONCE(!pages)) 3300 return -EINVAL; 3301 3302 gup_flags |= FOLL_PIN; 3303 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags); 3304 } 3305 EXPORT_SYMBOL(pin_user_pages_unlocked); 3306