Loading...
1#include <linux/kernel.h>
2#include <linux/errno.h>
3#include <linux/err.h>
4#include <linux/spinlock.h>
5
6#include <linux/mm.h>
7#include <linux/memremap.h>
8#include <linux/pagemap.h>
9#include <linux/rmap.h>
10#include <linux/swap.h>
11#include <linux/swapops.h>
12
13#include <linux/sched.h>
14#include <linux/rwsem.h>
15#include <linux/hugetlb.h>
16
17#include <asm/mmu_context.h>
18#include <asm/pgtable.h>
19#include <asm/tlbflush.h>
20
21#include "internal.h"
22
23static struct page *no_page_table(struct vm_area_struct *vma,
24 unsigned int flags)
25{
26 /*
27 * When core dumping an enormous anonymous area that nobody
28 * has touched so far, we don't want to allocate unnecessary pages or
29 * page tables. Return error instead of NULL to skip handle_mm_fault,
30 * then get_dump_page() will return NULL to leave a hole in the dump.
31 * But we can only make this optimization where a hole would surely
32 * be zero-filled if handle_mm_fault() actually did handle it.
33 */
34 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
35 return ERR_PTR(-EFAULT);
36 return NULL;
37}
38
39static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
40 pte_t *pte, unsigned int flags)
41{
42 /* No page to get reference */
43 if (flags & FOLL_GET)
44 return -EFAULT;
45
46 if (flags & FOLL_TOUCH) {
47 pte_t entry = *pte;
48
49 if (flags & FOLL_WRITE)
50 entry = pte_mkdirty(entry);
51 entry = pte_mkyoung(entry);
52
53 if (!pte_same(*pte, entry)) {
54 set_pte_at(vma->vm_mm, address, pte, entry);
55 update_mmu_cache(vma, address, pte);
56 }
57 }
58
59 /* Proper page table entry exists, but no corresponding struct page */
60 return -EEXIST;
61}
62
63static struct page *follow_page_pte(struct vm_area_struct *vma,
64 unsigned long address, pmd_t *pmd, unsigned int flags)
65{
66 struct mm_struct *mm = vma->vm_mm;
67 struct dev_pagemap *pgmap = NULL;
68 struct page *page;
69 spinlock_t *ptl;
70 pte_t *ptep, pte;
71
72retry:
73 if (unlikely(pmd_bad(*pmd)))
74 return no_page_table(vma, flags);
75
76 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
77 pte = *ptep;
78 if (!pte_present(pte)) {
79 swp_entry_t entry;
80 /*
81 * KSM's break_ksm() relies upon recognizing a ksm page
82 * even while it is being migrated, so for that case we
83 * need migration_entry_wait().
84 */
85 if (likely(!(flags & FOLL_MIGRATION)))
86 goto no_page;
87 if (pte_none(pte))
88 goto no_page;
89 entry = pte_to_swp_entry(pte);
90 if (!is_migration_entry(entry))
91 goto no_page;
92 pte_unmap_unlock(ptep, ptl);
93 migration_entry_wait(mm, pmd, address);
94 goto retry;
95 }
96 if ((flags & FOLL_NUMA) && pte_protnone(pte))
97 goto no_page;
98 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
99 pte_unmap_unlock(ptep, ptl);
100 return NULL;
101 }
102
103 page = vm_normal_page(vma, address, pte);
104 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
105 /*
106 * Only return device mapping pages in the FOLL_GET case since
107 * they are only valid while holding the pgmap reference.
108 */
109 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
110 if (pgmap)
111 page = pte_page(pte);
112 else
113 goto no_page;
114 } else if (unlikely(!page)) {
115 if (flags & FOLL_DUMP) {
116 /* Avoid special (like zero) pages in core dumps */
117 page = ERR_PTR(-EFAULT);
118 goto out;
119 }
120
121 if (is_zero_pfn(pte_pfn(pte))) {
122 page = pte_page(pte);
123 } else {
124 int ret;
125
126 ret = follow_pfn_pte(vma, address, ptep, flags);
127 page = ERR_PTR(ret);
128 goto out;
129 }
130 }
131
132 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
133 int ret;
134 get_page(page);
135 pte_unmap_unlock(ptep, ptl);
136 lock_page(page);
137 ret = split_huge_page(page);
138 unlock_page(page);
139 put_page(page);
140 if (ret)
141 return ERR_PTR(ret);
142 goto retry;
143 }
144
145 if (flags & FOLL_GET) {
146 get_page(page);
147
148 /* drop the pgmap reference now that we hold the page */
149 if (pgmap) {
150 put_dev_pagemap(pgmap);
151 pgmap = NULL;
152 }
153 }
154 if (flags & FOLL_TOUCH) {
155 if ((flags & FOLL_WRITE) &&
156 !pte_dirty(pte) && !PageDirty(page))
157 set_page_dirty(page);
158 /*
159 * pte_mkyoung() would be more correct here, but atomic care
160 * is needed to avoid losing the dirty bit: it is easier to use
161 * mark_page_accessed().
162 */
163 mark_page_accessed(page);
164 }
165 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
166 /* Do not mlock pte-mapped THP */
167 if (PageTransCompound(page))
168 goto out;
169
170 /*
171 * The preliminary mapping check is mainly to avoid the
172 * pointless overhead of lock_page on the ZERO_PAGE
173 * which might bounce very badly if there is contention.
174 *
175 * If the page is already locked, we don't need to
176 * handle it now - vmscan will handle it later if and
177 * when it attempts to reclaim the page.
178 */
179 if (page->mapping && trylock_page(page)) {
180 lru_add_drain(); /* push cached pages to LRU */
181 /*
182 * Because we lock page here, and migration is
183 * blocked by the pte's page reference, and we
184 * know the page is still mapped, we don't even
185 * need to check for file-cache page truncation.
186 */
187 mlock_vma_page(page);
188 unlock_page(page);
189 }
190 }
191out:
192 pte_unmap_unlock(ptep, ptl);
193 return page;
194no_page:
195 pte_unmap_unlock(ptep, ptl);
196 if (!pte_none(pte))
197 return NULL;
198 return no_page_table(vma, flags);
199}
200
201/**
202 * follow_page_mask - look up a page descriptor from a user-virtual address
203 * @vma: vm_area_struct mapping @address
204 * @address: virtual address to look up
205 * @flags: flags modifying lookup behaviour
206 * @page_mask: on output, *page_mask is set according to the size of the page
207 *
208 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
209 *
210 * Returns the mapped (struct page *), %NULL if no mapping exists, or
211 * an error pointer if there is a mapping to something not represented
212 * by a page descriptor (see also vm_normal_page()).
213 */
214struct page *follow_page_mask(struct vm_area_struct *vma,
215 unsigned long address, unsigned int flags,
216 unsigned int *page_mask)
217{
218 pgd_t *pgd;
219 pud_t *pud;
220 pmd_t *pmd;
221 spinlock_t *ptl;
222 struct page *page;
223 struct mm_struct *mm = vma->vm_mm;
224
225 *page_mask = 0;
226
227 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
228 if (!IS_ERR(page)) {
229 BUG_ON(flags & FOLL_GET);
230 return page;
231 }
232
233 pgd = pgd_offset(mm, address);
234 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
235 return no_page_table(vma, flags);
236
237 pud = pud_offset(pgd, address);
238 if (pud_none(*pud))
239 return no_page_table(vma, flags);
240 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
241 page = follow_huge_pud(mm, address, pud, flags);
242 if (page)
243 return page;
244 return no_page_table(vma, flags);
245 }
246 if (unlikely(pud_bad(*pud)))
247 return no_page_table(vma, flags);
248
249 pmd = pmd_offset(pud, address);
250 if (pmd_none(*pmd))
251 return no_page_table(vma, flags);
252 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
253 page = follow_huge_pmd(mm, address, pmd, flags);
254 if (page)
255 return page;
256 return no_page_table(vma, flags);
257 }
258 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
259 return no_page_table(vma, flags);
260 if (pmd_devmap(*pmd)) {
261 ptl = pmd_lock(mm, pmd);
262 page = follow_devmap_pmd(vma, address, pmd, flags);
263 spin_unlock(ptl);
264 if (page)
265 return page;
266 }
267 if (likely(!pmd_trans_huge(*pmd)))
268 return follow_page_pte(vma, address, pmd, flags);
269
270 ptl = pmd_lock(mm, pmd);
271 if (unlikely(!pmd_trans_huge(*pmd))) {
272 spin_unlock(ptl);
273 return follow_page_pte(vma, address, pmd, flags);
274 }
275 if (flags & FOLL_SPLIT) {
276 int ret;
277 page = pmd_page(*pmd);
278 if (is_huge_zero_page(page)) {
279 spin_unlock(ptl);
280 ret = 0;
281 split_huge_pmd(vma, pmd, address);
282 } else {
283 get_page(page);
284 spin_unlock(ptl);
285 lock_page(page);
286 ret = split_huge_page(page);
287 unlock_page(page);
288 put_page(page);
289 }
290
291 return ret ? ERR_PTR(ret) :
292 follow_page_pte(vma, address, pmd, flags);
293 }
294
295 page = follow_trans_huge_pmd(vma, address, pmd, flags);
296 spin_unlock(ptl);
297 *page_mask = HPAGE_PMD_NR - 1;
298 return page;
299}
300
301static int get_gate_page(struct mm_struct *mm, unsigned long address,
302 unsigned int gup_flags, struct vm_area_struct **vma,
303 struct page **page)
304{
305 pgd_t *pgd;
306 pud_t *pud;
307 pmd_t *pmd;
308 pte_t *pte;
309 int ret = -EFAULT;
310
311 /* user gate pages are read-only */
312 if (gup_flags & FOLL_WRITE)
313 return -EFAULT;
314 if (address > TASK_SIZE)
315 pgd = pgd_offset_k(address);
316 else
317 pgd = pgd_offset_gate(mm, address);
318 BUG_ON(pgd_none(*pgd));
319 pud = pud_offset(pgd, address);
320 BUG_ON(pud_none(*pud));
321 pmd = pmd_offset(pud, address);
322 if (pmd_none(*pmd))
323 return -EFAULT;
324 VM_BUG_ON(pmd_trans_huge(*pmd));
325 pte = pte_offset_map(pmd, address);
326 if (pte_none(*pte))
327 goto unmap;
328 *vma = get_gate_vma(mm);
329 if (!page)
330 goto out;
331 *page = vm_normal_page(*vma, address, *pte);
332 if (!*page) {
333 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
334 goto unmap;
335 *page = pte_page(*pte);
336 }
337 get_page(*page);
338out:
339 ret = 0;
340unmap:
341 pte_unmap(pte);
342 return ret;
343}
344
345/*
346 * mmap_sem must be held on entry. If @nonblocking != NULL and
347 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
348 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
349 */
350static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
351 unsigned long address, unsigned int *flags, int *nonblocking)
352{
353 struct mm_struct *mm = vma->vm_mm;
354 unsigned int fault_flags = 0;
355 int ret;
356
357 /* mlock all present pages, but do not fault in new pages */
358 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
359 return -ENOENT;
360 /* For mm_populate(), just skip the stack guard page. */
361 if ((*flags & FOLL_POPULATE) &&
362 (stack_guard_page_start(vma, address) ||
363 stack_guard_page_end(vma, address + PAGE_SIZE)))
364 return -ENOENT;
365 if (*flags & FOLL_WRITE)
366 fault_flags |= FAULT_FLAG_WRITE;
367 if (*flags & FOLL_REMOTE)
368 fault_flags |= FAULT_FLAG_REMOTE;
369 if (nonblocking)
370 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
371 if (*flags & FOLL_NOWAIT)
372 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
373 if (*flags & FOLL_TRIED) {
374 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
375 fault_flags |= FAULT_FLAG_TRIED;
376 }
377
378 ret = handle_mm_fault(mm, vma, address, fault_flags);
379 if (ret & VM_FAULT_ERROR) {
380 if (ret & VM_FAULT_OOM)
381 return -ENOMEM;
382 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
383 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
384 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
385 return -EFAULT;
386 BUG();
387 }
388
389 if (tsk) {
390 if (ret & VM_FAULT_MAJOR)
391 tsk->maj_flt++;
392 else
393 tsk->min_flt++;
394 }
395
396 if (ret & VM_FAULT_RETRY) {
397 if (nonblocking)
398 *nonblocking = 0;
399 return -EBUSY;
400 }
401
402 /*
403 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
404 * necessary, even if maybe_mkwrite decided not to set pte_write. We
405 * can thus safely do subsequent page lookups as if they were reads.
406 * But only do so when looping for pte_write is futile: in some cases
407 * userspace may also be wanting to write to the gotten user page,
408 * which a read fault here might prevent (a readonly page might get
409 * reCOWed by userspace write).
410 */
411 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
412 *flags &= ~FOLL_WRITE;
413 return 0;
414}
415
416static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
417{
418 vm_flags_t vm_flags = vma->vm_flags;
419 int write = (gup_flags & FOLL_WRITE);
420 int foreign = (gup_flags & FOLL_REMOTE);
421
422 if (vm_flags & (VM_IO | VM_PFNMAP))
423 return -EFAULT;
424
425 if (write) {
426 if (!(vm_flags & VM_WRITE)) {
427 if (!(gup_flags & FOLL_FORCE))
428 return -EFAULT;
429 /*
430 * We used to let the write,force case do COW in a
431 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
432 * set a breakpoint in a read-only mapping of an
433 * executable, without corrupting the file (yet only
434 * when that file had been opened for writing!).
435 * Anon pages in shared mappings are surprising: now
436 * just reject it.
437 */
438 if (!is_cow_mapping(vm_flags))
439 return -EFAULT;
440 }
441 } else if (!(vm_flags & VM_READ)) {
442 if (!(gup_flags & FOLL_FORCE))
443 return -EFAULT;
444 /*
445 * Is there actually any vma we can reach here which does not
446 * have VM_MAYREAD set?
447 */
448 if (!(vm_flags & VM_MAYREAD))
449 return -EFAULT;
450 }
451 /*
452 * gups are always data accesses, not instruction
453 * fetches, so execute=false here
454 */
455 if (!arch_vma_access_permitted(vma, write, false, foreign))
456 return -EFAULT;
457 return 0;
458}
459
460/**
461 * __get_user_pages() - pin user pages in memory
462 * @tsk: task_struct of target task
463 * @mm: mm_struct of target mm
464 * @start: starting user address
465 * @nr_pages: number of pages from start to pin
466 * @gup_flags: flags modifying pin behaviour
467 * @pages: array that receives pointers to the pages pinned.
468 * Should be at least nr_pages long. Or NULL, if caller
469 * only intends to ensure the pages are faulted in.
470 * @vmas: array of pointers to vmas corresponding to each page.
471 * Or NULL if the caller does not require them.
472 * @nonblocking: whether waiting for disk IO or mmap_sem contention
473 *
474 * Returns number of pages pinned. This may be fewer than the number
475 * requested. If nr_pages is 0 or negative, returns 0. If no pages
476 * were pinned, returns -errno. Each page returned must be released
477 * with a put_page() call when it is finished with. vmas will only
478 * remain valid while mmap_sem is held.
479 *
480 * Must be called with mmap_sem held. It may be released. See below.
481 *
482 * __get_user_pages walks a process's page tables and takes a reference to
483 * each struct page that each user address corresponds to at a given
484 * instant. That is, it takes the page that would be accessed if a user
485 * thread accesses the given user virtual address at that instant.
486 *
487 * This does not guarantee that the page exists in the user mappings when
488 * __get_user_pages returns, and there may even be a completely different
489 * page there in some cases (eg. if mmapped pagecache has been invalidated
490 * and subsequently re faulted). However it does guarantee that the page
491 * won't be freed completely. And mostly callers simply care that the page
492 * contains data that was valid *at some point in time*. Typically, an IO
493 * or similar operation cannot guarantee anything stronger anyway because
494 * locks can't be held over the syscall boundary.
495 *
496 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
497 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
498 * appropriate) must be called after the page is finished with, and
499 * before put_page is called.
500 *
501 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
502 * or mmap_sem contention, and if waiting is needed to pin all pages,
503 * *@nonblocking will be set to 0. Further, if @gup_flags does not
504 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
505 * this case.
506 *
507 * A caller using such a combination of @nonblocking and @gup_flags
508 * must therefore hold the mmap_sem for reading only, and recognize
509 * when it's been released. Otherwise, it must be held for either
510 * reading or writing and will not be released.
511 *
512 * In most cases, get_user_pages or get_user_pages_fast should be used
513 * instead of __get_user_pages. __get_user_pages should be used only if
514 * you need some special @gup_flags.
515 */
516long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
517 unsigned long start, unsigned long nr_pages,
518 unsigned int gup_flags, struct page **pages,
519 struct vm_area_struct **vmas, int *nonblocking)
520{
521 long i = 0;
522 unsigned int page_mask;
523 struct vm_area_struct *vma = NULL;
524
525 if (!nr_pages)
526 return 0;
527
528 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
529
530 /*
531 * If FOLL_FORCE is set then do not force a full fault as the hinting
532 * fault information is unrelated to the reference behaviour of a task
533 * using the address space
534 */
535 if (!(gup_flags & FOLL_FORCE))
536 gup_flags |= FOLL_NUMA;
537
538 do {
539 struct page *page;
540 unsigned int foll_flags = gup_flags;
541 unsigned int page_increm;
542
543 /* first iteration or cross vma bound */
544 if (!vma || start >= vma->vm_end) {
545 vma = find_extend_vma(mm, start);
546 if (!vma && in_gate_area(mm, start)) {
547 int ret;
548 ret = get_gate_page(mm, start & PAGE_MASK,
549 gup_flags, &vma,
550 pages ? &pages[i] : NULL);
551 if (ret)
552 return i ? : ret;
553 page_mask = 0;
554 goto next_page;
555 }
556
557 if (!vma || check_vma_flags(vma, gup_flags))
558 return i ? : -EFAULT;
559 if (is_vm_hugetlb_page(vma)) {
560 i = follow_hugetlb_page(mm, vma, pages, vmas,
561 &start, &nr_pages, i,
562 gup_flags);
563 continue;
564 }
565 }
566retry:
567 /*
568 * If we have a pending SIGKILL, don't keep faulting pages and
569 * potentially allocating memory.
570 */
571 if (unlikely(fatal_signal_pending(current)))
572 return i ? i : -ERESTARTSYS;
573 cond_resched();
574 page = follow_page_mask(vma, start, foll_flags, &page_mask);
575 if (!page) {
576 int ret;
577 ret = faultin_page(tsk, vma, start, &foll_flags,
578 nonblocking);
579 switch (ret) {
580 case 0:
581 goto retry;
582 case -EFAULT:
583 case -ENOMEM:
584 case -EHWPOISON:
585 return i ? i : ret;
586 case -EBUSY:
587 return i;
588 case -ENOENT:
589 goto next_page;
590 }
591 BUG();
592 } else if (PTR_ERR(page) == -EEXIST) {
593 /*
594 * Proper page table entry exists, but no corresponding
595 * struct page.
596 */
597 goto next_page;
598 } else if (IS_ERR(page)) {
599 return i ? i : PTR_ERR(page);
600 }
601 if (pages) {
602 pages[i] = page;
603 flush_anon_page(vma, page, start);
604 flush_dcache_page(page);
605 page_mask = 0;
606 }
607next_page:
608 if (vmas) {
609 vmas[i] = vma;
610 page_mask = 0;
611 }
612 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
613 if (page_increm > nr_pages)
614 page_increm = nr_pages;
615 i += page_increm;
616 start += page_increm * PAGE_SIZE;
617 nr_pages -= page_increm;
618 } while (nr_pages);
619 return i;
620}
621EXPORT_SYMBOL(__get_user_pages);
622
623bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
624{
625 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
626 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
627 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
628
629 if (!(vm_flags & vma->vm_flags))
630 return false;
631
632 /*
633 * The architecture might have a hardware protection
634 * mechanism other than read/write that can deny access.
635 *
636 * gup always represents data access, not instruction
637 * fetches, so execute=false here:
638 */
639 if (!arch_vma_access_permitted(vma, write, false, foreign))
640 return false;
641
642 return true;
643}
644
645/*
646 * fixup_user_fault() - manually resolve a user page fault
647 * @tsk: the task_struct to use for page fault accounting, or
648 * NULL if faults are not to be recorded.
649 * @mm: mm_struct of target mm
650 * @address: user address
651 * @fault_flags:flags to pass down to handle_mm_fault()
652 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
653 * does not allow retry
654 *
655 * This is meant to be called in the specific scenario where for locking reasons
656 * we try to access user memory in atomic context (within a pagefault_disable()
657 * section), this returns -EFAULT, and we want to resolve the user fault before
658 * trying again.
659 *
660 * Typically this is meant to be used by the futex code.
661 *
662 * The main difference with get_user_pages() is that this function will
663 * unconditionally call handle_mm_fault() which will in turn perform all the
664 * necessary SW fixup of the dirty and young bits in the PTE, while
665 * get_user_pages() only guarantees to update these in the struct page.
666 *
667 * This is important for some architectures where those bits also gate the
668 * access permission to the page because they are maintained in software. On
669 * such architectures, gup() will not be enough to make a subsequent access
670 * succeed.
671 *
672 * This function will not return with an unlocked mmap_sem. So it has not the
673 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
674 */
675int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
676 unsigned long address, unsigned int fault_flags,
677 bool *unlocked)
678{
679 struct vm_area_struct *vma;
680 int ret, major = 0;
681
682 if (unlocked)
683 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
684
685retry:
686 vma = find_extend_vma(mm, address);
687 if (!vma || address < vma->vm_start)
688 return -EFAULT;
689
690 if (!vma_permits_fault(vma, fault_flags))
691 return -EFAULT;
692
693 ret = handle_mm_fault(mm, vma, address, fault_flags);
694 major |= ret & VM_FAULT_MAJOR;
695 if (ret & VM_FAULT_ERROR) {
696 if (ret & VM_FAULT_OOM)
697 return -ENOMEM;
698 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
699 return -EHWPOISON;
700 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
701 return -EFAULT;
702 BUG();
703 }
704
705 if (ret & VM_FAULT_RETRY) {
706 down_read(&mm->mmap_sem);
707 if (!(fault_flags & FAULT_FLAG_TRIED)) {
708 *unlocked = true;
709 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
710 fault_flags |= FAULT_FLAG_TRIED;
711 goto retry;
712 }
713 }
714
715 if (tsk) {
716 if (major)
717 tsk->maj_flt++;
718 else
719 tsk->min_flt++;
720 }
721 return 0;
722}
723
724static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
725 struct mm_struct *mm,
726 unsigned long start,
727 unsigned long nr_pages,
728 int write, int force,
729 struct page **pages,
730 struct vm_area_struct **vmas,
731 int *locked, bool notify_drop,
732 unsigned int flags)
733{
734 long ret, pages_done;
735 bool lock_dropped;
736
737 if (locked) {
738 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
739 BUG_ON(vmas);
740 /* check caller initialized locked */
741 BUG_ON(*locked != 1);
742 }
743
744 if (pages)
745 flags |= FOLL_GET;
746 if (write)
747 flags |= FOLL_WRITE;
748 if (force)
749 flags |= FOLL_FORCE;
750
751 pages_done = 0;
752 lock_dropped = false;
753 for (;;) {
754 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
755 vmas, locked);
756 if (!locked)
757 /* VM_FAULT_RETRY couldn't trigger, bypass */
758 return ret;
759
760 /* VM_FAULT_RETRY cannot return errors */
761 if (!*locked) {
762 BUG_ON(ret < 0);
763 BUG_ON(ret >= nr_pages);
764 }
765
766 if (!pages)
767 /* If it's a prefault don't insist harder */
768 return ret;
769
770 if (ret > 0) {
771 nr_pages -= ret;
772 pages_done += ret;
773 if (!nr_pages)
774 break;
775 }
776 if (*locked) {
777 /* VM_FAULT_RETRY didn't trigger */
778 if (!pages_done)
779 pages_done = ret;
780 break;
781 }
782 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
783 pages += ret;
784 start += ret << PAGE_SHIFT;
785
786 /*
787 * Repeat on the address that fired VM_FAULT_RETRY
788 * without FAULT_FLAG_ALLOW_RETRY but with
789 * FAULT_FLAG_TRIED.
790 */
791 *locked = 1;
792 lock_dropped = true;
793 down_read(&mm->mmap_sem);
794 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
795 pages, NULL, NULL);
796 if (ret != 1) {
797 BUG_ON(ret > 1);
798 if (!pages_done)
799 pages_done = ret;
800 break;
801 }
802 nr_pages--;
803 pages_done++;
804 if (!nr_pages)
805 break;
806 pages++;
807 start += PAGE_SIZE;
808 }
809 if (notify_drop && lock_dropped && *locked) {
810 /*
811 * We must let the caller know we temporarily dropped the lock
812 * and so the critical section protected by it was lost.
813 */
814 up_read(&mm->mmap_sem);
815 *locked = 0;
816 }
817 return pages_done;
818}
819
820/*
821 * We can leverage the VM_FAULT_RETRY functionality in the page fault
822 * paths better by using either get_user_pages_locked() or
823 * get_user_pages_unlocked().
824 *
825 * get_user_pages_locked() is suitable to replace the form:
826 *
827 * down_read(&mm->mmap_sem);
828 * do_something()
829 * get_user_pages(tsk, mm, ..., pages, NULL);
830 * up_read(&mm->mmap_sem);
831 *
832 * to:
833 *
834 * int locked = 1;
835 * down_read(&mm->mmap_sem);
836 * do_something()
837 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
838 * if (locked)
839 * up_read(&mm->mmap_sem);
840 */
841long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
842 int write, int force, struct page **pages,
843 int *locked)
844{
845 return __get_user_pages_locked(current, current->mm, start, nr_pages,
846 write, force, pages, NULL, locked, true,
847 FOLL_TOUCH);
848}
849EXPORT_SYMBOL(get_user_pages_locked);
850
851/*
852 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
853 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
854 *
855 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
856 * caller if required (just like with __get_user_pages). "FOLL_GET",
857 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
858 * according to the parameters "pages", "write", "force"
859 * respectively.
860 */
861__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
862 unsigned long start, unsigned long nr_pages,
863 int write, int force, struct page **pages,
864 unsigned int gup_flags)
865{
866 long ret;
867 int locked = 1;
868 down_read(&mm->mmap_sem);
869 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
870 pages, NULL, &locked, false, gup_flags);
871 if (locked)
872 up_read(&mm->mmap_sem);
873 return ret;
874}
875EXPORT_SYMBOL(__get_user_pages_unlocked);
876
877/*
878 * get_user_pages_unlocked() is suitable to replace the form:
879 *
880 * down_read(&mm->mmap_sem);
881 * get_user_pages(tsk, mm, ..., pages, NULL);
882 * up_read(&mm->mmap_sem);
883 *
884 * with:
885 *
886 * get_user_pages_unlocked(tsk, mm, ..., pages);
887 *
888 * It is functionally equivalent to get_user_pages_fast so
889 * get_user_pages_fast should be used instead, if the two parameters
890 * "tsk" and "mm" are respectively equal to current and current->mm,
891 * or if "force" shall be set to 1 (get_user_pages_fast misses the
892 * "force" parameter).
893 */
894long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
895 int write, int force, struct page **pages)
896{
897 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
898 write, force, pages, FOLL_TOUCH);
899}
900EXPORT_SYMBOL(get_user_pages_unlocked);
901
902/*
903 * get_user_pages_remote() - pin user pages in memory
904 * @tsk: the task_struct to use for page fault accounting, or
905 * NULL if faults are not to be recorded.
906 * @mm: mm_struct of target mm
907 * @start: starting user address
908 * @nr_pages: number of pages from start to pin
909 * @write: whether pages will be written to by the caller
910 * @force: whether to force access even when user mapping is currently
911 * protected (but never forces write access to shared mapping).
912 * @pages: array that receives pointers to the pages pinned.
913 * Should be at least nr_pages long. Or NULL, if caller
914 * only intends to ensure the pages are faulted in.
915 * @vmas: array of pointers to vmas corresponding to each page.
916 * Or NULL if the caller does not require them.
917 *
918 * Returns number of pages pinned. This may be fewer than the number
919 * requested. If nr_pages is 0 or negative, returns 0. If no pages
920 * were pinned, returns -errno. Each page returned must be released
921 * with a put_page() call when it is finished with. vmas will only
922 * remain valid while mmap_sem is held.
923 *
924 * Must be called with mmap_sem held for read or write.
925 *
926 * get_user_pages walks a process's page tables and takes a reference to
927 * each struct page that each user address corresponds to at a given
928 * instant. That is, it takes the page that would be accessed if a user
929 * thread accesses the given user virtual address at that instant.
930 *
931 * This does not guarantee that the page exists in the user mappings when
932 * get_user_pages returns, and there may even be a completely different
933 * page there in some cases (eg. if mmapped pagecache has been invalidated
934 * and subsequently re faulted). However it does guarantee that the page
935 * won't be freed completely. And mostly callers simply care that the page
936 * contains data that was valid *at some point in time*. Typically, an IO
937 * or similar operation cannot guarantee anything stronger anyway because
938 * locks can't be held over the syscall boundary.
939 *
940 * If write=0, the page must not be written to. If the page is written to,
941 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
942 * after the page is finished with, and before put_page is called.
943 *
944 * get_user_pages is typically used for fewer-copy IO operations, to get a
945 * handle on the memory by some means other than accesses via the user virtual
946 * addresses. The pages may be submitted for DMA to devices or accessed via
947 * their kernel linear mapping (via the kmap APIs). Care should be taken to
948 * use the correct cache flushing APIs.
949 *
950 * See also get_user_pages_fast, for performance critical applications.
951 *
952 * get_user_pages should be phased out in favor of
953 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
954 * should use get_user_pages because it cannot pass
955 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
956 */
957long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
958 unsigned long start, unsigned long nr_pages,
959 int write, int force, struct page **pages,
960 struct vm_area_struct **vmas)
961{
962 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
963 pages, vmas, NULL, false,
964 FOLL_TOUCH | FOLL_REMOTE);
965}
966EXPORT_SYMBOL(get_user_pages_remote);
967
968/*
969 * This is the same as get_user_pages_remote(), just with a
970 * less-flexible calling convention where we assume that the task
971 * and mm being operated on are the current task's. We also
972 * obviously don't pass FOLL_REMOTE in here.
973 */
974long get_user_pages(unsigned long start, unsigned long nr_pages,
975 int write, int force, struct page **pages,
976 struct vm_area_struct **vmas)
977{
978 return __get_user_pages_locked(current, current->mm, start, nr_pages,
979 write, force, pages, vmas, NULL, false,
980 FOLL_TOUCH);
981}
982EXPORT_SYMBOL(get_user_pages);
983
984/**
985 * populate_vma_page_range() - populate a range of pages in the vma.
986 * @vma: target vma
987 * @start: start address
988 * @end: end address
989 * @nonblocking:
990 *
991 * This takes care of mlocking the pages too if VM_LOCKED is set.
992 *
993 * return 0 on success, negative error code on error.
994 *
995 * vma->vm_mm->mmap_sem must be held.
996 *
997 * If @nonblocking is NULL, it may be held for read or write and will
998 * be unperturbed.
999 *
1000 * If @nonblocking is non-NULL, it must held for read only and may be
1001 * released. If it's released, *@nonblocking will be set to 0.
1002 */
1003long populate_vma_page_range(struct vm_area_struct *vma,
1004 unsigned long start, unsigned long end, int *nonblocking)
1005{
1006 struct mm_struct *mm = vma->vm_mm;
1007 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1008 int gup_flags;
1009
1010 VM_BUG_ON(start & ~PAGE_MASK);
1011 VM_BUG_ON(end & ~PAGE_MASK);
1012 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1013 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1014 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1015
1016 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1017 if (vma->vm_flags & VM_LOCKONFAULT)
1018 gup_flags &= ~FOLL_POPULATE;
1019 /*
1020 * We want to touch writable mappings with a write fault in order
1021 * to break COW, except for shared mappings because these don't COW
1022 * and we would not want to dirty them for nothing.
1023 */
1024 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1025 gup_flags |= FOLL_WRITE;
1026
1027 /*
1028 * We want mlock to succeed for regions that have any permissions
1029 * other than PROT_NONE.
1030 */
1031 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1032 gup_flags |= FOLL_FORCE;
1033
1034 /*
1035 * We made sure addr is within a VMA, so the following will
1036 * not result in a stack expansion that recurses back here.
1037 */
1038 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1039 NULL, NULL, nonblocking);
1040}
1041
1042/*
1043 * __mm_populate - populate and/or mlock pages within a range of address space.
1044 *
1045 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1046 * flags. VMAs must be already marked with the desired vm_flags, and
1047 * mmap_sem must not be held.
1048 */
1049int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1050{
1051 struct mm_struct *mm = current->mm;
1052 unsigned long end, nstart, nend;
1053 struct vm_area_struct *vma = NULL;
1054 int locked = 0;
1055 long ret = 0;
1056
1057 VM_BUG_ON(start & ~PAGE_MASK);
1058 VM_BUG_ON(len != PAGE_ALIGN(len));
1059 end = start + len;
1060
1061 for (nstart = start; nstart < end; nstart = nend) {
1062 /*
1063 * We want to fault in pages for [nstart; end) address range.
1064 * Find first corresponding VMA.
1065 */
1066 if (!locked) {
1067 locked = 1;
1068 down_read(&mm->mmap_sem);
1069 vma = find_vma(mm, nstart);
1070 } else if (nstart >= vma->vm_end)
1071 vma = vma->vm_next;
1072 if (!vma || vma->vm_start >= end)
1073 break;
1074 /*
1075 * Set [nstart; nend) to intersection of desired address
1076 * range with the first VMA. Also, skip undesirable VMA types.
1077 */
1078 nend = min(end, vma->vm_end);
1079 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1080 continue;
1081 if (nstart < vma->vm_start)
1082 nstart = vma->vm_start;
1083 /*
1084 * Now fault in a range of pages. populate_vma_page_range()
1085 * double checks the vma flags, so that it won't mlock pages
1086 * if the vma was already munlocked.
1087 */
1088 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1089 if (ret < 0) {
1090 if (ignore_errors) {
1091 ret = 0;
1092 continue; /* continue at next VMA */
1093 }
1094 break;
1095 }
1096 nend = nstart + ret * PAGE_SIZE;
1097 ret = 0;
1098 }
1099 if (locked)
1100 up_read(&mm->mmap_sem);
1101 return ret; /* 0 or negative error code */
1102}
1103
1104/**
1105 * get_dump_page() - pin user page in memory while writing it to core dump
1106 * @addr: user address
1107 *
1108 * Returns struct page pointer of user page pinned for dump,
1109 * to be freed afterwards by put_page().
1110 *
1111 * Returns NULL on any kind of failure - a hole must then be inserted into
1112 * the corefile, to preserve alignment with its headers; and also returns
1113 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1114 * allowing a hole to be left in the corefile to save diskspace.
1115 *
1116 * Called without mmap_sem, but after all other threads have been killed.
1117 */
1118#ifdef CONFIG_ELF_CORE
1119struct page *get_dump_page(unsigned long addr)
1120{
1121 struct vm_area_struct *vma;
1122 struct page *page;
1123
1124 if (__get_user_pages(current, current->mm, addr, 1,
1125 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1126 NULL) < 1)
1127 return NULL;
1128 flush_cache_page(vma, addr, page_to_pfn(page));
1129 return page;
1130}
1131#endif /* CONFIG_ELF_CORE */
1132
1133/*
1134 * Generic RCU Fast GUP
1135 *
1136 * get_user_pages_fast attempts to pin user pages by walking the page
1137 * tables directly and avoids taking locks. Thus the walker needs to be
1138 * protected from page table pages being freed from under it, and should
1139 * block any THP splits.
1140 *
1141 * One way to achieve this is to have the walker disable interrupts, and
1142 * rely on IPIs from the TLB flushing code blocking before the page table
1143 * pages are freed. This is unsuitable for architectures that do not need
1144 * to broadcast an IPI when invalidating TLBs.
1145 *
1146 * Another way to achieve this is to batch up page table containing pages
1147 * belonging to more than one mm_user, then rcu_sched a callback to free those
1148 * pages. Disabling interrupts will allow the fast_gup walker to both block
1149 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1150 * (which is a relatively rare event). The code below adopts this strategy.
1151 *
1152 * Before activating this code, please be aware that the following assumptions
1153 * are currently made:
1154 *
1155 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1156 * pages containing page tables.
1157 *
1158 * *) ptes can be read atomically by the architecture.
1159 *
1160 * *) access_ok is sufficient to validate userspace address ranges.
1161 *
1162 * The last two assumptions can be relaxed by the addition of helper functions.
1163 *
1164 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1165 */
1166#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1167
1168#ifdef __HAVE_ARCH_PTE_SPECIAL
1169static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1170 int write, struct page **pages, int *nr)
1171{
1172 pte_t *ptep, *ptem;
1173 int ret = 0;
1174
1175 ptem = ptep = pte_offset_map(&pmd, addr);
1176 do {
1177 /*
1178 * In the line below we are assuming that the pte can be read
1179 * atomically. If this is not the case for your architecture,
1180 * please wrap this in a helper function!
1181 *
1182 * for an example see gup_get_pte in arch/x86/mm/gup.c
1183 */
1184 pte_t pte = READ_ONCE(*ptep);
1185 struct page *head, *page;
1186
1187 /*
1188 * Similar to the PMD case below, NUMA hinting must take slow
1189 * path using the pte_protnone check.
1190 */
1191 if (!pte_present(pte) || pte_special(pte) ||
1192 pte_protnone(pte) || (write && !pte_write(pte)))
1193 goto pte_unmap;
1194
1195 if (!arch_pte_access_permitted(pte, write))
1196 goto pte_unmap;
1197
1198 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1199 page = pte_page(pte);
1200 head = compound_head(page);
1201
1202 if (!page_cache_get_speculative(head))
1203 goto pte_unmap;
1204
1205 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1206 put_page(head);
1207 goto pte_unmap;
1208 }
1209
1210 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1211 pages[*nr] = page;
1212 (*nr)++;
1213
1214 } while (ptep++, addr += PAGE_SIZE, addr != end);
1215
1216 ret = 1;
1217
1218pte_unmap:
1219 pte_unmap(ptem);
1220 return ret;
1221}
1222#else
1223
1224/*
1225 * If we can't determine whether or not a pte is special, then fail immediately
1226 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1227 * to be special.
1228 *
1229 * For a futex to be placed on a THP tail page, get_futex_key requires a
1230 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1231 * useful to have gup_huge_pmd even if we can't operate on ptes.
1232 */
1233static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1234 int write, struct page **pages, int *nr)
1235{
1236 return 0;
1237}
1238#endif /* __HAVE_ARCH_PTE_SPECIAL */
1239
1240static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1241 unsigned long end, int write, struct page **pages, int *nr)
1242{
1243 struct page *head, *page;
1244 int refs;
1245
1246 if (write && !pmd_write(orig))
1247 return 0;
1248
1249 refs = 0;
1250 head = pmd_page(orig);
1251 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1252 do {
1253 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1254 pages[*nr] = page;
1255 (*nr)++;
1256 page++;
1257 refs++;
1258 } while (addr += PAGE_SIZE, addr != end);
1259
1260 if (!page_cache_add_speculative(head, refs)) {
1261 *nr -= refs;
1262 return 0;
1263 }
1264
1265 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1266 *nr -= refs;
1267 while (refs--)
1268 put_page(head);
1269 return 0;
1270 }
1271
1272 return 1;
1273}
1274
1275static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1276 unsigned long end, int write, struct page **pages, int *nr)
1277{
1278 struct page *head, *page;
1279 int refs;
1280
1281 if (write && !pud_write(orig))
1282 return 0;
1283
1284 refs = 0;
1285 head = pud_page(orig);
1286 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1287 do {
1288 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1289 pages[*nr] = page;
1290 (*nr)++;
1291 page++;
1292 refs++;
1293 } while (addr += PAGE_SIZE, addr != end);
1294
1295 if (!page_cache_add_speculative(head, refs)) {
1296 *nr -= refs;
1297 return 0;
1298 }
1299
1300 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1301 *nr -= refs;
1302 while (refs--)
1303 put_page(head);
1304 return 0;
1305 }
1306
1307 return 1;
1308}
1309
1310static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1311 unsigned long end, int write,
1312 struct page **pages, int *nr)
1313{
1314 int refs;
1315 struct page *head, *page;
1316
1317 if (write && !pgd_write(orig))
1318 return 0;
1319
1320 refs = 0;
1321 head = pgd_page(orig);
1322 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1323 do {
1324 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1325 pages[*nr] = page;
1326 (*nr)++;
1327 page++;
1328 refs++;
1329 } while (addr += PAGE_SIZE, addr != end);
1330
1331 if (!page_cache_add_speculative(head, refs)) {
1332 *nr -= refs;
1333 return 0;
1334 }
1335
1336 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1337 *nr -= refs;
1338 while (refs--)
1339 put_page(head);
1340 return 0;
1341 }
1342
1343 return 1;
1344}
1345
1346static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1347 int write, struct page **pages, int *nr)
1348{
1349 unsigned long next;
1350 pmd_t *pmdp;
1351
1352 pmdp = pmd_offset(&pud, addr);
1353 do {
1354 pmd_t pmd = READ_ONCE(*pmdp);
1355
1356 next = pmd_addr_end(addr, end);
1357 if (pmd_none(pmd))
1358 return 0;
1359
1360 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1361 /*
1362 * NUMA hinting faults need to be handled in the GUP
1363 * slowpath for accounting purposes and so that they
1364 * can be serialised against THP migration.
1365 */
1366 if (pmd_protnone(pmd))
1367 return 0;
1368
1369 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1370 pages, nr))
1371 return 0;
1372
1373 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1374 /*
1375 * architecture have different format for hugetlbfs
1376 * pmd format and THP pmd format
1377 */
1378 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1379 PMD_SHIFT, next, write, pages, nr))
1380 return 0;
1381 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1382 return 0;
1383 } while (pmdp++, addr = next, addr != end);
1384
1385 return 1;
1386}
1387
1388static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1389 int write, struct page **pages, int *nr)
1390{
1391 unsigned long next;
1392 pud_t *pudp;
1393
1394 pudp = pud_offset(&pgd, addr);
1395 do {
1396 pud_t pud = READ_ONCE(*pudp);
1397
1398 next = pud_addr_end(addr, end);
1399 if (pud_none(pud))
1400 return 0;
1401 if (unlikely(pud_huge(pud))) {
1402 if (!gup_huge_pud(pud, pudp, addr, next, write,
1403 pages, nr))
1404 return 0;
1405 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1406 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1407 PUD_SHIFT, next, write, pages, nr))
1408 return 0;
1409 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1410 return 0;
1411 } while (pudp++, addr = next, addr != end);
1412
1413 return 1;
1414}
1415
1416/*
1417 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1418 * the regular GUP. It will only return non-negative values.
1419 */
1420int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1421 struct page **pages)
1422{
1423 struct mm_struct *mm = current->mm;
1424 unsigned long addr, len, end;
1425 unsigned long next, flags;
1426 pgd_t *pgdp;
1427 int nr = 0;
1428
1429 start &= PAGE_MASK;
1430 addr = start;
1431 len = (unsigned long) nr_pages << PAGE_SHIFT;
1432 end = start + len;
1433
1434 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1435 start, len)))
1436 return 0;
1437
1438 /*
1439 * Disable interrupts. We use the nested form as we can already have
1440 * interrupts disabled by get_futex_key.
1441 *
1442 * With interrupts disabled, we block page table pages from being
1443 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1444 * for more details.
1445 *
1446 * We do not adopt an rcu_read_lock(.) here as we also want to
1447 * block IPIs that come from THPs splitting.
1448 */
1449
1450 local_irq_save(flags);
1451 pgdp = pgd_offset(mm, addr);
1452 do {
1453 pgd_t pgd = READ_ONCE(*pgdp);
1454
1455 next = pgd_addr_end(addr, end);
1456 if (pgd_none(pgd))
1457 break;
1458 if (unlikely(pgd_huge(pgd))) {
1459 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1460 pages, &nr))
1461 break;
1462 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1463 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1464 PGDIR_SHIFT, next, write, pages, &nr))
1465 break;
1466 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1467 break;
1468 } while (pgdp++, addr = next, addr != end);
1469 local_irq_restore(flags);
1470
1471 return nr;
1472}
1473
1474/**
1475 * get_user_pages_fast() - pin user pages in memory
1476 * @start: starting user address
1477 * @nr_pages: number of pages from start to pin
1478 * @write: whether pages will be written to
1479 * @pages: array that receives pointers to the pages pinned.
1480 * Should be at least nr_pages long.
1481 *
1482 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1483 * If not successful, it will fall back to taking the lock and
1484 * calling get_user_pages().
1485 *
1486 * Returns number of pages pinned. This may be fewer than the number
1487 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1488 * were pinned, returns -errno.
1489 */
1490int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1491 struct page **pages)
1492{
1493 int nr, ret;
1494
1495 start &= PAGE_MASK;
1496 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1497 ret = nr;
1498
1499 if (nr < nr_pages) {
1500 /* Try to get the remaining pages with get_user_pages */
1501 start += nr << PAGE_SHIFT;
1502 pages += nr;
1503
1504 ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
1505
1506 /* Have to be a bit careful with return values */
1507 if (nr > 0) {
1508 if (ret < 0)
1509 ret = nr;
1510 else
1511 ret += nr;
1512 }
1513 }
1514
1515 return ret;
1516}
1517
1518#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */
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
27struct follow_page_context {
28 struct dev_pagemap *pgmap;
29 unsigned int page_mask;
30};
31
32static void hpage_pincount_add(struct page *page, int refs)
33{
34 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
35 VM_BUG_ON_PAGE(page != compound_head(page), page);
36
37 atomic_add(refs, compound_pincount_ptr(page));
38}
39
40static void hpage_pincount_sub(struct page *page, int refs)
41{
42 VM_BUG_ON_PAGE(!hpage_pincount_available(page), page);
43 VM_BUG_ON_PAGE(page != compound_head(page), page);
44
45 atomic_sub(refs, compound_pincount_ptr(page));
46}
47
48/* Equivalent to calling put_page() @refs times. */
49static void put_page_refs(struct page *page, int refs)
50{
51#ifdef CONFIG_DEBUG_VM
52 if (VM_WARN_ON_ONCE_PAGE(page_ref_count(page) < refs, page))
53 return;
54#endif
55
56 /*
57 * Calling put_page() for each ref is unnecessarily slow. Only the last
58 * ref needs a put_page().
59 */
60 if (refs > 1)
61 page_ref_sub(page, refs - 1);
62 put_page(page);
63}
64
65/*
66 * Return the compound head page with ref appropriately incremented,
67 * or NULL if that failed.
68 */
69static inline struct page *try_get_compound_head(struct page *page, int refs)
70{
71 struct page *head = compound_head(page);
72
73 if (WARN_ON_ONCE(page_ref_count(head) < 0))
74 return NULL;
75 if (unlikely(!page_cache_add_speculative(head, refs)))
76 return NULL;
77
78 /*
79 * At this point we have a stable reference to the head page; but it
80 * could be that between the compound_head() lookup and the refcount
81 * increment, the compound page was split, in which case we'd end up
82 * holding a reference on a page that has nothing to do with the page
83 * we were given anymore.
84 * So now that the head page is stable, recheck that the pages still
85 * belong together.
86 */
87 if (unlikely(compound_head(page) != head)) {
88 put_page_refs(head, refs);
89 return NULL;
90 }
91
92 return head;
93}
94
95/*
96 * try_grab_compound_head() - attempt to elevate a page's refcount, by a
97 * flags-dependent amount.
98 *
99 * "grab" names in this file mean, "look at flags to decide whether to use
100 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
101 *
102 * Either FOLL_PIN or FOLL_GET (or neither) must be set, but not both at the
103 * same time. (That's true throughout the get_user_pages*() and
104 * pin_user_pages*() APIs.) Cases:
105 *
106 * FOLL_GET: page's refcount will be incremented by 1.
107 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
108 *
109 * Return: head page (with refcount appropriately incremented) for success, or
110 * NULL upon failure. If neither FOLL_GET nor FOLL_PIN was set, that's
111 * considered failure, and furthermore, a likely bug in the caller, so a warning
112 * is also emitted.
113 */
114__maybe_unused struct page *try_grab_compound_head(struct page *page,
115 int refs, unsigned int flags)
116{
117 if (flags & FOLL_GET)
118 return try_get_compound_head(page, refs);
119 else if (flags & FOLL_PIN) {
120 int orig_refs = refs;
121
122 /*
123 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
124 * right zone, so fail and let the caller fall back to the slow
125 * path.
126 */
127 if (unlikely((flags & FOLL_LONGTERM) &&
128 !is_pinnable_page(page)))
129 return NULL;
130
131 /*
132 * CAUTION: Don't use compound_head() on the page before this
133 * point, the result won't be stable.
134 */
135 page = try_get_compound_head(page, refs);
136 if (!page)
137 return NULL;
138
139 /*
140 * When pinning a compound page of order > 1 (which is what
141 * hpage_pincount_available() checks for), use an exact count to
142 * track it, via hpage_pincount_add/_sub().
143 *
144 * However, be sure to *also* increment the normal page refcount
145 * field at least once, so that the page really is pinned.
146 */
147 if (hpage_pincount_available(page))
148 hpage_pincount_add(page, refs);
149 else
150 page_ref_add(page, refs * (GUP_PIN_COUNTING_BIAS - 1));
151
152 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED,
153 orig_refs);
154
155 return page;
156 }
157
158 WARN_ON_ONCE(1);
159 return NULL;
160}
161
162static void put_compound_head(struct page *page, int refs, unsigned int flags)
163{
164 if (flags & FOLL_PIN) {
165 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_RELEASED,
166 refs);
167
168 if (hpage_pincount_available(page))
169 hpage_pincount_sub(page, refs);
170 else
171 refs *= GUP_PIN_COUNTING_BIAS;
172 }
173
174 put_page_refs(page, refs);
175}
176
177/**
178 * try_grab_page() - elevate a page's refcount by a flag-dependent amount
179 *
180 * This might not do anything at all, depending on the flags argument.
181 *
182 * "grab" names in this file mean, "look at flags to decide whether to use
183 * FOLL_PIN or FOLL_GET behavior, when incrementing the page's refcount.
184 *
185 * @page: pointer to page to be grabbed
186 * @flags: gup flags: these are the FOLL_* flag values.
187 *
188 * Either FOLL_PIN or FOLL_GET (or neither) may be set, but not both at the same
189 * time. Cases:
190 *
191 * FOLL_GET: page's refcount will be incremented by 1.
192 * FOLL_PIN: page's refcount will be incremented by GUP_PIN_COUNTING_BIAS.
193 *
194 * Return: true for success, or if no action was required (if neither FOLL_PIN
195 * nor FOLL_GET was set, nothing is done). False for failure: FOLL_GET or
196 * FOLL_PIN was set, but the page could not be grabbed.
197 */
198bool __must_check try_grab_page(struct page *page, unsigned int flags)
199{
200 WARN_ON_ONCE((flags & (FOLL_GET | FOLL_PIN)) == (FOLL_GET | FOLL_PIN));
201
202 if (flags & FOLL_GET)
203 return try_get_page(page);
204 else if (flags & FOLL_PIN) {
205 int refs = 1;
206
207 page = compound_head(page);
208
209 if (WARN_ON_ONCE(page_ref_count(page) <= 0))
210 return false;
211
212 if (hpage_pincount_available(page))
213 hpage_pincount_add(page, 1);
214 else
215 refs = GUP_PIN_COUNTING_BIAS;
216
217 /*
218 * Similar to try_grab_compound_head(): even if using the
219 * hpage_pincount_add/_sub() routines, be sure to
220 * *also* increment the normal page refcount field at least
221 * once, so that the page really is pinned.
222 */
223 page_ref_add(page, refs);
224
225 mod_node_page_state(page_pgdat(page), NR_FOLL_PIN_ACQUIRED, 1);
226 }
227
228 return true;
229}
230
231/**
232 * unpin_user_page() - release a dma-pinned page
233 * @page: pointer to page to be released
234 *
235 * Pages that were pinned via pin_user_pages*() must be released via either
236 * unpin_user_page(), or one of the unpin_user_pages*() routines. This is so
237 * that such pages can be separately tracked and uniquely handled. In
238 * particular, interactions with RDMA and filesystems need special handling.
239 */
240void unpin_user_page(struct page *page)
241{
242 put_compound_head(compound_head(page), 1, FOLL_PIN);
243}
244EXPORT_SYMBOL(unpin_user_page);
245
246static inline void compound_range_next(unsigned long i, unsigned long npages,
247 struct page **list, struct page **head,
248 unsigned int *ntails)
249{
250 struct page *next, *page;
251 unsigned int nr = 1;
252
253 if (i >= npages)
254 return;
255
256 next = *list + i;
257 page = compound_head(next);
258 if (PageCompound(page) && compound_order(page) >= 1)
259 nr = min_t(unsigned int,
260 page + compound_nr(page) - next, npages - i);
261
262 *head = page;
263 *ntails = nr;
264}
265
266#define for_each_compound_range(__i, __list, __npages, __head, __ntails) \
267 for (__i = 0, \
268 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)); \
269 __i < __npages; __i += __ntails, \
270 compound_range_next(__i, __npages, __list, &(__head), &(__ntails)))
271
272static inline void compound_next(unsigned long i, unsigned long npages,
273 struct page **list, struct page **head,
274 unsigned int *ntails)
275{
276 struct page *page;
277 unsigned int nr;
278
279 if (i >= npages)
280 return;
281
282 page = compound_head(list[i]);
283 for (nr = i + 1; nr < npages; nr++) {
284 if (compound_head(list[nr]) != page)
285 break;
286 }
287
288 *head = page;
289 *ntails = nr - i;
290}
291
292#define for_each_compound_head(__i, __list, __npages, __head, __ntails) \
293 for (__i = 0, \
294 compound_next(__i, __npages, __list, &(__head), &(__ntails)); \
295 __i < __npages; __i += __ntails, \
296 compound_next(__i, __npages, __list, &(__head), &(__ntails)))
297
298/**
299 * unpin_user_pages_dirty_lock() - release and optionally dirty gup-pinned pages
300 * @pages: array of pages to be maybe marked dirty, and definitely released.
301 * @npages: number of pages in the @pages array.
302 * @make_dirty: whether to mark the pages dirty
303 *
304 * "gup-pinned page" refers to a page that has had one of the get_user_pages()
305 * variants called on that page.
306 *
307 * For each page in the @pages array, make that page (or its head page, if a
308 * compound page) dirty, if @make_dirty is true, and if the page was previously
309 * listed as clean. In any case, releases all pages using unpin_user_page(),
310 * possibly via unpin_user_pages(), for the non-dirty case.
311 *
312 * Please see the unpin_user_page() documentation for details.
313 *
314 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
315 * required, then the caller should a) verify that this is really correct,
316 * because _lock() is usually required, and b) hand code it:
317 * set_page_dirty_lock(), unpin_user_page().
318 *
319 */
320void unpin_user_pages_dirty_lock(struct page **pages, unsigned long npages,
321 bool make_dirty)
322{
323 unsigned long index;
324 struct page *head;
325 unsigned int ntails;
326
327 if (!make_dirty) {
328 unpin_user_pages(pages, npages);
329 return;
330 }
331
332 for_each_compound_head(index, pages, npages, head, ntails) {
333 /*
334 * Checking PageDirty at this point may race with
335 * clear_page_dirty_for_io(), but that's OK. Two key
336 * cases:
337 *
338 * 1) This code sees the page as already dirty, so it
339 * skips the call to set_page_dirty(). That could happen
340 * because clear_page_dirty_for_io() called
341 * page_mkclean(), followed by set_page_dirty().
342 * However, now the page is going to get written back,
343 * which meets the original intention of setting it
344 * dirty, so all is well: clear_page_dirty_for_io() goes
345 * on to call TestClearPageDirty(), and write the page
346 * back.
347 *
348 * 2) This code sees the page as clean, so it calls
349 * set_page_dirty(). The page stays dirty, despite being
350 * written back, so it gets written back again in the
351 * next writeback cycle. This is harmless.
352 */
353 if (!PageDirty(head))
354 set_page_dirty_lock(head);
355 put_compound_head(head, ntails, FOLL_PIN);
356 }
357}
358EXPORT_SYMBOL(unpin_user_pages_dirty_lock);
359
360/**
361 * unpin_user_page_range_dirty_lock() - release and optionally dirty
362 * gup-pinned page range
363 *
364 * @page: the starting page of a range maybe marked dirty, and definitely released.
365 * @npages: number of consecutive pages to release.
366 * @make_dirty: whether to mark the pages dirty
367 *
368 * "gup-pinned page range" refers to a range of pages that has had one of the
369 * pin_user_pages() variants called on that page.
370 *
371 * For the page ranges defined by [page .. page+npages], make that range (or
372 * its head pages, if a compound page) dirty, if @make_dirty is true, and if the
373 * page range was previously listed as clean.
374 *
375 * set_page_dirty_lock() is used internally. If instead, set_page_dirty() is
376 * required, then the caller should a) verify that this is really correct,
377 * because _lock() is usually required, and b) hand code it:
378 * set_page_dirty_lock(), unpin_user_page().
379 *
380 */
381void unpin_user_page_range_dirty_lock(struct page *page, unsigned long npages,
382 bool make_dirty)
383{
384 unsigned long index;
385 struct page *head;
386 unsigned int ntails;
387
388 for_each_compound_range(index, &page, npages, head, ntails) {
389 if (make_dirty && !PageDirty(head))
390 set_page_dirty_lock(head);
391 put_compound_head(head, ntails, FOLL_PIN);
392 }
393}
394EXPORT_SYMBOL(unpin_user_page_range_dirty_lock);
395
396/**
397 * unpin_user_pages() - release an array of gup-pinned pages.
398 * @pages: array of pages to be marked dirty and released.
399 * @npages: number of pages in the @pages array.
400 *
401 * For each page in the @pages array, release the page using unpin_user_page().
402 *
403 * Please see the unpin_user_page() documentation for details.
404 */
405void unpin_user_pages(struct page **pages, unsigned long npages)
406{
407 unsigned long index;
408 struct page *head;
409 unsigned int ntails;
410
411 /*
412 * If this WARN_ON() fires, then the system *might* be leaking pages (by
413 * leaving them pinned), but probably not. More likely, gup/pup returned
414 * a hard -ERRNO error to the caller, who erroneously passed it here.
415 */
416 if (WARN_ON(IS_ERR_VALUE(npages)))
417 return;
418
419 for_each_compound_head(index, pages, npages, head, ntails)
420 put_compound_head(head, ntails, FOLL_PIN);
421}
422EXPORT_SYMBOL(unpin_user_pages);
423
424/*
425 * Set the MMF_HAS_PINNED if not set yet; after set it'll be there for the mm's
426 * lifecycle. Avoid setting the bit unless necessary, or it might cause write
427 * cache bouncing on large SMP machines for concurrent pinned gups.
428 */
429static inline void mm_set_has_pinned_flag(unsigned long *mm_flags)
430{
431 if (!test_bit(MMF_HAS_PINNED, mm_flags))
432 set_bit(MMF_HAS_PINNED, mm_flags);
433}
434
435#ifdef CONFIG_MMU
436static struct page *no_page_table(struct vm_area_struct *vma,
437 unsigned int flags)
438{
439 /*
440 * When core dumping an enormous anonymous area that nobody
441 * has touched so far, we don't want to allocate unnecessary pages or
442 * page tables. Return error instead of NULL to skip handle_mm_fault,
443 * then get_dump_page() will return NULL to leave a hole in the dump.
444 * But we can only make this optimization where a hole would surely
445 * be zero-filled if handle_mm_fault() actually did handle it.
446 */
447 if ((flags & FOLL_DUMP) &&
448 (vma_is_anonymous(vma) || !vma->vm_ops->fault))
449 return ERR_PTR(-EFAULT);
450 return NULL;
451}
452
453static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
454 pte_t *pte, unsigned int flags)
455{
456 /* No page to get reference */
457 if (flags & FOLL_GET)
458 return -EFAULT;
459
460 if (flags & FOLL_TOUCH) {
461 pte_t entry = *pte;
462
463 if (flags & FOLL_WRITE)
464 entry = pte_mkdirty(entry);
465 entry = pte_mkyoung(entry);
466
467 if (!pte_same(*pte, entry)) {
468 set_pte_at(vma->vm_mm, address, pte, entry);
469 update_mmu_cache(vma, address, pte);
470 }
471 }
472
473 /* Proper page table entry exists, but no corresponding struct page */
474 return -EEXIST;
475}
476
477/*
478 * FOLL_FORCE can write to even unwritable pte's, but only
479 * after we've gone through a COW cycle and they are dirty.
480 */
481static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
482{
483 return pte_write(pte) ||
484 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
485}
486
487static struct page *follow_page_pte(struct vm_area_struct *vma,
488 unsigned long address, pmd_t *pmd, unsigned int flags,
489 struct dev_pagemap **pgmap)
490{
491 struct mm_struct *mm = vma->vm_mm;
492 struct page *page;
493 spinlock_t *ptl;
494 pte_t *ptep, pte;
495 int ret;
496
497 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
498 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
499 (FOLL_PIN | FOLL_GET)))
500 return ERR_PTR(-EINVAL);
501retry:
502 if (unlikely(pmd_bad(*pmd)))
503 return no_page_table(vma, flags);
504
505 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
506 pte = *ptep;
507 if (!pte_present(pte)) {
508 swp_entry_t entry;
509 /*
510 * KSM's break_ksm() relies upon recognizing a ksm page
511 * even while it is being migrated, so for that case we
512 * need migration_entry_wait().
513 */
514 if (likely(!(flags & FOLL_MIGRATION)))
515 goto no_page;
516 if (pte_none(pte))
517 goto no_page;
518 entry = pte_to_swp_entry(pte);
519 if (!is_migration_entry(entry))
520 goto no_page;
521 pte_unmap_unlock(ptep, ptl);
522 migration_entry_wait(mm, pmd, address);
523 goto retry;
524 }
525 if ((flags & FOLL_NUMA) && pte_protnone(pte))
526 goto no_page;
527 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
528 pte_unmap_unlock(ptep, ptl);
529 return NULL;
530 }
531
532 page = vm_normal_page(vma, address, pte);
533 if (!page && pte_devmap(pte) && (flags & (FOLL_GET | FOLL_PIN))) {
534 /*
535 * Only return device mapping pages in the FOLL_GET or FOLL_PIN
536 * case since they are only valid while holding the pgmap
537 * reference.
538 */
539 *pgmap = get_dev_pagemap(pte_pfn(pte), *pgmap);
540 if (*pgmap)
541 page = pte_page(pte);
542 else
543 goto no_page;
544 } else if (unlikely(!page)) {
545 if (flags & FOLL_DUMP) {
546 /* Avoid special (like zero) pages in core dumps */
547 page = ERR_PTR(-EFAULT);
548 goto out;
549 }
550
551 if (is_zero_pfn(pte_pfn(pte))) {
552 page = pte_page(pte);
553 } else {
554 ret = follow_pfn_pte(vma, address, ptep, flags);
555 page = ERR_PTR(ret);
556 goto out;
557 }
558 }
559
560 /* try_grab_page() does nothing unless FOLL_GET or FOLL_PIN is set. */
561 if (unlikely(!try_grab_page(page, flags))) {
562 page = ERR_PTR(-ENOMEM);
563 goto out;
564 }
565 /*
566 * We need to make the page accessible if and only if we are going
567 * to access its content (the FOLL_PIN case). Please see
568 * Documentation/core-api/pin_user_pages.rst for details.
569 */
570 if (flags & FOLL_PIN) {
571 ret = arch_make_page_accessible(page);
572 if (ret) {
573 unpin_user_page(page);
574 page = ERR_PTR(ret);
575 goto out;
576 }
577 }
578 if (flags & FOLL_TOUCH) {
579 if ((flags & FOLL_WRITE) &&
580 !pte_dirty(pte) && !PageDirty(page))
581 set_page_dirty(page);
582 /*
583 * pte_mkyoung() would be more correct here, but atomic care
584 * is needed to avoid losing the dirty bit: it is easier to use
585 * mark_page_accessed().
586 */
587 mark_page_accessed(page);
588 }
589 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
590 /* Do not mlock pte-mapped THP */
591 if (PageTransCompound(page))
592 goto out;
593
594 /*
595 * The preliminary mapping check is mainly to avoid the
596 * pointless overhead of lock_page on the ZERO_PAGE
597 * which might bounce very badly if there is contention.
598 *
599 * If the page is already locked, we don't need to
600 * handle it now - vmscan will handle it later if and
601 * when it attempts to reclaim the page.
602 */
603 if (page->mapping && trylock_page(page)) {
604 lru_add_drain(); /* push cached pages to LRU */
605 /*
606 * Because we lock page here, and migration is
607 * blocked by the pte's page reference, and we
608 * know the page is still mapped, we don't even
609 * need to check for file-cache page truncation.
610 */
611 mlock_vma_page(page);
612 unlock_page(page);
613 }
614 }
615out:
616 pte_unmap_unlock(ptep, ptl);
617 return page;
618no_page:
619 pte_unmap_unlock(ptep, ptl);
620 if (!pte_none(pte))
621 return NULL;
622 return no_page_table(vma, flags);
623}
624
625static struct page *follow_pmd_mask(struct vm_area_struct *vma,
626 unsigned long address, pud_t *pudp,
627 unsigned int flags,
628 struct follow_page_context *ctx)
629{
630 pmd_t *pmd, pmdval;
631 spinlock_t *ptl;
632 struct page *page;
633 struct mm_struct *mm = vma->vm_mm;
634
635 pmd = pmd_offset(pudp, address);
636 /*
637 * The READ_ONCE() will stabilize the pmdval in a register or
638 * on the stack so that it will stop changing under the code.
639 */
640 pmdval = READ_ONCE(*pmd);
641 if (pmd_none(pmdval))
642 return no_page_table(vma, flags);
643 if (pmd_huge(pmdval) && is_vm_hugetlb_page(vma)) {
644 page = follow_huge_pmd(mm, address, pmd, flags);
645 if (page)
646 return page;
647 return no_page_table(vma, flags);
648 }
649 if (is_hugepd(__hugepd(pmd_val(pmdval)))) {
650 page = follow_huge_pd(vma, address,
651 __hugepd(pmd_val(pmdval)), flags,
652 PMD_SHIFT);
653 if (page)
654 return page;
655 return no_page_table(vma, flags);
656 }
657retry:
658 if (!pmd_present(pmdval)) {
659 if (likely(!(flags & FOLL_MIGRATION)))
660 return no_page_table(vma, flags);
661 VM_BUG_ON(thp_migration_supported() &&
662 !is_pmd_migration_entry(pmdval));
663 if (is_pmd_migration_entry(pmdval))
664 pmd_migration_entry_wait(mm, pmd);
665 pmdval = READ_ONCE(*pmd);
666 /*
667 * MADV_DONTNEED may convert the pmd to null because
668 * mmap_lock is held in read mode
669 */
670 if (pmd_none(pmdval))
671 return no_page_table(vma, flags);
672 goto retry;
673 }
674 if (pmd_devmap(pmdval)) {
675 ptl = pmd_lock(mm, pmd);
676 page = follow_devmap_pmd(vma, address, pmd, flags, &ctx->pgmap);
677 spin_unlock(ptl);
678 if (page)
679 return page;
680 }
681 if (likely(!pmd_trans_huge(pmdval)))
682 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
683
684 if ((flags & FOLL_NUMA) && pmd_protnone(pmdval))
685 return no_page_table(vma, flags);
686
687retry_locked:
688 ptl = pmd_lock(mm, pmd);
689 if (unlikely(pmd_none(*pmd))) {
690 spin_unlock(ptl);
691 return no_page_table(vma, flags);
692 }
693 if (unlikely(!pmd_present(*pmd))) {
694 spin_unlock(ptl);
695 if (likely(!(flags & FOLL_MIGRATION)))
696 return no_page_table(vma, flags);
697 pmd_migration_entry_wait(mm, pmd);
698 goto retry_locked;
699 }
700 if (unlikely(!pmd_trans_huge(*pmd))) {
701 spin_unlock(ptl);
702 return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
703 }
704 if (flags & FOLL_SPLIT_PMD) {
705 int ret;
706 page = pmd_page(*pmd);
707 if (is_huge_zero_page(page)) {
708 spin_unlock(ptl);
709 ret = 0;
710 split_huge_pmd(vma, pmd, address);
711 if (pmd_trans_unstable(pmd))
712 ret = -EBUSY;
713 } else {
714 spin_unlock(ptl);
715 split_huge_pmd(vma, pmd, address);
716 ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
717 }
718
719 return ret ? ERR_PTR(ret) :
720 follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
721 }
722 page = follow_trans_huge_pmd(vma, address, pmd, flags);
723 spin_unlock(ptl);
724 ctx->page_mask = HPAGE_PMD_NR - 1;
725 return page;
726}
727
728static struct page *follow_pud_mask(struct vm_area_struct *vma,
729 unsigned long address, p4d_t *p4dp,
730 unsigned int flags,
731 struct follow_page_context *ctx)
732{
733 pud_t *pud;
734 spinlock_t *ptl;
735 struct page *page;
736 struct mm_struct *mm = vma->vm_mm;
737
738 pud = pud_offset(p4dp, address);
739 if (pud_none(*pud))
740 return no_page_table(vma, flags);
741 if (pud_huge(*pud) && is_vm_hugetlb_page(vma)) {
742 page = follow_huge_pud(mm, address, pud, flags);
743 if (page)
744 return page;
745 return no_page_table(vma, flags);
746 }
747 if (is_hugepd(__hugepd(pud_val(*pud)))) {
748 page = follow_huge_pd(vma, address,
749 __hugepd(pud_val(*pud)), flags,
750 PUD_SHIFT);
751 if (page)
752 return page;
753 return no_page_table(vma, flags);
754 }
755 if (pud_devmap(*pud)) {
756 ptl = pud_lock(mm, pud);
757 page = follow_devmap_pud(vma, address, pud, flags, &ctx->pgmap);
758 spin_unlock(ptl);
759 if (page)
760 return page;
761 }
762 if (unlikely(pud_bad(*pud)))
763 return no_page_table(vma, flags);
764
765 return follow_pmd_mask(vma, address, pud, flags, ctx);
766}
767
768static struct page *follow_p4d_mask(struct vm_area_struct *vma,
769 unsigned long address, pgd_t *pgdp,
770 unsigned int flags,
771 struct follow_page_context *ctx)
772{
773 p4d_t *p4d;
774 struct page *page;
775
776 p4d = p4d_offset(pgdp, address);
777 if (p4d_none(*p4d))
778 return no_page_table(vma, flags);
779 BUILD_BUG_ON(p4d_huge(*p4d));
780 if (unlikely(p4d_bad(*p4d)))
781 return no_page_table(vma, flags);
782
783 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
784 page = follow_huge_pd(vma, address,
785 __hugepd(p4d_val(*p4d)), flags,
786 P4D_SHIFT);
787 if (page)
788 return page;
789 return no_page_table(vma, flags);
790 }
791 return follow_pud_mask(vma, address, p4d, flags, ctx);
792}
793
794/**
795 * follow_page_mask - look up a page descriptor from a user-virtual address
796 * @vma: vm_area_struct mapping @address
797 * @address: virtual address to look up
798 * @flags: flags modifying lookup behaviour
799 * @ctx: contains dev_pagemap for %ZONE_DEVICE memory pinning and a
800 * pointer to output page_mask
801 *
802 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
803 *
804 * When getting pages from ZONE_DEVICE memory, the @ctx->pgmap caches
805 * the device's dev_pagemap metadata to avoid repeating expensive lookups.
806 *
807 * On output, the @ctx->page_mask is set according to the size of the page.
808 *
809 * Return: the mapped (struct page *), %NULL if no mapping exists, or
810 * an error pointer if there is a mapping to something not represented
811 * by a page descriptor (see also vm_normal_page()).
812 */
813static struct page *follow_page_mask(struct vm_area_struct *vma,
814 unsigned long address, unsigned int flags,
815 struct follow_page_context *ctx)
816{
817 pgd_t *pgd;
818 struct page *page;
819 struct mm_struct *mm = vma->vm_mm;
820
821 ctx->page_mask = 0;
822
823 /* make this handle hugepd */
824 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
825 if (!IS_ERR(page)) {
826 WARN_ON_ONCE(flags & (FOLL_GET | FOLL_PIN));
827 return page;
828 }
829
830 pgd = pgd_offset(mm, address);
831
832 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
833 return no_page_table(vma, flags);
834
835 if (pgd_huge(*pgd)) {
836 page = follow_huge_pgd(mm, address, pgd, flags);
837 if (page)
838 return page;
839 return no_page_table(vma, flags);
840 }
841 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
842 page = follow_huge_pd(vma, address,
843 __hugepd(pgd_val(*pgd)), flags,
844 PGDIR_SHIFT);
845 if (page)
846 return page;
847 return no_page_table(vma, flags);
848 }
849
850 return follow_p4d_mask(vma, address, pgd, flags, ctx);
851}
852
853struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
854 unsigned int foll_flags)
855{
856 struct follow_page_context ctx = { NULL };
857 struct page *page;
858
859 if (vma_is_secretmem(vma))
860 return NULL;
861
862 page = follow_page_mask(vma, address, foll_flags, &ctx);
863 if (ctx.pgmap)
864 put_dev_pagemap(ctx.pgmap);
865 return page;
866}
867
868static int get_gate_page(struct mm_struct *mm, unsigned long address,
869 unsigned int gup_flags, struct vm_area_struct **vma,
870 struct page **page)
871{
872 pgd_t *pgd;
873 p4d_t *p4d;
874 pud_t *pud;
875 pmd_t *pmd;
876 pte_t *pte;
877 int ret = -EFAULT;
878
879 /* user gate pages are read-only */
880 if (gup_flags & FOLL_WRITE)
881 return -EFAULT;
882 if (address > TASK_SIZE)
883 pgd = pgd_offset_k(address);
884 else
885 pgd = pgd_offset_gate(mm, address);
886 if (pgd_none(*pgd))
887 return -EFAULT;
888 p4d = p4d_offset(pgd, address);
889 if (p4d_none(*p4d))
890 return -EFAULT;
891 pud = pud_offset(p4d, address);
892 if (pud_none(*pud))
893 return -EFAULT;
894 pmd = pmd_offset(pud, address);
895 if (!pmd_present(*pmd))
896 return -EFAULT;
897 VM_BUG_ON(pmd_trans_huge(*pmd));
898 pte = pte_offset_map(pmd, address);
899 if (pte_none(*pte))
900 goto unmap;
901 *vma = get_gate_vma(mm);
902 if (!page)
903 goto out;
904 *page = vm_normal_page(*vma, address, *pte);
905 if (!*page) {
906 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
907 goto unmap;
908 *page = pte_page(*pte);
909 }
910 if (unlikely(!try_grab_page(*page, gup_flags))) {
911 ret = -ENOMEM;
912 goto unmap;
913 }
914out:
915 ret = 0;
916unmap:
917 pte_unmap(pte);
918 return ret;
919}
920
921/*
922 * mmap_lock must be held on entry. If @locked != NULL and *@flags
923 * does not include FOLL_NOWAIT, the mmap_lock may be released. If it
924 * is, *@locked will be set to 0 and -EBUSY returned.
925 */
926static int faultin_page(struct vm_area_struct *vma,
927 unsigned long address, unsigned int *flags, int *locked)
928{
929 unsigned int fault_flags = 0;
930 vm_fault_t ret;
931
932 /* mlock all present pages, but do not fault in new pages */
933 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
934 return -ENOENT;
935 if (*flags & FOLL_WRITE)
936 fault_flags |= FAULT_FLAG_WRITE;
937 if (*flags & FOLL_REMOTE)
938 fault_flags |= FAULT_FLAG_REMOTE;
939 if (locked)
940 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
941 if (*flags & FOLL_NOWAIT)
942 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
943 if (*flags & FOLL_TRIED) {
944 /*
945 * Note: FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_TRIED
946 * can co-exist
947 */
948 fault_flags |= FAULT_FLAG_TRIED;
949 }
950
951 ret = handle_mm_fault(vma, address, fault_flags, NULL);
952 if (ret & VM_FAULT_ERROR) {
953 int err = vm_fault_to_errno(ret, *flags);
954
955 if (err)
956 return err;
957 BUG();
958 }
959
960 if (ret & VM_FAULT_RETRY) {
961 if (locked && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
962 *locked = 0;
963 return -EBUSY;
964 }
965
966 /*
967 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
968 * necessary, even if maybe_mkwrite decided not to set pte_write. We
969 * can thus safely do subsequent page lookups as if they were reads.
970 * But only do so when looping for pte_write is futile: in some cases
971 * userspace may also be wanting to write to the gotten user page,
972 * which a read fault here might prevent (a readonly page might get
973 * reCOWed by userspace write).
974 */
975 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
976 *flags |= FOLL_COW;
977 return 0;
978}
979
980static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
981{
982 vm_flags_t vm_flags = vma->vm_flags;
983 int write = (gup_flags & FOLL_WRITE);
984 int foreign = (gup_flags & FOLL_REMOTE);
985
986 if (vm_flags & (VM_IO | VM_PFNMAP))
987 return -EFAULT;
988
989 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
990 return -EFAULT;
991
992 if ((gup_flags & FOLL_LONGTERM) && vma_is_fsdax(vma))
993 return -EOPNOTSUPP;
994
995 if (vma_is_secretmem(vma))
996 return -EFAULT;
997
998 if (write) {
999 if (!(vm_flags & VM_WRITE)) {
1000 if (!(gup_flags & FOLL_FORCE))
1001 return -EFAULT;
1002 /*
1003 * We used to let the write,force case do COW in a
1004 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
1005 * set a breakpoint in a read-only mapping of an
1006 * executable, without corrupting the file (yet only
1007 * when that file had been opened for writing!).
1008 * Anon pages in shared mappings are surprising: now
1009 * just reject it.
1010 */
1011 if (!is_cow_mapping(vm_flags))
1012 return -EFAULT;
1013 }
1014 } else if (!(vm_flags & VM_READ)) {
1015 if (!(gup_flags & FOLL_FORCE))
1016 return -EFAULT;
1017 /*
1018 * Is there actually any vma we can reach here which does not
1019 * have VM_MAYREAD set?
1020 */
1021 if (!(vm_flags & VM_MAYREAD))
1022 return -EFAULT;
1023 }
1024 /*
1025 * gups are always data accesses, not instruction
1026 * fetches, so execute=false here
1027 */
1028 if (!arch_vma_access_permitted(vma, write, false, foreign))
1029 return -EFAULT;
1030 return 0;
1031}
1032
1033/**
1034 * __get_user_pages() - pin user pages in memory
1035 * @mm: mm_struct of target mm
1036 * @start: starting user address
1037 * @nr_pages: number of pages from start to pin
1038 * @gup_flags: flags modifying pin behaviour
1039 * @pages: array that receives pointers to the pages pinned.
1040 * Should be at least nr_pages long. Or NULL, if caller
1041 * only intends to ensure the pages are faulted in.
1042 * @vmas: array of pointers to vmas corresponding to each page.
1043 * Or NULL if the caller does not require them.
1044 * @locked: whether we're still with the mmap_lock held
1045 *
1046 * Returns either number of pages pinned (which may be less than the
1047 * number requested), or an error. Details about the return value:
1048 *
1049 * -- If nr_pages is 0, returns 0.
1050 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1051 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1052 * pages pinned. Again, this may be less than nr_pages.
1053 * -- 0 return value is possible when the fault would need to be retried.
1054 *
1055 * The caller is responsible for releasing returned @pages, via put_page().
1056 *
1057 * @vmas are valid only as long as mmap_lock is held.
1058 *
1059 * Must be called with mmap_lock held. It may be released. See below.
1060 *
1061 * __get_user_pages walks a process's page tables and takes a reference to
1062 * each struct page that each user address corresponds to at a given
1063 * instant. That is, it takes the page that would be accessed if a user
1064 * thread accesses the given user virtual address at that instant.
1065 *
1066 * This does not guarantee that the page exists in the user mappings when
1067 * __get_user_pages returns, and there may even be a completely different
1068 * page there in some cases (eg. if mmapped pagecache has been invalidated
1069 * and subsequently re faulted). However it does guarantee that the page
1070 * won't be freed completely. And mostly callers simply care that the page
1071 * contains data that was valid *at some point in time*. Typically, an IO
1072 * or similar operation cannot guarantee anything stronger anyway because
1073 * locks can't be held over the syscall boundary.
1074 *
1075 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1076 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1077 * appropriate) must be called after the page is finished with, and
1078 * before put_page is called.
1079 *
1080 * If @locked != NULL, *@locked will be set to 0 when mmap_lock is
1081 * released by an up_read(). That can happen if @gup_flags does not
1082 * have FOLL_NOWAIT.
1083 *
1084 * A caller using such a combination of @locked and @gup_flags
1085 * must therefore hold the mmap_lock for reading only, and recognize
1086 * when it's been released. Otherwise, it must be held for either
1087 * reading or writing and will not be released.
1088 *
1089 * In most cases, get_user_pages or get_user_pages_fast should be used
1090 * instead of __get_user_pages. __get_user_pages should be used only if
1091 * you need some special @gup_flags.
1092 */
1093static long __get_user_pages(struct mm_struct *mm,
1094 unsigned long start, unsigned long nr_pages,
1095 unsigned int gup_flags, struct page **pages,
1096 struct vm_area_struct **vmas, int *locked)
1097{
1098 long ret = 0, i = 0;
1099 struct vm_area_struct *vma = NULL;
1100 struct follow_page_context ctx = { NULL };
1101
1102 if (!nr_pages)
1103 return 0;
1104
1105 start = untagged_addr(start);
1106
1107 VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
1108
1109 /*
1110 * If FOLL_FORCE is set then do not force a full fault as the hinting
1111 * fault information is unrelated to the reference behaviour of a task
1112 * using the address space
1113 */
1114 if (!(gup_flags & FOLL_FORCE))
1115 gup_flags |= FOLL_NUMA;
1116
1117 do {
1118 struct page *page;
1119 unsigned int foll_flags = gup_flags;
1120 unsigned int page_increm;
1121
1122 /* first iteration or cross vma bound */
1123 if (!vma || start >= vma->vm_end) {
1124 vma = find_extend_vma(mm, start);
1125 if (!vma && in_gate_area(mm, start)) {
1126 ret = get_gate_page(mm, start & PAGE_MASK,
1127 gup_flags, &vma,
1128 pages ? &pages[i] : NULL);
1129 if (ret)
1130 goto out;
1131 ctx.page_mask = 0;
1132 goto next_page;
1133 }
1134
1135 if (!vma) {
1136 ret = -EFAULT;
1137 goto out;
1138 }
1139 ret = check_vma_flags(vma, gup_flags);
1140 if (ret)
1141 goto out;
1142
1143 if (is_vm_hugetlb_page(vma)) {
1144 i = follow_hugetlb_page(mm, vma, pages, vmas,
1145 &start, &nr_pages, i,
1146 gup_flags, locked);
1147 if (locked && *locked == 0) {
1148 /*
1149 * We've got a VM_FAULT_RETRY
1150 * and we've lost mmap_lock.
1151 * We must stop here.
1152 */
1153 BUG_ON(gup_flags & FOLL_NOWAIT);
1154 BUG_ON(ret != 0);
1155 goto out;
1156 }
1157 continue;
1158 }
1159 }
1160retry:
1161 /*
1162 * If we have a pending SIGKILL, don't keep faulting pages and
1163 * potentially allocating memory.
1164 */
1165 if (fatal_signal_pending(current)) {
1166 ret = -EINTR;
1167 goto out;
1168 }
1169 cond_resched();
1170
1171 page = follow_page_mask(vma, start, foll_flags, &ctx);
1172 if (!page) {
1173 ret = faultin_page(vma, start, &foll_flags, locked);
1174 switch (ret) {
1175 case 0:
1176 goto retry;
1177 case -EBUSY:
1178 ret = 0;
1179 fallthrough;
1180 case -EFAULT:
1181 case -ENOMEM:
1182 case -EHWPOISON:
1183 goto out;
1184 case -ENOENT:
1185 goto next_page;
1186 }
1187 BUG();
1188 } else if (PTR_ERR(page) == -EEXIST) {
1189 /*
1190 * Proper page table entry exists, but no corresponding
1191 * struct page.
1192 */
1193 goto next_page;
1194 } else if (IS_ERR(page)) {
1195 ret = PTR_ERR(page);
1196 goto out;
1197 }
1198 if (pages) {
1199 pages[i] = page;
1200 flush_anon_page(vma, page, start);
1201 flush_dcache_page(page);
1202 ctx.page_mask = 0;
1203 }
1204next_page:
1205 if (vmas) {
1206 vmas[i] = vma;
1207 ctx.page_mask = 0;
1208 }
1209 page_increm = 1 + (~(start >> PAGE_SHIFT) & ctx.page_mask);
1210 if (page_increm > nr_pages)
1211 page_increm = nr_pages;
1212 i += page_increm;
1213 start += page_increm * PAGE_SIZE;
1214 nr_pages -= page_increm;
1215 } while (nr_pages);
1216out:
1217 if (ctx.pgmap)
1218 put_dev_pagemap(ctx.pgmap);
1219 return i ? i : ret;
1220}
1221
1222static bool vma_permits_fault(struct vm_area_struct *vma,
1223 unsigned int fault_flags)
1224{
1225 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
1226 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
1227 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
1228
1229 if (!(vm_flags & vma->vm_flags))
1230 return false;
1231
1232 /*
1233 * The architecture might have a hardware protection
1234 * mechanism other than read/write that can deny access.
1235 *
1236 * gup always represents data access, not instruction
1237 * fetches, so execute=false here:
1238 */
1239 if (!arch_vma_access_permitted(vma, write, false, foreign))
1240 return false;
1241
1242 return true;
1243}
1244
1245/**
1246 * fixup_user_fault() - manually resolve a user page fault
1247 * @mm: mm_struct of target mm
1248 * @address: user address
1249 * @fault_flags:flags to pass down to handle_mm_fault()
1250 * @unlocked: did we unlock the mmap_lock while retrying, maybe NULL if caller
1251 * does not allow retry. If NULL, the caller must guarantee
1252 * that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY.
1253 *
1254 * This is meant to be called in the specific scenario where for locking reasons
1255 * we try to access user memory in atomic context (within a pagefault_disable()
1256 * section), this returns -EFAULT, and we want to resolve the user fault before
1257 * trying again.
1258 *
1259 * Typically this is meant to be used by the futex code.
1260 *
1261 * The main difference with get_user_pages() is that this function will
1262 * unconditionally call handle_mm_fault() which will in turn perform all the
1263 * necessary SW fixup of the dirty and young bits in the PTE, while
1264 * get_user_pages() only guarantees to update these in the struct page.
1265 *
1266 * This is important for some architectures where those bits also gate the
1267 * access permission to the page because they are maintained in software. On
1268 * such architectures, gup() will not be enough to make a subsequent access
1269 * succeed.
1270 *
1271 * This function will not return with an unlocked mmap_lock. So it has not the
1272 * same semantics wrt the @mm->mmap_lock as does filemap_fault().
1273 */
1274int fixup_user_fault(struct mm_struct *mm,
1275 unsigned long address, unsigned int fault_flags,
1276 bool *unlocked)
1277{
1278 struct vm_area_struct *vma;
1279 vm_fault_t ret, major = 0;
1280
1281 address = untagged_addr(address);
1282
1283 if (unlocked)
1284 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1285
1286retry:
1287 vma = find_extend_vma(mm, address);
1288 if (!vma || address < vma->vm_start)
1289 return -EFAULT;
1290
1291 if (!vma_permits_fault(vma, fault_flags))
1292 return -EFAULT;
1293
1294 if ((fault_flags & FAULT_FLAG_KILLABLE) &&
1295 fatal_signal_pending(current))
1296 return -EINTR;
1297
1298 ret = handle_mm_fault(vma, address, fault_flags, NULL);
1299 major |= ret & VM_FAULT_MAJOR;
1300 if (ret & VM_FAULT_ERROR) {
1301 int err = vm_fault_to_errno(ret, 0);
1302
1303 if (err)
1304 return err;
1305 BUG();
1306 }
1307
1308 if (ret & VM_FAULT_RETRY) {
1309 mmap_read_lock(mm);
1310 *unlocked = true;
1311 fault_flags |= FAULT_FLAG_TRIED;
1312 goto retry;
1313 }
1314
1315 return 0;
1316}
1317EXPORT_SYMBOL_GPL(fixup_user_fault);
1318
1319/*
1320 * Please note that this function, unlike __get_user_pages will not
1321 * return 0 for nr_pages > 0 without FOLL_NOWAIT
1322 */
1323static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
1324 unsigned long start,
1325 unsigned long nr_pages,
1326 struct page **pages,
1327 struct vm_area_struct **vmas,
1328 int *locked,
1329 unsigned int flags)
1330{
1331 long ret, pages_done;
1332 bool lock_dropped;
1333
1334 if (locked) {
1335 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
1336 BUG_ON(vmas);
1337 /* check caller initialized locked */
1338 BUG_ON(*locked != 1);
1339 }
1340
1341 if (flags & FOLL_PIN)
1342 mm_set_has_pinned_flag(&mm->flags);
1343
1344 /*
1345 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
1346 * is to set FOLL_GET if the caller wants pages[] filled in (but has
1347 * carelessly failed to specify FOLL_GET), so keep doing that, but only
1348 * for FOLL_GET, not for the newer FOLL_PIN.
1349 *
1350 * FOLL_PIN always expects pages to be non-null, but no need to assert
1351 * that here, as any failures will be obvious enough.
1352 */
1353 if (pages && !(flags & FOLL_PIN))
1354 flags |= FOLL_GET;
1355
1356 pages_done = 0;
1357 lock_dropped = false;
1358 for (;;) {
1359 ret = __get_user_pages(mm, start, nr_pages, flags, pages,
1360 vmas, locked);
1361 if (!locked)
1362 /* VM_FAULT_RETRY couldn't trigger, bypass */
1363 return ret;
1364
1365 /* VM_FAULT_RETRY cannot return errors */
1366 if (!*locked) {
1367 BUG_ON(ret < 0);
1368 BUG_ON(ret >= nr_pages);
1369 }
1370
1371 if (ret > 0) {
1372 nr_pages -= ret;
1373 pages_done += ret;
1374 if (!nr_pages)
1375 break;
1376 }
1377 if (*locked) {
1378 /*
1379 * VM_FAULT_RETRY didn't trigger or it was a
1380 * FOLL_NOWAIT.
1381 */
1382 if (!pages_done)
1383 pages_done = ret;
1384 break;
1385 }
1386 /*
1387 * VM_FAULT_RETRY triggered, so seek to the faulting offset.
1388 * For the prefault case (!pages) we only update counts.
1389 */
1390 if (likely(pages))
1391 pages += ret;
1392 start += ret << PAGE_SHIFT;
1393 lock_dropped = true;
1394
1395retry:
1396 /*
1397 * Repeat on the address that fired VM_FAULT_RETRY
1398 * with both FAULT_FLAG_ALLOW_RETRY and
1399 * FAULT_FLAG_TRIED. Note that GUP can be interrupted
1400 * by fatal signals, so we need to check it before we
1401 * start trying again otherwise it can loop forever.
1402 */
1403
1404 if (fatal_signal_pending(current)) {
1405 if (!pages_done)
1406 pages_done = -EINTR;
1407 break;
1408 }
1409
1410 ret = mmap_read_lock_killable(mm);
1411 if (ret) {
1412 BUG_ON(ret > 0);
1413 if (!pages_done)
1414 pages_done = ret;
1415 break;
1416 }
1417
1418 *locked = 1;
1419 ret = __get_user_pages(mm, start, 1, flags | FOLL_TRIED,
1420 pages, NULL, locked);
1421 if (!*locked) {
1422 /* Continue to retry until we succeeded */
1423 BUG_ON(ret != 0);
1424 goto retry;
1425 }
1426 if (ret != 1) {
1427 BUG_ON(ret > 1);
1428 if (!pages_done)
1429 pages_done = ret;
1430 break;
1431 }
1432 nr_pages--;
1433 pages_done++;
1434 if (!nr_pages)
1435 break;
1436 if (likely(pages))
1437 pages++;
1438 start += PAGE_SIZE;
1439 }
1440 if (lock_dropped && *locked) {
1441 /*
1442 * We must let the caller know we temporarily dropped the lock
1443 * and so the critical section protected by it was lost.
1444 */
1445 mmap_read_unlock(mm);
1446 *locked = 0;
1447 }
1448 return pages_done;
1449}
1450
1451/**
1452 * populate_vma_page_range() - populate a range of pages in the vma.
1453 * @vma: target vma
1454 * @start: start address
1455 * @end: end address
1456 * @locked: whether the mmap_lock is still held
1457 *
1458 * This takes care of mlocking the pages too if VM_LOCKED is set.
1459 *
1460 * Return either number of pages pinned in the vma, or a negative error
1461 * code on error.
1462 *
1463 * vma->vm_mm->mmap_lock must be held.
1464 *
1465 * If @locked is NULL, it may be held for read or write and will
1466 * be unperturbed.
1467 *
1468 * If @locked is non-NULL, it must held for read only and may be
1469 * released. If it's released, *@locked will be set to 0.
1470 */
1471long populate_vma_page_range(struct vm_area_struct *vma,
1472 unsigned long start, unsigned long end, int *locked)
1473{
1474 struct mm_struct *mm = vma->vm_mm;
1475 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1476 int gup_flags;
1477
1478 VM_BUG_ON(start & ~PAGE_MASK);
1479 VM_BUG_ON(end & ~PAGE_MASK);
1480 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1481 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1482 mmap_assert_locked(mm);
1483
1484 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1485 if (vma->vm_flags & VM_LOCKONFAULT)
1486 gup_flags &= ~FOLL_POPULATE;
1487 /*
1488 * We want to touch writable mappings with a write fault in order
1489 * to break COW, except for shared mappings because these don't COW
1490 * and we would not want to dirty them for nothing.
1491 */
1492 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1493 gup_flags |= FOLL_WRITE;
1494
1495 /*
1496 * We want mlock to succeed for regions that have any permissions
1497 * other than PROT_NONE.
1498 */
1499 if (vma_is_accessible(vma))
1500 gup_flags |= FOLL_FORCE;
1501
1502 /*
1503 * We made sure addr is within a VMA, so the following will
1504 * not result in a stack expansion that recurses back here.
1505 */
1506 return __get_user_pages(mm, start, nr_pages, gup_flags,
1507 NULL, NULL, locked);
1508}
1509
1510/*
1511 * faultin_vma_page_range() - populate (prefault) page tables inside the
1512 * given VMA range readable/writable
1513 *
1514 * This takes care of mlocking the pages, too, if VM_LOCKED is set.
1515 *
1516 * @vma: target vma
1517 * @start: start address
1518 * @end: end address
1519 * @write: whether to prefault readable or writable
1520 * @locked: whether the mmap_lock is still held
1521 *
1522 * Returns either number of processed pages in the vma, or a negative error
1523 * code on error (see __get_user_pages()).
1524 *
1525 * vma->vm_mm->mmap_lock must be held. The range must be page-aligned and
1526 * covered by the VMA.
1527 *
1528 * If @locked is NULL, it may be held for read or write and will be unperturbed.
1529 *
1530 * If @locked is non-NULL, it must held for read only and may be released. If
1531 * it's released, *@locked will be set to 0.
1532 */
1533long faultin_vma_page_range(struct vm_area_struct *vma, unsigned long start,
1534 unsigned long end, bool write, int *locked)
1535{
1536 struct mm_struct *mm = vma->vm_mm;
1537 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1538 int gup_flags;
1539
1540 VM_BUG_ON(!PAGE_ALIGNED(start));
1541 VM_BUG_ON(!PAGE_ALIGNED(end));
1542 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1543 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1544 mmap_assert_locked(mm);
1545
1546 /*
1547 * FOLL_TOUCH: Mark page accessed and thereby young; will also mark
1548 * the page dirty with FOLL_WRITE -- which doesn't make a
1549 * difference with !FOLL_FORCE, because the page is writable
1550 * in the page table.
1551 * FOLL_HWPOISON: Return -EHWPOISON instead of -EFAULT when we hit
1552 * a poisoned page.
1553 * FOLL_POPULATE: Always populate memory with VM_LOCKONFAULT.
1554 * !FOLL_FORCE: Require proper access permissions.
1555 */
1556 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK | FOLL_HWPOISON;
1557 if (write)
1558 gup_flags |= FOLL_WRITE;
1559
1560 /*
1561 * We want to report -EINVAL instead of -EFAULT for any permission
1562 * problems or incompatible mappings.
1563 */
1564 if (check_vma_flags(vma, gup_flags))
1565 return -EINVAL;
1566
1567 return __get_user_pages(mm, start, nr_pages, gup_flags,
1568 NULL, NULL, locked);
1569}
1570
1571/*
1572 * __mm_populate - populate and/or mlock pages within a range of address space.
1573 *
1574 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1575 * flags. VMAs must be already marked with the desired vm_flags, and
1576 * mmap_lock must not be held.
1577 */
1578int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1579{
1580 struct mm_struct *mm = current->mm;
1581 unsigned long end, nstart, nend;
1582 struct vm_area_struct *vma = NULL;
1583 int locked = 0;
1584 long ret = 0;
1585
1586 end = start + len;
1587
1588 for (nstart = start; nstart < end; nstart = nend) {
1589 /*
1590 * We want to fault in pages for [nstart; end) address range.
1591 * Find first corresponding VMA.
1592 */
1593 if (!locked) {
1594 locked = 1;
1595 mmap_read_lock(mm);
1596 vma = find_vma(mm, nstart);
1597 } else if (nstart >= vma->vm_end)
1598 vma = vma->vm_next;
1599 if (!vma || vma->vm_start >= end)
1600 break;
1601 /*
1602 * Set [nstart; nend) to intersection of desired address
1603 * range with the first VMA. Also, skip undesirable VMA types.
1604 */
1605 nend = min(end, vma->vm_end);
1606 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1607 continue;
1608 if (nstart < vma->vm_start)
1609 nstart = vma->vm_start;
1610 /*
1611 * Now fault in a range of pages. populate_vma_page_range()
1612 * double checks the vma flags, so that it won't mlock pages
1613 * if the vma was already munlocked.
1614 */
1615 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1616 if (ret < 0) {
1617 if (ignore_errors) {
1618 ret = 0;
1619 continue; /* continue at next VMA */
1620 }
1621 break;
1622 }
1623 nend = nstart + ret * PAGE_SIZE;
1624 ret = 0;
1625 }
1626 if (locked)
1627 mmap_read_unlock(mm);
1628 return ret; /* 0 or negative error code */
1629}
1630#else /* CONFIG_MMU */
1631static long __get_user_pages_locked(struct mm_struct *mm, unsigned long start,
1632 unsigned long nr_pages, struct page **pages,
1633 struct vm_area_struct **vmas, int *locked,
1634 unsigned int foll_flags)
1635{
1636 struct vm_area_struct *vma;
1637 unsigned long vm_flags;
1638 long i;
1639
1640 /* calculate required read or write permissions.
1641 * If FOLL_FORCE is set, we only require the "MAY" flags.
1642 */
1643 vm_flags = (foll_flags & FOLL_WRITE) ?
1644 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1645 vm_flags &= (foll_flags & FOLL_FORCE) ?
1646 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1647
1648 for (i = 0; i < nr_pages; i++) {
1649 vma = find_vma(mm, start);
1650 if (!vma)
1651 goto finish_or_fault;
1652
1653 /* protect what we can, including chardevs */
1654 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1655 !(vm_flags & vma->vm_flags))
1656 goto finish_or_fault;
1657
1658 if (pages) {
1659 pages[i] = virt_to_page(start);
1660 if (pages[i])
1661 get_page(pages[i]);
1662 }
1663 if (vmas)
1664 vmas[i] = vma;
1665 start = (start + PAGE_SIZE) & PAGE_MASK;
1666 }
1667
1668 return i;
1669
1670finish_or_fault:
1671 return i ? : -EFAULT;
1672}
1673#endif /* !CONFIG_MMU */
1674
1675/**
1676 * get_dump_page() - pin user page in memory while writing it to core dump
1677 * @addr: user address
1678 *
1679 * Returns struct page pointer of user page pinned for dump,
1680 * to be freed afterwards by put_page().
1681 *
1682 * Returns NULL on any kind of failure - a hole must then be inserted into
1683 * the corefile, to preserve alignment with its headers; and also returns
1684 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1685 * allowing a hole to be left in the corefile to save disk space.
1686 *
1687 * Called without mmap_lock (takes and releases the mmap_lock by itself).
1688 */
1689#ifdef CONFIG_ELF_CORE
1690struct page *get_dump_page(unsigned long addr)
1691{
1692 struct mm_struct *mm = current->mm;
1693 struct page *page;
1694 int locked = 1;
1695 int ret;
1696
1697 if (mmap_read_lock_killable(mm))
1698 return NULL;
1699 ret = __get_user_pages_locked(mm, addr, 1, &page, NULL, &locked,
1700 FOLL_FORCE | FOLL_DUMP | FOLL_GET);
1701 if (locked)
1702 mmap_read_unlock(mm);
1703 return (ret == 1) ? page : NULL;
1704}
1705#endif /* CONFIG_ELF_CORE */
1706
1707#ifdef CONFIG_MIGRATION
1708/*
1709 * Check whether all pages are pinnable, if so return number of pages. If some
1710 * pages are not pinnable, migrate them, and unpin all pages. Return zero if
1711 * pages were migrated, or if some pages were not successfully isolated.
1712 * Return negative error if migration fails.
1713 */
1714static long check_and_migrate_movable_pages(unsigned long nr_pages,
1715 struct page **pages,
1716 unsigned int gup_flags)
1717{
1718 unsigned long i;
1719 unsigned long isolation_error_count = 0;
1720 bool drain_allow = true;
1721 LIST_HEAD(movable_page_list);
1722 long ret = 0;
1723 struct page *prev_head = NULL;
1724 struct page *head;
1725 struct migration_target_control mtc = {
1726 .nid = NUMA_NO_NODE,
1727 .gfp_mask = GFP_USER | __GFP_NOWARN,
1728 };
1729
1730 for (i = 0; i < nr_pages; i++) {
1731 head = compound_head(pages[i]);
1732 if (head == prev_head)
1733 continue;
1734 prev_head = head;
1735 /*
1736 * If we get a movable page, since we are going to be pinning
1737 * these entries, try to move them out if possible.
1738 */
1739 if (!is_pinnable_page(head)) {
1740 if (PageHuge(head)) {
1741 if (!isolate_huge_page(head, &movable_page_list))
1742 isolation_error_count++;
1743 } else {
1744 if (!PageLRU(head) && drain_allow) {
1745 lru_add_drain_all();
1746 drain_allow = false;
1747 }
1748
1749 if (isolate_lru_page(head)) {
1750 isolation_error_count++;
1751 continue;
1752 }
1753 list_add_tail(&head->lru, &movable_page_list);
1754 mod_node_page_state(page_pgdat(head),
1755 NR_ISOLATED_ANON +
1756 page_is_file_lru(head),
1757 thp_nr_pages(head));
1758 }
1759 }
1760 }
1761
1762 /*
1763 * If list is empty, and no isolation errors, means that all pages are
1764 * in the correct zone.
1765 */
1766 if (list_empty(&movable_page_list) && !isolation_error_count)
1767 return nr_pages;
1768
1769 if (gup_flags & FOLL_PIN) {
1770 unpin_user_pages(pages, nr_pages);
1771 } else {
1772 for (i = 0; i < nr_pages; i++)
1773 put_page(pages[i]);
1774 }
1775 if (!list_empty(&movable_page_list)) {
1776 ret = migrate_pages(&movable_page_list, alloc_migration_target,
1777 NULL, (unsigned long)&mtc, MIGRATE_SYNC,
1778 MR_LONGTERM_PIN);
1779 if (ret && !list_empty(&movable_page_list))
1780 putback_movable_pages(&movable_page_list);
1781 }
1782
1783 return ret > 0 ? -ENOMEM : ret;
1784}
1785#else
1786static long check_and_migrate_movable_pages(unsigned long nr_pages,
1787 struct page **pages,
1788 unsigned int gup_flags)
1789{
1790 return nr_pages;
1791}
1792#endif /* CONFIG_MIGRATION */
1793
1794/*
1795 * __gup_longterm_locked() is a wrapper for __get_user_pages_locked which
1796 * allows us to process the FOLL_LONGTERM flag.
1797 */
1798static long __gup_longterm_locked(struct mm_struct *mm,
1799 unsigned long start,
1800 unsigned long nr_pages,
1801 struct page **pages,
1802 struct vm_area_struct **vmas,
1803 unsigned int gup_flags)
1804{
1805 unsigned int flags;
1806 long rc;
1807
1808 if (!(gup_flags & FOLL_LONGTERM))
1809 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1810 NULL, gup_flags);
1811 flags = memalloc_pin_save();
1812 do {
1813 rc = __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1814 NULL, gup_flags);
1815 if (rc <= 0)
1816 break;
1817 rc = check_and_migrate_movable_pages(rc, pages, gup_flags);
1818 } while (!rc);
1819 memalloc_pin_restore(flags);
1820
1821 return rc;
1822}
1823
1824static bool is_valid_gup_flags(unsigned int gup_flags)
1825{
1826 /*
1827 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
1828 * never directly by the caller, so enforce that with an assertion:
1829 */
1830 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
1831 return false;
1832 /*
1833 * FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying
1834 * that is, FOLL_LONGTERM is a specific case, more restrictive case of
1835 * FOLL_PIN.
1836 */
1837 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
1838 return false;
1839
1840 return true;
1841}
1842
1843#ifdef CONFIG_MMU
1844static long __get_user_pages_remote(struct mm_struct *mm,
1845 unsigned long start, unsigned long nr_pages,
1846 unsigned int gup_flags, struct page **pages,
1847 struct vm_area_struct **vmas, int *locked)
1848{
1849 /*
1850 * Parts of FOLL_LONGTERM behavior are incompatible with
1851 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
1852 * vmas. However, this only comes up if locked is set, and there are
1853 * callers that do request FOLL_LONGTERM, but do not set locked. So,
1854 * allow what we can.
1855 */
1856 if (gup_flags & FOLL_LONGTERM) {
1857 if (WARN_ON_ONCE(locked))
1858 return -EINVAL;
1859 /*
1860 * This will check the vmas (even if our vmas arg is NULL)
1861 * and return -ENOTSUPP if DAX isn't allowed in this case:
1862 */
1863 return __gup_longterm_locked(mm, start, nr_pages, pages,
1864 vmas, gup_flags | FOLL_TOUCH |
1865 FOLL_REMOTE);
1866 }
1867
1868 return __get_user_pages_locked(mm, start, nr_pages, pages, vmas,
1869 locked,
1870 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1871}
1872
1873/**
1874 * get_user_pages_remote() - pin user pages in memory
1875 * @mm: mm_struct of target mm
1876 * @start: starting user address
1877 * @nr_pages: number of pages from start to pin
1878 * @gup_flags: flags modifying lookup behaviour
1879 * @pages: array that receives pointers to the pages pinned.
1880 * Should be at least nr_pages long. Or NULL, if caller
1881 * only intends to ensure the pages are faulted in.
1882 * @vmas: array of pointers to vmas corresponding to each page.
1883 * Or NULL if the caller does not require them.
1884 * @locked: pointer to lock flag indicating whether lock is held and
1885 * subsequently whether VM_FAULT_RETRY functionality can be
1886 * utilised. Lock must initially be held.
1887 *
1888 * Returns either number of pages pinned (which may be less than the
1889 * number requested), or an error. Details about the return value:
1890 *
1891 * -- If nr_pages is 0, returns 0.
1892 * -- If nr_pages is >0, but no pages were pinned, returns -errno.
1893 * -- If nr_pages is >0, and some pages were pinned, returns the number of
1894 * pages pinned. Again, this may be less than nr_pages.
1895 *
1896 * The caller is responsible for releasing returned @pages, via put_page().
1897 *
1898 * @vmas are valid only as long as mmap_lock is held.
1899 *
1900 * Must be called with mmap_lock held for read or write.
1901 *
1902 * get_user_pages_remote walks a process's page tables and takes a reference
1903 * to each struct page that each user address corresponds to at a given
1904 * instant. That is, it takes the page that would be accessed if a user
1905 * thread accesses the given user virtual address at that instant.
1906 *
1907 * This does not guarantee that the page exists in the user mappings when
1908 * get_user_pages_remote returns, and there may even be a completely different
1909 * page there in some cases (eg. if mmapped pagecache has been invalidated
1910 * and subsequently re faulted). However it does guarantee that the page
1911 * won't be freed completely. And mostly callers simply care that the page
1912 * contains data that was valid *at some point in time*. Typically, an IO
1913 * or similar operation cannot guarantee anything stronger anyway because
1914 * locks can't be held over the syscall boundary.
1915 *
1916 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1917 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1918 * be called after the page is finished with, and before put_page is called.
1919 *
1920 * get_user_pages_remote is typically used for fewer-copy IO operations,
1921 * to get a handle on the memory by some means other than accesses
1922 * via the user virtual addresses. The pages may be submitted for
1923 * DMA to devices or accessed via their kernel linear mapping (via the
1924 * kmap APIs). Care should be taken to use the correct cache flushing APIs.
1925 *
1926 * See also get_user_pages_fast, for performance critical applications.
1927 *
1928 * get_user_pages_remote should be phased out in favor of
1929 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1930 * should use get_user_pages_remote because it cannot pass
1931 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1932 */
1933long get_user_pages_remote(struct mm_struct *mm,
1934 unsigned long start, unsigned long nr_pages,
1935 unsigned int gup_flags, struct page **pages,
1936 struct vm_area_struct **vmas, int *locked)
1937{
1938 if (!is_valid_gup_flags(gup_flags))
1939 return -EINVAL;
1940
1941 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
1942 pages, vmas, locked);
1943}
1944EXPORT_SYMBOL(get_user_pages_remote);
1945
1946#else /* CONFIG_MMU */
1947long get_user_pages_remote(struct mm_struct *mm,
1948 unsigned long start, unsigned long nr_pages,
1949 unsigned int gup_flags, struct page **pages,
1950 struct vm_area_struct **vmas, int *locked)
1951{
1952 return 0;
1953}
1954
1955static long __get_user_pages_remote(struct mm_struct *mm,
1956 unsigned long start, unsigned long nr_pages,
1957 unsigned int gup_flags, struct page **pages,
1958 struct vm_area_struct **vmas, int *locked)
1959{
1960 return 0;
1961}
1962#endif /* !CONFIG_MMU */
1963
1964/**
1965 * get_user_pages() - pin user pages in memory
1966 * @start: starting user address
1967 * @nr_pages: number of pages from start to pin
1968 * @gup_flags: flags modifying lookup behaviour
1969 * @pages: array that receives pointers to the pages pinned.
1970 * Should be at least nr_pages long. Or NULL, if caller
1971 * only intends to ensure the pages are faulted in.
1972 * @vmas: array of pointers to vmas corresponding to each page.
1973 * Or NULL if the caller does not require them.
1974 *
1975 * This is the same as get_user_pages_remote(), just with a less-flexible
1976 * calling convention where we assume that the mm being operated on belongs to
1977 * the current task, and doesn't allow passing of a locked parameter. We also
1978 * obviously don't pass FOLL_REMOTE in here.
1979 */
1980long get_user_pages(unsigned long start, unsigned long nr_pages,
1981 unsigned int gup_flags, struct page **pages,
1982 struct vm_area_struct **vmas)
1983{
1984 if (!is_valid_gup_flags(gup_flags))
1985 return -EINVAL;
1986
1987 return __gup_longterm_locked(current->mm, start, nr_pages,
1988 pages, vmas, gup_flags | FOLL_TOUCH);
1989}
1990EXPORT_SYMBOL(get_user_pages);
1991
1992/**
1993 * get_user_pages_locked() - variant of get_user_pages()
1994 *
1995 * @start: starting user address
1996 * @nr_pages: number of pages from start to pin
1997 * @gup_flags: flags modifying lookup behaviour
1998 * @pages: array that receives pointers to the pages pinned.
1999 * Should be at least nr_pages long. Or NULL, if caller
2000 * only intends to ensure the pages are faulted in.
2001 * @locked: pointer to lock flag indicating whether lock is held and
2002 * subsequently whether VM_FAULT_RETRY functionality can be
2003 * utilised. Lock must initially be held.
2004 *
2005 * It is suitable to replace the form:
2006 *
2007 * mmap_read_lock(mm);
2008 * do_something()
2009 * get_user_pages(mm, ..., pages, NULL);
2010 * mmap_read_unlock(mm);
2011 *
2012 * to:
2013 *
2014 * int locked = 1;
2015 * mmap_read_lock(mm);
2016 * do_something()
2017 * get_user_pages_locked(mm, ..., pages, &locked);
2018 * if (locked)
2019 * mmap_read_unlock(mm);
2020 *
2021 * We can leverage the VM_FAULT_RETRY functionality in the page fault
2022 * paths better by using either get_user_pages_locked() or
2023 * get_user_pages_unlocked().
2024 *
2025 */
2026long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
2027 unsigned int gup_flags, struct page **pages,
2028 int *locked)
2029{
2030 /*
2031 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2032 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2033 * vmas. As there are no users of this flag in this call we simply
2034 * disallow this option for now.
2035 */
2036 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2037 return -EINVAL;
2038 /*
2039 * FOLL_PIN must only be set internally by the pin_user_pages*() APIs,
2040 * never directly by the caller, so enforce that:
2041 */
2042 if (WARN_ON_ONCE(gup_flags & FOLL_PIN))
2043 return -EINVAL;
2044
2045 return __get_user_pages_locked(current->mm, start, nr_pages,
2046 pages, NULL, locked,
2047 gup_flags | FOLL_TOUCH);
2048}
2049EXPORT_SYMBOL(get_user_pages_locked);
2050
2051/*
2052 * get_user_pages_unlocked() is suitable to replace the form:
2053 *
2054 * mmap_read_lock(mm);
2055 * get_user_pages(mm, ..., pages, NULL);
2056 * mmap_read_unlock(mm);
2057 *
2058 * with:
2059 *
2060 * get_user_pages_unlocked(mm, ..., pages);
2061 *
2062 * It is functionally equivalent to get_user_pages_fast so
2063 * get_user_pages_fast should be used instead if specific gup_flags
2064 * (e.g. FOLL_FORCE) are not required.
2065 */
2066long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2067 struct page **pages, unsigned int gup_flags)
2068{
2069 struct mm_struct *mm = current->mm;
2070 int locked = 1;
2071 long ret;
2072
2073 /*
2074 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
2075 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
2076 * vmas. As there are no users of this flag in this call we simply
2077 * disallow this option for now.
2078 */
2079 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
2080 return -EINVAL;
2081
2082 mmap_read_lock(mm);
2083 ret = __get_user_pages_locked(mm, start, nr_pages, pages, NULL,
2084 &locked, gup_flags | FOLL_TOUCH);
2085 if (locked)
2086 mmap_read_unlock(mm);
2087 return ret;
2088}
2089EXPORT_SYMBOL(get_user_pages_unlocked);
2090
2091/*
2092 * Fast GUP
2093 *
2094 * get_user_pages_fast attempts to pin user pages by walking the page
2095 * tables directly and avoids taking locks. Thus the walker needs to be
2096 * protected from page table pages being freed from under it, and should
2097 * block any THP splits.
2098 *
2099 * One way to achieve this is to have the walker disable interrupts, and
2100 * rely on IPIs from the TLB flushing code blocking before the page table
2101 * pages are freed. This is unsuitable for architectures that do not need
2102 * to broadcast an IPI when invalidating TLBs.
2103 *
2104 * Another way to achieve this is to batch up page table containing pages
2105 * belonging to more than one mm_user, then rcu_sched a callback to free those
2106 * pages. Disabling interrupts will allow the fast_gup walker to both block
2107 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
2108 * (which is a relatively rare event). The code below adopts this strategy.
2109 *
2110 * Before activating this code, please be aware that the following assumptions
2111 * are currently made:
2112 *
2113 * *) Either MMU_GATHER_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
2114 * free pages containing page tables or TLB flushing requires IPI broadcast.
2115 *
2116 * *) ptes can be read atomically by the architecture.
2117 *
2118 * *) access_ok is sufficient to validate userspace address ranges.
2119 *
2120 * The last two assumptions can be relaxed by the addition of helper functions.
2121 *
2122 * This code is based heavily on the PowerPC implementation by Nick Piggin.
2123 */
2124#ifdef CONFIG_HAVE_FAST_GUP
2125
2126static void __maybe_unused undo_dev_pagemap(int *nr, int nr_start,
2127 unsigned int flags,
2128 struct page **pages)
2129{
2130 while ((*nr) - nr_start) {
2131 struct page *page = pages[--(*nr)];
2132
2133 ClearPageReferenced(page);
2134 if (flags & FOLL_PIN)
2135 unpin_user_page(page);
2136 else
2137 put_page(page);
2138 }
2139}
2140
2141#ifdef CONFIG_ARCH_HAS_PTE_SPECIAL
2142static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2143 unsigned int flags, struct page **pages, int *nr)
2144{
2145 struct dev_pagemap *pgmap = NULL;
2146 int nr_start = *nr, ret = 0;
2147 pte_t *ptep, *ptem;
2148
2149 ptem = ptep = pte_offset_map(&pmd, addr);
2150 do {
2151 pte_t pte = ptep_get_lockless(ptep);
2152 struct page *head, *page;
2153
2154 /*
2155 * Similar to the PMD case below, NUMA hinting must take slow
2156 * path using the pte_protnone check.
2157 */
2158 if (pte_protnone(pte))
2159 goto pte_unmap;
2160
2161 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2162 goto pte_unmap;
2163
2164 if (pte_devmap(pte)) {
2165 if (unlikely(flags & FOLL_LONGTERM))
2166 goto pte_unmap;
2167
2168 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
2169 if (unlikely(!pgmap)) {
2170 undo_dev_pagemap(nr, nr_start, flags, pages);
2171 goto pte_unmap;
2172 }
2173 } else if (pte_special(pte))
2174 goto pte_unmap;
2175
2176 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2177 page = pte_page(pte);
2178
2179 head = try_grab_compound_head(page, 1, flags);
2180 if (!head)
2181 goto pte_unmap;
2182
2183 if (unlikely(page_is_secretmem(page))) {
2184 put_compound_head(head, 1, flags);
2185 goto pte_unmap;
2186 }
2187
2188 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2189 put_compound_head(head, 1, flags);
2190 goto pte_unmap;
2191 }
2192
2193 VM_BUG_ON_PAGE(compound_head(page) != head, page);
2194
2195 /*
2196 * We need to make the page accessible if and only if we are
2197 * going to access its content (the FOLL_PIN case). Please
2198 * see Documentation/core-api/pin_user_pages.rst for
2199 * details.
2200 */
2201 if (flags & FOLL_PIN) {
2202 ret = arch_make_page_accessible(page);
2203 if (ret) {
2204 unpin_user_page(page);
2205 goto pte_unmap;
2206 }
2207 }
2208 SetPageReferenced(page);
2209 pages[*nr] = page;
2210 (*nr)++;
2211
2212 } while (ptep++, addr += PAGE_SIZE, addr != end);
2213
2214 ret = 1;
2215
2216pte_unmap:
2217 if (pgmap)
2218 put_dev_pagemap(pgmap);
2219 pte_unmap(ptem);
2220 return ret;
2221}
2222#else
2223
2224/*
2225 * If we can't determine whether or not a pte is special, then fail immediately
2226 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
2227 * to be special.
2228 *
2229 * For a futex to be placed on a THP tail page, get_futex_key requires a
2230 * get_user_pages_fast_only implementation that can pin pages. Thus it's still
2231 * useful to have gup_huge_pmd even if we can't operate on ptes.
2232 */
2233static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
2234 unsigned int flags, struct page **pages, int *nr)
2235{
2236 return 0;
2237}
2238#endif /* CONFIG_ARCH_HAS_PTE_SPECIAL */
2239
2240#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2241static int __gup_device_huge(unsigned long pfn, unsigned long addr,
2242 unsigned long end, unsigned int flags,
2243 struct page **pages, int *nr)
2244{
2245 int nr_start = *nr;
2246 struct dev_pagemap *pgmap = NULL;
2247
2248 do {
2249 struct page *page = pfn_to_page(pfn);
2250
2251 pgmap = get_dev_pagemap(pfn, pgmap);
2252 if (unlikely(!pgmap)) {
2253 undo_dev_pagemap(nr, nr_start, flags, pages);
2254 return 0;
2255 }
2256 SetPageReferenced(page);
2257 pages[*nr] = page;
2258 if (unlikely(!try_grab_page(page, flags))) {
2259 undo_dev_pagemap(nr, nr_start, flags, pages);
2260 return 0;
2261 }
2262 (*nr)++;
2263 pfn++;
2264 } while (addr += PAGE_SIZE, addr != end);
2265
2266 if (pgmap)
2267 put_dev_pagemap(pgmap);
2268 return 1;
2269}
2270
2271static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2272 unsigned long end, unsigned int flags,
2273 struct page **pages, int *nr)
2274{
2275 unsigned long fault_pfn;
2276 int nr_start = *nr;
2277
2278 fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2279 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2280 return 0;
2281
2282 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2283 undo_dev_pagemap(nr, nr_start, flags, pages);
2284 return 0;
2285 }
2286 return 1;
2287}
2288
2289static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2290 unsigned long end, unsigned int flags,
2291 struct page **pages, int *nr)
2292{
2293 unsigned long fault_pfn;
2294 int nr_start = *nr;
2295
2296 fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2297 if (!__gup_device_huge(fault_pfn, addr, end, flags, pages, nr))
2298 return 0;
2299
2300 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2301 undo_dev_pagemap(nr, nr_start, flags, pages);
2302 return 0;
2303 }
2304 return 1;
2305}
2306#else
2307static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2308 unsigned long end, unsigned int flags,
2309 struct page **pages, int *nr)
2310{
2311 BUILD_BUG();
2312 return 0;
2313}
2314
2315static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
2316 unsigned long end, unsigned int flags,
2317 struct page **pages, int *nr)
2318{
2319 BUILD_BUG();
2320 return 0;
2321}
2322#endif
2323
2324static int record_subpages(struct page *page, unsigned long addr,
2325 unsigned long end, struct page **pages)
2326{
2327 int nr;
2328
2329 for (nr = 0; addr != end; addr += PAGE_SIZE)
2330 pages[nr++] = page++;
2331
2332 return nr;
2333}
2334
2335#ifdef CONFIG_ARCH_HAS_HUGEPD
2336static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
2337 unsigned long sz)
2338{
2339 unsigned long __boundary = (addr + sz) & ~(sz-1);
2340 return (__boundary - 1 < end - 1) ? __boundary : end;
2341}
2342
2343static int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
2344 unsigned long end, unsigned int flags,
2345 struct page **pages, int *nr)
2346{
2347 unsigned long pte_end;
2348 struct page *head, *page;
2349 pte_t pte;
2350 int refs;
2351
2352 pte_end = (addr + sz) & ~(sz-1);
2353 if (pte_end < end)
2354 end = pte_end;
2355
2356 pte = huge_ptep_get(ptep);
2357
2358 if (!pte_access_permitted(pte, flags & FOLL_WRITE))
2359 return 0;
2360
2361 /* hugepages are never "special" */
2362 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
2363
2364 head = pte_page(pte);
2365 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
2366 refs = record_subpages(page, addr, end, pages + *nr);
2367
2368 head = try_grab_compound_head(head, refs, flags);
2369 if (!head)
2370 return 0;
2371
2372 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
2373 put_compound_head(head, refs, flags);
2374 return 0;
2375 }
2376
2377 *nr += refs;
2378 SetPageReferenced(head);
2379 return 1;
2380}
2381
2382static int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2383 unsigned int pdshift, unsigned long end, unsigned int flags,
2384 struct page **pages, int *nr)
2385{
2386 pte_t *ptep;
2387 unsigned long sz = 1UL << hugepd_shift(hugepd);
2388 unsigned long next;
2389
2390 ptep = hugepte_offset(hugepd, addr, pdshift);
2391 do {
2392 next = hugepte_addr_end(addr, end, sz);
2393 if (!gup_hugepte(ptep, sz, addr, end, flags, pages, nr))
2394 return 0;
2395 } while (ptep++, addr = next, addr != end);
2396
2397 return 1;
2398}
2399#else
2400static inline int gup_huge_pd(hugepd_t hugepd, unsigned long addr,
2401 unsigned int pdshift, unsigned long end, unsigned int flags,
2402 struct page **pages, int *nr)
2403{
2404 return 0;
2405}
2406#endif /* CONFIG_ARCH_HAS_HUGEPD */
2407
2408static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
2409 unsigned long end, unsigned int flags,
2410 struct page **pages, int *nr)
2411{
2412 struct page *head, *page;
2413 int refs;
2414
2415 if (!pmd_access_permitted(orig, flags & FOLL_WRITE))
2416 return 0;
2417
2418 if (pmd_devmap(orig)) {
2419 if (unlikely(flags & FOLL_LONGTERM))
2420 return 0;
2421 return __gup_device_huge_pmd(orig, pmdp, addr, end, flags,
2422 pages, nr);
2423 }
2424
2425 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
2426 refs = record_subpages(page, addr, end, pages + *nr);
2427
2428 head = try_grab_compound_head(pmd_page(orig), refs, flags);
2429 if (!head)
2430 return 0;
2431
2432 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
2433 put_compound_head(head, refs, flags);
2434 return 0;
2435 }
2436
2437 *nr += refs;
2438 SetPageReferenced(head);
2439 return 1;
2440}
2441
2442static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
2443 unsigned long end, unsigned int flags,
2444 struct page **pages, int *nr)
2445{
2446 struct page *head, *page;
2447 int refs;
2448
2449 if (!pud_access_permitted(orig, flags & FOLL_WRITE))
2450 return 0;
2451
2452 if (pud_devmap(orig)) {
2453 if (unlikely(flags & FOLL_LONGTERM))
2454 return 0;
2455 return __gup_device_huge_pud(orig, pudp, addr, end, flags,
2456 pages, nr);
2457 }
2458
2459 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
2460 refs = record_subpages(page, addr, end, pages + *nr);
2461
2462 head = try_grab_compound_head(pud_page(orig), refs, flags);
2463 if (!head)
2464 return 0;
2465
2466 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
2467 put_compound_head(head, refs, flags);
2468 return 0;
2469 }
2470
2471 *nr += refs;
2472 SetPageReferenced(head);
2473 return 1;
2474}
2475
2476static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
2477 unsigned long end, unsigned int flags,
2478 struct page **pages, int *nr)
2479{
2480 int refs;
2481 struct page *head, *page;
2482
2483 if (!pgd_access_permitted(orig, flags & FOLL_WRITE))
2484 return 0;
2485
2486 BUILD_BUG_ON(pgd_devmap(orig));
2487
2488 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
2489 refs = record_subpages(page, addr, end, pages + *nr);
2490
2491 head = try_grab_compound_head(pgd_page(orig), refs, flags);
2492 if (!head)
2493 return 0;
2494
2495 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
2496 put_compound_head(head, refs, flags);
2497 return 0;
2498 }
2499
2500 *nr += refs;
2501 SetPageReferenced(head);
2502 return 1;
2503}
2504
2505static int gup_pmd_range(pud_t *pudp, pud_t pud, unsigned long addr, unsigned long end,
2506 unsigned int flags, struct page **pages, int *nr)
2507{
2508 unsigned long next;
2509 pmd_t *pmdp;
2510
2511 pmdp = pmd_offset_lockless(pudp, pud, addr);
2512 do {
2513 pmd_t pmd = READ_ONCE(*pmdp);
2514
2515 next = pmd_addr_end(addr, end);
2516 if (!pmd_present(pmd))
2517 return 0;
2518
2519 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
2520 pmd_devmap(pmd))) {
2521 /*
2522 * NUMA hinting faults need to be handled in the GUP
2523 * slowpath for accounting purposes and so that they
2524 * can be serialised against THP migration.
2525 */
2526 if (pmd_protnone(pmd))
2527 return 0;
2528
2529 if (!gup_huge_pmd(pmd, pmdp, addr, next, flags,
2530 pages, nr))
2531 return 0;
2532
2533 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
2534 /*
2535 * architecture have different format for hugetlbfs
2536 * pmd format and THP pmd format
2537 */
2538 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
2539 PMD_SHIFT, next, flags, pages, nr))
2540 return 0;
2541 } else if (!gup_pte_range(pmd, addr, next, flags, pages, nr))
2542 return 0;
2543 } while (pmdp++, addr = next, addr != end);
2544
2545 return 1;
2546}
2547
2548static int gup_pud_range(p4d_t *p4dp, p4d_t p4d, unsigned long addr, unsigned long end,
2549 unsigned int flags, struct page **pages, int *nr)
2550{
2551 unsigned long next;
2552 pud_t *pudp;
2553
2554 pudp = pud_offset_lockless(p4dp, p4d, addr);
2555 do {
2556 pud_t pud = READ_ONCE(*pudp);
2557
2558 next = pud_addr_end(addr, end);
2559 if (unlikely(!pud_present(pud)))
2560 return 0;
2561 if (unlikely(pud_huge(pud))) {
2562 if (!gup_huge_pud(pud, pudp, addr, next, flags,
2563 pages, nr))
2564 return 0;
2565 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
2566 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
2567 PUD_SHIFT, next, flags, pages, nr))
2568 return 0;
2569 } else if (!gup_pmd_range(pudp, pud, addr, next, flags, pages, nr))
2570 return 0;
2571 } while (pudp++, addr = next, addr != end);
2572
2573 return 1;
2574}
2575
2576static int gup_p4d_range(pgd_t *pgdp, pgd_t pgd, unsigned long addr, unsigned long end,
2577 unsigned int flags, struct page **pages, int *nr)
2578{
2579 unsigned long next;
2580 p4d_t *p4dp;
2581
2582 p4dp = p4d_offset_lockless(pgdp, pgd, addr);
2583 do {
2584 p4d_t p4d = READ_ONCE(*p4dp);
2585
2586 next = p4d_addr_end(addr, end);
2587 if (p4d_none(p4d))
2588 return 0;
2589 BUILD_BUG_ON(p4d_huge(p4d));
2590 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
2591 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
2592 P4D_SHIFT, next, flags, pages, nr))
2593 return 0;
2594 } else if (!gup_pud_range(p4dp, p4d, addr, next, flags, pages, nr))
2595 return 0;
2596 } while (p4dp++, addr = next, addr != end);
2597
2598 return 1;
2599}
2600
2601static void gup_pgd_range(unsigned long addr, unsigned long end,
2602 unsigned int flags, struct page **pages, int *nr)
2603{
2604 unsigned long next;
2605 pgd_t *pgdp;
2606
2607 pgdp = pgd_offset(current->mm, addr);
2608 do {
2609 pgd_t pgd = READ_ONCE(*pgdp);
2610
2611 next = pgd_addr_end(addr, end);
2612 if (pgd_none(pgd))
2613 return;
2614 if (unlikely(pgd_huge(pgd))) {
2615 if (!gup_huge_pgd(pgd, pgdp, addr, next, flags,
2616 pages, nr))
2617 return;
2618 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
2619 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
2620 PGDIR_SHIFT, next, flags, pages, nr))
2621 return;
2622 } else if (!gup_p4d_range(pgdp, pgd, addr, next, flags, pages, nr))
2623 return;
2624 } while (pgdp++, addr = next, addr != end);
2625}
2626#else
2627static inline void gup_pgd_range(unsigned long addr, unsigned long end,
2628 unsigned int flags, struct page **pages, int *nr)
2629{
2630}
2631#endif /* CONFIG_HAVE_FAST_GUP */
2632
2633#ifndef gup_fast_permitted
2634/*
2635 * Check if it's allowed to use get_user_pages_fast_only() for the range, or
2636 * we need to fall back to the slow version:
2637 */
2638static bool gup_fast_permitted(unsigned long start, unsigned long end)
2639{
2640 return true;
2641}
2642#endif
2643
2644static int __gup_longterm_unlocked(unsigned long start, int nr_pages,
2645 unsigned int gup_flags, struct page **pages)
2646{
2647 int ret;
2648
2649 /*
2650 * FIXME: FOLL_LONGTERM does not work with
2651 * get_user_pages_unlocked() (see comments in that function)
2652 */
2653 if (gup_flags & FOLL_LONGTERM) {
2654 mmap_read_lock(current->mm);
2655 ret = __gup_longterm_locked(current->mm,
2656 start, nr_pages,
2657 pages, NULL, gup_flags);
2658 mmap_read_unlock(current->mm);
2659 } else {
2660 ret = get_user_pages_unlocked(start, nr_pages,
2661 pages, gup_flags);
2662 }
2663
2664 return ret;
2665}
2666
2667static unsigned long lockless_pages_from_mm(unsigned long start,
2668 unsigned long end,
2669 unsigned int gup_flags,
2670 struct page **pages)
2671{
2672 unsigned long flags;
2673 int nr_pinned = 0;
2674 unsigned seq;
2675
2676 if (!IS_ENABLED(CONFIG_HAVE_FAST_GUP) ||
2677 !gup_fast_permitted(start, end))
2678 return 0;
2679
2680 if (gup_flags & FOLL_PIN) {
2681 seq = raw_read_seqcount(¤t->mm->write_protect_seq);
2682 if (seq & 1)
2683 return 0;
2684 }
2685
2686 /*
2687 * Disable interrupts. The nested form is used, in order to allow full,
2688 * general purpose use of this routine.
2689 *
2690 * With interrupts disabled, we block page table pages from being freed
2691 * from under us. See struct mmu_table_batch comments in
2692 * include/asm-generic/tlb.h for more details.
2693 *
2694 * We do not adopt an rcu_read_lock() here as we also want to block IPIs
2695 * that come from THPs splitting.
2696 */
2697 local_irq_save(flags);
2698 gup_pgd_range(start, end, gup_flags, pages, &nr_pinned);
2699 local_irq_restore(flags);
2700
2701 /*
2702 * When pinning pages for DMA there could be a concurrent write protect
2703 * from fork() via copy_page_range(), in this case always fail fast GUP.
2704 */
2705 if (gup_flags & FOLL_PIN) {
2706 if (read_seqcount_retry(¤t->mm->write_protect_seq, seq)) {
2707 unpin_user_pages(pages, nr_pinned);
2708 return 0;
2709 }
2710 }
2711 return nr_pinned;
2712}
2713
2714static int internal_get_user_pages_fast(unsigned long start,
2715 unsigned long nr_pages,
2716 unsigned int gup_flags,
2717 struct page **pages)
2718{
2719 unsigned long len, end;
2720 unsigned long nr_pinned;
2721 int ret;
2722
2723 if (WARN_ON_ONCE(gup_flags & ~(FOLL_WRITE | FOLL_LONGTERM |
2724 FOLL_FORCE | FOLL_PIN | FOLL_GET |
2725 FOLL_FAST_ONLY)))
2726 return -EINVAL;
2727
2728 if (gup_flags & FOLL_PIN)
2729 mm_set_has_pinned_flag(¤t->mm->flags);
2730
2731 if (!(gup_flags & FOLL_FAST_ONLY))
2732 might_lock_read(¤t->mm->mmap_lock);
2733
2734 start = untagged_addr(start) & PAGE_MASK;
2735 len = nr_pages << PAGE_SHIFT;
2736 if (check_add_overflow(start, len, &end))
2737 return 0;
2738 if (unlikely(!access_ok((void __user *)start, len)))
2739 return -EFAULT;
2740
2741 nr_pinned = lockless_pages_from_mm(start, end, gup_flags, pages);
2742 if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
2743 return nr_pinned;
2744
2745 /* Slow path: try to get the remaining pages with get_user_pages */
2746 start += nr_pinned << PAGE_SHIFT;
2747 pages += nr_pinned;
2748 ret = __gup_longterm_unlocked(start, nr_pages - nr_pinned, gup_flags,
2749 pages);
2750 if (ret < 0) {
2751 /*
2752 * The caller has to unpin the pages we already pinned so
2753 * returning -errno is not an option
2754 */
2755 if (nr_pinned)
2756 return nr_pinned;
2757 return ret;
2758 }
2759 return ret + nr_pinned;
2760}
2761
2762/**
2763 * get_user_pages_fast_only() - pin user pages in memory
2764 * @start: starting user address
2765 * @nr_pages: number of pages from start to pin
2766 * @gup_flags: flags modifying pin behaviour
2767 * @pages: array that receives pointers to the pages pinned.
2768 * Should be at least nr_pages long.
2769 *
2770 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
2771 * the regular GUP.
2772 * Note a difference with get_user_pages_fast: this always returns the
2773 * number of pages pinned, 0 if no pages were pinned.
2774 *
2775 * If the architecture does not support this function, simply return with no
2776 * pages pinned.
2777 *
2778 * Careful, careful! COW breaking can go either way, so a non-write
2779 * access can get ambiguous page results. If you call this function without
2780 * 'write' set, you'd better be sure that you're ok with that ambiguity.
2781 */
2782int get_user_pages_fast_only(unsigned long start, int nr_pages,
2783 unsigned int gup_flags, struct page **pages)
2784{
2785 int nr_pinned;
2786 /*
2787 * Internally (within mm/gup.c), gup fast variants must set FOLL_GET,
2788 * because gup fast is always a "pin with a +1 page refcount" request.
2789 *
2790 * FOLL_FAST_ONLY is required in order to match the API description of
2791 * this routine: no fall back to regular ("slow") GUP.
2792 */
2793 gup_flags |= FOLL_GET | FOLL_FAST_ONLY;
2794
2795 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2796 pages);
2797
2798 /*
2799 * As specified in the API description above, this routine is not
2800 * allowed to return negative values. However, the common core
2801 * routine internal_get_user_pages_fast() *can* return -errno.
2802 * Therefore, correct for that here:
2803 */
2804 if (nr_pinned < 0)
2805 nr_pinned = 0;
2806
2807 return nr_pinned;
2808}
2809EXPORT_SYMBOL_GPL(get_user_pages_fast_only);
2810
2811/**
2812 * get_user_pages_fast() - pin user pages in memory
2813 * @start: starting user address
2814 * @nr_pages: number of pages from start to pin
2815 * @gup_flags: flags modifying pin behaviour
2816 * @pages: array that receives pointers to the pages pinned.
2817 * Should be at least nr_pages long.
2818 *
2819 * Attempt to pin user pages in memory without taking mm->mmap_lock.
2820 * If not successful, it will fall back to taking the lock and
2821 * calling get_user_pages().
2822 *
2823 * Returns number of pages pinned. This may be fewer than the number requested.
2824 * If nr_pages is 0 or negative, returns 0. If no pages were pinned, returns
2825 * -errno.
2826 */
2827int get_user_pages_fast(unsigned long start, int nr_pages,
2828 unsigned int gup_flags, struct page **pages)
2829{
2830 if (!is_valid_gup_flags(gup_flags))
2831 return -EINVAL;
2832
2833 /*
2834 * The caller may or may not have explicitly set FOLL_GET; either way is
2835 * OK. However, internally (within mm/gup.c), gup fast variants must set
2836 * FOLL_GET, because gup fast is always a "pin with a +1 page refcount"
2837 * request.
2838 */
2839 gup_flags |= FOLL_GET;
2840 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2841}
2842EXPORT_SYMBOL_GPL(get_user_pages_fast);
2843
2844/**
2845 * pin_user_pages_fast() - pin user pages in memory without taking locks
2846 *
2847 * @start: starting user address
2848 * @nr_pages: number of pages from start to pin
2849 * @gup_flags: flags modifying pin behaviour
2850 * @pages: array that receives pointers to the pages pinned.
2851 * Should be at least nr_pages long.
2852 *
2853 * Nearly the same as get_user_pages_fast(), except that FOLL_PIN is set. See
2854 * get_user_pages_fast() for documentation on the function arguments, because
2855 * the arguments here are identical.
2856 *
2857 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2858 * see Documentation/core-api/pin_user_pages.rst for further details.
2859 */
2860int pin_user_pages_fast(unsigned long start, int nr_pages,
2861 unsigned int gup_flags, struct page **pages)
2862{
2863 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2864 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2865 return -EINVAL;
2866
2867 gup_flags |= FOLL_PIN;
2868 return internal_get_user_pages_fast(start, nr_pages, gup_flags, pages);
2869}
2870EXPORT_SYMBOL_GPL(pin_user_pages_fast);
2871
2872/*
2873 * This is the FOLL_PIN equivalent of get_user_pages_fast_only(). Behavior
2874 * is the same, except that this one sets FOLL_PIN instead of FOLL_GET.
2875 *
2876 * The API rules are the same, too: no negative values may be returned.
2877 */
2878int pin_user_pages_fast_only(unsigned long start, int nr_pages,
2879 unsigned int gup_flags, struct page **pages)
2880{
2881 int nr_pinned;
2882
2883 /*
2884 * FOLL_GET and FOLL_PIN are mutually exclusive. Note that the API
2885 * rules require returning 0, rather than -errno:
2886 */
2887 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2888 return 0;
2889 /*
2890 * FOLL_FAST_ONLY is required in order to match the API description of
2891 * this routine: no fall back to regular ("slow") GUP.
2892 */
2893 gup_flags |= (FOLL_PIN | FOLL_FAST_ONLY);
2894 nr_pinned = internal_get_user_pages_fast(start, nr_pages, gup_flags,
2895 pages);
2896 /*
2897 * This routine is not allowed to return negative values. However,
2898 * internal_get_user_pages_fast() *can* return -errno. Therefore,
2899 * correct for that here:
2900 */
2901 if (nr_pinned < 0)
2902 nr_pinned = 0;
2903
2904 return nr_pinned;
2905}
2906EXPORT_SYMBOL_GPL(pin_user_pages_fast_only);
2907
2908/**
2909 * pin_user_pages_remote() - pin pages of a remote process
2910 *
2911 * @mm: mm_struct of target mm
2912 * @start: starting user address
2913 * @nr_pages: number of pages from start to pin
2914 * @gup_flags: flags modifying lookup behaviour
2915 * @pages: array that receives pointers to the pages pinned.
2916 * Should be at least nr_pages long. Or NULL, if caller
2917 * only intends to ensure the pages are faulted in.
2918 * @vmas: array of pointers to vmas corresponding to each page.
2919 * Or NULL if the caller does not require them.
2920 * @locked: pointer to lock flag indicating whether lock is held and
2921 * subsequently whether VM_FAULT_RETRY functionality can be
2922 * utilised. Lock must initially be held.
2923 *
2924 * Nearly the same as get_user_pages_remote(), except that FOLL_PIN is set. See
2925 * get_user_pages_remote() for documentation on the function arguments, because
2926 * the arguments here are identical.
2927 *
2928 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2929 * see Documentation/core-api/pin_user_pages.rst for details.
2930 */
2931long pin_user_pages_remote(struct mm_struct *mm,
2932 unsigned long start, unsigned long nr_pages,
2933 unsigned int gup_flags, struct page **pages,
2934 struct vm_area_struct **vmas, int *locked)
2935{
2936 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2937 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2938 return -EINVAL;
2939
2940 gup_flags |= FOLL_PIN;
2941 return __get_user_pages_remote(mm, start, nr_pages, gup_flags,
2942 pages, vmas, locked);
2943}
2944EXPORT_SYMBOL(pin_user_pages_remote);
2945
2946/**
2947 * pin_user_pages() - pin user pages in memory for use by other devices
2948 *
2949 * @start: starting user address
2950 * @nr_pages: number of pages from start to pin
2951 * @gup_flags: flags modifying lookup behaviour
2952 * @pages: array that receives pointers to the pages pinned.
2953 * Should be at least nr_pages long. Or NULL, if caller
2954 * only intends to ensure the pages are faulted in.
2955 * @vmas: array of pointers to vmas corresponding to each page.
2956 * Or NULL if the caller does not require them.
2957 *
2958 * Nearly the same as get_user_pages(), except that FOLL_TOUCH is not set, and
2959 * FOLL_PIN is set.
2960 *
2961 * FOLL_PIN means that the pages must be released via unpin_user_page(). Please
2962 * see Documentation/core-api/pin_user_pages.rst for details.
2963 */
2964long pin_user_pages(unsigned long start, unsigned long nr_pages,
2965 unsigned int gup_flags, struct page **pages,
2966 struct vm_area_struct **vmas)
2967{
2968 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2969 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2970 return -EINVAL;
2971
2972 gup_flags |= FOLL_PIN;
2973 return __gup_longterm_locked(current->mm, start, nr_pages,
2974 pages, vmas, gup_flags);
2975}
2976EXPORT_SYMBOL(pin_user_pages);
2977
2978/*
2979 * pin_user_pages_unlocked() is the FOLL_PIN variant of
2980 * get_user_pages_unlocked(). Behavior is the same, except that this one sets
2981 * FOLL_PIN and rejects FOLL_GET.
2982 */
2983long pin_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
2984 struct page **pages, unsigned int gup_flags)
2985{
2986 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
2987 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
2988 return -EINVAL;
2989
2990 gup_flags |= FOLL_PIN;
2991 return get_user_pages_unlocked(start, nr_pages, pages, gup_flags);
2992}
2993EXPORT_SYMBOL(pin_user_pages_unlocked);
2994
2995/*
2996 * pin_user_pages_locked() is the FOLL_PIN variant of get_user_pages_locked().
2997 * Behavior is the same, except that this one sets FOLL_PIN and rejects
2998 * FOLL_GET.
2999 */
3000long pin_user_pages_locked(unsigned long start, unsigned long nr_pages,
3001 unsigned int gup_flags, struct page **pages,
3002 int *locked)
3003{
3004 /*
3005 * FIXME: Current FOLL_LONGTERM behavior is incompatible with
3006 * FAULT_FLAG_ALLOW_RETRY because of the FS DAX check requirement on
3007 * vmas. As there are no users of this flag in this call we simply
3008 * disallow this option for now.
3009 */
3010 if (WARN_ON_ONCE(gup_flags & FOLL_LONGTERM))
3011 return -EINVAL;
3012
3013 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
3014 if (WARN_ON_ONCE(gup_flags & FOLL_GET))
3015 return -EINVAL;
3016
3017 gup_flags |= FOLL_PIN;
3018 return __get_user_pages_locked(current->mm, start, nr_pages,
3019 pages, NULL, locked,
3020 gup_flags | FOLL_TOUCH);
3021}
3022EXPORT_SYMBOL(pin_user_pages_locked);