]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/gup.c
Merge tag 'upstream-4.12-rc1' of git://git.infradead.org/linux-ubifs
[mirror_ubuntu-bionic-kernel.git] / mm / gup.c
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/signal.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
23 static 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
39 static 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
63 /*
64 * FOLL_FORCE can write to even unwritable pte's, but only
65 * after we've gone through a COW cycle and they are dirty.
66 */
67 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
68 {
69 return pte_write(pte) ||
70 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
71 }
72
73 static struct page *follow_page_pte(struct vm_area_struct *vma,
74 unsigned long address, pmd_t *pmd, unsigned int flags)
75 {
76 struct mm_struct *mm = vma->vm_mm;
77 struct dev_pagemap *pgmap = NULL;
78 struct page *page;
79 spinlock_t *ptl;
80 pte_t *ptep, pte;
81
82 retry:
83 if (unlikely(pmd_bad(*pmd)))
84 return no_page_table(vma, flags);
85
86 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
87 pte = *ptep;
88 if (!pte_present(pte)) {
89 swp_entry_t entry;
90 /*
91 * KSM's break_ksm() relies upon recognizing a ksm page
92 * even while it is being migrated, so for that case we
93 * need migration_entry_wait().
94 */
95 if (likely(!(flags & FOLL_MIGRATION)))
96 goto no_page;
97 if (pte_none(pte))
98 goto no_page;
99 entry = pte_to_swp_entry(pte);
100 if (!is_migration_entry(entry))
101 goto no_page;
102 pte_unmap_unlock(ptep, ptl);
103 migration_entry_wait(mm, pmd, address);
104 goto retry;
105 }
106 if ((flags & FOLL_NUMA) && pte_protnone(pte))
107 goto no_page;
108 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
109 pte_unmap_unlock(ptep, ptl);
110 return NULL;
111 }
112
113 page = vm_normal_page(vma, address, pte);
114 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
115 /*
116 * Only return device mapping pages in the FOLL_GET case since
117 * they are only valid while holding the pgmap reference.
118 */
119 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
120 if (pgmap)
121 page = pte_page(pte);
122 else
123 goto no_page;
124 } else if (unlikely(!page)) {
125 if (flags & FOLL_DUMP) {
126 /* Avoid special (like zero) pages in core dumps */
127 page = ERR_PTR(-EFAULT);
128 goto out;
129 }
130
131 if (is_zero_pfn(pte_pfn(pte))) {
132 page = pte_page(pte);
133 } else {
134 int ret;
135
136 ret = follow_pfn_pte(vma, address, ptep, flags);
137 page = ERR_PTR(ret);
138 goto out;
139 }
140 }
141
142 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
143 int ret;
144 get_page(page);
145 pte_unmap_unlock(ptep, ptl);
146 lock_page(page);
147 ret = split_huge_page(page);
148 unlock_page(page);
149 put_page(page);
150 if (ret)
151 return ERR_PTR(ret);
152 goto retry;
153 }
154
155 if (flags & FOLL_GET) {
156 get_page(page);
157
158 /* drop the pgmap reference now that we hold the page */
159 if (pgmap) {
160 put_dev_pagemap(pgmap);
161 pgmap = NULL;
162 }
163 }
164 if (flags & FOLL_TOUCH) {
165 if ((flags & FOLL_WRITE) &&
166 !pte_dirty(pte) && !PageDirty(page))
167 set_page_dirty(page);
168 /*
169 * pte_mkyoung() would be more correct here, but atomic care
170 * is needed to avoid losing the dirty bit: it is easier to use
171 * mark_page_accessed().
172 */
173 mark_page_accessed(page);
174 }
175 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page))
178 goto out;
179
180 /*
181 * The preliminary mapping check is mainly to avoid the
182 * pointless overhead of lock_page on the ZERO_PAGE
183 * which might bounce very badly if there is contention.
184 *
185 * If the page is already locked, we don't need to
186 * handle it now - vmscan will handle it later if and
187 * when it attempts to reclaim the page.
188 */
189 if (page->mapping && trylock_page(page)) {
190 lru_add_drain(); /* push cached pages to LRU */
191 /*
192 * Because we lock page here, and migration is
193 * blocked by the pte's page reference, and we
194 * know the page is still mapped, we don't even
195 * need to check for file-cache page truncation.
196 */
197 mlock_vma_page(page);
198 unlock_page(page);
199 }
200 }
201 out:
202 pte_unmap_unlock(ptep, ptl);
203 return page;
204 no_page:
205 pte_unmap_unlock(ptep, ptl);
206 if (!pte_none(pte))
207 return NULL;
208 return no_page_table(vma, flags);
209 }
210
211 /**
212 * follow_page_mask - look up a page descriptor from a user-virtual address
213 * @vma: vm_area_struct mapping @address
214 * @address: virtual address to look up
215 * @flags: flags modifying lookup behaviour
216 * @page_mask: on output, *page_mask is set according to the size of the page
217 *
218 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
219 *
220 * Returns the mapped (struct page *), %NULL if no mapping exists, or
221 * an error pointer if there is a mapping to something not represented
222 * by a page descriptor (see also vm_normal_page()).
223 */
224 struct page *follow_page_mask(struct vm_area_struct *vma,
225 unsigned long address, unsigned int flags,
226 unsigned int *page_mask)
227 {
228 pgd_t *pgd;
229 p4d_t *p4d;
230 pud_t *pud;
231 pmd_t *pmd;
232 spinlock_t *ptl;
233 struct page *page;
234 struct mm_struct *mm = vma->vm_mm;
235
236 *page_mask = 0;
237
238 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
239 if (!IS_ERR(page)) {
240 BUG_ON(flags & FOLL_GET);
241 return page;
242 }
243
244 pgd = pgd_offset(mm, address);
245 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
246 return no_page_table(vma, flags);
247 p4d = p4d_offset(pgd, address);
248 if (p4d_none(*p4d))
249 return no_page_table(vma, flags);
250 BUILD_BUG_ON(p4d_huge(*p4d));
251 if (unlikely(p4d_bad(*p4d)))
252 return no_page_table(vma, flags);
253 pud = pud_offset(p4d, address);
254 if (pud_none(*pud))
255 return no_page_table(vma, flags);
256 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
257 page = follow_huge_pud(mm, address, pud, flags);
258 if (page)
259 return page;
260 return no_page_table(vma, flags);
261 }
262 if (pud_devmap(*pud)) {
263 ptl = pud_lock(mm, pud);
264 page = follow_devmap_pud(vma, address, pud, flags);
265 spin_unlock(ptl);
266 if (page)
267 return page;
268 }
269 if (unlikely(pud_bad(*pud)))
270 return no_page_table(vma, flags);
271
272 pmd = pmd_offset(pud, address);
273 if (pmd_none(*pmd))
274 return no_page_table(vma, flags);
275 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
276 page = follow_huge_pmd(mm, address, pmd, flags);
277 if (page)
278 return page;
279 return no_page_table(vma, flags);
280 }
281 if (pmd_devmap(*pmd)) {
282 ptl = pmd_lock(mm, pmd);
283 page = follow_devmap_pmd(vma, address, pmd, flags);
284 spin_unlock(ptl);
285 if (page)
286 return page;
287 }
288 if (likely(!pmd_trans_huge(*pmd)))
289 return follow_page_pte(vma, address, pmd, flags);
290
291 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
292 return no_page_table(vma, flags);
293
294 ptl = pmd_lock(mm, pmd);
295 if (unlikely(!pmd_trans_huge(*pmd))) {
296 spin_unlock(ptl);
297 return follow_page_pte(vma, address, pmd, flags);
298 }
299 if (flags & FOLL_SPLIT) {
300 int ret;
301 page = pmd_page(*pmd);
302 if (is_huge_zero_page(page)) {
303 spin_unlock(ptl);
304 ret = 0;
305 split_huge_pmd(vma, pmd, address);
306 if (pmd_trans_unstable(pmd))
307 ret = -EBUSY;
308 } else {
309 get_page(page);
310 spin_unlock(ptl);
311 lock_page(page);
312 ret = split_huge_page(page);
313 unlock_page(page);
314 put_page(page);
315 if (pmd_none(*pmd))
316 return no_page_table(vma, flags);
317 }
318
319 return ret ? ERR_PTR(ret) :
320 follow_page_pte(vma, address, pmd, flags);
321 }
322
323 page = follow_trans_huge_pmd(vma, address, pmd, flags);
324 spin_unlock(ptl);
325 *page_mask = HPAGE_PMD_NR - 1;
326 return page;
327 }
328
329 static int get_gate_page(struct mm_struct *mm, unsigned long address,
330 unsigned int gup_flags, struct vm_area_struct **vma,
331 struct page **page)
332 {
333 pgd_t *pgd;
334 p4d_t *p4d;
335 pud_t *pud;
336 pmd_t *pmd;
337 pte_t *pte;
338 int ret = -EFAULT;
339
340 /* user gate pages are read-only */
341 if (gup_flags & FOLL_WRITE)
342 return -EFAULT;
343 if (address > TASK_SIZE)
344 pgd = pgd_offset_k(address);
345 else
346 pgd = pgd_offset_gate(mm, address);
347 BUG_ON(pgd_none(*pgd));
348 p4d = p4d_offset(pgd, address);
349 BUG_ON(p4d_none(*p4d));
350 pud = pud_offset(p4d, address);
351 BUG_ON(pud_none(*pud));
352 pmd = pmd_offset(pud, address);
353 if (pmd_none(*pmd))
354 return -EFAULT;
355 VM_BUG_ON(pmd_trans_huge(*pmd));
356 pte = pte_offset_map(pmd, address);
357 if (pte_none(*pte))
358 goto unmap;
359 *vma = get_gate_vma(mm);
360 if (!page)
361 goto out;
362 *page = vm_normal_page(*vma, address, *pte);
363 if (!*page) {
364 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
365 goto unmap;
366 *page = pte_page(*pte);
367 }
368 get_page(*page);
369 out:
370 ret = 0;
371 unmap:
372 pte_unmap(pte);
373 return ret;
374 }
375
376 /*
377 * mmap_sem must be held on entry. If @nonblocking != NULL and
378 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
379 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
380 */
381 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
382 unsigned long address, unsigned int *flags, int *nonblocking)
383 {
384 unsigned int fault_flags = 0;
385 int ret;
386
387 /* mlock all present pages, but do not fault in new pages */
388 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
389 return -ENOENT;
390 /* For mm_populate(), just skip the stack guard page. */
391 if ((*flags & FOLL_POPULATE) &&
392 (stack_guard_page_start(vma, address) ||
393 stack_guard_page_end(vma, address + PAGE_SIZE)))
394 return -ENOENT;
395 if (*flags & FOLL_WRITE)
396 fault_flags |= FAULT_FLAG_WRITE;
397 if (*flags & FOLL_REMOTE)
398 fault_flags |= FAULT_FLAG_REMOTE;
399 if (nonblocking)
400 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
401 if (*flags & FOLL_NOWAIT)
402 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
403 if (*flags & FOLL_TRIED) {
404 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
405 fault_flags |= FAULT_FLAG_TRIED;
406 }
407
408 ret = handle_mm_fault(vma, address, fault_flags);
409 if (ret & VM_FAULT_ERROR) {
410 if (ret & VM_FAULT_OOM)
411 return -ENOMEM;
412 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
413 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
414 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
415 return -EFAULT;
416 BUG();
417 }
418
419 if (tsk) {
420 if (ret & VM_FAULT_MAJOR)
421 tsk->maj_flt++;
422 else
423 tsk->min_flt++;
424 }
425
426 if (ret & VM_FAULT_RETRY) {
427 if (nonblocking)
428 *nonblocking = 0;
429 return -EBUSY;
430 }
431
432 /*
433 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
434 * necessary, even if maybe_mkwrite decided not to set pte_write. We
435 * can thus safely do subsequent page lookups as if they were reads.
436 * But only do so when looping for pte_write is futile: in some cases
437 * userspace may also be wanting to write to the gotten user page,
438 * which a read fault here might prevent (a readonly page might get
439 * reCOWed by userspace write).
440 */
441 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
442 *flags |= FOLL_COW;
443 return 0;
444 }
445
446 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
447 {
448 vm_flags_t vm_flags = vma->vm_flags;
449 int write = (gup_flags & FOLL_WRITE);
450 int foreign = (gup_flags & FOLL_REMOTE);
451
452 if (vm_flags & (VM_IO | VM_PFNMAP))
453 return -EFAULT;
454
455 if (write) {
456 if (!(vm_flags & VM_WRITE)) {
457 if (!(gup_flags & FOLL_FORCE))
458 return -EFAULT;
459 /*
460 * We used to let the write,force case do COW in a
461 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
462 * set a breakpoint in a read-only mapping of an
463 * executable, without corrupting the file (yet only
464 * when that file had been opened for writing!).
465 * Anon pages in shared mappings are surprising: now
466 * just reject it.
467 */
468 if (!is_cow_mapping(vm_flags))
469 return -EFAULT;
470 }
471 } else if (!(vm_flags & VM_READ)) {
472 if (!(gup_flags & FOLL_FORCE))
473 return -EFAULT;
474 /*
475 * Is there actually any vma we can reach here which does not
476 * have VM_MAYREAD set?
477 */
478 if (!(vm_flags & VM_MAYREAD))
479 return -EFAULT;
480 }
481 /*
482 * gups are always data accesses, not instruction
483 * fetches, so execute=false here
484 */
485 if (!arch_vma_access_permitted(vma, write, false, foreign))
486 return -EFAULT;
487 return 0;
488 }
489
490 /**
491 * __get_user_pages() - pin user pages in memory
492 * @tsk: task_struct of target task
493 * @mm: mm_struct of target mm
494 * @start: starting user address
495 * @nr_pages: number of pages from start to pin
496 * @gup_flags: flags modifying pin behaviour
497 * @pages: array that receives pointers to the pages pinned.
498 * Should be at least nr_pages long. Or NULL, if caller
499 * only intends to ensure the pages are faulted in.
500 * @vmas: array of pointers to vmas corresponding to each page.
501 * Or NULL if the caller does not require them.
502 * @nonblocking: whether waiting for disk IO or mmap_sem contention
503 *
504 * Returns number of pages pinned. This may be fewer than the number
505 * requested. If nr_pages is 0 or negative, returns 0. If no pages
506 * were pinned, returns -errno. Each page returned must be released
507 * with a put_page() call when it is finished with. vmas will only
508 * remain valid while mmap_sem is held.
509 *
510 * Must be called with mmap_sem held. It may be released. See below.
511 *
512 * __get_user_pages walks a process's page tables and takes a reference to
513 * each struct page that each user address corresponds to at a given
514 * instant. That is, it takes the page that would be accessed if a user
515 * thread accesses the given user virtual address at that instant.
516 *
517 * This does not guarantee that the page exists in the user mappings when
518 * __get_user_pages returns, and there may even be a completely different
519 * page there in some cases (eg. if mmapped pagecache has been invalidated
520 * and subsequently re faulted). However it does guarantee that the page
521 * won't be freed completely. And mostly callers simply care that the page
522 * contains data that was valid *at some point in time*. Typically, an IO
523 * or similar operation cannot guarantee anything stronger anyway because
524 * locks can't be held over the syscall boundary.
525 *
526 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
527 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
528 * appropriate) must be called after the page is finished with, and
529 * before put_page is called.
530 *
531 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
532 * or mmap_sem contention, and if waiting is needed to pin all pages,
533 * *@nonblocking will be set to 0. Further, if @gup_flags does not
534 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
535 * this case.
536 *
537 * A caller using such a combination of @nonblocking and @gup_flags
538 * must therefore hold the mmap_sem for reading only, and recognize
539 * when it's been released. Otherwise, it must be held for either
540 * reading or writing and will not be released.
541 *
542 * In most cases, get_user_pages or get_user_pages_fast should be used
543 * instead of __get_user_pages. __get_user_pages should be used only if
544 * you need some special @gup_flags.
545 */
546 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
547 unsigned long start, unsigned long nr_pages,
548 unsigned int gup_flags, struct page **pages,
549 struct vm_area_struct **vmas, int *nonblocking)
550 {
551 long i = 0;
552 unsigned int page_mask;
553 struct vm_area_struct *vma = NULL;
554
555 if (!nr_pages)
556 return 0;
557
558 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
559
560 /*
561 * If FOLL_FORCE is set then do not force a full fault as the hinting
562 * fault information is unrelated to the reference behaviour of a task
563 * using the address space
564 */
565 if (!(gup_flags & FOLL_FORCE))
566 gup_flags |= FOLL_NUMA;
567
568 do {
569 struct page *page;
570 unsigned int foll_flags = gup_flags;
571 unsigned int page_increm;
572
573 /* first iteration or cross vma bound */
574 if (!vma || start >= vma->vm_end) {
575 vma = find_extend_vma(mm, start);
576 if (!vma && in_gate_area(mm, start)) {
577 int ret;
578 ret = get_gate_page(mm, start & PAGE_MASK,
579 gup_flags, &vma,
580 pages ? &pages[i] : NULL);
581 if (ret)
582 return i ? : ret;
583 page_mask = 0;
584 goto next_page;
585 }
586
587 if (!vma || check_vma_flags(vma, gup_flags))
588 return i ? : -EFAULT;
589 if (is_vm_hugetlb_page(vma)) {
590 i = follow_hugetlb_page(mm, vma, pages, vmas,
591 &start, &nr_pages, i,
592 gup_flags, nonblocking);
593 continue;
594 }
595 }
596 retry:
597 /*
598 * If we have a pending SIGKILL, don't keep faulting pages and
599 * potentially allocating memory.
600 */
601 if (unlikely(fatal_signal_pending(current)))
602 return i ? i : -ERESTARTSYS;
603 cond_resched();
604 page = follow_page_mask(vma, start, foll_flags, &page_mask);
605 if (!page) {
606 int ret;
607 ret = faultin_page(tsk, vma, start, &foll_flags,
608 nonblocking);
609 switch (ret) {
610 case 0:
611 goto retry;
612 case -EFAULT:
613 case -ENOMEM:
614 case -EHWPOISON:
615 return i ? i : ret;
616 case -EBUSY:
617 return i;
618 case -ENOENT:
619 goto next_page;
620 }
621 BUG();
622 } else if (PTR_ERR(page) == -EEXIST) {
623 /*
624 * Proper page table entry exists, but no corresponding
625 * struct page.
626 */
627 goto next_page;
628 } else if (IS_ERR(page)) {
629 return i ? i : PTR_ERR(page);
630 }
631 if (pages) {
632 pages[i] = page;
633 flush_anon_page(vma, page, start);
634 flush_dcache_page(page);
635 page_mask = 0;
636 }
637 next_page:
638 if (vmas) {
639 vmas[i] = vma;
640 page_mask = 0;
641 }
642 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
643 if (page_increm > nr_pages)
644 page_increm = nr_pages;
645 i += page_increm;
646 start += page_increm * PAGE_SIZE;
647 nr_pages -= page_increm;
648 } while (nr_pages);
649 return i;
650 }
651
652 static bool vma_permits_fault(struct vm_area_struct *vma,
653 unsigned int fault_flags)
654 {
655 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
656 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
657 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
658
659 if (!(vm_flags & vma->vm_flags))
660 return false;
661
662 /*
663 * The architecture might have a hardware protection
664 * mechanism other than read/write that can deny access.
665 *
666 * gup always represents data access, not instruction
667 * fetches, so execute=false here:
668 */
669 if (!arch_vma_access_permitted(vma, write, false, foreign))
670 return false;
671
672 return true;
673 }
674
675 /*
676 * fixup_user_fault() - manually resolve a user page fault
677 * @tsk: the task_struct to use for page fault accounting, or
678 * NULL if faults are not to be recorded.
679 * @mm: mm_struct of target mm
680 * @address: user address
681 * @fault_flags:flags to pass down to handle_mm_fault()
682 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
683 * does not allow retry
684 *
685 * This is meant to be called in the specific scenario where for locking reasons
686 * we try to access user memory in atomic context (within a pagefault_disable()
687 * section), this returns -EFAULT, and we want to resolve the user fault before
688 * trying again.
689 *
690 * Typically this is meant to be used by the futex code.
691 *
692 * The main difference with get_user_pages() is that this function will
693 * unconditionally call handle_mm_fault() which will in turn perform all the
694 * necessary SW fixup of the dirty and young bits in the PTE, while
695 * get_user_pages() only guarantees to update these in the struct page.
696 *
697 * This is important for some architectures where those bits also gate the
698 * access permission to the page because they are maintained in software. On
699 * such architectures, gup() will not be enough to make a subsequent access
700 * succeed.
701 *
702 * This function will not return with an unlocked mmap_sem. So it has not the
703 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
704 */
705 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
706 unsigned long address, unsigned int fault_flags,
707 bool *unlocked)
708 {
709 struct vm_area_struct *vma;
710 int ret, major = 0;
711
712 if (unlocked)
713 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
714
715 retry:
716 vma = find_extend_vma(mm, address);
717 if (!vma || address < vma->vm_start)
718 return -EFAULT;
719
720 if (!vma_permits_fault(vma, fault_flags))
721 return -EFAULT;
722
723 ret = handle_mm_fault(vma, address, fault_flags);
724 major |= ret & VM_FAULT_MAJOR;
725 if (ret & VM_FAULT_ERROR) {
726 if (ret & VM_FAULT_OOM)
727 return -ENOMEM;
728 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
729 return -EHWPOISON;
730 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
731 return -EFAULT;
732 BUG();
733 }
734
735 if (ret & VM_FAULT_RETRY) {
736 down_read(&mm->mmap_sem);
737 if (!(fault_flags & FAULT_FLAG_TRIED)) {
738 *unlocked = true;
739 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
740 fault_flags |= FAULT_FLAG_TRIED;
741 goto retry;
742 }
743 }
744
745 if (tsk) {
746 if (major)
747 tsk->maj_flt++;
748 else
749 tsk->min_flt++;
750 }
751 return 0;
752 }
753 EXPORT_SYMBOL_GPL(fixup_user_fault);
754
755 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
756 struct mm_struct *mm,
757 unsigned long start,
758 unsigned long nr_pages,
759 struct page **pages,
760 struct vm_area_struct **vmas,
761 int *locked, bool notify_drop,
762 unsigned int flags)
763 {
764 long ret, pages_done;
765 bool lock_dropped;
766
767 if (locked) {
768 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
769 BUG_ON(vmas);
770 /* check caller initialized locked */
771 BUG_ON(*locked != 1);
772 }
773
774 if (pages)
775 flags |= FOLL_GET;
776
777 pages_done = 0;
778 lock_dropped = false;
779 for (;;) {
780 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
781 vmas, locked);
782 if (!locked)
783 /* VM_FAULT_RETRY couldn't trigger, bypass */
784 return ret;
785
786 /* VM_FAULT_RETRY cannot return errors */
787 if (!*locked) {
788 BUG_ON(ret < 0);
789 BUG_ON(ret >= nr_pages);
790 }
791
792 if (!pages)
793 /* If it's a prefault don't insist harder */
794 return ret;
795
796 if (ret > 0) {
797 nr_pages -= ret;
798 pages_done += ret;
799 if (!nr_pages)
800 break;
801 }
802 if (*locked) {
803 /* VM_FAULT_RETRY didn't trigger */
804 if (!pages_done)
805 pages_done = ret;
806 break;
807 }
808 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
809 pages += ret;
810 start += ret << PAGE_SHIFT;
811
812 /*
813 * Repeat on the address that fired VM_FAULT_RETRY
814 * without FAULT_FLAG_ALLOW_RETRY but with
815 * FAULT_FLAG_TRIED.
816 */
817 *locked = 1;
818 lock_dropped = true;
819 down_read(&mm->mmap_sem);
820 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
821 pages, NULL, NULL);
822 if (ret != 1) {
823 BUG_ON(ret > 1);
824 if (!pages_done)
825 pages_done = ret;
826 break;
827 }
828 nr_pages--;
829 pages_done++;
830 if (!nr_pages)
831 break;
832 pages++;
833 start += PAGE_SIZE;
834 }
835 if (notify_drop && lock_dropped && *locked) {
836 /*
837 * We must let the caller know we temporarily dropped the lock
838 * and so the critical section protected by it was lost.
839 */
840 up_read(&mm->mmap_sem);
841 *locked = 0;
842 }
843 return pages_done;
844 }
845
846 /*
847 * We can leverage the VM_FAULT_RETRY functionality in the page fault
848 * paths better by using either get_user_pages_locked() or
849 * get_user_pages_unlocked().
850 *
851 * get_user_pages_locked() is suitable to replace the form:
852 *
853 * down_read(&mm->mmap_sem);
854 * do_something()
855 * get_user_pages(tsk, mm, ..., pages, NULL);
856 * up_read(&mm->mmap_sem);
857 *
858 * to:
859 *
860 * int locked = 1;
861 * down_read(&mm->mmap_sem);
862 * do_something()
863 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
864 * if (locked)
865 * up_read(&mm->mmap_sem);
866 */
867 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
868 unsigned int gup_flags, struct page **pages,
869 int *locked)
870 {
871 return __get_user_pages_locked(current, current->mm, start, nr_pages,
872 pages, NULL, locked, true,
873 gup_flags | FOLL_TOUCH);
874 }
875 EXPORT_SYMBOL(get_user_pages_locked);
876
877 /*
878 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
879 * tsk, mm to be specified.
880 *
881 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
882 * caller if required (just like with __get_user_pages). "FOLL_GET"
883 * is set implicitly if "pages" is non-NULL.
884 */
885 static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
886 struct mm_struct *mm, unsigned long start,
887 unsigned long nr_pages, struct page **pages,
888 unsigned int gup_flags)
889 {
890 long ret;
891 int locked = 1;
892
893 down_read(&mm->mmap_sem);
894 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
895 &locked, false, gup_flags);
896 if (locked)
897 up_read(&mm->mmap_sem);
898 return ret;
899 }
900
901 /*
902 * get_user_pages_unlocked() is suitable to replace the form:
903 *
904 * down_read(&mm->mmap_sem);
905 * get_user_pages(tsk, mm, ..., pages, NULL);
906 * up_read(&mm->mmap_sem);
907 *
908 * with:
909 *
910 * get_user_pages_unlocked(tsk, mm, ..., pages);
911 *
912 * It is functionally equivalent to get_user_pages_fast so
913 * get_user_pages_fast should be used instead if specific gup_flags
914 * (e.g. FOLL_FORCE) are not required.
915 */
916 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
917 struct page **pages, unsigned int gup_flags)
918 {
919 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
920 pages, gup_flags | FOLL_TOUCH);
921 }
922 EXPORT_SYMBOL(get_user_pages_unlocked);
923
924 /*
925 * get_user_pages_remote() - pin user pages in memory
926 * @tsk: the task_struct to use for page fault accounting, or
927 * NULL if faults are not to be recorded.
928 * @mm: mm_struct of target mm
929 * @start: starting user address
930 * @nr_pages: number of pages from start to pin
931 * @gup_flags: flags modifying lookup behaviour
932 * @pages: array that receives pointers to the pages pinned.
933 * Should be at least nr_pages long. Or NULL, if caller
934 * only intends to ensure the pages are faulted in.
935 * @vmas: array of pointers to vmas corresponding to each page.
936 * Or NULL if the caller does not require them.
937 * @locked: pointer to lock flag indicating whether lock is held and
938 * subsequently whether VM_FAULT_RETRY functionality can be
939 * utilised. Lock must initially be held.
940 *
941 * Returns number of pages pinned. This may be fewer than the number
942 * requested. If nr_pages is 0 or negative, returns 0. If no pages
943 * were pinned, returns -errno. Each page returned must be released
944 * with a put_page() call when it is finished with. vmas will only
945 * remain valid while mmap_sem is held.
946 *
947 * Must be called with mmap_sem held for read or write.
948 *
949 * get_user_pages walks a process's page tables and takes a reference to
950 * each struct page that each user address corresponds to at a given
951 * instant. That is, it takes the page that would be accessed if a user
952 * thread accesses the given user virtual address at that instant.
953 *
954 * This does not guarantee that the page exists in the user mappings when
955 * get_user_pages returns, and there may even be a completely different
956 * page there in some cases (eg. if mmapped pagecache has been invalidated
957 * and subsequently re faulted). However it does guarantee that the page
958 * won't be freed completely. And mostly callers simply care that the page
959 * contains data that was valid *at some point in time*. Typically, an IO
960 * or similar operation cannot guarantee anything stronger anyway because
961 * locks can't be held over the syscall boundary.
962 *
963 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
964 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
965 * be called after the page is finished with, and before put_page is called.
966 *
967 * get_user_pages is typically used for fewer-copy IO operations, to get a
968 * handle on the memory by some means other than accesses via the user virtual
969 * addresses. The pages may be submitted for DMA to devices or accessed via
970 * their kernel linear mapping (via the kmap APIs). Care should be taken to
971 * use the correct cache flushing APIs.
972 *
973 * See also get_user_pages_fast, for performance critical applications.
974 *
975 * get_user_pages should be phased out in favor of
976 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
977 * should use get_user_pages because it cannot pass
978 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
979 */
980 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
981 unsigned long start, unsigned long nr_pages,
982 unsigned int gup_flags, struct page **pages,
983 struct vm_area_struct **vmas, int *locked)
984 {
985 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
986 locked, true,
987 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
988 }
989 EXPORT_SYMBOL(get_user_pages_remote);
990
991 /*
992 * This is the same as get_user_pages_remote(), just with a
993 * less-flexible calling convention where we assume that the task
994 * and mm being operated on are the current task's and don't allow
995 * passing of a locked parameter. We also obviously don't pass
996 * FOLL_REMOTE in here.
997 */
998 long get_user_pages(unsigned long start, unsigned long nr_pages,
999 unsigned int gup_flags, struct page **pages,
1000 struct vm_area_struct **vmas)
1001 {
1002 return __get_user_pages_locked(current, current->mm, start, nr_pages,
1003 pages, vmas, NULL, false,
1004 gup_flags | FOLL_TOUCH);
1005 }
1006 EXPORT_SYMBOL(get_user_pages);
1007
1008 /**
1009 * populate_vma_page_range() - populate a range of pages in the vma.
1010 * @vma: target vma
1011 * @start: start address
1012 * @end: end address
1013 * @nonblocking:
1014 *
1015 * This takes care of mlocking the pages too if VM_LOCKED is set.
1016 *
1017 * return 0 on success, negative error code on error.
1018 *
1019 * vma->vm_mm->mmap_sem must be held.
1020 *
1021 * If @nonblocking is NULL, it may be held for read or write and will
1022 * be unperturbed.
1023 *
1024 * If @nonblocking is non-NULL, it must held for read only and may be
1025 * released. If it's released, *@nonblocking will be set to 0.
1026 */
1027 long populate_vma_page_range(struct vm_area_struct *vma,
1028 unsigned long start, unsigned long end, int *nonblocking)
1029 {
1030 struct mm_struct *mm = vma->vm_mm;
1031 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1032 int gup_flags;
1033
1034 VM_BUG_ON(start & ~PAGE_MASK);
1035 VM_BUG_ON(end & ~PAGE_MASK);
1036 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1037 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1038 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1039
1040 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1041 if (vma->vm_flags & VM_LOCKONFAULT)
1042 gup_flags &= ~FOLL_POPULATE;
1043 /*
1044 * We want to touch writable mappings with a write fault in order
1045 * to break COW, except for shared mappings because these don't COW
1046 * and we would not want to dirty them for nothing.
1047 */
1048 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1049 gup_flags |= FOLL_WRITE;
1050
1051 /*
1052 * We want mlock to succeed for regions that have any permissions
1053 * other than PROT_NONE.
1054 */
1055 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1056 gup_flags |= FOLL_FORCE;
1057
1058 /*
1059 * We made sure addr is within a VMA, so the following will
1060 * not result in a stack expansion that recurses back here.
1061 */
1062 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1063 NULL, NULL, nonblocking);
1064 }
1065
1066 /*
1067 * __mm_populate - populate and/or mlock pages within a range of address space.
1068 *
1069 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1070 * flags. VMAs must be already marked with the desired vm_flags, and
1071 * mmap_sem must not be held.
1072 */
1073 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1074 {
1075 struct mm_struct *mm = current->mm;
1076 unsigned long end, nstart, nend;
1077 struct vm_area_struct *vma = NULL;
1078 int locked = 0;
1079 long ret = 0;
1080
1081 VM_BUG_ON(start & ~PAGE_MASK);
1082 VM_BUG_ON(len != PAGE_ALIGN(len));
1083 end = start + len;
1084
1085 for (nstart = start; nstart < end; nstart = nend) {
1086 /*
1087 * We want to fault in pages for [nstart; end) address range.
1088 * Find first corresponding VMA.
1089 */
1090 if (!locked) {
1091 locked = 1;
1092 down_read(&mm->mmap_sem);
1093 vma = find_vma(mm, nstart);
1094 } else if (nstart >= vma->vm_end)
1095 vma = vma->vm_next;
1096 if (!vma || vma->vm_start >= end)
1097 break;
1098 /*
1099 * Set [nstart; nend) to intersection of desired address
1100 * range with the first VMA. Also, skip undesirable VMA types.
1101 */
1102 nend = min(end, vma->vm_end);
1103 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1104 continue;
1105 if (nstart < vma->vm_start)
1106 nstart = vma->vm_start;
1107 /*
1108 * Now fault in a range of pages. populate_vma_page_range()
1109 * double checks the vma flags, so that it won't mlock pages
1110 * if the vma was already munlocked.
1111 */
1112 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1113 if (ret < 0) {
1114 if (ignore_errors) {
1115 ret = 0;
1116 continue; /* continue at next VMA */
1117 }
1118 break;
1119 }
1120 nend = nstart + ret * PAGE_SIZE;
1121 ret = 0;
1122 }
1123 if (locked)
1124 up_read(&mm->mmap_sem);
1125 return ret; /* 0 or negative error code */
1126 }
1127
1128 /**
1129 * get_dump_page() - pin user page in memory while writing it to core dump
1130 * @addr: user address
1131 *
1132 * Returns struct page pointer of user page pinned for dump,
1133 * to be freed afterwards by put_page().
1134 *
1135 * Returns NULL on any kind of failure - a hole must then be inserted into
1136 * the corefile, to preserve alignment with its headers; and also returns
1137 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1138 * allowing a hole to be left in the corefile to save diskspace.
1139 *
1140 * Called without mmap_sem, but after all other threads have been killed.
1141 */
1142 #ifdef CONFIG_ELF_CORE
1143 struct page *get_dump_page(unsigned long addr)
1144 {
1145 struct vm_area_struct *vma;
1146 struct page *page;
1147
1148 if (__get_user_pages(current, current->mm, addr, 1,
1149 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1150 NULL) < 1)
1151 return NULL;
1152 flush_cache_page(vma, addr, page_to_pfn(page));
1153 return page;
1154 }
1155 #endif /* CONFIG_ELF_CORE */
1156
1157 /*
1158 * Generic RCU Fast GUP
1159 *
1160 * get_user_pages_fast attempts to pin user pages by walking the page
1161 * tables directly and avoids taking locks. Thus the walker needs to be
1162 * protected from page table pages being freed from under it, and should
1163 * block any THP splits.
1164 *
1165 * One way to achieve this is to have the walker disable interrupts, and
1166 * rely on IPIs from the TLB flushing code blocking before the page table
1167 * pages are freed. This is unsuitable for architectures that do not need
1168 * to broadcast an IPI when invalidating TLBs.
1169 *
1170 * Another way to achieve this is to batch up page table containing pages
1171 * belonging to more than one mm_user, then rcu_sched a callback to free those
1172 * pages. Disabling interrupts will allow the fast_gup walker to both block
1173 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1174 * (which is a relatively rare event). The code below adopts this strategy.
1175 *
1176 * Before activating this code, please be aware that the following assumptions
1177 * are currently made:
1178 *
1179 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1180 * pages containing page tables.
1181 *
1182 * *) ptes can be read atomically by the architecture.
1183 *
1184 * *) access_ok is sufficient to validate userspace address ranges.
1185 *
1186 * The last two assumptions can be relaxed by the addition of helper functions.
1187 *
1188 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1189 */
1190 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1191
1192 #ifndef gup_get_pte
1193 /*
1194 * We assume that the PTE can be read atomically. If this is not the case for
1195 * your architecture, please provide the helper.
1196 */
1197 static inline pte_t gup_get_pte(pte_t *ptep)
1198 {
1199 return READ_ONCE(*ptep);
1200 }
1201 #endif
1202
1203 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1204 {
1205 while ((*nr) - nr_start) {
1206 struct page *page = pages[--(*nr)];
1207
1208 ClearPageReferenced(page);
1209 put_page(page);
1210 }
1211 }
1212
1213 #ifdef __HAVE_ARCH_PTE_SPECIAL
1214 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1215 int write, struct page **pages, int *nr)
1216 {
1217 struct dev_pagemap *pgmap = NULL;
1218 int nr_start = *nr, ret = 0;
1219 pte_t *ptep, *ptem;
1220
1221 ptem = ptep = pte_offset_map(&pmd, addr);
1222 do {
1223 pte_t pte = gup_get_pte(ptep);
1224 struct page *head, *page;
1225
1226 /*
1227 * Similar to the PMD case below, NUMA hinting must take slow
1228 * path using the pte_protnone check.
1229 */
1230 if (pte_protnone(pte))
1231 goto pte_unmap;
1232
1233 if (!pte_access_permitted(pte, write))
1234 goto pte_unmap;
1235
1236 if (pte_devmap(pte)) {
1237 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1238 if (unlikely(!pgmap)) {
1239 undo_dev_pagemap(nr, nr_start, pages);
1240 goto pte_unmap;
1241 }
1242 } else if (pte_special(pte))
1243 goto pte_unmap;
1244
1245 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1246 page = pte_page(pte);
1247 head = compound_head(page);
1248
1249 if (!page_cache_get_speculative(head))
1250 goto pte_unmap;
1251
1252 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1253 put_page(head);
1254 goto pte_unmap;
1255 }
1256
1257 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1258
1259 put_dev_pagemap(pgmap);
1260 SetPageReferenced(page);
1261 pages[*nr] = page;
1262 (*nr)++;
1263
1264 } while (ptep++, addr += PAGE_SIZE, addr != end);
1265
1266 ret = 1;
1267
1268 pte_unmap:
1269 pte_unmap(ptem);
1270 return ret;
1271 }
1272 #else
1273
1274 /*
1275 * If we can't determine whether or not a pte is special, then fail immediately
1276 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1277 * to be special.
1278 *
1279 * For a futex to be placed on a THP tail page, get_futex_key requires a
1280 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1281 * useful to have gup_huge_pmd even if we can't operate on ptes.
1282 */
1283 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1284 int write, struct page **pages, int *nr)
1285 {
1286 return 0;
1287 }
1288 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1289
1290 #ifdef __HAVE_ARCH_PTE_DEVMAP
1291 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1292 unsigned long end, struct page **pages, int *nr)
1293 {
1294 int nr_start = *nr;
1295 struct dev_pagemap *pgmap = NULL;
1296
1297 do {
1298 struct page *page = pfn_to_page(pfn);
1299
1300 pgmap = get_dev_pagemap(pfn, pgmap);
1301 if (unlikely(!pgmap)) {
1302 undo_dev_pagemap(nr, nr_start, pages);
1303 return 0;
1304 }
1305 SetPageReferenced(page);
1306 pages[*nr] = page;
1307 get_page(page);
1308 put_dev_pagemap(pgmap);
1309 (*nr)++;
1310 pfn++;
1311 } while (addr += PAGE_SIZE, addr != end);
1312 return 1;
1313 }
1314
1315 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1316 unsigned long end, struct page **pages, int *nr)
1317 {
1318 unsigned long fault_pfn;
1319
1320 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1321 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1322 }
1323
1324 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1325 unsigned long end, struct page **pages, int *nr)
1326 {
1327 unsigned long fault_pfn;
1328
1329 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1330 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1331 }
1332 #else
1333 static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1334 unsigned long end, struct page **pages, int *nr)
1335 {
1336 BUILD_BUG();
1337 return 0;
1338 }
1339
1340 static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1341 unsigned long end, struct page **pages, int *nr)
1342 {
1343 BUILD_BUG();
1344 return 0;
1345 }
1346 #endif
1347
1348 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1349 unsigned long end, int write, struct page **pages, int *nr)
1350 {
1351 struct page *head, *page;
1352 int refs;
1353
1354 if (!pmd_access_permitted(orig, write))
1355 return 0;
1356
1357 if (pmd_devmap(orig))
1358 return __gup_device_huge_pmd(orig, addr, end, pages, nr);
1359
1360 refs = 0;
1361 head = pmd_page(orig);
1362 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1363 do {
1364 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1365 pages[*nr] = page;
1366 (*nr)++;
1367 page++;
1368 refs++;
1369 } while (addr += PAGE_SIZE, addr != end);
1370
1371 if (!page_cache_add_speculative(head, refs)) {
1372 *nr -= refs;
1373 return 0;
1374 }
1375
1376 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1377 *nr -= refs;
1378 while (refs--)
1379 put_page(head);
1380 return 0;
1381 }
1382
1383 SetPageReferenced(head);
1384 return 1;
1385 }
1386
1387 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1388 unsigned long end, int write, struct page **pages, int *nr)
1389 {
1390 struct page *head, *page;
1391 int refs;
1392
1393 if (!pud_access_permitted(orig, write))
1394 return 0;
1395
1396 if (pud_devmap(orig))
1397 return __gup_device_huge_pud(orig, addr, end, pages, nr);
1398
1399 refs = 0;
1400 head = pud_page(orig);
1401 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1402 do {
1403 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1404 pages[*nr] = page;
1405 (*nr)++;
1406 page++;
1407 refs++;
1408 } while (addr += PAGE_SIZE, addr != end);
1409
1410 if (!page_cache_add_speculative(head, refs)) {
1411 *nr -= refs;
1412 return 0;
1413 }
1414
1415 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1416 *nr -= refs;
1417 while (refs--)
1418 put_page(head);
1419 return 0;
1420 }
1421
1422 SetPageReferenced(head);
1423 return 1;
1424 }
1425
1426 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1427 unsigned long end, int write,
1428 struct page **pages, int *nr)
1429 {
1430 int refs;
1431 struct page *head, *page;
1432
1433 if (!pgd_access_permitted(orig, write))
1434 return 0;
1435
1436 BUILD_BUG_ON(pgd_devmap(orig));
1437 refs = 0;
1438 head = pgd_page(orig);
1439 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1440 do {
1441 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1442 pages[*nr] = page;
1443 (*nr)++;
1444 page++;
1445 refs++;
1446 } while (addr += PAGE_SIZE, addr != end);
1447
1448 if (!page_cache_add_speculative(head, refs)) {
1449 *nr -= refs;
1450 return 0;
1451 }
1452
1453 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1454 *nr -= refs;
1455 while (refs--)
1456 put_page(head);
1457 return 0;
1458 }
1459
1460 SetPageReferenced(head);
1461 return 1;
1462 }
1463
1464 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1465 int write, struct page **pages, int *nr)
1466 {
1467 unsigned long next;
1468 pmd_t *pmdp;
1469
1470 pmdp = pmd_offset(&pud, addr);
1471 do {
1472 pmd_t pmd = READ_ONCE(*pmdp);
1473
1474 next = pmd_addr_end(addr, end);
1475 if (pmd_none(pmd))
1476 return 0;
1477
1478 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1479 /*
1480 * NUMA hinting faults need to be handled in the GUP
1481 * slowpath for accounting purposes and so that they
1482 * can be serialised against THP migration.
1483 */
1484 if (pmd_protnone(pmd))
1485 return 0;
1486
1487 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1488 pages, nr))
1489 return 0;
1490
1491 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1492 /*
1493 * architecture have different format for hugetlbfs
1494 * pmd format and THP pmd format
1495 */
1496 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1497 PMD_SHIFT, next, write, pages, nr))
1498 return 0;
1499 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1500 return 0;
1501 } while (pmdp++, addr = next, addr != end);
1502
1503 return 1;
1504 }
1505
1506 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1507 int write, struct page **pages, int *nr)
1508 {
1509 unsigned long next;
1510 pud_t *pudp;
1511
1512 pudp = pud_offset(&p4d, addr);
1513 do {
1514 pud_t pud = READ_ONCE(*pudp);
1515
1516 next = pud_addr_end(addr, end);
1517 if (pud_none(pud))
1518 return 0;
1519 if (unlikely(pud_huge(pud))) {
1520 if (!gup_huge_pud(pud, pudp, addr, next, write,
1521 pages, nr))
1522 return 0;
1523 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1524 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1525 PUD_SHIFT, next, write, pages, nr))
1526 return 0;
1527 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1528 return 0;
1529 } while (pudp++, addr = next, addr != end);
1530
1531 return 1;
1532 }
1533
1534 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1535 int write, struct page **pages, int *nr)
1536 {
1537 unsigned long next;
1538 p4d_t *p4dp;
1539
1540 p4dp = p4d_offset(&pgd, addr);
1541 do {
1542 p4d_t p4d = READ_ONCE(*p4dp);
1543
1544 next = p4d_addr_end(addr, end);
1545 if (p4d_none(p4d))
1546 return 0;
1547 BUILD_BUG_ON(p4d_huge(p4d));
1548 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1549 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1550 P4D_SHIFT, next, write, pages, nr))
1551 return 0;
1552 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1553 return 0;
1554 } while (p4dp++, addr = next, addr != end);
1555
1556 return 1;
1557 }
1558
1559 /*
1560 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1561 * the regular GUP. It will only return non-negative values.
1562 */
1563 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1564 struct page **pages)
1565 {
1566 struct mm_struct *mm = current->mm;
1567 unsigned long addr, len, end;
1568 unsigned long next, flags;
1569 pgd_t *pgdp;
1570 int nr = 0;
1571
1572 start &= PAGE_MASK;
1573 addr = start;
1574 len = (unsigned long) nr_pages << PAGE_SHIFT;
1575 end = start + len;
1576
1577 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1578 (void __user *)start, len)))
1579 return 0;
1580
1581 /*
1582 * Disable interrupts. We use the nested form as we can already have
1583 * interrupts disabled by get_futex_key.
1584 *
1585 * With interrupts disabled, we block page table pages from being
1586 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1587 * for more details.
1588 *
1589 * We do not adopt an rcu_read_lock(.) here as we also want to
1590 * block IPIs that come from THPs splitting.
1591 */
1592
1593 local_irq_save(flags);
1594 pgdp = pgd_offset(mm, addr);
1595 do {
1596 pgd_t pgd = READ_ONCE(*pgdp);
1597
1598 next = pgd_addr_end(addr, end);
1599 if (pgd_none(pgd))
1600 break;
1601 if (unlikely(pgd_huge(pgd))) {
1602 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1603 pages, &nr))
1604 break;
1605 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1606 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1607 PGDIR_SHIFT, next, write, pages, &nr))
1608 break;
1609 } else if (!gup_p4d_range(pgd, addr, next, write, pages, &nr))
1610 break;
1611 } while (pgdp++, addr = next, addr != end);
1612 local_irq_restore(flags);
1613
1614 return nr;
1615 }
1616
1617 #ifndef gup_fast_permitted
1618 /*
1619 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1620 * we need to fall back to the slow version:
1621 */
1622 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1623 {
1624 unsigned long len, end;
1625
1626 len = (unsigned long) nr_pages << PAGE_SHIFT;
1627 end = start + len;
1628 return end >= start;
1629 }
1630 #endif
1631
1632 /**
1633 * get_user_pages_fast() - pin user pages in memory
1634 * @start: starting user address
1635 * @nr_pages: number of pages from start to pin
1636 * @write: whether pages will be written to
1637 * @pages: array that receives pointers to the pages pinned.
1638 * Should be at least nr_pages long.
1639 *
1640 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1641 * If not successful, it will fall back to taking the lock and
1642 * calling get_user_pages().
1643 *
1644 * Returns number of pages pinned. This may be fewer than the number
1645 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1646 * were pinned, returns -errno.
1647 */
1648 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1649 struct page **pages)
1650 {
1651 int nr = 0, ret = 0;
1652
1653 start &= PAGE_MASK;
1654
1655 if (gup_fast_permitted(start, nr_pages, write)) {
1656 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1657 ret = nr;
1658 }
1659
1660 if (nr < nr_pages) {
1661 /* Try to get the remaining pages with get_user_pages */
1662 start += nr << PAGE_SHIFT;
1663 pages += nr;
1664
1665 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1666 write ? FOLL_WRITE : 0);
1667
1668 /* Have to be a bit careful with return values */
1669 if (nr > 0) {
1670 if (ret < 0)
1671 ret = nr;
1672 else
1673 ret += nr;
1674 }
1675 }
1676
1677 return ret;
1678 }
1679
1680 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */