blob: b3c7214d710d5ea8bab8648b5182c53d882f3c31 [file] [log] [blame]
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001#include <linux/kernel.h>
2#include <linux/errno.h>
3#include <linux/err.h>
4#include <linux/spinlock.h>
5
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07006#include <linux/mm.h>
Dan Williams3565fce2016-01-15 16:56:55 -08007#include <linux/memremap.h>
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07008#include <linux/pagemap.h>
9#include <linux/rmap.h>
10#include <linux/swap.h>
11#include <linux/swapops.h>
12
Ingo Molnar174cd4b2017-02-02 19:15:33 +010013#include <linux/sched/signal.h>
Steve Capper2667f502014-10-09 15:29:14 -070014#include <linux/rwsem.h>
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +053015#include <linux/hugetlb.h>
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070016
Dave Hansen33a709b2016-02-12 13:02:19 -080017#include <asm/mmu_context.h>
Steve Capper2667f502014-10-09 15:29:14 -070018#include <asm/pgtable.h>
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070019#include <asm/tlbflush.h>
Steve Capper2667f502014-10-09 15:29:14 -070020
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070021#include "internal.h"
22
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070023static struct page *no_page_table(struct vm_area_struct *vma,
24 unsigned int flags)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070025{
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070026 /*
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
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070039static 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
Linus Torvalds19be0ea2016-10-13 13:07:36 -070063/*
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 */
67static 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
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070073static struct page *follow_page_pte(struct vm_area_struct *vma,
74 unsigned long address, pmd_t *pmd, unsigned int flags)
75{
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070076 struct mm_struct *mm = vma->vm_mm;
Dan Williams3565fce2016-01-15 16:56:55 -080077 struct dev_pagemap *pgmap = NULL;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070078 struct page *page;
79 spinlock_t *ptl;
80 pte_t *ptep, pte;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070081
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070082retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070083 if (unlikely(pmd_bad(*pmd)))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070084 return no_page_table(vma, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070085
86 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070087 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;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -080097 if (pte_none(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070098 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);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700104 goto retry;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700105 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800106 if ((flags & FOLL_NUMA) && pte_protnone(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700107 goto no_page;
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700108 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700109 pte_unmap_unlock(ptep, ptl);
110 return NULL;
111 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700112
113 page = vm_normal_page(vma, address, pte);
Dan Williams3565fce2016-01-15 16:56:55 -0800114 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)) {
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700125 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 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700140 }
141
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800142 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
Dan Williams3565fce2016-01-15 16:56:55 -0800155 if (flags & FOLL_GET) {
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800156 get_page(page);
Dan Williams3565fce2016-01-15 16:56:55 -0800157
158 /* drop the pgmap reference now that we hold the page */
159 if (pgmap) {
160 put_dev_pagemap(pgmap);
161 pgmap = NULL;
162 }
163 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700164 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 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800175 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -0800176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page))
178 goto out;
179
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700180 /*
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 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700201out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700202 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700203 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700204no_page:
205 pte_unmap_unlock(ptep, ptl);
206 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700207 return NULL;
208 return no_page_table(vma, flags);
209}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700210
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700211/**
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 */
224struct 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;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300229 p4d_t *p4d;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700230 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);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300247 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);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700254 if (pud_none(*pud))
255 return no_page_table(vma, flags);
256 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800257 page = follow_huge_pud(mm, address, pud, flags);
258 if (page)
259 return page;
260 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700261 }
Matthew Wilcoxa00cc7d2017-02-24 14:57:02 -0800262 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 }
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700269 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) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800276 page = follow_huge_pmd(mm, address, pmd, flags);
277 if (page)
278 return page;
279 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700280 }
Dan Williams3565fce2016-01-15 16:56:55 -0800281 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 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800288 if (likely(!pmd_trans_huge(*pmd)))
289 return follow_page_pte(vma, address, pmd, flags);
290
Aneesh Kumar K.Vdb08f202017-02-24 14:59:53 -0800291 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
292 return no_page_table(vma, flags);
293
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800294 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);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700298 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800299 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;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800305 split_huge_pmd(vma, pmd, address);
Naoya Horiguchi337d9ab2016-07-26 15:24:03 -0700306 if (pmd_trans_unstable(pmd))
307 ret = -EBUSY;
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800308 } 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);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -0700315 if (pmd_none(*pmd))
316 return no_page_table(vma, flags);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800317 }
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;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700327}
328
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700329static 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;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300334 p4d_t *p4d;
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700335 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));
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300348 p4d = p4d_offset(pgd, address);
349 BUG_ON(p4d_none(*p4d));
350 pud = pud_offset(p4d, address);
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700351 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);
369out:
370 ret = 0;
371unmap:
372 pte_unmap(pte);
373 return ret;
374}
375
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700376/*
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 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700381static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
382 unsigned long address, unsigned int *flags, int *nonblocking)
383{
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700384 unsigned int fault_flags = 0;
385 int ret;
386
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800387 /* mlock all present pages, but do not fault in new pages */
388 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
389 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700390 /* For mm_populate(), just skip the stack guard page. */
391 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700392 (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;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800397 if (*flags & FOLL_REMOTE)
398 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700399 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;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700403 if (*flags & FOLL_TRIED) {
404 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
405 fault_flags |= FAULT_FLAG_TRIED;
406 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700407
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700408 ret = handle_mm_fault(vma, address, fault_flags);
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700409 if (ret & VM_FAULT_ERROR) {
James Morse9a291a72017-06-02 14:46:46 -0700410 int err = vm_fault_to_errno(ret, *flags);
411
412 if (err)
413 return err;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700414 BUG();
415 }
416
417 if (tsk) {
418 if (ret & VM_FAULT_MAJOR)
419 tsk->maj_flt++;
420 else
421 tsk->min_flt++;
422 }
423
424 if (ret & VM_FAULT_RETRY) {
425 if (nonblocking)
426 *nonblocking = 0;
427 return -EBUSY;
428 }
429
430 /*
431 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
432 * necessary, even if maybe_mkwrite decided not to set pte_write. We
433 * can thus safely do subsequent page lookups as if they were reads.
434 * But only do so when looping for pte_write is futile: in some cases
435 * userspace may also be wanting to write to the gotten user page,
436 * which a read fault here might prevent (a readonly page might get
437 * reCOWed by userspace write).
438 */
439 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700440 *flags |= FOLL_COW;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700441 return 0;
442}
443
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700444static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
445{
446 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800447 int write = (gup_flags & FOLL_WRITE);
448 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700449
450 if (vm_flags & (VM_IO | VM_PFNMAP))
451 return -EFAULT;
452
Dave Hansen1b2ee122016-02-12 13:02:21 -0800453 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700454 if (!(vm_flags & VM_WRITE)) {
455 if (!(gup_flags & FOLL_FORCE))
456 return -EFAULT;
457 /*
458 * We used to let the write,force case do COW in a
459 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
460 * set a breakpoint in a read-only mapping of an
461 * executable, without corrupting the file (yet only
462 * when that file had been opened for writing!).
463 * Anon pages in shared mappings are surprising: now
464 * just reject it.
465 */
Hugh Dickins46435362016-01-30 18:03:16 -0800466 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700467 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700468 }
469 } else if (!(vm_flags & VM_READ)) {
470 if (!(gup_flags & FOLL_FORCE))
471 return -EFAULT;
472 /*
473 * Is there actually any vma we can reach here which does not
474 * have VM_MAYREAD set?
475 */
476 if (!(vm_flags & VM_MAYREAD))
477 return -EFAULT;
478 }
Dave Hansend61172b2016-02-12 13:02:24 -0800479 /*
480 * gups are always data accesses, not instruction
481 * fetches, so execute=false here
482 */
483 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800484 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700485 return 0;
486}
487
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700488/**
489 * __get_user_pages() - pin user pages in memory
490 * @tsk: task_struct of target task
491 * @mm: mm_struct of target mm
492 * @start: starting user address
493 * @nr_pages: number of pages from start to pin
494 * @gup_flags: flags modifying pin behaviour
495 * @pages: array that receives pointers to the pages pinned.
496 * Should be at least nr_pages long. Or NULL, if caller
497 * only intends to ensure the pages are faulted in.
498 * @vmas: array of pointers to vmas corresponding to each page.
499 * Or NULL if the caller does not require them.
500 * @nonblocking: whether waiting for disk IO or mmap_sem contention
501 *
502 * Returns number of pages pinned. This may be fewer than the number
503 * requested. If nr_pages is 0 or negative, returns 0. If no pages
504 * were pinned, returns -errno. Each page returned must be released
505 * with a put_page() call when it is finished with. vmas will only
506 * remain valid while mmap_sem is held.
507 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700508 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700509 *
510 * __get_user_pages walks a process's page tables and takes a reference to
511 * each struct page that each user address corresponds to at a given
512 * instant. That is, it takes the page that would be accessed if a user
513 * thread accesses the given user virtual address at that instant.
514 *
515 * This does not guarantee that the page exists in the user mappings when
516 * __get_user_pages returns, and there may even be a completely different
517 * page there in some cases (eg. if mmapped pagecache has been invalidated
518 * and subsequently re faulted). However it does guarantee that the page
519 * won't be freed completely. And mostly callers simply care that the page
520 * contains data that was valid *at some point in time*. Typically, an IO
521 * or similar operation cannot guarantee anything stronger anyway because
522 * locks can't be held over the syscall boundary.
523 *
524 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
525 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
526 * appropriate) must be called after the page is finished with, and
527 * before put_page is called.
528 *
529 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
530 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700531 * *@nonblocking will be set to 0. Further, if @gup_flags does not
532 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
533 * this case.
534 *
535 * A caller using such a combination of @nonblocking and @gup_flags
536 * must therefore hold the mmap_sem for reading only, and recognize
537 * when it's been released. Otherwise, it must be held for either
538 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700539 *
540 * In most cases, get_user_pages or get_user_pages_fast should be used
541 * instead of __get_user_pages. __get_user_pages should be used only if
542 * you need some special @gup_flags.
543 */
Lorenzo Stoakes0d731752016-10-24 10:57:25 +0100544static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700545 unsigned long start, unsigned long nr_pages,
546 unsigned int gup_flags, struct page **pages,
547 struct vm_area_struct **vmas, int *nonblocking)
548{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700549 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700550 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700551 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700552
553 if (!nr_pages)
554 return 0;
555
556 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
557
558 /*
559 * If FOLL_FORCE is set then do not force a full fault as the hinting
560 * fault information is unrelated to the reference behaviour of a task
561 * using the address space
562 */
563 if (!(gup_flags & FOLL_FORCE))
564 gup_flags |= FOLL_NUMA;
565
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700566 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700567 struct page *page;
568 unsigned int foll_flags = gup_flags;
569 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700570
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700571 /* first iteration or cross vma bound */
572 if (!vma || start >= vma->vm_end) {
573 vma = find_extend_vma(mm, start);
574 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700575 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700576 ret = get_gate_page(mm, start & PAGE_MASK,
577 gup_flags, &vma,
578 pages ? &pages[i] : NULL);
579 if (ret)
580 return i ? : ret;
581 page_mask = 0;
582 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700583 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700584
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700585 if (!vma || check_vma_flags(vma, gup_flags))
586 return i ? : -EFAULT;
587 if (is_vm_hugetlb_page(vma)) {
588 i = follow_hugetlb_page(mm, vma, pages, vmas,
589 &start, &nr_pages, i,
Andrea Arcangeli87ffc112017-02-22 15:43:13 -0800590 gup_flags, nonblocking);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700591 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700592 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700593 }
594retry:
595 /*
596 * If we have a pending SIGKILL, don't keep faulting pages and
597 * potentially allocating memory.
598 */
599 if (unlikely(fatal_signal_pending(current)))
600 return i ? i : -ERESTARTSYS;
601 cond_resched();
602 page = follow_page_mask(vma, start, foll_flags, &page_mask);
603 if (!page) {
604 int ret;
605 ret = faultin_page(tsk, vma, start, &foll_flags,
606 nonblocking);
607 switch (ret) {
608 case 0:
609 goto retry;
610 case -EFAULT:
611 case -ENOMEM:
612 case -EHWPOISON:
613 return i ? i : ret;
614 case -EBUSY:
615 return i;
616 case -ENOENT:
617 goto next_page;
618 }
619 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700620 } else if (PTR_ERR(page) == -EEXIST) {
621 /*
622 * Proper page table entry exists, but no corresponding
623 * struct page.
624 */
625 goto next_page;
626 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700627 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700628 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700629 if (pages) {
630 pages[i] = page;
631 flush_anon_page(vma, page, start);
632 flush_dcache_page(page);
633 page_mask = 0;
634 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700635next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700636 if (vmas) {
637 vmas[i] = vma;
638 page_mask = 0;
639 }
640 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
641 if (page_increm > nr_pages)
642 page_increm = nr_pages;
643 i += page_increm;
644 start += page_increm * PAGE_SIZE;
645 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700646 } while (nr_pages);
647 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700648}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700649
Tobias Klauser771ab432016-12-12 16:41:53 -0800650static bool vma_permits_fault(struct vm_area_struct *vma,
651 unsigned int fault_flags)
Dave Hansend4925e02016-02-12 13:02:16 -0800652{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800653 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
654 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800655 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800656
657 if (!(vm_flags & vma->vm_flags))
658 return false;
659
Dave Hansen33a709b2016-02-12 13:02:19 -0800660 /*
661 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800662 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800663 *
664 * gup always represents data access, not instruction
665 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800666 */
Dave Hansend61172b2016-02-12 13:02:24 -0800667 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800668 return false;
669
Dave Hansend4925e02016-02-12 13:02:16 -0800670 return true;
671}
672
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700673/*
674 * fixup_user_fault() - manually resolve a user page fault
675 * @tsk: the task_struct to use for page fault accounting, or
676 * NULL if faults are not to be recorded.
677 * @mm: mm_struct of target mm
678 * @address: user address
679 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800680 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
681 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700682 *
683 * This is meant to be called in the specific scenario where for locking reasons
684 * we try to access user memory in atomic context (within a pagefault_disable()
685 * section), this returns -EFAULT, and we want to resolve the user fault before
686 * trying again.
687 *
688 * Typically this is meant to be used by the futex code.
689 *
690 * The main difference with get_user_pages() is that this function will
691 * unconditionally call handle_mm_fault() which will in turn perform all the
692 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800693 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700694 *
695 * This is important for some architectures where those bits also gate the
696 * access permission to the page because they are maintained in software. On
697 * such architectures, gup() will not be enough to make a subsequent access
698 * succeed.
699 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800700 * This function will not return with an unlocked mmap_sem. So it has not the
701 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700702 */
703int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800704 unsigned long address, unsigned int fault_flags,
705 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700706{
707 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800708 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700709
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800710 if (unlocked)
711 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
712
713retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700714 vma = find_extend_vma(mm, address);
715 if (!vma || address < vma->vm_start)
716 return -EFAULT;
717
Dave Hansend4925e02016-02-12 13:02:16 -0800718 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700719 return -EFAULT;
720
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700721 ret = handle_mm_fault(vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800722 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700723 if (ret & VM_FAULT_ERROR) {
James Morse9a291a72017-06-02 14:46:46 -0700724 int err = vm_fault_to_errno(ret, 0);
725
726 if (err)
727 return err;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700728 BUG();
729 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800730
731 if (ret & VM_FAULT_RETRY) {
732 down_read(&mm->mmap_sem);
733 if (!(fault_flags & FAULT_FLAG_TRIED)) {
734 *unlocked = true;
735 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
736 fault_flags |= FAULT_FLAG_TRIED;
737 goto retry;
738 }
739 }
740
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700741 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800742 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700743 tsk->maj_flt++;
744 else
745 tsk->min_flt++;
746 }
747 return 0;
748}
Paolo Bonziniadd6a0c2016-06-07 17:51:18 +0200749EXPORT_SYMBOL_GPL(fixup_user_fault);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700750
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800751static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
752 struct mm_struct *mm,
753 unsigned long start,
754 unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800755 struct page **pages,
756 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800757 int *locked, bool notify_drop,
758 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800759{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800760 long ret, pages_done;
761 bool lock_dropped;
762
763 if (locked) {
764 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
765 BUG_ON(vmas);
766 /* check caller initialized locked */
767 BUG_ON(*locked != 1);
768 }
769
770 if (pages)
771 flags |= FOLL_GET;
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800772
773 pages_done = 0;
774 lock_dropped = false;
775 for (;;) {
776 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
777 vmas, locked);
778 if (!locked)
779 /* VM_FAULT_RETRY couldn't trigger, bypass */
780 return ret;
781
782 /* VM_FAULT_RETRY cannot return errors */
783 if (!*locked) {
784 BUG_ON(ret < 0);
785 BUG_ON(ret >= nr_pages);
786 }
787
788 if (!pages)
789 /* If it's a prefault don't insist harder */
790 return ret;
791
792 if (ret > 0) {
793 nr_pages -= ret;
794 pages_done += ret;
795 if (!nr_pages)
796 break;
797 }
798 if (*locked) {
799 /* VM_FAULT_RETRY didn't trigger */
800 if (!pages_done)
801 pages_done = ret;
802 break;
803 }
804 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
805 pages += ret;
806 start += ret << PAGE_SHIFT;
807
808 /*
809 * Repeat on the address that fired VM_FAULT_RETRY
810 * without FAULT_FLAG_ALLOW_RETRY but with
811 * FAULT_FLAG_TRIED.
812 */
813 *locked = 1;
814 lock_dropped = true;
815 down_read(&mm->mmap_sem);
816 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
817 pages, NULL, NULL);
818 if (ret != 1) {
819 BUG_ON(ret > 1);
820 if (!pages_done)
821 pages_done = ret;
822 break;
823 }
824 nr_pages--;
825 pages_done++;
826 if (!nr_pages)
827 break;
828 pages++;
829 start += PAGE_SIZE;
830 }
831 if (notify_drop && lock_dropped && *locked) {
832 /*
833 * We must let the caller know we temporarily dropped the lock
834 * and so the critical section protected by it was lost.
835 */
836 up_read(&mm->mmap_sem);
837 *locked = 0;
838 }
839 return pages_done;
840}
841
842/*
843 * We can leverage the VM_FAULT_RETRY functionality in the page fault
844 * paths better by using either get_user_pages_locked() or
845 * get_user_pages_unlocked().
846 *
847 * get_user_pages_locked() is suitable to replace the form:
848 *
849 * down_read(&mm->mmap_sem);
850 * do_something()
851 * get_user_pages(tsk, mm, ..., pages, NULL);
852 * up_read(&mm->mmap_sem);
853 *
854 * to:
855 *
856 * int locked = 1;
857 * down_read(&mm->mmap_sem);
858 * do_something()
859 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
860 * if (locked)
861 * up_read(&mm->mmap_sem);
862 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200863long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100864 unsigned int gup_flags, struct page **pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800865 int *locked)
866{
Dave Hansencde70142016-02-12 13:01:55 -0800867 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100868 pages, NULL, locked, true,
869 gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800870}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200871EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800872
873/*
Lorenzo Stoakes80a79512016-12-12 16:42:46 -0800874 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
875 * tsk, mm to be specified.
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800876 *
877 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
Lorenzo Stoakes80a79512016-12-12 16:42:46 -0800878 * caller if required (just like with __get_user_pages). "FOLL_GET"
879 * is set implicitly if "pages" is non-NULL.
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800880 */
Lorenzo Stoakes8b7457e2016-12-14 15:06:55 -0800881static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
882 struct mm_struct *mm, unsigned long start,
883 unsigned long nr_pages, struct page **pages,
884 unsigned int gup_flags)
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800885{
886 long ret;
887 int locked = 1;
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100888
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800889 down_read(&mm->mmap_sem);
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100890 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
891 &locked, false, gup_flags);
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800892 if (locked)
893 up_read(&mm->mmap_sem);
894 return ret;
895}
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800896
897/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800898 * get_user_pages_unlocked() is suitable to replace the form:
899 *
900 * down_read(&mm->mmap_sem);
901 * get_user_pages(tsk, mm, ..., pages, NULL);
902 * up_read(&mm->mmap_sem);
903 *
904 * with:
905 *
906 * get_user_pages_unlocked(tsk, mm, ..., pages);
907 *
908 * It is functionally equivalent to get_user_pages_fast so
Lorenzo Stoakes80a79512016-12-12 16:42:46 -0800909 * get_user_pages_fast should be used instead if specific gup_flags
910 * (e.g. FOLL_FORCE) are not required.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800911 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200912long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100913 struct page **pages, unsigned int gup_flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800914{
Dave Hansencde70142016-02-12 13:01:55 -0800915 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100916 pages, gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800917}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200918EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800919
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700920/*
Dave Hansen1e987792016-02-12 13:01:54 -0800921 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700922 * @tsk: the task_struct to use for page fault accounting, or
923 * NULL if faults are not to be recorded.
924 * @mm: mm_struct of target mm
925 * @start: starting user address
926 * @nr_pages: number of pages from start to pin
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100927 * @gup_flags: flags modifying lookup behaviour
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700928 * @pages: array that receives pointers to the pages pinned.
929 * Should be at least nr_pages long. Or NULL, if caller
930 * only intends to ensure the pages are faulted in.
931 * @vmas: array of pointers to vmas corresponding to each page.
932 * Or NULL if the caller does not require them.
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -0800933 * @locked: pointer to lock flag indicating whether lock is held and
934 * subsequently whether VM_FAULT_RETRY functionality can be
935 * utilised. Lock must initially be held.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700936 *
937 * Returns number of pages pinned. This may be fewer than the number
938 * requested. If nr_pages is 0 or negative, returns 0. If no pages
939 * were pinned, returns -errno. Each page returned must be released
940 * with a put_page() call when it is finished with. vmas will only
941 * remain valid while mmap_sem is held.
942 *
943 * Must be called with mmap_sem held for read or write.
944 *
945 * get_user_pages walks a process's page tables and takes a reference to
946 * each struct page that each user address corresponds to at a given
947 * instant. That is, it takes the page that would be accessed if a user
948 * thread accesses the given user virtual address at that instant.
949 *
950 * This does not guarantee that the page exists in the user mappings when
951 * get_user_pages returns, and there may even be a completely different
952 * page there in some cases (eg. if mmapped pagecache has been invalidated
953 * and subsequently re faulted). However it does guarantee that the page
954 * won't be freed completely. And mostly callers simply care that the page
955 * contains data that was valid *at some point in time*. Typically, an IO
956 * or similar operation cannot guarantee anything stronger anyway because
957 * locks can't be held over the syscall boundary.
958 *
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100959 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
960 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
961 * be called after the page is finished with, and before put_page is called.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700962 *
963 * get_user_pages is typically used for fewer-copy IO operations, to get a
964 * handle on the memory by some means other than accesses via the user virtual
965 * addresses. The pages may be submitted for DMA to devices or accessed via
966 * their kernel linear mapping (via the kmap APIs). Care should be taken to
967 * use the correct cache flushing APIs.
968 *
969 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800970 *
971 * get_user_pages should be phased out in favor of
972 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
973 * should use get_user_pages because it cannot pass
974 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700975 */
Dave Hansen1e987792016-02-12 13:01:54 -0800976long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
977 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100978 unsigned int gup_flags, struct page **pages,
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -0800979 struct vm_area_struct **vmas, int *locked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700980{
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100981 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -0800982 locked, true,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100983 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
Dave Hansen1e987792016-02-12 13:01:54 -0800984}
985EXPORT_SYMBOL(get_user_pages_remote);
986
987/*
Dave Hansend4edcf02016-02-12 13:01:56 -0800988 * This is the same as get_user_pages_remote(), just with a
989 * less-flexible calling convention where we assume that the task
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -0800990 * and mm being operated on are the current task's and don't allow
991 * passing of a locked parameter. We also obviously don't pass
992 * FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -0800993 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200994long get_user_pages(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +0100995 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800996 struct vm_area_struct **vmas)
997{
Dave Hansencde70142016-02-12 13:01:55 -0800998 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +0100999 pages, vmas, NULL, false,
1000 gup_flags | FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001001}
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001002EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001003
1004/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001005 * populate_vma_page_range() - populate a range of pages in the vma.
1006 * @vma: target vma
1007 * @start: start address
1008 * @end: end address
1009 * @nonblocking:
1010 *
1011 * This takes care of mlocking the pages too if VM_LOCKED is set.
1012 *
1013 * return 0 on success, negative error code on error.
1014 *
1015 * vma->vm_mm->mmap_sem must be held.
1016 *
1017 * If @nonblocking is NULL, it may be held for read or write and will
1018 * be unperturbed.
1019 *
1020 * If @nonblocking is non-NULL, it must held for read only and may be
1021 * released. If it's released, *@nonblocking will be set to 0.
1022 */
1023long populate_vma_page_range(struct vm_area_struct *vma,
1024 unsigned long start, unsigned long end, int *nonblocking)
1025{
1026 struct mm_struct *mm = vma->vm_mm;
1027 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1028 int gup_flags;
1029
1030 VM_BUG_ON(start & ~PAGE_MASK);
1031 VM_BUG_ON(end & ~PAGE_MASK);
1032 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1033 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1034 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1035
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001036 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1037 if (vma->vm_flags & VM_LOCKONFAULT)
1038 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001039 /*
1040 * We want to touch writable mappings with a write fault in order
1041 * to break COW, except for shared mappings because these don't COW
1042 * and we would not want to dirty them for nothing.
1043 */
1044 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1045 gup_flags |= FOLL_WRITE;
1046
1047 /*
1048 * We want mlock to succeed for regions that have any permissions
1049 * other than PROT_NONE.
1050 */
1051 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1052 gup_flags |= FOLL_FORCE;
1053
1054 /*
1055 * We made sure addr is within a VMA, so the following will
1056 * not result in a stack expansion that recurses back here.
1057 */
1058 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1059 NULL, NULL, nonblocking);
1060}
1061
1062/*
1063 * __mm_populate - populate and/or mlock pages within a range of address space.
1064 *
1065 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1066 * flags. VMAs must be already marked with the desired vm_flags, and
1067 * mmap_sem must not be held.
1068 */
1069int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1070{
1071 struct mm_struct *mm = current->mm;
1072 unsigned long end, nstart, nend;
1073 struct vm_area_struct *vma = NULL;
1074 int locked = 0;
1075 long ret = 0;
1076
1077 VM_BUG_ON(start & ~PAGE_MASK);
1078 VM_BUG_ON(len != PAGE_ALIGN(len));
1079 end = start + len;
1080
1081 for (nstart = start; nstart < end; nstart = nend) {
1082 /*
1083 * We want to fault in pages for [nstart; end) address range.
1084 * Find first corresponding VMA.
1085 */
1086 if (!locked) {
1087 locked = 1;
1088 down_read(&mm->mmap_sem);
1089 vma = find_vma(mm, nstart);
1090 } else if (nstart >= vma->vm_end)
1091 vma = vma->vm_next;
1092 if (!vma || vma->vm_start >= end)
1093 break;
1094 /*
1095 * Set [nstart; nend) to intersection of desired address
1096 * range with the first VMA. Also, skip undesirable VMA types.
1097 */
1098 nend = min(end, vma->vm_end);
1099 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1100 continue;
1101 if (nstart < vma->vm_start)
1102 nstart = vma->vm_start;
1103 /*
1104 * Now fault in a range of pages. populate_vma_page_range()
1105 * double checks the vma flags, so that it won't mlock pages
1106 * if the vma was already munlocked.
1107 */
1108 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1109 if (ret < 0) {
1110 if (ignore_errors) {
1111 ret = 0;
1112 continue; /* continue at next VMA */
1113 }
1114 break;
1115 }
1116 nend = nstart + ret * PAGE_SIZE;
1117 ret = 0;
1118 }
1119 if (locked)
1120 up_read(&mm->mmap_sem);
1121 return ret; /* 0 or negative error code */
1122}
1123
1124/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001125 * get_dump_page() - pin user page in memory while writing it to core dump
1126 * @addr: user address
1127 *
1128 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001129 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001130 *
1131 * Returns NULL on any kind of failure - a hole must then be inserted into
1132 * the corefile, to preserve alignment with its headers; and also returns
1133 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1134 * allowing a hole to be left in the corefile to save diskspace.
1135 *
1136 * Called without mmap_sem, but after all other threads have been killed.
1137 */
1138#ifdef CONFIG_ELF_CORE
1139struct page *get_dump_page(unsigned long addr)
1140{
1141 struct vm_area_struct *vma;
1142 struct page *page;
1143
1144 if (__get_user_pages(current, current->mm, addr, 1,
1145 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1146 NULL) < 1)
1147 return NULL;
1148 flush_cache_page(vma, addr, page_to_pfn(page));
1149 return page;
1150}
1151#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001152
1153/*
Ingo Molnar6dd29b32017-04-23 11:37:17 +02001154 * Generic RCU Fast GUP
Steve Capper2667f502014-10-09 15:29:14 -07001155 *
1156 * get_user_pages_fast attempts to pin user pages by walking the page
1157 * tables directly and avoids taking locks. Thus the walker needs to be
1158 * protected from page table pages being freed from under it, and should
1159 * block any THP splits.
1160 *
1161 * One way to achieve this is to have the walker disable interrupts, and
1162 * rely on IPIs from the TLB flushing code blocking before the page table
1163 * pages are freed. This is unsuitable for architectures that do not need
1164 * to broadcast an IPI when invalidating TLBs.
1165 *
1166 * Another way to achieve this is to batch up page table containing pages
1167 * belonging to more than one mm_user, then rcu_sched a callback to free those
1168 * pages. Disabling interrupts will allow the fast_gup walker to both block
1169 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1170 * (which is a relatively rare event). The code below adopts this strategy.
1171 *
1172 * Before activating this code, please be aware that the following assumptions
1173 * are currently made:
1174 *
Ingo Molnar6dd29b32017-04-23 11:37:17 +02001175 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1176 * pages containing page tables.
Steve Capper2667f502014-10-09 15:29:14 -07001177 *
Steve Capper2667f502014-10-09 15:29:14 -07001178 * *) ptes can be read atomically by the architecture.
1179 *
1180 * *) access_ok is sufficient to validate userspace address ranges.
1181 *
1182 * The last two assumptions can be relaxed by the addition of helper functions.
1183 *
1184 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1185 */
Ingo Molnar6dd29b32017-04-23 11:37:17 +02001186#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
Steve Capper2667f502014-10-09 15:29:14 -07001187
Kirill A. Shutemov0005d202017-03-16 18:26:51 +03001188#ifndef gup_get_pte
1189/*
1190 * We assume that the PTE can be read atomically. If this is not the case for
1191 * your architecture, please provide the helper.
1192 */
1193static inline pte_t gup_get_pte(pte_t *ptep)
1194{
1195 return READ_ONCE(*ptep);
1196}
1197#endif
1198
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001199static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1200{
1201 while ((*nr) - nr_start) {
1202 struct page *page = pages[--(*nr)];
1203
1204 ClearPageReferenced(page);
1205 put_page(page);
1206 }
1207}
1208
Steve Capper2667f502014-10-09 15:29:14 -07001209#ifdef __HAVE_ARCH_PTE_SPECIAL
1210static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1211 int write, struct page **pages, int *nr)
1212{
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001213 struct dev_pagemap *pgmap = NULL;
1214 int nr_start = *nr, ret = 0;
Steve Capper2667f502014-10-09 15:29:14 -07001215 pte_t *ptep, *ptem;
Steve Capper2667f502014-10-09 15:29:14 -07001216
1217 ptem = ptep = pte_offset_map(&pmd, addr);
1218 do {
Kirill A. Shutemov0005d202017-03-16 18:26:51 +03001219 pte_t pte = gup_get_pte(ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001220 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001221
1222 /*
1223 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001224 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001225 */
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001226 if (pte_protnone(pte))
1227 goto pte_unmap;
1228
1229 if (!pte_access_permitted(pte, write))
1230 goto pte_unmap;
1231
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001232 if (pte_devmap(pte)) {
1233 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1234 if (unlikely(!pgmap)) {
1235 undo_dev_pagemap(nr, nr_start, pages);
1236 goto pte_unmap;
1237 }
1238 } else if (pte_special(pte))
Steve Capper2667f502014-10-09 15:29:14 -07001239 goto pte_unmap;
1240
1241 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1242 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001243 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001244
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001245 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001246 goto pte_unmap;
1247
1248 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001249 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001250 goto pte_unmap;
1251 }
1252
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001253 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001254
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001255 put_dev_pagemap(pgmap);
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001256 SetPageReferenced(page);
Steve Capper2667f502014-10-09 15:29:14 -07001257 pages[*nr] = page;
1258 (*nr)++;
1259
1260 } while (ptep++, addr += PAGE_SIZE, addr != end);
1261
1262 ret = 1;
1263
1264pte_unmap:
1265 pte_unmap(ptem);
1266 return ret;
1267}
1268#else
1269
1270/*
1271 * If we can't determine whether or not a pte is special, then fail immediately
1272 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1273 * to be special.
1274 *
1275 * For a futex to be placed on a THP tail page, get_futex_key requires a
1276 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1277 * useful to have gup_huge_pmd even if we can't operate on ptes.
1278 */
1279static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1280 int write, struct page **pages, int *nr)
1281{
1282 return 0;
1283}
1284#endif /* __HAVE_ARCH_PTE_SPECIAL */
1285
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001286#ifdef __HAVE_ARCH_PTE_DEVMAP
1287static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1288 unsigned long end, struct page **pages, int *nr)
1289{
1290 int nr_start = *nr;
1291 struct dev_pagemap *pgmap = NULL;
1292
1293 do {
1294 struct page *page = pfn_to_page(pfn);
1295
1296 pgmap = get_dev_pagemap(pfn, pgmap);
1297 if (unlikely(!pgmap)) {
1298 undo_dev_pagemap(nr, nr_start, pages);
1299 return 0;
1300 }
1301 SetPageReferenced(page);
1302 pages[*nr] = page;
1303 get_page(page);
1304 put_dev_pagemap(pgmap);
1305 (*nr)++;
1306 pfn++;
1307 } while (addr += PAGE_SIZE, addr != end);
1308 return 1;
1309}
1310
1311static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1312 unsigned long end, struct page **pages, int *nr)
1313{
1314 unsigned long fault_pfn;
1315
1316 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1317 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1318}
1319
1320static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1321 unsigned long end, struct page **pages, int *nr)
1322{
1323 unsigned long fault_pfn;
1324
1325 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1326 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1327}
1328#else
1329static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1330 unsigned long end, struct page **pages, int *nr)
1331{
1332 BUILD_BUG();
1333 return 0;
1334}
1335
1336static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1337 unsigned long end, struct page **pages, int *nr)
1338{
1339 BUILD_BUG();
1340 return 0;
1341}
1342#endif
1343
Steve Capper2667f502014-10-09 15:29:14 -07001344static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1345 unsigned long end, int write, struct page **pages, int *nr)
1346{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001347 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001348 int refs;
1349
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001350 if (!pmd_access_permitted(orig, write))
Steve Capper2667f502014-10-09 15:29:14 -07001351 return 0;
1352
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001353 if (pmd_devmap(orig))
1354 return __gup_device_huge_pmd(orig, addr, end, pages, nr);
1355
Steve Capper2667f502014-10-09 15:29:14 -07001356 refs = 0;
1357 head = pmd_page(orig);
1358 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001359 do {
1360 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1361 pages[*nr] = page;
1362 (*nr)++;
1363 page++;
1364 refs++;
1365 } while (addr += PAGE_SIZE, addr != end);
1366
1367 if (!page_cache_add_speculative(head, refs)) {
1368 *nr -= refs;
1369 return 0;
1370 }
1371
1372 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1373 *nr -= refs;
1374 while (refs--)
1375 put_page(head);
1376 return 0;
1377 }
1378
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001379 SetPageReferenced(head);
Steve Capper2667f502014-10-09 15:29:14 -07001380 return 1;
1381}
1382
1383static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1384 unsigned long end, int write, struct page **pages, int *nr)
1385{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001386 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001387 int refs;
1388
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001389 if (!pud_access_permitted(orig, write))
Steve Capper2667f502014-10-09 15:29:14 -07001390 return 0;
1391
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001392 if (pud_devmap(orig))
1393 return __gup_device_huge_pud(orig, addr, end, pages, nr);
1394
Steve Capper2667f502014-10-09 15:29:14 -07001395 refs = 0;
1396 head = pud_page(orig);
1397 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001398 do {
1399 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1400 pages[*nr] = page;
1401 (*nr)++;
1402 page++;
1403 refs++;
1404 } while (addr += PAGE_SIZE, addr != end);
1405
1406 if (!page_cache_add_speculative(head, refs)) {
1407 *nr -= refs;
1408 return 0;
1409 }
1410
1411 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1412 *nr -= refs;
1413 while (refs--)
1414 put_page(head);
1415 return 0;
1416 }
1417
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001418 SetPageReferenced(head);
Steve Capper2667f502014-10-09 15:29:14 -07001419 return 1;
1420}
1421
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301422static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1423 unsigned long end, int write,
1424 struct page **pages, int *nr)
1425{
1426 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001427 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301428
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001429 if (!pgd_access_permitted(orig, write))
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301430 return 0;
1431
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001432 BUILD_BUG_ON(pgd_devmap(orig));
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301433 refs = 0;
1434 head = pgd_page(orig);
1435 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301436 do {
1437 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1438 pages[*nr] = page;
1439 (*nr)++;
1440 page++;
1441 refs++;
1442 } while (addr += PAGE_SIZE, addr != end);
1443
1444 if (!page_cache_add_speculative(head, refs)) {
1445 *nr -= refs;
1446 return 0;
1447 }
1448
1449 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1450 *nr -= refs;
1451 while (refs--)
1452 put_page(head);
1453 return 0;
1454 }
1455
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001456 SetPageReferenced(head);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301457 return 1;
1458}
1459
Steve Capper2667f502014-10-09 15:29:14 -07001460static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1461 int write, struct page **pages, int *nr)
1462{
1463 unsigned long next;
1464 pmd_t *pmdp;
1465
1466 pmdp = pmd_offset(&pud, addr);
1467 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001468 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001469
1470 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001471 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001472 return 0;
1473
1474 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1475 /*
1476 * NUMA hinting faults need to be handled in the GUP
1477 * slowpath for accounting purposes and so that they
1478 * can be serialised against THP migration.
1479 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001480 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001481 return 0;
1482
1483 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1484 pages, nr))
1485 return 0;
1486
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301487 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1488 /*
1489 * architecture have different format for hugetlbfs
1490 * pmd format and THP pmd format
1491 */
1492 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1493 PMD_SHIFT, next, write, pages, nr))
1494 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001495 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1496 return 0;
1497 } while (pmdp++, addr = next, addr != end);
1498
1499 return 1;
1500}
1501
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001502static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301503 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001504{
1505 unsigned long next;
1506 pud_t *pudp;
1507
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001508 pudp = pud_offset(&p4d, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001509 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001510 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001511
1512 next = pud_addr_end(addr, end);
1513 if (pud_none(pud))
1514 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301515 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001516 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301517 pages, nr))
1518 return 0;
1519 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1520 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1521 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001522 return 0;
1523 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1524 return 0;
1525 } while (pudp++, addr = next, addr != end);
1526
1527 return 1;
1528}
1529
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001530static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1531 int write, struct page **pages, int *nr)
1532{
1533 unsigned long next;
1534 p4d_t *p4dp;
1535
1536 p4dp = p4d_offset(&pgd, addr);
1537 do {
1538 p4d_t p4d = READ_ONCE(*p4dp);
1539
1540 next = p4d_addr_end(addr, end);
1541 if (p4d_none(p4d))
1542 return 0;
1543 BUILD_BUG_ON(p4d_huge(p4d));
1544 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1545 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1546 P4D_SHIFT, next, write, pages, nr))
1547 return 0;
Kirill A. Shutemovce70df02017-03-13 08:22:13 +03001548 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001549 return 0;
1550 } while (p4dp++, addr = next, addr != end);
1551
1552 return 1;
1553}
1554
Steve Capper2667f502014-10-09 15:29:14 -07001555/*
1556 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1557 * the regular GUP. It will only return non-negative values.
1558 */
1559int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1560 struct page **pages)
1561{
1562 struct mm_struct *mm = current->mm;
1563 unsigned long addr, len, end;
1564 unsigned long next, flags;
1565 pgd_t *pgdp;
1566 int nr = 0;
1567
1568 start &= PAGE_MASK;
1569 addr = start;
1570 len = (unsigned long) nr_pages << PAGE_SHIFT;
1571 end = start + len;
1572
1573 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
Arnd Bergmannaa2369f2017-05-03 14:56:12 -07001574 (void __user *)start, len)))
Steve Capper2667f502014-10-09 15:29:14 -07001575 return 0;
1576
1577 /*
1578 * Disable interrupts. We use the nested form as we can already have
1579 * interrupts disabled by get_futex_key.
1580 *
1581 * With interrupts disabled, we block page table pages from being
1582 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1583 * for more details.
1584 *
1585 * We do not adopt an rcu_read_lock(.) here as we also want to
1586 * block IPIs that come from THPs splitting.
1587 */
1588
1589 local_irq_save(flags);
1590 pgdp = pgd_offset(mm, addr);
1591 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001592 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301593
Steve Capper2667f502014-10-09 15:29:14 -07001594 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301595 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001596 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301597 if (unlikely(pgd_huge(pgd))) {
1598 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1599 pages, &nr))
1600 break;
1601 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1602 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1603 PGDIR_SHIFT, next, write, pages, &nr))
1604 break;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001605 } else if (!gup_p4d_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001606 break;
1607 } while (pgdp++, addr = next, addr != end);
1608 local_irq_restore(flags);
1609
1610 return nr;
1611}
1612
Kirill A. Shutemov73e10a62017-03-16 18:26:54 +03001613#ifndef gup_fast_permitted
1614/*
1615 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1616 * we need to fall back to the slow version:
1617 */
1618bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1619{
1620 unsigned long len, end;
1621
1622 len = (unsigned long) nr_pages << PAGE_SHIFT;
1623 end = start + len;
1624 return end >= start;
1625}
1626#endif
1627
Steve Capper2667f502014-10-09 15:29:14 -07001628/**
1629 * get_user_pages_fast() - pin user pages in memory
1630 * @start: starting user address
1631 * @nr_pages: number of pages from start to pin
1632 * @write: whether pages will be written to
1633 * @pages: array that receives pointers to the pages pinned.
1634 * Should be at least nr_pages long.
1635 *
1636 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1637 * If not successful, it will fall back to taking the lock and
1638 * calling get_user_pages().
1639 *
1640 * Returns number of pages pinned. This may be fewer than the number
1641 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1642 * were pinned, returns -errno.
1643 */
1644int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1645 struct page **pages)
1646{
Kirill A. Shutemov73e10a62017-03-16 18:26:54 +03001647 int nr = 0, ret = 0;
Steve Capper2667f502014-10-09 15:29:14 -07001648
1649 start &= PAGE_MASK;
Kirill A. Shutemov73e10a62017-03-16 18:26:54 +03001650
1651 if (gup_fast_permitted(start, nr_pages, write)) {
1652 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1653 ret = nr;
1654 }
Steve Capper2667f502014-10-09 15:29:14 -07001655
1656 if (nr < nr_pages) {
1657 /* Try to get the remaining pages with get_user_pages */
1658 start += nr << PAGE_SHIFT;
1659 pages += nr;
1660
Lorenzo Stoakesc1641542016-10-13 01:20:13 +01001661 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1662 write ? FOLL_WRITE : 0);
Steve Capper2667f502014-10-09 15:29:14 -07001663
1664 /* Have to be a bit careful with return values */
1665 if (nr > 0) {
1666 if (ret < 0)
1667 ret = nr;
1668 else
1669 ret += nr;
1670 }
1671 }
1672
1673 return ret;
1674}
1675
Ingo Molnar6dd29b32017-04-23 11:37:17 +02001676#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */