blob: 23f01c40c88f63cc88ff62548a4d120b0c406a8e [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
Aneesh Kumar K.V080dbb62017-07-06 15:38:44 -0700211static struct page *follow_pmd_mask(struct vm_area_struct *vma,
212 unsigned long address, pud_t *pudp,
213 unsigned int flags, unsigned int *page_mask)
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700214{
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700215 pmd_t *pmd;
216 spinlock_t *ptl;
217 struct page *page;
218 struct mm_struct *mm = vma->vm_mm;
219
Aneesh Kumar K.V080dbb62017-07-06 15:38:44 -0700220 pmd = pmd_offset(pudp, address);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700221 if (pmd_none(*pmd))
222 return no_page_table(vma, flags);
223 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800224 page = follow_huge_pmd(mm, address, pmd, flags);
225 if (page)
226 return page;
227 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700228 }
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -0700229 if (is_hugepd(__hugepd(pmd_val(*pmd)))) {
230 page = follow_huge_pd(vma, address,
231 __hugepd(pmd_val(*pmd)), flags,
232 PMD_SHIFT);
233 if (page)
234 return page;
235 return no_page_table(vma, flags);
236 }
Dan Williams3565fce2016-01-15 16:56:55 -0800237 if (pmd_devmap(*pmd)) {
238 ptl = pmd_lock(mm, pmd);
239 page = follow_devmap_pmd(vma, address, pmd, flags);
240 spin_unlock(ptl);
241 if (page)
242 return page;
243 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800244 if (likely(!pmd_trans_huge(*pmd)))
245 return follow_page_pte(vma, address, pmd, flags);
246
Aneesh Kumar K.Vdb08f202017-02-24 14:59:53 -0800247 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
248 return no_page_table(vma, flags);
249
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800250 ptl = pmd_lock(mm, pmd);
251 if (unlikely(!pmd_trans_huge(*pmd))) {
252 spin_unlock(ptl);
253 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700254 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800255 if (flags & FOLL_SPLIT) {
256 int ret;
257 page = pmd_page(*pmd);
258 if (is_huge_zero_page(page)) {
259 spin_unlock(ptl);
260 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800261 split_huge_pmd(vma, pmd, address);
Naoya Horiguchi337d9ab2016-07-26 15:24:03 -0700262 if (pmd_trans_unstable(pmd))
263 ret = -EBUSY;
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800264 } else {
265 get_page(page);
266 spin_unlock(ptl);
267 lock_page(page);
268 ret = split_huge_page(page);
269 unlock_page(page);
270 put_page(page);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -0700271 if (pmd_none(*pmd))
272 return no_page_table(vma, flags);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800273 }
274
275 return ret ? ERR_PTR(ret) :
276 follow_page_pte(vma, address, pmd, flags);
277 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800278 page = follow_trans_huge_pmd(vma, address, pmd, flags);
279 spin_unlock(ptl);
280 *page_mask = HPAGE_PMD_NR - 1;
281 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700282}
283
Aneesh Kumar K.V080dbb62017-07-06 15:38:44 -0700284
285static struct page *follow_pud_mask(struct vm_area_struct *vma,
286 unsigned long address, p4d_t *p4dp,
287 unsigned int flags, unsigned int *page_mask)
288{
289 pud_t *pud;
290 spinlock_t *ptl;
291 struct page *page;
292 struct mm_struct *mm = vma->vm_mm;
293
294 pud = pud_offset(p4dp, address);
295 if (pud_none(*pud))
296 return no_page_table(vma, flags);
297 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
298 page = follow_huge_pud(mm, address, pud, flags);
299 if (page)
300 return page;
301 return no_page_table(vma, flags);
302 }
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -0700303 if (is_hugepd(__hugepd(pud_val(*pud)))) {
304 page = follow_huge_pd(vma, address,
305 __hugepd(pud_val(*pud)), flags,
306 PUD_SHIFT);
307 if (page)
308 return page;
309 return no_page_table(vma, flags);
310 }
Aneesh Kumar K.V080dbb62017-07-06 15:38:44 -0700311 if (pud_devmap(*pud)) {
312 ptl = pud_lock(mm, pud);
313 page = follow_devmap_pud(vma, address, pud, flags);
314 spin_unlock(ptl);
315 if (page)
316 return page;
317 }
318 if (unlikely(pud_bad(*pud)))
319 return no_page_table(vma, flags);
320
321 return follow_pmd_mask(vma, address, pud, flags, page_mask);
322}
323
324
325static struct page *follow_p4d_mask(struct vm_area_struct *vma,
326 unsigned long address, pgd_t *pgdp,
327 unsigned int flags, unsigned int *page_mask)
328{
329 p4d_t *p4d;
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -0700330 struct page *page;
Aneesh Kumar K.V080dbb62017-07-06 15:38:44 -0700331
332 p4d = p4d_offset(pgdp, address);
333 if (p4d_none(*p4d))
334 return no_page_table(vma, flags);
335 BUILD_BUG_ON(p4d_huge(*p4d));
336 if (unlikely(p4d_bad(*p4d)))
337 return no_page_table(vma, flags);
338
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -0700339 if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
340 page = follow_huge_pd(vma, address,
341 __hugepd(p4d_val(*p4d)), flags,
342 P4D_SHIFT);
343 if (page)
344 return page;
345 return no_page_table(vma, flags);
346 }
Aneesh Kumar K.V080dbb62017-07-06 15:38:44 -0700347 return follow_pud_mask(vma, address, p4d, flags, page_mask);
348}
349
350/**
351 * follow_page_mask - look up a page descriptor from a user-virtual address
352 * @vma: vm_area_struct mapping @address
353 * @address: virtual address to look up
354 * @flags: flags modifying lookup behaviour
355 * @page_mask: on output, *page_mask is set according to the size of the page
356 *
357 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
358 *
359 * Returns the mapped (struct page *), %NULL if no mapping exists, or
360 * an error pointer if there is a mapping to something not represented
361 * by a page descriptor (see also vm_normal_page()).
362 */
363struct page *follow_page_mask(struct vm_area_struct *vma,
364 unsigned long address, unsigned int flags,
365 unsigned int *page_mask)
366{
367 pgd_t *pgd;
368 struct page *page;
369 struct mm_struct *mm = vma->vm_mm;
370
371 *page_mask = 0;
372
373 /* make this handle hugepd */
374 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
375 if (!IS_ERR(page)) {
376 BUG_ON(flags & FOLL_GET);
377 return page;
378 }
379
380 pgd = pgd_offset(mm, address);
381
382 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
383 return no_page_table(vma, flags);
384
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -0700385 if (pgd_huge(*pgd)) {
386 page = follow_huge_pgd(mm, address, pgd, flags);
387 if (page)
388 return page;
389 return no_page_table(vma, flags);
390 }
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -0700391 if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
392 page = follow_huge_pd(vma, address,
393 __hugepd(pgd_val(*pgd)), flags,
394 PGDIR_SHIFT);
395 if (page)
396 return page;
397 return no_page_table(vma, flags);
398 }
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -0700399
Aneesh Kumar K.V080dbb62017-07-06 15:38:44 -0700400 return follow_p4d_mask(vma, address, pgd, flags, page_mask);
401}
402
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700403static int get_gate_page(struct mm_struct *mm, unsigned long address,
404 unsigned int gup_flags, struct vm_area_struct **vma,
405 struct page **page)
406{
407 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300408 p4d_t *p4d;
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700409 pud_t *pud;
410 pmd_t *pmd;
411 pte_t *pte;
412 int ret = -EFAULT;
413
414 /* user gate pages are read-only */
415 if (gup_flags & FOLL_WRITE)
416 return -EFAULT;
417 if (address > TASK_SIZE)
418 pgd = pgd_offset_k(address);
419 else
420 pgd = pgd_offset_gate(mm, address);
421 BUG_ON(pgd_none(*pgd));
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +0300422 p4d = p4d_offset(pgd, address);
423 BUG_ON(p4d_none(*p4d));
424 pud = pud_offset(p4d, address);
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700425 BUG_ON(pud_none(*pud));
426 pmd = pmd_offset(pud, address);
427 if (pmd_none(*pmd))
428 return -EFAULT;
429 VM_BUG_ON(pmd_trans_huge(*pmd));
430 pte = pte_offset_map(pmd, address);
431 if (pte_none(*pte))
432 goto unmap;
433 *vma = get_gate_vma(mm);
434 if (!page)
435 goto out;
436 *page = vm_normal_page(*vma, address, *pte);
437 if (!*page) {
438 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
439 goto unmap;
440 *page = pte_page(*pte);
441 }
442 get_page(*page);
443out:
444 ret = 0;
445unmap:
446 pte_unmap(pte);
447 return ret;
448}
449
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700450/*
451 * mmap_sem must be held on entry. If @nonblocking != NULL and
452 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
453 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
454 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700455static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
456 unsigned long address, unsigned int *flags, int *nonblocking)
457{
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700458 unsigned int fault_flags = 0;
459 int ret;
460
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800461 /* mlock all present pages, but do not fault in new pages */
462 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
463 return -ENOENT;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700464 if (*flags & FOLL_WRITE)
465 fault_flags |= FAULT_FLAG_WRITE;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800466 if (*flags & FOLL_REMOTE)
467 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700468 if (nonblocking)
469 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
470 if (*flags & FOLL_NOWAIT)
471 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700472 if (*flags & FOLL_TRIED) {
473 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
474 fault_flags |= FAULT_FLAG_TRIED;
475 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700476
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700477 ret = handle_mm_fault(vma, address, fault_flags);
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700478 if (ret & VM_FAULT_ERROR) {
James Morse9a291a72017-06-02 14:46:46 -0700479 int err = vm_fault_to_errno(ret, *flags);
480
481 if (err)
482 return err;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700483 BUG();
484 }
485
486 if (tsk) {
487 if (ret & VM_FAULT_MAJOR)
488 tsk->maj_flt++;
489 else
490 tsk->min_flt++;
491 }
492
493 if (ret & VM_FAULT_RETRY) {
494 if (nonblocking)
495 *nonblocking = 0;
496 return -EBUSY;
497 }
498
499 /*
500 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
501 * necessary, even if maybe_mkwrite decided not to set pte_write. We
502 * can thus safely do subsequent page lookups as if they were reads.
503 * But only do so when looping for pte_write is futile: in some cases
504 * userspace may also be wanting to write to the gotten user page,
505 * which a read fault here might prevent (a readonly page might get
506 * reCOWed by userspace write).
507 */
508 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700509 *flags |= FOLL_COW;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700510 return 0;
511}
512
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700513static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
514{
515 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800516 int write = (gup_flags & FOLL_WRITE);
517 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700518
519 if (vm_flags & (VM_IO | VM_PFNMAP))
520 return -EFAULT;
521
Dave Hansen1b2ee122016-02-12 13:02:21 -0800522 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700523 if (!(vm_flags & VM_WRITE)) {
524 if (!(gup_flags & FOLL_FORCE))
525 return -EFAULT;
526 /*
527 * We used to let the write,force case do COW in a
528 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
529 * set a breakpoint in a read-only mapping of an
530 * executable, without corrupting the file (yet only
531 * when that file had been opened for writing!).
532 * Anon pages in shared mappings are surprising: now
533 * just reject it.
534 */
Hugh Dickins46435362016-01-30 18:03:16 -0800535 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700536 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700537 }
538 } else if (!(vm_flags & VM_READ)) {
539 if (!(gup_flags & FOLL_FORCE))
540 return -EFAULT;
541 /*
542 * Is there actually any vma we can reach here which does not
543 * have VM_MAYREAD set?
544 */
545 if (!(vm_flags & VM_MAYREAD))
546 return -EFAULT;
547 }
Dave Hansend61172b2016-02-12 13:02:24 -0800548 /*
549 * gups are always data accesses, not instruction
550 * fetches, so execute=false here
551 */
552 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800553 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700554 return 0;
555}
556
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700557/**
558 * __get_user_pages() - pin user pages in memory
559 * @tsk: task_struct of target task
560 * @mm: mm_struct of target mm
561 * @start: starting user address
562 * @nr_pages: number of pages from start to pin
563 * @gup_flags: flags modifying pin behaviour
564 * @pages: array that receives pointers to the pages pinned.
565 * Should be at least nr_pages long. Or NULL, if caller
566 * only intends to ensure the pages are faulted in.
567 * @vmas: array of pointers to vmas corresponding to each page.
568 * Or NULL if the caller does not require them.
569 * @nonblocking: whether waiting for disk IO or mmap_sem contention
570 *
571 * Returns number of pages pinned. This may be fewer than the number
572 * requested. If nr_pages is 0 or negative, returns 0. If no pages
573 * were pinned, returns -errno. Each page returned must be released
574 * with a put_page() call when it is finished with. vmas will only
575 * remain valid while mmap_sem is held.
576 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700577 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700578 *
579 * __get_user_pages walks a process's page tables and takes a reference to
580 * each struct page that each user address corresponds to at a given
581 * instant. That is, it takes the page that would be accessed if a user
582 * thread accesses the given user virtual address at that instant.
583 *
584 * This does not guarantee that the page exists in the user mappings when
585 * __get_user_pages returns, and there may even be a completely different
586 * page there in some cases (eg. if mmapped pagecache has been invalidated
587 * and subsequently re faulted). However it does guarantee that the page
588 * won't be freed completely. And mostly callers simply care that the page
589 * contains data that was valid *at some point in time*. Typically, an IO
590 * or similar operation cannot guarantee anything stronger anyway because
591 * locks can't be held over the syscall boundary.
592 *
593 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
594 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
595 * appropriate) must be called after the page is finished with, and
596 * before put_page is called.
597 *
598 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
599 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700600 * *@nonblocking will be set to 0. Further, if @gup_flags does not
601 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
602 * this case.
603 *
604 * A caller using such a combination of @nonblocking and @gup_flags
605 * must therefore hold the mmap_sem for reading only, and recognize
606 * when it's been released. Otherwise, it must be held for either
607 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700608 *
609 * In most cases, get_user_pages or get_user_pages_fast should be used
610 * instead of __get_user_pages. __get_user_pages should be used only if
611 * you need some special @gup_flags.
612 */
Lorenzo Stoakes0d731752016-10-24 10:57:25 +0100613static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700614 unsigned long start, unsigned long nr_pages,
615 unsigned int gup_flags, struct page **pages,
616 struct vm_area_struct **vmas, int *nonblocking)
617{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700618 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700619 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700620 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700621
622 if (!nr_pages)
623 return 0;
624
625 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
626
627 /*
628 * If FOLL_FORCE is set then do not force a full fault as the hinting
629 * fault information is unrelated to the reference behaviour of a task
630 * using the address space
631 */
632 if (!(gup_flags & FOLL_FORCE))
633 gup_flags |= FOLL_NUMA;
634
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700635 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700636 struct page *page;
637 unsigned int foll_flags = gup_flags;
638 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700639
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700640 /* first iteration or cross vma bound */
641 if (!vma || start >= vma->vm_end) {
642 vma = find_extend_vma(mm, start);
643 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700644 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700645 ret = get_gate_page(mm, start & PAGE_MASK,
646 gup_flags, &vma,
647 pages ? &pages[i] : NULL);
648 if (ret)
649 return i ? : ret;
650 page_mask = 0;
651 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700652 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700653
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700654 if (!vma || check_vma_flags(vma, gup_flags))
655 return i ? : -EFAULT;
656 if (is_vm_hugetlb_page(vma)) {
657 i = follow_hugetlb_page(mm, vma, pages, vmas,
658 &start, &nr_pages, i,
Andrea Arcangeli87ffc112017-02-22 15:43:13 -0800659 gup_flags, nonblocking);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700660 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700661 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700662 }
663retry:
664 /*
665 * If we have a pending SIGKILL, don't keep faulting pages and
666 * potentially allocating memory.
667 */
668 if (unlikely(fatal_signal_pending(current)))
669 return i ? i : -ERESTARTSYS;
670 cond_resched();
671 page = follow_page_mask(vma, start, foll_flags, &page_mask);
672 if (!page) {
673 int ret;
674 ret = faultin_page(tsk, vma, start, &foll_flags,
675 nonblocking);
676 switch (ret) {
677 case 0:
678 goto retry;
679 case -EFAULT:
680 case -ENOMEM:
681 case -EHWPOISON:
682 return i ? i : ret;
683 case -EBUSY:
684 return i;
685 case -ENOENT:
686 goto next_page;
687 }
688 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700689 } else if (PTR_ERR(page) == -EEXIST) {
690 /*
691 * Proper page table entry exists, but no corresponding
692 * struct page.
693 */
694 goto next_page;
695 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700696 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700697 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700698 if (pages) {
699 pages[i] = page;
700 flush_anon_page(vma, page, start);
701 flush_dcache_page(page);
702 page_mask = 0;
703 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700704next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700705 if (vmas) {
706 vmas[i] = vma;
707 page_mask = 0;
708 }
709 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
710 if (page_increm > nr_pages)
711 page_increm = nr_pages;
712 i += page_increm;
713 start += page_increm * PAGE_SIZE;
714 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700715 } while (nr_pages);
716 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700717}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700718
Tobias Klauser771ab432016-12-12 16:41:53 -0800719static bool vma_permits_fault(struct vm_area_struct *vma,
720 unsigned int fault_flags)
Dave Hansend4925e02016-02-12 13:02:16 -0800721{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800722 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
723 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800724 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800725
726 if (!(vm_flags & vma->vm_flags))
727 return false;
728
Dave Hansen33a709b2016-02-12 13:02:19 -0800729 /*
730 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800731 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800732 *
733 * gup always represents data access, not instruction
734 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800735 */
Dave Hansend61172b2016-02-12 13:02:24 -0800736 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800737 return false;
738
Dave Hansend4925e02016-02-12 13:02:16 -0800739 return true;
740}
741
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700742/*
743 * fixup_user_fault() - manually resolve a user page fault
744 * @tsk: the task_struct to use for page fault accounting, or
745 * NULL if faults are not to be recorded.
746 * @mm: mm_struct of target mm
747 * @address: user address
748 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800749 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
750 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700751 *
752 * This is meant to be called in the specific scenario where for locking reasons
753 * we try to access user memory in atomic context (within a pagefault_disable()
754 * section), this returns -EFAULT, and we want to resolve the user fault before
755 * trying again.
756 *
757 * Typically this is meant to be used by the futex code.
758 *
759 * The main difference with get_user_pages() is that this function will
760 * unconditionally call handle_mm_fault() which will in turn perform all the
761 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800762 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700763 *
764 * This is important for some architectures where those bits also gate the
765 * access permission to the page because they are maintained in software. On
766 * such architectures, gup() will not be enough to make a subsequent access
767 * succeed.
768 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800769 * This function will not return with an unlocked mmap_sem. So it has not the
770 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700771 */
772int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800773 unsigned long address, unsigned int fault_flags,
774 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700775{
776 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800777 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700778
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800779 if (unlocked)
780 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
781
782retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700783 vma = find_extend_vma(mm, address);
784 if (!vma || address < vma->vm_start)
785 return -EFAULT;
786
Dave Hansend4925e02016-02-12 13:02:16 -0800787 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700788 return -EFAULT;
789
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700790 ret = handle_mm_fault(vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800791 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700792 if (ret & VM_FAULT_ERROR) {
James Morse9a291a72017-06-02 14:46:46 -0700793 int err = vm_fault_to_errno(ret, 0);
794
795 if (err)
796 return err;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700797 BUG();
798 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800799
800 if (ret & VM_FAULT_RETRY) {
801 down_read(&mm->mmap_sem);
802 if (!(fault_flags & FAULT_FLAG_TRIED)) {
803 *unlocked = true;
804 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
805 fault_flags |= FAULT_FLAG_TRIED;
806 goto retry;
807 }
808 }
809
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700810 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800811 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700812 tsk->maj_flt++;
813 else
814 tsk->min_flt++;
815 }
816 return 0;
817}
Paolo Bonziniadd6a0c2016-06-07 17:51:18 +0200818EXPORT_SYMBOL_GPL(fixup_user_fault);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700819
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800820static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
821 struct mm_struct *mm,
822 unsigned long start,
823 unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800824 struct page **pages,
825 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800826 int *locked, bool notify_drop,
827 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800828{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800829 long ret, pages_done;
830 bool lock_dropped;
831
832 if (locked) {
833 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
834 BUG_ON(vmas);
835 /* check caller initialized locked */
836 BUG_ON(*locked != 1);
837 }
838
839 if (pages)
840 flags |= FOLL_GET;
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800841
842 pages_done = 0;
843 lock_dropped = false;
844 for (;;) {
845 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
846 vmas, locked);
847 if (!locked)
848 /* VM_FAULT_RETRY couldn't trigger, bypass */
849 return ret;
850
851 /* VM_FAULT_RETRY cannot return errors */
852 if (!*locked) {
853 BUG_ON(ret < 0);
854 BUG_ON(ret >= nr_pages);
855 }
856
857 if (!pages)
858 /* If it's a prefault don't insist harder */
859 return ret;
860
861 if (ret > 0) {
862 nr_pages -= ret;
863 pages_done += ret;
864 if (!nr_pages)
865 break;
866 }
867 if (*locked) {
868 /* VM_FAULT_RETRY didn't trigger */
869 if (!pages_done)
870 pages_done = ret;
871 break;
872 }
873 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
874 pages += ret;
875 start += ret << PAGE_SHIFT;
876
877 /*
878 * Repeat on the address that fired VM_FAULT_RETRY
879 * without FAULT_FLAG_ALLOW_RETRY but with
880 * FAULT_FLAG_TRIED.
881 */
882 *locked = 1;
883 lock_dropped = true;
884 down_read(&mm->mmap_sem);
885 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
886 pages, NULL, NULL);
887 if (ret != 1) {
888 BUG_ON(ret > 1);
889 if (!pages_done)
890 pages_done = ret;
891 break;
892 }
893 nr_pages--;
894 pages_done++;
895 if (!nr_pages)
896 break;
897 pages++;
898 start += PAGE_SIZE;
899 }
900 if (notify_drop && lock_dropped && *locked) {
901 /*
902 * We must let the caller know we temporarily dropped the lock
903 * and so the critical section protected by it was lost.
904 */
905 up_read(&mm->mmap_sem);
906 *locked = 0;
907 }
908 return pages_done;
909}
910
911/*
912 * We can leverage the VM_FAULT_RETRY functionality in the page fault
913 * paths better by using either get_user_pages_locked() or
914 * get_user_pages_unlocked().
915 *
916 * get_user_pages_locked() is suitable to replace the form:
917 *
918 * down_read(&mm->mmap_sem);
919 * do_something()
920 * get_user_pages(tsk, mm, ..., pages, NULL);
921 * up_read(&mm->mmap_sem);
922 *
923 * to:
924 *
925 * int locked = 1;
926 * down_read(&mm->mmap_sem);
927 * do_something()
928 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
929 * if (locked)
930 * up_read(&mm->mmap_sem);
931 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200932long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100933 unsigned int gup_flags, struct page **pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800934 int *locked)
935{
Dave Hansencde70142016-02-12 13:01:55 -0800936 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100937 pages, NULL, locked, true,
938 gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800939}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200940EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800941
942/*
Lorenzo Stoakes80a79512016-12-12 16:42:46 -0800943 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
944 * tsk, mm to be specified.
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800945 *
946 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
Lorenzo Stoakes80a79512016-12-12 16:42:46 -0800947 * caller if required (just like with __get_user_pages). "FOLL_GET"
948 * is set implicitly if "pages" is non-NULL.
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800949 */
Lorenzo Stoakes8b7457e2016-12-14 15:06:55 -0800950static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
951 struct mm_struct *mm, unsigned long start,
952 unsigned long nr_pages, struct page **pages,
953 unsigned int gup_flags)
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800954{
955 long ret;
956 int locked = 1;
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100957
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800958 down_read(&mm->mmap_sem);
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100959 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
960 &locked, false, gup_flags);
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800961 if (locked)
962 up_read(&mm->mmap_sem);
963 return ret;
964}
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800965
966/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800967 * get_user_pages_unlocked() is suitable to replace the form:
968 *
969 * down_read(&mm->mmap_sem);
970 * get_user_pages(tsk, mm, ..., pages, NULL);
971 * up_read(&mm->mmap_sem);
972 *
973 * with:
974 *
975 * get_user_pages_unlocked(tsk, mm, ..., pages);
976 *
977 * It is functionally equivalent to get_user_pages_fast so
Lorenzo Stoakes80a79512016-12-12 16:42:46 -0800978 * get_user_pages_fast should be used instead if specific gup_flags
979 * (e.g. FOLL_FORCE) are not required.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800980 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200981long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100982 struct page **pages, unsigned int gup_flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800983{
Dave Hansencde70142016-02-12 13:01:55 -0800984 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100985 pages, gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800986}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200987EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800988
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700989/*
Dave Hansen1e987792016-02-12 13:01:54 -0800990 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700991 * @tsk: the task_struct to use for page fault accounting, or
992 * NULL if faults are not to be recorded.
993 * @mm: mm_struct of target mm
994 * @start: starting user address
995 * @nr_pages: number of pages from start to pin
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100996 * @gup_flags: flags modifying lookup behaviour
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700997 * @pages: array that receives pointers to the pages pinned.
998 * Should be at least nr_pages long. Or NULL, if caller
999 * only intends to ensure the pages are faulted in.
1000 * @vmas: array of pointers to vmas corresponding to each page.
1001 * Or NULL if the caller does not require them.
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -08001002 * @locked: pointer to lock flag indicating whether lock is held and
1003 * subsequently whether VM_FAULT_RETRY functionality can be
1004 * utilised. Lock must initially be held.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001005 *
1006 * Returns number of pages pinned. This may be fewer than the number
1007 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1008 * were pinned, returns -errno. Each page returned must be released
1009 * with a put_page() call when it is finished with. vmas will only
1010 * remain valid while mmap_sem is held.
1011 *
1012 * Must be called with mmap_sem held for read or write.
1013 *
1014 * get_user_pages walks a process's page tables and takes a reference to
1015 * each struct page that each user address corresponds to at a given
1016 * instant. That is, it takes the page that would be accessed if a user
1017 * thread accesses the given user virtual address at that instant.
1018 *
1019 * This does not guarantee that the page exists in the user mappings when
1020 * get_user_pages returns, and there may even be a completely different
1021 * page there in some cases (eg. if mmapped pagecache has been invalidated
1022 * and subsequently re faulted). However it does guarantee that the page
1023 * won't be freed completely. And mostly callers simply care that the page
1024 * contains data that was valid *at some point in time*. Typically, an IO
1025 * or similar operation cannot guarantee anything stronger anyway because
1026 * locks can't be held over the syscall boundary.
1027 *
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +01001028 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1029 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1030 * be called after the page is finished with, and before put_page is called.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001031 *
1032 * get_user_pages is typically used for fewer-copy IO operations, to get a
1033 * handle on the memory by some means other than accesses via the user virtual
1034 * addresses. The pages may be submitted for DMA to devices or accessed via
1035 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1036 * use the correct cache flushing APIs.
1037 *
1038 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -08001039 *
1040 * get_user_pages should be phased out in favor of
1041 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1042 * should use get_user_pages because it cannot pass
1043 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001044 */
Dave Hansen1e987792016-02-12 13:01:54 -08001045long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1046 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +01001047 unsigned int gup_flags, struct page **pages,
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -08001048 struct vm_area_struct **vmas, int *locked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001049{
Lorenzo Stoakes859110d2016-10-13 01:20:11 +01001050 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -08001051 locked, true,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +01001052 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
Dave Hansen1e987792016-02-12 13:01:54 -08001053}
1054EXPORT_SYMBOL(get_user_pages_remote);
1055
1056/*
Dave Hansend4edcf02016-02-12 13:01:56 -08001057 * This is the same as get_user_pages_remote(), just with a
1058 * less-flexible calling convention where we assume that the task
Lorenzo Stoakes5b56d492016-12-14 15:06:52 -08001059 * and mm being operated on are the current task's and don't allow
1060 * passing of a locked parameter. We also obviously don't pass
1061 * FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -08001062 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001063long get_user_pages(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +01001064 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -08001065 struct vm_area_struct **vmas)
1066{
Dave Hansencde70142016-02-12 13:01:55 -08001067 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +01001068 pages, vmas, NULL, false,
1069 gup_flags | FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001070}
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001071EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001072
1073/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001074 * populate_vma_page_range() - populate a range of pages in the vma.
1075 * @vma: target vma
1076 * @start: start address
1077 * @end: end address
1078 * @nonblocking:
1079 *
1080 * This takes care of mlocking the pages too if VM_LOCKED is set.
1081 *
1082 * return 0 on success, negative error code on error.
1083 *
1084 * vma->vm_mm->mmap_sem must be held.
1085 *
1086 * If @nonblocking is NULL, it may be held for read or write and will
1087 * be unperturbed.
1088 *
1089 * If @nonblocking is non-NULL, it must held for read only and may be
1090 * released. If it's released, *@nonblocking will be set to 0.
1091 */
1092long populate_vma_page_range(struct vm_area_struct *vma,
1093 unsigned long start, unsigned long end, int *nonblocking)
1094{
1095 struct mm_struct *mm = vma->vm_mm;
1096 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1097 int gup_flags;
1098
1099 VM_BUG_ON(start & ~PAGE_MASK);
1100 VM_BUG_ON(end & ~PAGE_MASK);
1101 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1102 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1103 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1104
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001105 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1106 if (vma->vm_flags & VM_LOCKONFAULT)
1107 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001108 /*
1109 * We want to touch writable mappings with a write fault in order
1110 * to break COW, except for shared mappings because these don't COW
1111 * and we would not want to dirty them for nothing.
1112 */
1113 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1114 gup_flags |= FOLL_WRITE;
1115
1116 /*
1117 * We want mlock to succeed for regions that have any permissions
1118 * other than PROT_NONE.
1119 */
1120 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1121 gup_flags |= FOLL_FORCE;
1122
1123 /*
1124 * We made sure addr is within a VMA, so the following will
1125 * not result in a stack expansion that recurses back here.
1126 */
1127 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1128 NULL, NULL, nonblocking);
1129}
1130
1131/*
1132 * __mm_populate - populate and/or mlock pages within a range of address space.
1133 *
1134 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1135 * flags. VMAs must be already marked with the desired vm_flags, and
1136 * mmap_sem must not be held.
1137 */
1138int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1139{
1140 struct mm_struct *mm = current->mm;
1141 unsigned long end, nstart, nend;
1142 struct vm_area_struct *vma = NULL;
1143 int locked = 0;
1144 long ret = 0;
1145
1146 VM_BUG_ON(start & ~PAGE_MASK);
1147 VM_BUG_ON(len != PAGE_ALIGN(len));
1148 end = start + len;
1149
1150 for (nstart = start; nstart < end; nstart = nend) {
1151 /*
1152 * We want to fault in pages for [nstart; end) address range.
1153 * Find first corresponding VMA.
1154 */
1155 if (!locked) {
1156 locked = 1;
1157 down_read(&mm->mmap_sem);
1158 vma = find_vma(mm, nstart);
1159 } else if (nstart >= vma->vm_end)
1160 vma = vma->vm_next;
1161 if (!vma || vma->vm_start >= end)
1162 break;
1163 /*
1164 * Set [nstart; nend) to intersection of desired address
1165 * range with the first VMA. Also, skip undesirable VMA types.
1166 */
1167 nend = min(end, vma->vm_end);
1168 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1169 continue;
1170 if (nstart < vma->vm_start)
1171 nstart = vma->vm_start;
1172 /*
1173 * Now fault in a range of pages. populate_vma_page_range()
1174 * double checks the vma flags, so that it won't mlock pages
1175 * if the vma was already munlocked.
1176 */
1177 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1178 if (ret < 0) {
1179 if (ignore_errors) {
1180 ret = 0;
1181 continue; /* continue at next VMA */
1182 }
1183 break;
1184 }
1185 nend = nstart + ret * PAGE_SIZE;
1186 ret = 0;
1187 }
1188 if (locked)
1189 up_read(&mm->mmap_sem);
1190 return ret; /* 0 or negative error code */
1191}
1192
1193/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001194 * get_dump_page() - pin user page in memory while writing it to core dump
1195 * @addr: user address
1196 *
1197 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001198 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001199 *
1200 * Returns NULL on any kind of failure - a hole must then be inserted into
1201 * the corefile, to preserve alignment with its headers; and also returns
1202 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1203 * allowing a hole to be left in the corefile to save diskspace.
1204 *
1205 * Called without mmap_sem, but after all other threads have been killed.
1206 */
1207#ifdef CONFIG_ELF_CORE
1208struct page *get_dump_page(unsigned long addr)
1209{
1210 struct vm_area_struct *vma;
1211 struct page *page;
1212
1213 if (__get_user_pages(current, current->mm, addr, 1,
1214 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1215 NULL) < 1)
1216 return NULL;
1217 flush_cache_page(vma, addr, page_to_pfn(page));
1218 return page;
1219}
1220#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001221
1222/*
Kirill A. Shutemove5855132017-06-06 14:31:20 +03001223 * Generic Fast GUP
Steve Capper2667f502014-10-09 15:29:14 -07001224 *
1225 * get_user_pages_fast attempts to pin user pages by walking the page
1226 * tables directly and avoids taking locks. Thus the walker needs to be
1227 * protected from page table pages being freed from under it, and should
1228 * block any THP splits.
1229 *
1230 * One way to achieve this is to have the walker disable interrupts, and
1231 * rely on IPIs from the TLB flushing code blocking before the page table
1232 * pages are freed. This is unsuitable for architectures that do not need
1233 * to broadcast an IPI when invalidating TLBs.
1234 *
1235 * Another way to achieve this is to batch up page table containing pages
1236 * belonging to more than one mm_user, then rcu_sched a callback to free those
1237 * pages. Disabling interrupts will allow the fast_gup walker to both block
1238 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1239 * (which is a relatively rare event). The code below adopts this strategy.
1240 *
1241 * Before activating this code, please be aware that the following assumptions
1242 * are currently made:
1243 *
Kirill A. Shutemove5855132017-06-06 14:31:20 +03001244 * *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1245 * free pages containing page tables or TLB flushing requires IPI broadcast.
Steve Capper2667f502014-10-09 15:29:14 -07001246 *
Steve Capper2667f502014-10-09 15:29:14 -07001247 * *) ptes can be read atomically by the architecture.
1248 *
1249 * *) access_ok is sufficient to validate userspace address ranges.
1250 *
1251 * The last two assumptions can be relaxed by the addition of helper functions.
1252 *
1253 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1254 */
Kirill A. Shutemove5855132017-06-06 14:31:20 +03001255#ifdef CONFIG_HAVE_GENERIC_GUP
Steve Capper2667f502014-10-09 15:29:14 -07001256
Kirill A. Shutemov0005d202017-03-16 18:26:51 +03001257#ifndef gup_get_pte
1258/*
1259 * We assume that the PTE can be read atomically. If this is not the case for
1260 * your architecture, please provide the helper.
1261 */
1262static inline pte_t gup_get_pte(pte_t *ptep)
1263{
1264 return READ_ONCE(*ptep);
1265}
1266#endif
1267
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001268static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1269{
1270 while ((*nr) - nr_start) {
1271 struct page *page = pages[--(*nr)];
1272
1273 ClearPageReferenced(page);
1274 put_page(page);
1275 }
1276}
1277
Steve Capper2667f502014-10-09 15:29:14 -07001278#ifdef __HAVE_ARCH_PTE_SPECIAL
1279static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1280 int write, struct page **pages, int *nr)
1281{
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001282 struct dev_pagemap *pgmap = NULL;
1283 int nr_start = *nr, ret = 0;
Steve Capper2667f502014-10-09 15:29:14 -07001284 pte_t *ptep, *ptem;
Steve Capper2667f502014-10-09 15:29:14 -07001285
1286 ptem = ptep = pte_offset_map(&pmd, addr);
1287 do {
Kirill A. Shutemov0005d202017-03-16 18:26:51 +03001288 pte_t pte = gup_get_pte(ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001289 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001290
1291 /*
1292 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001293 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001294 */
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001295 if (pte_protnone(pte))
1296 goto pte_unmap;
1297
1298 if (!pte_access_permitted(pte, write))
1299 goto pte_unmap;
1300
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001301 if (pte_devmap(pte)) {
1302 pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1303 if (unlikely(!pgmap)) {
1304 undo_dev_pagemap(nr, nr_start, pages);
1305 goto pte_unmap;
1306 }
1307 } else if (pte_special(pte))
Steve Capper2667f502014-10-09 15:29:14 -07001308 goto pte_unmap;
1309
1310 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1311 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001312 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001313
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001314 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001315 goto pte_unmap;
1316
1317 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001318 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001319 goto pte_unmap;
1320 }
1321
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001322 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001323
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001324 put_dev_pagemap(pgmap);
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001325 SetPageReferenced(page);
Steve Capper2667f502014-10-09 15:29:14 -07001326 pages[*nr] = page;
1327 (*nr)++;
1328
1329 } while (ptep++, addr += PAGE_SIZE, addr != end);
1330
1331 ret = 1;
1332
1333pte_unmap:
1334 pte_unmap(ptem);
1335 return ret;
1336}
1337#else
1338
1339/*
1340 * If we can't determine whether or not a pte is special, then fail immediately
1341 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1342 * to be special.
1343 *
1344 * For a futex to be placed on a THP tail page, get_futex_key requires a
1345 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1346 * useful to have gup_huge_pmd even if we can't operate on ptes.
1347 */
1348static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1349 int write, struct page **pages, int *nr)
1350{
1351 return 0;
1352}
1353#endif /* __HAVE_ARCH_PTE_SPECIAL */
1354
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001355#ifdef __HAVE_ARCH_PTE_DEVMAP
1356static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1357 unsigned long end, struct page **pages, int *nr)
1358{
1359 int nr_start = *nr;
1360 struct dev_pagemap *pgmap = NULL;
1361
1362 do {
1363 struct page *page = pfn_to_page(pfn);
1364
1365 pgmap = get_dev_pagemap(pfn, pgmap);
1366 if (unlikely(!pgmap)) {
1367 undo_dev_pagemap(nr, nr_start, pages);
1368 return 0;
1369 }
1370 SetPageReferenced(page);
1371 pages[*nr] = page;
1372 get_page(page);
1373 put_dev_pagemap(pgmap);
1374 (*nr)++;
1375 pfn++;
1376 } while (addr += PAGE_SIZE, addr != end);
1377 return 1;
1378}
1379
1380static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1381 unsigned long end, struct page **pages, int *nr)
1382{
1383 unsigned long fault_pfn;
1384
1385 fault_pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1386 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1387}
1388
1389static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1390 unsigned long end, struct page **pages, int *nr)
1391{
1392 unsigned long fault_pfn;
1393
1394 fault_pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1395 return __gup_device_huge(fault_pfn, addr, end, pages, nr);
1396}
1397#else
1398static int __gup_device_huge_pmd(pmd_t pmd, unsigned long addr,
1399 unsigned long end, struct page **pages, int *nr)
1400{
1401 BUILD_BUG();
1402 return 0;
1403}
1404
1405static int __gup_device_huge_pud(pud_t pud, unsigned long addr,
1406 unsigned long end, struct page **pages, int *nr)
1407{
1408 BUILD_BUG();
1409 return 0;
1410}
1411#endif
1412
Steve Capper2667f502014-10-09 15:29:14 -07001413static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1414 unsigned long end, int write, struct page **pages, int *nr)
1415{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001416 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001417 int refs;
1418
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001419 if (!pmd_access_permitted(orig, write))
Steve Capper2667f502014-10-09 15:29:14 -07001420 return 0;
1421
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001422 if (pmd_devmap(orig))
1423 return __gup_device_huge_pmd(orig, addr, end, pages, nr);
1424
Steve Capper2667f502014-10-09 15:29:14 -07001425 refs = 0;
Punit Agrawald63206e2017-07-06 15:39:39 -07001426 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001427 do {
Steve Capper2667f502014-10-09 15:29:14 -07001428 pages[*nr] = page;
1429 (*nr)++;
1430 page++;
1431 refs++;
1432 } while (addr += PAGE_SIZE, addr != end);
1433
Punit Agrawald63206e2017-07-06 15:39:39 -07001434 head = compound_head(pmd_page(orig));
Steve Capper2667f502014-10-09 15:29:14 -07001435 if (!page_cache_add_speculative(head, refs)) {
1436 *nr -= refs;
1437 return 0;
1438 }
1439
1440 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1441 *nr -= refs;
1442 while (refs--)
1443 put_page(head);
1444 return 0;
1445 }
1446
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001447 SetPageReferenced(head);
Steve Capper2667f502014-10-09 15:29:14 -07001448 return 1;
1449}
1450
1451static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1452 unsigned long end, int write, struct page **pages, int *nr)
1453{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001454 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001455 int refs;
1456
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001457 if (!pud_access_permitted(orig, write))
Steve Capper2667f502014-10-09 15:29:14 -07001458 return 0;
1459
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001460 if (pud_devmap(orig))
1461 return __gup_device_huge_pud(orig, addr, end, pages, nr);
1462
Steve Capper2667f502014-10-09 15:29:14 -07001463 refs = 0;
Punit Agrawald63206e2017-07-06 15:39:39 -07001464 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001465 do {
Steve Capper2667f502014-10-09 15:29:14 -07001466 pages[*nr] = page;
1467 (*nr)++;
1468 page++;
1469 refs++;
1470 } while (addr += PAGE_SIZE, addr != end);
1471
Punit Agrawald63206e2017-07-06 15:39:39 -07001472 head = compound_head(pud_page(orig));
Steve Capper2667f502014-10-09 15:29:14 -07001473 if (!page_cache_add_speculative(head, refs)) {
1474 *nr -= refs;
1475 return 0;
1476 }
1477
1478 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1479 *nr -= refs;
1480 while (refs--)
1481 put_page(head);
1482 return 0;
1483 }
1484
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001485 SetPageReferenced(head);
Steve Capper2667f502014-10-09 15:29:14 -07001486 return 1;
1487}
1488
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301489static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1490 unsigned long end, int write,
1491 struct page **pages, int *nr)
1492{
1493 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001494 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301495
Kirill A. Shutemove7884f82017-03-16 18:26:50 +03001496 if (!pgd_access_permitted(orig, write))
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301497 return 0;
1498
Kirill A. Shutemovb59f65f2017-03-16 18:26:53 +03001499 BUILD_BUG_ON(pgd_devmap(orig));
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301500 refs = 0;
Punit Agrawald63206e2017-07-06 15:39:39 -07001501 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301502 do {
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301503 pages[*nr] = page;
1504 (*nr)++;
1505 page++;
1506 refs++;
1507 } while (addr += PAGE_SIZE, addr != end);
1508
Punit Agrawald63206e2017-07-06 15:39:39 -07001509 head = compound_head(pgd_page(orig));
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301510 if (!page_cache_add_speculative(head, refs)) {
1511 *nr -= refs;
1512 return 0;
1513 }
1514
1515 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1516 *nr -= refs;
1517 while (refs--)
1518 put_page(head);
1519 return 0;
1520 }
1521
Kirill A. Shutemove9348052017-03-16 18:26:52 +03001522 SetPageReferenced(head);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301523 return 1;
1524}
1525
Steve Capper2667f502014-10-09 15:29:14 -07001526static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1527 int write, struct page **pages, int *nr)
1528{
1529 unsigned long next;
1530 pmd_t *pmdp;
1531
1532 pmdp = pmd_offset(&pud, addr);
1533 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001534 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001535
1536 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001537 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001538 return 0;
1539
1540 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1541 /*
1542 * NUMA hinting faults need to be handled in the GUP
1543 * slowpath for accounting purposes and so that they
1544 * can be serialised against THP migration.
1545 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001546 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001547 return 0;
1548
1549 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1550 pages, nr))
1551 return 0;
1552
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301553 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1554 /*
1555 * architecture have different format for hugetlbfs
1556 * pmd format and THP pmd format
1557 */
1558 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1559 PMD_SHIFT, next, write, pages, nr))
1560 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001561 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1562 return 0;
1563 } while (pmdp++, addr = next, addr != end);
1564
1565 return 1;
1566}
1567
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001568static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301569 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001570{
1571 unsigned long next;
1572 pud_t *pudp;
1573
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001574 pudp = pud_offset(&p4d, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001575 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001576 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001577
1578 next = pud_addr_end(addr, end);
1579 if (pud_none(pud))
1580 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301581 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001582 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301583 pages, nr))
1584 return 0;
1585 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1586 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1587 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001588 return 0;
1589 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1590 return 0;
1591 } while (pudp++, addr = next, addr != end);
1592
1593 return 1;
1594}
1595
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001596static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1597 int write, struct page **pages, int *nr)
1598{
1599 unsigned long next;
1600 p4d_t *p4dp;
1601
1602 p4dp = p4d_offset(&pgd, addr);
1603 do {
1604 p4d_t p4d = READ_ONCE(*p4dp);
1605
1606 next = p4d_addr_end(addr, end);
1607 if (p4d_none(p4d))
1608 return 0;
1609 BUILD_BUG_ON(p4d_huge(p4d));
1610 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1611 if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1612 P4D_SHIFT, next, write, pages, nr))
1613 return 0;
Kirill A. Shutemovce70df02017-03-13 08:22:13 +03001614 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001615 return 0;
1616 } while (p4dp++, addr = next, addr != end);
1617
1618 return 1;
1619}
1620
Steve Capper2667f502014-10-09 15:29:14 -07001621/*
1622 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1623 * the regular GUP. It will only return non-negative values.
1624 */
1625int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1626 struct page **pages)
1627{
1628 struct mm_struct *mm = current->mm;
1629 unsigned long addr, len, end;
1630 unsigned long next, flags;
1631 pgd_t *pgdp;
1632 int nr = 0;
1633
1634 start &= PAGE_MASK;
1635 addr = start;
1636 len = (unsigned long) nr_pages << PAGE_SHIFT;
1637 end = start + len;
1638
1639 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
Arnd Bergmannaa2369f2017-05-03 14:56:12 -07001640 (void __user *)start, len)))
Steve Capper2667f502014-10-09 15:29:14 -07001641 return 0;
1642
1643 /*
1644 * Disable interrupts. We use the nested form as we can already have
1645 * interrupts disabled by get_futex_key.
1646 *
1647 * With interrupts disabled, we block page table pages from being
1648 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1649 * for more details.
1650 *
1651 * We do not adopt an rcu_read_lock(.) here as we also want to
1652 * block IPIs that come from THPs splitting.
1653 */
1654
1655 local_irq_save(flags);
1656 pgdp = pgd_offset(mm, addr);
1657 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001658 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301659
Steve Capper2667f502014-10-09 15:29:14 -07001660 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301661 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001662 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301663 if (unlikely(pgd_huge(pgd))) {
1664 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1665 pages, &nr))
1666 break;
1667 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1668 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1669 PGDIR_SHIFT, next, write, pages, &nr))
1670 break;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03001671 } else if (!gup_p4d_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001672 break;
1673 } while (pgdp++, addr = next, addr != end);
1674 local_irq_restore(flags);
1675
1676 return nr;
1677}
1678
Kirill A. Shutemov73e10a62017-03-16 18:26:54 +03001679#ifndef gup_fast_permitted
1680/*
1681 * Check if it's allowed to use __get_user_pages_fast() for the range, or
1682 * we need to fall back to the slow version:
1683 */
1684bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1685{
1686 unsigned long len, end;
1687
1688 len = (unsigned long) nr_pages << PAGE_SHIFT;
1689 end = start + len;
1690 return end >= start;
1691}
1692#endif
1693
Steve Capper2667f502014-10-09 15:29:14 -07001694/**
1695 * get_user_pages_fast() - pin user pages in memory
1696 * @start: starting user address
1697 * @nr_pages: number of pages from start to pin
1698 * @write: whether pages will be written to
1699 * @pages: array that receives pointers to the pages pinned.
1700 * Should be at least nr_pages long.
1701 *
1702 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1703 * If not successful, it will fall back to taking the lock and
1704 * calling get_user_pages().
1705 *
1706 * Returns number of pages pinned. This may be fewer than the number
1707 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1708 * were pinned, returns -errno.
1709 */
1710int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1711 struct page **pages)
1712{
Kirill A. Shutemov73e10a62017-03-16 18:26:54 +03001713 int nr = 0, ret = 0;
Steve Capper2667f502014-10-09 15:29:14 -07001714
1715 start &= PAGE_MASK;
Kirill A. Shutemov73e10a62017-03-16 18:26:54 +03001716
1717 if (gup_fast_permitted(start, nr_pages, write)) {
1718 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1719 ret = nr;
1720 }
Steve Capper2667f502014-10-09 15:29:14 -07001721
1722 if (nr < nr_pages) {
1723 /* Try to get the remaining pages with get_user_pages */
1724 start += nr << PAGE_SHIFT;
1725 pages += nr;
1726
Lorenzo Stoakesc1641542016-10-13 01:20:13 +01001727 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1728 write ? FOLL_WRITE : 0);
Steve Capper2667f502014-10-09 15:29:14 -07001729
1730 /* Have to be a bit careful with return values */
1731 if (nr > 0) {
1732 if (ret < 0)
1733 ret = nr;
1734 else
1735 ret += nr;
1736 }
1737 }
1738
1739 return ret;
1740}
1741
Kirill A. Shutemove5855132017-06-06 14:31:20 +03001742#endif /* CONFIG_HAVE_GENERIC_GUP */