blob: be4ccddac26f8e075fcda887a9f07677e4ee1d1b [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
Steve Capper2667f502014-10-09 15:29:14 -070013#include <linux/sched.h>
14#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;
229 pud_t *pud;
230 pmd_t *pmd;
231 spinlock_t *ptl;
232 struct page *page;
233 struct mm_struct *mm = vma->vm_mm;
234
235 *page_mask = 0;
236
237 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
238 if (!IS_ERR(page)) {
239 BUG_ON(flags & FOLL_GET);
240 return page;
241 }
242
243 pgd = pgd_offset(mm, address);
244 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
245 return no_page_table(vma, flags);
246
247 pud = pud_offset(pgd, address);
248 if (pud_none(*pud))
249 return no_page_table(vma, flags);
250 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800251 page = follow_huge_pud(mm, address, pud, flags);
252 if (page)
253 return page;
254 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700255 }
256 if (unlikely(pud_bad(*pud)))
257 return no_page_table(vma, flags);
258
259 pmd = pmd_offset(pud, address);
260 if (pmd_none(*pmd))
261 return no_page_table(vma, flags);
262 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800263 page = follow_huge_pmd(mm, address, pmd, flags);
264 if (page)
265 return page;
266 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700267 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800268 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700269 return no_page_table(vma, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800270 if (pmd_devmap(*pmd)) {
271 ptl = pmd_lock(mm, pmd);
272 page = follow_devmap_pmd(vma, address, pmd, flags);
273 spin_unlock(ptl);
274 if (page)
275 return page;
276 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800277 if (likely(!pmd_trans_huge(*pmd)))
278 return follow_page_pte(vma, address, pmd, flags);
279
280 ptl = pmd_lock(mm, pmd);
281 if (unlikely(!pmd_trans_huge(*pmd))) {
282 spin_unlock(ptl);
283 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700284 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800285 if (flags & FOLL_SPLIT) {
286 int ret;
287 page = pmd_page(*pmd);
288 if (is_huge_zero_page(page)) {
289 spin_unlock(ptl);
290 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800291 split_huge_pmd(vma, pmd, address);
Naoya Horiguchi337d9ab2016-07-26 15:24:03 -0700292 if (pmd_trans_unstable(pmd))
293 ret = -EBUSY;
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800294 } else {
295 get_page(page);
296 spin_unlock(ptl);
297 lock_page(page);
298 ret = split_huge_page(page);
299 unlock_page(page);
300 put_page(page);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -0700301 if (pmd_none(*pmd))
302 return no_page_table(vma, flags);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800303 }
304
305 return ret ? ERR_PTR(ret) :
306 follow_page_pte(vma, address, pmd, flags);
307 }
308
309 page = follow_trans_huge_pmd(vma, address, pmd, flags);
310 spin_unlock(ptl);
311 *page_mask = HPAGE_PMD_NR - 1;
312 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700313}
314
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700315static int get_gate_page(struct mm_struct *mm, unsigned long address,
316 unsigned int gup_flags, struct vm_area_struct **vma,
317 struct page **page)
318{
319 pgd_t *pgd;
320 pud_t *pud;
321 pmd_t *pmd;
322 pte_t *pte;
323 int ret = -EFAULT;
324
325 /* user gate pages are read-only */
326 if (gup_flags & FOLL_WRITE)
327 return -EFAULT;
328 if (address > TASK_SIZE)
329 pgd = pgd_offset_k(address);
330 else
331 pgd = pgd_offset_gate(mm, address);
332 BUG_ON(pgd_none(*pgd));
333 pud = pud_offset(pgd, address);
334 BUG_ON(pud_none(*pud));
335 pmd = pmd_offset(pud, address);
336 if (pmd_none(*pmd))
337 return -EFAULT;
338 VM_BUG_ON(pmd_trans_huge(*pmd));
339 pte = pte_offset_map(pmd, address);
340 if (pte_none(*pte))
341 goto unmap;
342 *vma = get_gate_vma(mm);
343 if (!page)
344 goto out;
345 *page = vm_normal_page(*vma, address, *pte);
346 if (!*page) {
347 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
348 goto unmap;
349 *page = pte_page(*pte);
350 }
351 get_page(*page);
352out:
353 ret = 0;
354unmap:
355 pte_unmap(pte);
356 return ret;
357}
358
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700359/*
360 * mmap_sem must be held on entry. If @nonblocking != NULL and
361 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
362 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
363 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700364static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
365 unsigned long address, unsigned int *flags, int *nonblocking)
366{
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700367 unsigned int fault_flags = 0;
368 int ret;
369
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800370 /* mlock all present pages, but do not fault in new pages */
371 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
372 return -ENOENT;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700373 if (*flags & FOLL_WRITE)
374 fault_flags |= FAULT_FLAG_WRITE;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800375 if (*flags & FOLL_REMOTE)
376 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700377 if (nonblocking)
378 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
379 if (*flags & FOLL_NOWAIT)
380 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700381 if (*flags & FOLL_TRIED) {
382 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
383 fault_flags |= FAULT_FLAG_TRIED;
384 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700385
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700386 ret = handle_mm_fault(vma, address, fault_flags);
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700387 if (ret & VM_FAULT_ERROR) {
388 if (ret & VM_FAULT_OOM)
389 return -ENOMEM;
390 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
391 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800392 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700393 return -EFAULT;
394 BUG();
395 }
396
397 if (tsk) {
398 if (ret & VM_FAULT_MAJOR)
399 tsk->maj_flt++;
400 else
401 tsk->min_flt++;
402 }
403
404 if (ret & VM_FAULT_RETRY) {
405 if (nonblocking)
406 *nonblocking = 0;
407 return -EBUSY;
408 }
409
410 /*
411 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
412 * necessary, even if maybe_mkwrite decided not to set pte_write. We
413 * can thus safely do subsequent page lookups as if they were reads.
414 * But only do so when looping for pte_write is futile: in some cases
415 * userspace may also be wanting to write to the gotten user page,
416 * which a read fault here might prevent (a readonly page might get
417 * reCOWed by userspace write).
418 */
419 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700420 *flags |= FOLL_COW;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700421 return 0;
422}
423
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700424static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
425{
426 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800427 int write = (gup_flags & FOLL_WRITE);
428 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700429
430 if (vm_flags & (VM_IO | VM_PFNMAP))
431 return -EFAULT;
432
Willy Tarreau6f1abf82018-05-11 08:11:44 +0200433 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
434 return -EFAULT;
435
Dave Hansen1b2ee122016-02-12 13:02:21 -0800436 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700437 if (!(vm_flags & VM_WRITE)) {
438 if (!(gup_flags & FOLL_FORCE))
439 return -EFAULT;
440 /*
441 * We used to let the write,force case do COW in a
442 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
443 * set a breakpoint in a read-only mapping of an
444 * executable, without corrupting the file (yet only
445 * when that file had been opened for writing!).
446 * Anon pages in shared mappings are surprising: now
447 * just reject it.
448 */
Hugh Dickins46435362016-01-30 18:03:16 -0800449 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700450 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700451 }
452 } else if (!(vm_flags & VM_READ)) {
453 if (!(gup_flags & FOLL_FORCE))
454 return -EFAULT;
455 /*
456 * Is there actually any vma we can reach here which does not
457 * have VM_MAYREAD set?
458 */
459 if (!(vm_flags & VM_MAYREAD))
460 return -EFAULT;
461 }
Dave Hansend61172b2016-02-12 13:02:24 -0800462 /*
463 * gups are always data accesses, not instruction
464 * fetches, so execute=false here
465 */
466 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800467 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700468 return 0;
469}
470
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700471/**
472 * __get_user_pages() - pin user pages in memory
473 * @tsk: task_struct of target task
474 * @mm: mm_struct of target mm
475 * @start: starting user address
476 * @nr_pages: number of pages from start to pin
477 * @gup_flags: flags modifying pin behaviour
478 * @pages: array that receives pointers to the pages pinned.
479 * Should be at least nr_pages long. Or NULL, if caller
480 * only intends to ensure the pages are faulted in.
481 * @vmas: array of pointers to vmas corresponding to each page.
482 * Or NULL if the caller does not require them.
483 * @nonblocking: whether waiting for disk IO or mmap_sem contention
484 *
485 * Returns number of pages pinned. This may be fewer than the number
486 * requested. If nr_pages is 0 or negative, returns 0. If no pages
487 * were pinned, returns -errno. Each page returned must be released
488 * with a put_page() call when it is finished with. vmas will only
489 * remain valid while mmap_sem is held.
490 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700491 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700492 *
493 * __get_user_pages walks a process's page tables and takes a reference to
494 * each struct page that each user address corresponds to at a given
495 * instant. That is, it takes the page that would be accessed if a user
496 * thread accesses the given user virtual address at that instant.
497 *
498 * This does not guarantee that the page exists in the user mappings when
499 * __get_user_pages returns, and there may even be a completely different
500 * page there in some cases (eg. if mmapped pagecache has been invalidated
501 * and subsequently re faulted). However it does guarantee that the page
502 * won't be freed completely. And mostly callers simply care that the page
503 * contains data that was valid *at some point in time*. Typically, an IO
504 * or similar operation cannot guarantee anything stronger anyway because
505 * locks can't be held over the syscall boundary.
506 *
507 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
508 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
509 * appropriate) must be called after the page is finished with, and
510 * before put_page is called.
511 *
512 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
513 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700514 * *@nonblocking will be set to 0. Further, if @gup_flags does not
515 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
516 * this case.
517 *
518 * A caller using such a combination of @nonblocking and @gup_flags
519 * must therefore hold the mmap_sem for reading only, and recognize
520 * when it's been released. Otherwise, it must be held for either
521 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700522 *
523 * In most cases, get_user_pages or get_user_pages_fast should be used
524 * instead of __get_user_pages. __get_user_pages should be used only if
525 * you need some special @gup_flags.
526 */
Lorenzo Stoakes0d731752016-10-24 10:57:25 +0100527static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700528 unsigned long start, unsigned long nr_pages,
529 unsigned int gup_flags, struct page **pages,
530 struct vm_area_struct **vmas, int *nonblocking)
531{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700532 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700533 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700534 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700535
536 if (!nr_pages)
537 return 0;
538
539 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
540
541 /*
542 * If FOLL_FORCE is set then do not force a full fault as the hinting
543 * fault information is unrelated to the reference behaviour of a task
544 * using the address space
545 */
546 if (!(gup_flags & FOLL_FORCE))
547 gup_flags |= FOLL_NUMA;
548
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700549 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700550 struct page *page;
551 unsigned int foll_flags = gup_flags;
552 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700553
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700554 /* first iteration or cross vma bound */
555 if (!vma || start >= vma->vm_end) {
556 vma = find_extend_vma(mm, start);
557 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700558 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700559 ret = get_gate_page(mm, start & PAGE_MASK,
560 gup_flags, &vma,
561 pages ? &pages[i] : NULL);
562 if (ret)
563 return i ? : ret;
564 page_mask = 0;
565 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700566 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700567
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700568 if (!vma || check_vma_flags(vma, gup_flags))
569 return i ? : -EFAULT;
570 if (is_vm_hugetlb_page(vma)) {
571 i = follow_hugetlb_page(mm, vma, pages, vmas,
572 &start, &nr_pages, i,
573 gup_flags);
574 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700575 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700576 }
577retry:
578 /*
579 * If we have a pending SIGKILL, don't keep faulting pages and
580 * potentially allocating memory.
581 */
582 if (unlikely(fatal_signal_pending(current)))
583 return i ? i : -ERESTARTSYS;
584 cond_resched();
585 page = follow_page_mask(vma, start, foll_flags, &page_mask);
586 if (!page) {
587 int ret;
588 ret = faultin_page(tsk, vma, start, &foll_flags,
589 nonblocking);
590 switch (ret) {
591 case 0:
592 goto retry;
593 case -EFAULT:
594 case -ENOMEM:
595 case -EHWPOISON:
596 return i ? i : ret;
597 case -EBUSY:
598 return i;
599 case -ENOENT:
600 goto next_page;
601 }
602 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700603 } else if (PTR_ERR(page) == -EEXIST) {
604 /*
605 * Proper page table entry exists, but no corresponding
606 * struct page.
607 */
608 goto next_page;
609 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700610 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700611 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700612 if (pages) {
613 pages[i] = page;
614 flush_anon_page(vma, page, start);
615 flush_dcache_page(page);
616 page_mask = 0;
617 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700618next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700619 if (vmas) {
620 vmas[i] = vma;
621 page_mask = 0;
622 }
623 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
624 if (page_increm > nr_pages)
625 page_increm = nr_pages;
626 i += page_increm;
627 start += page_increm * PAGE_SIZE;
628 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700629 } while (nr_pages);
630 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700631}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700632
Dave Hansend4925e02016-02-12 13:02:16 -0800633bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
634{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800635 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
636 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800637 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800638
639 if (!(vm_flags & vma->vm_flags))
640 return false;
641
Dave Hansen33a709b2016-02-12 13:02:19 -0800642 /*
643 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800644 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800645 *
646 * gup always represents data access, not instruction
647 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800648 */
Dave Hansend61172b2016-02-12 13:02:24 -0800649 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800650 return false;
651
Dave Hansend4925e02016-02-12 13:02:16 -0800652 return true;
653}
654
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700655/*
656 * fixup_user_fault() - manually resolve a user page fault
657 * @tsk: the task_struct to use for page fault accounting, or
658 * NULL if faults are not to be recorded.
659 * @mm: mm_struct of target mm
660 * @address: user address
661 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800662 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
663 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700664 *
665 * This is meant to be called in the specific scenario where for locking reasons
666 * we try to access user memory in atomic context (within a pagefault_disable()
667 * section), this returns -EFAULT, and we want to resolve the user fault before
668 * trying again.
669 *
670 * Typically this is meant to be used by the futex code.
671 *
672 * The main difference with get_user_pages() is that this function will
673 * unconditionally call handle_mm_fault() which will in turn perform all the
674 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800675 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700676 *
677 * This is important for some architectures where those bits also gate the
678 * access permission to the page because they are maintained in software. On
679 * such architectures, gup() will not be enough to make a subsequent access
680 * succeed.
681 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800682 * This function will not return with an unlocked mmap_sem. So it has not the
683 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700684 */
685int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800686 unsigned long address, unsigned int fault_flags,
687 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700688{
689 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800690 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700691
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800692 if (unlocked)
693 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
694
695retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700696 vma = find_extend_vma(mm, address);
697 if (!vma || address < vma->vm_start)
698 return -EFAULT;
699
Dave Hansend4925e02016-02-12 13:02:16 -0800700 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700701 return -EFAULT;
702
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700703 ret = handle_mm_fault(vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800704 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700705 if (ret & VM_FAULT_ERROR) {
706 if (ret & VM_FAULT_OOM)
707 return -ENOMEM;
708 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
709 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800710 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700711 return -EFAULT;
712 BUG();
713 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800714
715 if (ret & VM_FAULT_RETRY) {
716 down_read(&mm->mmap_sem);
717 if (!(fault_flags & FAULT_FLAG_TRIED)) {
718 *unlocked = true;
719 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
720 fault_flags |= FAULT_FLAG_TRIED;
721 goto retry;
722 }
723 }
724
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700725 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800726 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700727 tsk->maj_flt++;
728 else
729 tsk->min_flt++;
730 }
731 return 0;
732}
Paolo Bonziniadd6a0c2016-06-07 17:51:18 +0200733EXPORT_SYMBOL_GPL(fixup_user_fault);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700734
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800735static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
736 struct mm_struct *mm,
737 unsigned long start,
738 unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800739 struct page **pages,
740 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800741 int *locked, bool notify_drop,
742 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800743{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800744 long ret, pages_done;
745 bool lock_dropped;
746
747 if (locked) {
748 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
749 BUG_ON(vmas);
750 /* check caller initialized locked */
751 BUG_ON(*locked != 1);
752 }
753
754 if (pages)
755 flags |= FOLL_GET;
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800756
757 pages_done = 0;
758 lock_dropped = false;
759 for (;;) {
760 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
761 vmas, locked);
762 if (!locked)
763 /* VM_FAULT_RETRY couldn't trigger, bypass */
764 return ret;
765
766 /* VM_FAULT_RETRY cannot return errors */
767 if (!*locked) {
768 BUG_ON(ret < 0);
769 BUG_ON(ret >= nr_pages);
770 }
771
772 if (!pages)
773 /* If it's a prefault don't insist harder */
774 return ret;
775
776 if (ret > 0) {
777 nr_pages -= ret;
778 pages_done += ret;
779 if (!nr_pages)
780 break;
781 }
782 if (*locked) {
783 /* VM_FAULT_RETRY didn't trigger */
784 if (!pages_done)
785 pages_done = ret;
786 break;
787 }
788 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
789 pages += ret;
790 start += ret << PAGE_SHIFT;
791
792 /*
793 * Repeat on the address that fired VM_FAULT_RETRY
794 * without FAULT_FLAG_ALLOW_RETRY but with
795 * FAULT_FLAG_TRIED.
796 */
797 *locked = 1;
798 lock_dropped = true;
799 down_read(&mm->mmap_sem);
800 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
801 pages, NULL, NULL);
802 if (ret != 1) {
803 BUG_ON(ret > 1);
804 if (!pages_done)
805 pages_done = ret;
806 break;
807 }
808 nr_pages--;
809 pages_done++;
810 if (!nr_pages)
811 break;
812 pages++;
813 start += PAGE_SIZE;
814 }
815 if (notify_drop && lock_dropped && *locked) {
816 /*
817 * We must let the caller know we temporarily dropped the lock
818 * and so the critical section protected by it was lost.
819 */
820 up_read(&mm->mmap_sem);
821 *locked = 0;
822 }
823 return pages_done;
824}
825
826/*
827 * We can leverage the VM_FAULT_RETRY functionality in the page fault
828 * paths better by using either get_user_pages_locked() or
829 * get_user_pages_unlocked().
830 *
831 * get_user_pages_locked() is suitable to replace the form:
832 *
833 * down_read(&mm->mmap_sem);
834 * do_something()
835 * get_user_pages(tsk, mm, ..., pages, NULL);
836 * up_read(&mm->mmap_sem);
837 *
838 * to:
839 *
840 * int locked = 1;
841 * down_read(&mm->mmap_sem);
842 * do_something()
843 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
844 * if (locked)
845 * up_read(&mm->mmap_sem);
846 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200847long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100848 unsigned int gup_flags, struct page **pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800849 int *locked)
850{
Dave Hansencde70142016-02-12 13:01:55 -0800851 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100852 pages, NULL, locked, true,
853 gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800854}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200855EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800856
857/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800858 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
859 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
860 *
861 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
862 * caller if required (just like with __get_user_pages). "FOLL_GET",
863 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
864 * according to the parameters "pages", "write", "force"
865 * respectively.
866 */
867__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
868 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesd4944b02016-10-13 01:20:12 +0100869 struct page **pages, unsigned int gup_flags)
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800870{
871 long ret;
872 int locked = 1;
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100873
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800874 down_read(&mm->mmap_sem);
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100875 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
876 &locked, false, gup_flags);
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800877 if (locked)
878 up_read(&mm->mmap_sem);
879 return ret;
880}
881EXPORT_SYMBOL(__get_user_pages_unlocked);
882
883/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800884 * get_user_pages_unlocked() is suitable to replace the form:
885 *
886 * down_read(&mm->mmap_sem);
887 * get_user_pages(tsk, mm, ..., pages, NULL);
888 * up_read(&mm->mmap_sem);
889 *
890 * with:
891 *
892 * get_user_pages_unlocked(tsk, mm, ..., pages);
893 *
894 * It is functionally equivalent to get_user_pages_fast so
895 * get_user_pages_fast should be used instead, if the two parameters
896 * "tsk" and "mm" are respectively equal to current and current->mm,
897 * or if "force" shall be set to 1 (get_user_pages_fast misses the
898 * "force" parameter).
899 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200900long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100901 struct page **pages, unsigned int gup_flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800902{
Dave Hansencde70142016-02-12 13:01:55 -0800903 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100904 pages, gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800905}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200906EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800907
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700908/*
Dave Hansen1e987792016-02-12 13:01:54 -0800909 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700910 * @tsk: the task_struct to use for page fault accounting, or
911 * NULL if faults are not to be recorded.
912 * @mm: mm_struct of target mm
913 * @start: starting user address
914 * @nr_pages: number of pages from start to pin
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100915 * @gup_flags: flags modifying lookup behaviour
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700916 * @pages: array that receives pointers to the pages pinned.
917 * Should be at least nr_pages long. Or NULL, if caller
918 * only intends to ensure the pages are faulted in.
919 * @vmas: array of pointers to vmas corresponding to each page.
920 * Or NULL if the caller does not require them.
921 *
922 * Returns number of pages pinned. This may be fewer than the number
923 * requested. If nr_pages is 0 or negative, returns 0. If no pages
924 * were pinned, returns -errno. Each page returned must be released
925 * with a put_page() call when it is finished with. vmas will only
926 * remain valid while mmap_sem is held.
927 *
928 * Must be called with mmap_sem held for read or write.
929 *
930 * get_user_pages walks a process's page tables and takes a reference to
931 * each struct page that each user address corresponds to at a given
932 * instant. That is, it takes the page that would be accessed if a user
933 * thread accesses the given user virtual address at that instant.
934 *
935 * This does not guarantee that the page exists in the user mappings when
936 * get_user_pages returns, and there may even be a completely different
937 * page there in some cases (eg. if mmapped pagecache has been invalidated
938 * and subsequently re faulted). However it does guarantee that the page
939 * won't be freed completely. And mostly callers simply care that the page
940 * contains data that was valid *at some point in time*. Typically, an IO
941 * or similar operation cannot guarantee anything stronger anyway because
942 * locks can't be held over the syscall boundary.
943 *
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100944 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
945 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
946 * be called after the page is finished with, and before put_page is called.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700947 *
948 * get_user_pages is typically used for fewer-copy IO operations, to get a
949 * handle on the memory by some means other than accesses via the user virtual
950 * addresses. The pages may be submitted for DMA to devices or accessed via
951 * their kernel linear mapping (via the kmap APIs). Care should be taken to
952 * use the correct cache flushing APIs.
953 *
954 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800955 *
956 * get_user_pages should be phased out in favor of
957 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
958 * should use get_user_pages because it cannot pass
959 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700960 */
Dave Hansen1e987792016-02-12 13:01:54 -0800961long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
962 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100963 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800964 struct vm_area_struct **vmas)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700965{
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100966 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100967 NULL, false,
968 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
Dave Hansen1e987792016-02-12 13:01:54 -0800969}
970EXPORT_SYMBOL(get_user_pages_remote);
971
972/*
Dave Hansend4edcf02016-02-12 13:01:56 -0800973 * This is the same as get_user_pages_remote(), just with a
974 * less-flexible calling convention where we assume that the task
975 * and mm being operated on are the current task's. We also
976 * obviously don't pass FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -0800977 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200978long get_user_pages(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +0100979 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800980 struct vm_area_struct **vmas)
981{
Dave Hansencde70142016-02-12 13:01:55 -0800982 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +0100983 pages, vmas, NULL, false,
984 gup_flags | FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700985}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200986EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700987
Dan Williamsb29ea3c2018-02-23 14:05:49 -0800988#ifdef CONFIG_FS_DAX
989/*
990 * This is the same as get_user_pages() in that it assumes we are
991 * operating on the current task's mm, but it goes further to validate
992 * that the vmas associated with the address range are suitable for
993 * longterm elevated page reference counts. For example, filesystem-dax
994 * mappings are subject to the lifetime enforced by the filesystem and
995 * we need guarantees that longterm users like RDMA and V4L2 only
996 * establish mappings that have a kernel enforced revocation mechanism.
997 *
998 * "longterm" == userspace controlled elevated page count lifetime.
999 * Contrast this to iov_iter_get_pages() usages which are transient.
1000 */
1001long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1002 unsigned int gup_flags, struct page **pages,
1003 struct vm_area_struct **vmas_arg)
1004{
1005 struct vm_area_struct **vmas = vmas_arg;
1006 struct vm_area_struct *vma_prev = NULL;
1007 long rc, i;
1008
1009 if (!pages)
1010 return -EINVAL;
1011
1012 if (!vmas) {
1013 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1014 GFP_KERNEL);
1015 if (!vmas)
1016 return -ENOMEM;
1017 }
1018
1019 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1020
1021 for (i = 0; i < rc; i++) {
1022 struct vm_area_struct *vma = vmas[i];
1023
1024 if (vma == vma_prev)
1025 continue;
1026
1027 vma_prev = vma;
1028
1029 if (vma_is_fsdax(vma))
1030 break;
1031 }
1032
1033 /*
1034 * Either get_user_pages() failed, or the vma validation
1035 * succeeded, in either case we don't need to put_page() before
1036 * returning.
1037 */
1038 if (i >= rc)
1039 goto out;
1040
1041 for (i = 0; i < rc; i++)
1042 put_page(pages[i]);
1043 rc = -EOPNOTSUPP;
1044out:
1045 if (vmas != vmas_arg)
1046 kfree(vmas);
1047 return rc;
1048}
1049EXPORT_SYMBOL(get_user_pages_longterm);
1050#endif /* CONFIG_FS_DAX */
1051
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001052/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001053 * populate_vma_page_range() - populate a range of pages in the vma.
1054 * @vma: target vma
1055 * @start: start address
1056 * @end: end address
1057 * @nonblocking:
1058 *
1059 * This takes care of mlocking the pages too if VM_LOCKED is set.
1060 *
1061 * return 0 on success, negative error code on error.
1062 *
1063 * vma->vm_mm->mmap_sem must be held.
1064 *
1065 * If @nonblocking is NULL, it may be held for read or write and will
1066 * be unperturbed.
1067 *
1068 * If @nonblocking is non-NULL, it must held for read only and may be
1069 * released. If it's released, *@nonblocking will be set to 0.
1070 */
1071long populate_vma_page_range(struct vm_area_struct *vma,
1072 unsigned long start, unsigned long end, int *nonblocking)
1073{
1074 struct mm_struct *mm = vma->vm_mm;
1075 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1076 int gup_flags;
1077
1078 VM_BUG_ON(start & ~PAGE_MASK);
1079 VM_BUG_ON(end & ~PAGE_MASK);
1080 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1081 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1082 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1083
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001084 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1085 if (vma->vm_flags & VM_LOCKONFAULT)
1086 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001087 /*
1088 * We want to touch writable mappings with a write fault in order
1089 * to break COW, except for shared mappings because these don't COW
1090 * and we would not want to dirty them for nothing.
1091 */
1092 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1093 gup_flags |= FOLL_WRITE;
1094
1095 /*
1096 * We want mlock to succeed for regions that have any permissions
1097 * other than PROT_NONE.
1098 */
1099 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1100 gup_flags |= FOLL_FORCE;
1101
1102 /*
1103 * We made sure addr is within a VMA, so the following will
1104 * not result in a stack expansion that recurses back here.
1105 */
1106 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1107 NULL, NULL, nonblocking);
1108}
1109
1110/*
1111 * __mm_populate - populate and/or mlock pages within a range of address space.
1112 *
1113 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1114 * flags. VMAs must be already marked with the desired vm_flags, and
1115 * mmap_sem must not be held.
1116 */
1117int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1118{
1119 struct mm_struct *mm = current->mm;
1120 unsigned long end, nstart, nend;
1121 struct vm_area_struct *vma = NULL;
1122 int locked = 0;
1123 long ret = 0;
1124
1125 VM_BUG_ON(start & ~PAGE_MASK);
1126 VM_BUG_ON(len != PAGE_ALIGN(len));
1127 end = start + len;
1128
1129 for (nstart = start; nstart < end; nstart = nend) {
1130 /*
1131 * We want to fault in pages for [nstart; end) address range.
1132 * Find first corresponding VMA.
1133 */
1134 if (!locked) {
1135 locked = 1;
1136 down_read(&mm->mmap_sem);
1137 vma = find_vma(mm, nstart);
1138 } else if (nstart >= vma->vm_end)
1139 vma = vma->vm_next;
1140 if (!vma || vma->vm_start >= end)
1141 break;
1142 /*
1143 * Set [nstart; nend) to intersection of desired address
1144 * range with the first VMA. Also, skip undesirable VMA types.
1145 */
1146 nend = min(end, vma->vm_end);
1147 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1148 continue;
1149 if (nstart < vma->vm_start)
1150 nstart = vma->vm_start;
1151 /*
1152 * Now fault in a range of pages. populate_vma_page_range()
1153 * double checks the vma flags, so that it won't mlock pages
1154 * if the vma was already munlocked.
1155 */
1156 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1157 if (ret < 0) {
1158 if (ignore_errors) {
1159 ret = 0;
1160 continue; /* continue at next VMA */
1161 }
1162 break;
1163 }
1164 nend = nstart + ret * PAGE_SIZE;
1165 ret = 0;
1166 }
1167 if (locked)
1168 up_read(&mm->mmap_sem);
1169 return ret; /* 0 or negative error code */
1170}
1171
1172/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001173 * get_dump_page() - pin user page in memory while writing it to core dump
1174 * @addr: user address
1175 *
1176 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001177 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001178 *
1179 * Returns NULL on any kind of failure - a hole must then be inserted into
1180 * the corefile, to preserve alignment with its headers; and also returns
1181 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1182 * allowing a hole to be left in the corefile to save diskspace.
1183 *
1184 * Called without mmap_sem, but after all other threads have been killed.
1185 */
1186#ifdef CONFIG_ELF_CORE
1187struct page *get_dump_page(unsigned long addr)
1188{
1189 struct vm_area_struct *vma;
1190 struct page *page;
1191
1192 if (__get_user_pages(current, current->mm, addr, 1,
1193 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1194 NULL) < 1)
1195 return NULL;
1196 flush_cache_page(vma, addr, page_to_pfn(page));
1197 return page;
1198}
1199#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001200
1201/*
1202 * Generic RCU Fast GUP
1203 *
1204 * get_user_pages_fast attempts to pin user pages by walking the page
1205 * tables directly and avoids taking locks. Thus the walker needs to be
1206 * protected from page table pages being freed from under it, and should
1207 * block any THP splits.
1208 *
1209 * One way to achieve this is to have the walker disable interrupts, and
1210 * rely on IPIs from the TLB flushing code blocking before the page table
1211 * pages are freed. This is unsuitable for architectures that do not need
1212 * to broadcast an IPI when invalidating TLBs.
1213 *
1214 * Another way to achieve this is to batch up page table containing pages
1215 * belonging to more than one mm_user, then rcu_sched a callback to free those
1216 * pages. Disabling interrupts will allow the fast_gup walker to both block
1217 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1218 * (which is a relatively rare event). The code below adopts this strategy.
1219 *
1220 * Before activating this code, please be aware that the following assumptions
1221 * are currently made:
1222 *
1223 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1224 * pages containing page tables.
1225 *
Steve Capper2667f502014-10-09 15:29:14 -07001226 * *) ptes can be read atomically by the architecture.
1227 *
1228 * *) access_ok is sufficient to validate userspace address ranges.
1229 *
1230 * The last two assumptions can be relaxed by the addition of helper functions.
1231 *
1232 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1233 */
1234#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1235
1236#ifdef __HAVE_ARCH_PTE_SPECIAL
1237static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1238 int write, struct page **pages, int *nr)
1239{
1240 pte_t *ptep, *ptem;
1241 int ret = 0;
1242
1243 ptem = ptep = pte_offset_map(&pmd, addr);
1244 do {
1245 /*
1246 * In the line below we are assuming that the pte can be read
1247 * atomically. If this is not the case for your architecture,
1248 * please wrap this in a helper function!
1249 *
1250 * for an example see gup_get_pte in arch/x86/mm/gup.c
1251 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001252 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001253 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001254
1255 /*
1256 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001257 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001258 */
1259 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001260 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001261 goto pte_unmap;
1262
Dave Hansen33a709b2016-02-12 13:02:19 -08001263 if (!arch_pte_access_permitted(pte, write))
1264 goto pte_unmap;
1265
Steve Capper2667f502014-10-09 15:29:14 -07001266 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1267 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001268 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001269
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001270 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001271 goto pte_unmap;
1272
1273 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001274 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001275 goto pte_unmap;
1276 }
1277
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001278 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001279 pages[*nr] = page;
1280 (*nr)++;
1281
1282 } while (ptep++, addr += PAGE_SIZE, addr != end);
1283
1284 ret = 1;
1285
1286pte_unmap:
1287 pte_unmap(ptem);
1288 return ret;
1289}
1290#else
1291
1292/*
1293 * If we can't determine whether or not a pte is special, then fail immediately
1294 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1295 * to be special.
1296 *
1297 * For a futex to be placed on a THP tail page, get_futex_key requires a
1298 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1299 * useful to have gup_huge_pmd even if we can't operate on ptes.
1300 */
1301static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1302 int write, struct page **pages, int *nr)
1303{
1304 return 0;
1305}
1306#endif /* __HAVE_ARCH_PTE_SPECIAL */
1307
1308static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1309 unsigned long end, int write, struct page **pages, int *nr)
1310{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001311 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001312 int refs;
1313
1314 if (write && !pmd_write(orig))
1315 return 0;
1316
1317 refs = 0;
1318 head = pmd_page(orig);
1319 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001320 do {
1321 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1322 pages[*nr] = page;
1323 (*nr)++;
1324 page++;
1325 refs++;
1326 } while (addr += PAGE_SIZE, addr != end);
1327
1328 if (!page_cache_add_speculative(head, refs)) {
1329 *nr -= refs;
1330 return 0;
1331 }
1332
1333 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1334 *nr -= refs;
1335 while (refs--)
1336 put_page(head);
1337 return 0;
1338 }
1339
Steve Capper2667f502014-10-09 15:29:14 -07001340 return 1;
1341}
1342
1343static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1344 unsigned long end, int write, struct page **pages, int *nr)
1345{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001346 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001347 int refs;
1348
1349 if (write && !pud_write(orig))
1350 return 0;
1351
1352 refs = 0;
1353 head = pud_page(orig);
1354 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001355 do {
1356 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1357 pages[*nr] = page;
1358 (*nr)++;
1359 page++;
1360 refs++;
1361 } while (addr += PAGE_SIZE, addr != end);
1362
1363 if (!page_cache_add_speculative(head, refs)) {
1364 *nr -= refs;
1365 return 0;
1366 }
1367
1368 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1369 *nr -= refs;
1370 while (refs--)
1371 put_page(head);
1372 return 0;
1373 }
1374
Steve Capper2667f502014-10-09 15:29:14 -07001375 return 1;
1376}
1377
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301378static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1379 unsigned long end, int write,
1380 struct page **pages, int *nr)
1381{
1382 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001383 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301384
1385 if (write && !pgd_write(orig))
1386 return 0;
1387
1388 refs = 0;
1389 head = pgd_page(orig);
1390 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301391 do {
1392 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1393 pages[*nr] = page;
1394 (*nr)++;
1395 page++;
1396 refs++;
1397 } while (addr += PAGE_SIZE, addr != end);
1398
1399 if (!page_cache_add_speculative(head, refs)) {
1400 *nr -= refs;
1401 return 0;
1402 }
1403
1404 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1405 *nr -= refs;
1406 while (refs--)
1407 put_page(head);
1408 return 0;
1409 }
1410
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301411 return 1;
1412}
1413
Steve Capper2667f502014-10-09 15:29:14 -07001414static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1415 int write, struct page **pages, int *nr)
1416{
1417 unsigned long next;
1418 pmd_t *pmdp;
1419
1420 pmdp = pmd_offset(&pud, addr);
1421 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001422 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001423
1424 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001425 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001426 return 0;
1427
1428 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1429 /*
1430 * NUMA hinting faults need to be handled in the GUP
1431 * slowpath for accounting purposes and so that they
1432 * can be serialised against THP migration.
1433 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001434 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001435 return 0;
1436
1437 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1438 pages, nr))
1439 return 0;
1440
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301441 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1442 /*
1443 * architecture have different format for hugetlbfs
1444 * pmd format and THP pmd format
1445 */
1446 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1447 PMD_SHIFT, next, write, pages, nr))
1448 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001449 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1450 return 0;
1451 } while (pmdp++, addr = next, addr != end);
1452
1453 return 1;
1454}
1455
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301456static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1457 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001458{
1459 unsigned long next;
1460 pud_t *pudp;
1461
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301462 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001463 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001464 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001465
1466 next = pud_addr_end(addr, end);
1467 if (pud_none(pud))
1468 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301469 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001470 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301471 pages, nr))
1472 return 0;
1473 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1474 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1475 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001476 return 0;
1477 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1478 return 0;
1479 } while (pudp++, addr = next, addr != end);
1480
1481 return 1;
1482}
1483
1484/*
1485 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1486 * the regular GUP. It will only return non-negative values.
1487 */
1488int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1489 struct page **pages)
1490{
1491 struct mm_struct *mm = current->mm;
1492 unsigned long addr, len, end;
1493 unsigned long next, flags;
1494 pgd_t *pgdp;
1495 int nr = 0;
1496
1497 start &= PAGE_MASK;
1498 addr = start;
1499 len = (unsigned long) nr_pages << PAGE_SHIFT;
1500 end = start + len;
1501
1502 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1503 start, len)))
1504 return 0;
1505
1506 /*
1507 * Disable interrupts. We use the nested form as we can already have
1508 * interrupts disabled by get_futex_key.
1509 *
1510 * With interrupts disabled, we block page table pages from being
1511 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1512 * for more details.
1513 *
1514 * We do not adopt an rcu_read_lock(.) here as we also want to
1515 * block IPIs that come from THPs splitting.
1516 */
1517
1518 local_irq_save(flags);
1519 pgdp = pgd_offset(mm, addr);
1520 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001521 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301522
Steve Capper2667f502014-10-09 15:29:14 -07001523 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301524 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001525 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301526 if (unlikely(pgd_huge(pgd))) {
1527 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1528 pages, &nr))
1529 break;
1530 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1531 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1532 PGDIR_SHIFT, next, write, pages, &nr))
1533 break;
1534 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001535 break;
1536 } while (pgdp++, addr = next, addr != end);
1537 local_irq_restore(flags);
1538
1539 return nr;
1540}
1541
1542/**
1543 * get_user_pages_fast() - pin user pages in memory
1544 * @start: starting user address
1545 * @nr_pages: number of pages from start to pin
1546 * @write: whether pages will be written to
1547 * @pages: array that receives pointers to the pages pinned.
1548 * Should be at least nr_pages long.
1549 *
1550 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1551 * If not successful, it will fall back to taking the lock and
1552 * calling get_user_pages().
1553 *
1554 * Returns number of pages pinned. This may be fewer than the number
1555 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1556 * were pinned, returns -errno.
1557 */
1558int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1559 struct page **pages)
1560{
Steve Capper2667f502014-10-09 15:29:14 -07001561 int nr, ret;
1562
1563 start &= PAGE_MASK;
1564 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1565 ret = nr;
1566
1567 if (nr < nr_pages) {
1568 /* Try to get the remaining pages with get_user_pages */
1569 start += nr << PAGE_SHIFT;
1570 pages += nr;
1571
Lorenzo Stoakesc1641542016-10-13 01:20:13 +01001572 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1573 write ? FOLL_WRITE : 0);
Steve Capper2667f502014-10-09 15:29:14 -07001574
1575 /* Have to be a bit careful with return values */
1576 if (nr > 0) {
1577 if (ret < 0)
1578 ret = nr;
1579 else
1580 ret += nr;
1581 }
1582 }
1583
1584 return ret;
1585}
1586
1587#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */