blob: 7bf19ffa21999c13fa1f24dc01a6bda77217688c [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
Steve Capper2667f502014-10-09 15:29:14 -070017#include <asm/pgtable.h>
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070018#include <asm/tlbflush.h>
Steve Capper2667f502014-10-09 15:29:14 -070019
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070020#include "internal.h"
21
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070022static struct page *no_page_table(struct vm_area_struct *vma,
23 unsigned int flags)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070024{
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070025 /*
26 * When core dumping an enormous anonymous area that nobody
27 * has touched so far, we don't want to allocate unnecessary pages or
28 * page tables. Return error instead of NULL to skip handle_mm_fault,
29 * then get_dump_page() will return NULL to leave a hole in the dump.
30 * But we can only make this optimization where a hole would surely
31 * be zero-filled if handle_mm_fault() actually did handle it.
32 */
33 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
34 return ERR_PTR(-EFAULT);
35 return NULL;
36}
37
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070038static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
39 pte_t *pte, unsigned int flags)
40{
41 /* No page to get reference */
42 if (flags & FOLL_GET)
43 return -EFAULT;
44
45 if (flags & FOLL_TOUCH) {
46 pte_t entry = *pte;
47
48 if (flags & FOLL_WRITE)
49 entry = pte_mkdirty(entry);
50 entry = pte_mkyoung(entry);
51
52 if (!pte_same(*pte, entry)) {
53 set_pte_at(vma->vm_mm, address, pte, entry);
54 update_mmu_cache(vma, address, pte);
55 }
56 }
57
58 /* Proper page table entry exists, but no corresponding struct page */
59 return -EEXIST;
60}
61
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070062static struct page *follow_page_pte(struct vm_area_struct *vma,
63 unsigned long address, pmd_t *pmd, unsigned int flags)
64{
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070065 struct mm_struct *mm = vma->vm_mm;
Dan Williams3565fce2016-01-15 16:56:55 -080066 struct dev_pagemap *pgmap = NULL;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070067 struct page *page;
68 spinlock_t *ptl;
69 pte_t *ptep, pte;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070070
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070071retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070072 if (unlikely(pmd_bad(*pmd)))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070073 return no_page_table(vma, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070074
75 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070076 pte = *ptep;
77 if (!pte_present(pte)) {
78 swp_entry_t entry;
79 /*
80 * KSM's break_ksm() relies upon recognizing a ksm page
81 * even while it is being migrated, so for that case we
82 * need migration_entry_wait().
83 */
84 if (likely(!(flags & FOLL_MIGRATION)))
85 goto no_page;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -080086 if (pte_none(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070087 goto no_page;
88 entry = pte_to_swp_entry(pte);
89 if (!is_migration_entry(entry))
90 goto no_page;
91 pte_unmap_unlock(ptep, ptl);
92 migration_entry_wait(mm, pmd, address);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070093 goto retry;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070094 }
Mel Gorman8a0516e2015-02-12 14:58:22 -080095 if ((flags & FOLL_NUMA) && pte_protnone(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070096 goto no_page;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070097 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
98 pte_unmap_unlock(ptep, ptl);
99 return NULL;
100 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700101
102 page = vm_normal_page(vma, address, pte);
Dan Williams3565fce2016-01-15 16:56:55 -0800103 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
104 /*
105 * Only return device mapping pages in the FOLL_GET case since
106 * they are only valid while holding the pgmap reference.
107 */
108 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
109 if (pgmap)
110 page = pte_page(pte);
111 else
112 goto no_page;
113 } else if (unlikely(!page)) {
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700114 if (flags & FOLL_DUMP) {
115 /* Avoid special (like zero) pages in core dumps */
116 page = ERR_PTR(-EFAULT);
117 goto out;
118 }
119
120 if (is_zero_pfn(pte_pfn(pte))) {
121 page = pte_page(pte);
122 } else {
123 int ret;
124
125 ret = follow_pfn_pte(vma, address, ptep, flags);
126 page = ERR_PTR(ret);
127 goto out;
128 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700129 }
130
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800131 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
132 int ret;
133 get_page(page);
134 pte_unmap_unlock(ptep, ptl);
135 lock_page(page);
136 ret = split_huge_page(page);
137 unlock_page(page);
138 put_page(page);
139 if (ret)
140 return ERR_PTR(ret);
141 goto retry;
142 }
143
Dan Williams3565fce2016-01-15 16:56:55 -0800144 if (flags & FOLL_GET) {
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800145 get_page(page);
Dan Williams3565fce2016-01-15 16:56:55 -0800146
147 /* drop the pgmap reference now that we hold the page */
148 if (pgmap) {
149 put_dev_pagemap(pgmap);
150 pgmap = NULL;
151 }
152 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700153 if (flags & FOLL_TOUCH) {
154 if ((flags & FOLL_WRITE) &&
155 !pte_dirty(pte) && !PageDirty(page))
156 set_page_dirty(page);
157 /*
158 * pte_mkyoung() would be more correct here, but atomic care
159 * is needed to avoid losing the dirty bit: it is easier to use
160 * mark_page_accessed().
161 */
162 mark_page_accessed(page);
163 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800164 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -0800165 /* Do not mlock pte-mapped THP */
166 if (PageTransCompound(page))
167 goto out;
168
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700169 /*
170 * The preliminary mapping check is mainly to avoid the
171 * pointless overhead of lock_page on the ZERO_PAGE
172 * which might bounce very badly if there is contention.
173 *
174 * If the page is already locked, we don't need to
175 * handle it now - vmscan will handle it later if and
176 * when it attempts to reclaim the page.
177 */
178 if (page->mapping && trylock_page(page)) {
179 lru_add_drain(); /* push cached pages to LRU */
180 /*
181 * Because we lock page here, and migration is
182 * blocked by the pte's page reference, and we
183 * know the page is still mapped, we don't even
184 * need to check for file-cache page truncation.
185 */
186 mlock_vma_page(page);
187 unlock_page(page);
188 }
189 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700190out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700191 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700192 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700193no_page:
194 pte_unmap_unlock(ptep, ptl);
195 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700196 return NULL;
197 return no_page_table(vma, flags);
198}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700199
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700200/**
201 * follow_page_mask - look up a page descriptor from a user-virtual address
202 * @vma: vm_area_struct mapping @address
203 * @address: virtual address to look up
204 * @flags: flags modifying lookup behaviour
205 * @page_mask: on output, *page_mask is set according to the size of the page
206 *
207 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
208 *
209 * Returns the mapped (struct page *), %NULL if no mapping exists, or
210 * an error pointer if there is a mapping to something not represented
211 * by a page descriptor (see also vm_normal_page()).
212 */
213struct page *follow_page_mask(struct vm_area_struct *vma,
214 unsigned long address, unsigned int flags,
215 unsigned int *page_mask)
216{
217 pgd_t *pgd;
218 pud_t *pud;
219 pmd_t *pmd;
220 spinlock_t *ptl;
221 struct page *page;
222 struct mm_struct *mm = vma->vm_mm;
223
224 *page_mask = 0;
225
226 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
227 if (!IS_ERR(page)) {
228 BUG_ON(flags & FOLL_GET);
229 return page;
230 }
231
232 pgd = pgd_offset(mm, address);
233 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
234 return no_page_table(vma, flags);
235
236 pud = pud_offset(pgd, address);
237 if (pud_none(*pud))
238 return no_page_table(vma, flags);
239 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800240 page = follow_huge_pud(mm, address, pud, flags);
241 if (page)
242 return page;
243 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700244 }
245 if (unlikely(pud_bad(*pud)))
246 return no_page_table(vma, flags);
247
248 pmd = pmd_offset(pud, address);
249 if (pmd_none(*pmd))
250 return no_page_table(vma, flags);
251 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800252 page = follow_huge_pmd(mm, address, pmd, flags);
253 if (page)
254 return page;
255 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700256 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800257 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700258 return no_page_table(vma, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800259 if (pmd_devmap(*pmd)) {
260 ptl = pmd_lock(mm, pmd);
261 page = follow_devmap_pmd(vma, address, pmd, flags);
262 spin_unlock(ptl);
263 if (page)
264 return page;
265 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800266 if (likely(!pmd_trans_huge(*pmd)))
267 return follow_page_pte(vma, address, pmd, flags);
268
269 ptl = pmd_lock(mm, pmd);
270 if (unlikely(!pmd_trans_huge(*pmd))) {
271 spin_unlock(ptl);
272 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700273 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800274 if (flags & FOLL_SPLIT) {
275 int ret;
276 page = pmd_page(*pmd);
277 if (is_huge_zero_page(page)) {
278 spin_unlock(ptl);
279 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800280 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800281 } else {
282 get_page(page);
283 spin_unlock(ptl);
284 lock_page(page);
285 ret = split_huge_page(page);
286 unlock_page(page);
287 put_page(page);
288 }
289
290 return ret ? ERR_PTR(ret) :
291 follow_page_pte(vma, address, pmd, flags);
292 }
293
294 page = follow_trans_huge_pmd(vma, address, pmd, flags);
295 spin_unlock(ptl);
296 *page_mask = HPAGE_PMD_NR - 1;
297 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700298}
299
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700300static int get_gate_page(struct mm_struct *mm, unsigned long address,
301 unsigned int gup_flags, struct vm_area_struct **vma,
302 struct page **page)
303{
304 pgd_t *pgd;
305 pud_t *pud;
306 pmd_t *pmd;
307 pte_t *pte;
308 int ret = -EFAULT;
309
310 /* user gate pages are read-only */
311 if (gup_flags & FOLL_WRITE)
312 return -EFAULT;
313 if (address > TASK_SIZE)
314 pgd = pgd_offset_k(address);
315 else
316 pgd = pgd_offset_gate(mm, address);
317 BUG_ON(pgd_none(*pgd));
318 pud = pud_offset(pgd, address);
319 BUG_ON(pud_none(*pud));
320 pmd = pmd_offset(pud, address);
321 if (pmd_none(*pmd))
322 return -EFAULT;
323 VM_BUG_ON(pmd_trans_huge(*pmd));
324 pte = pte_offset_map(pmd, address);
325 if (pte_none(*pte))
326 goto unmap;
327 *vma = get_gate_vma(mm);
328 if (!page)
329 goto out;
330 *page = vm_normal_page(*vma, address, *pte);
331 if (!*page) {
332 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
333 goto unmap;
334 *page = pte_page(*pte);
335 }
336 get_page(*page);
337out:
338 ret = 0;
339unmap:
340 pte_unmap(pte);
341 return ret;
342}
343
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700344/*
345 * mmap_sem must be held on entry. If @nonblocking != NULL and
346 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
347 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
348 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700349static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
350 unsigned long address, unsigned int *flags, int *nonblocking)
351{
352 struct mm_struct *mm = vma->vm_mm;
353 unsigned int fault_flags = 0;
354 int ret;
355
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800356 /* mlock all present pages, but do not fault in new pages */
357 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
358 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700359 /* For mm_populate(), just skip the stack guard page. */
360 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700361 (stack_guard_page_start(vma, address) ||
362 stack_guard_page_end(vma, address + PAGE_SIZE)))
363 return -ENOENT;
364 if (*flags & FOLL_WRITE)
365 fault_flags |= FAULT_FLAG_WRITE;
366 if (nonblocking)
367 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
368 if (*flags & FOLL_NOWAIT)
369 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700370 if (*flags & FOLL_TRIED) {
371 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
372 fault_flags |= FAULT_FLAG_TRIED;
373 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700374
375 ret = handle_mm_fault(mm, vma, address, fault_flags);
376 if (ret & VM_FAULT_ERROR) {
377 if (ret & VM_FAULT_OOM)
378 return -ENOMEM;
379 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
380 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800381 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700382 return -EFAULT;
383 BUG();
384 }
385
386 if (tsk) {
387 if (ret & VM_FAULT_MAJOR)
388 tsk->maj_flt++;
389 else
390 tsk->min_flt++;
391 }
392
393 if (ret & VM_FAULT_RETRY) {
394 if (nonblocking)
395 *nonblocking = 0;
396 return -EBUSY;
397 }
398
399 /*
400 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
401 * necessary, even if maybe_mkwrite decided not to set pte_write. We
402 * can thus safely do subsequent page lookups as if they were reads.
403 * But only do so when looping for pte_write is futile: in some cases
404 * userspace may also be wanting to write to the gotten user page,
405 * which a read fault here might prevent (a readonly page might get
406 * reCOWed by userspace write).
407 */
408 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
409 *flags &= ~FOLL_WRITE;
410 return 0;
411}
412
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700413static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
414{
415 vm_flags_t vm_flags = vma->vm_flags;
416
417 if (vm_flags & (VM_IO | VM_PFNMAP))
418 return -EFAULT;
419
420 if (gup_flags & FOLL_WRITE) {
421 if (!(vm_flags & VM_WRITE)) {
422 if (!(gup_flags & FOLL_FORCE))
423 return -EFAULT;
424 /*
425 * We used to let the write,force case do COW in a
426 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
427 * set a breakpoint in a read-only mapping of an
428 * executable, without corrupting the file (yet only
429 * when that file had been opened for writing!).
430 * Anon pages in shared mappings are surprising: now
431 * just reject it.
432 */
Hugh Dickins46435362016-01-30 18:03:16 -0800433 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700434 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700435 }
436 } else if (!(vm_flags & VM_READ)) {
437 if (!(gup_flags & FOLL_FORCE))
438 return -EFAULT;
439 /*
440 * Is there actually any vma we can reach here which does not
441 * have VM_MAYREAD set?
442 */
443 if (!(vm_flags & VM_MAYREAD))
444 return -EFAULT;
445 }
446 return 0;
447}
448
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700449/**
450 * __get_user_pages() - pin user pages in memory
451 * @tsk: task_struct of target task
452 * @mm: mm_struct of target mm
453 * @start: starting user address
454 * @nr_pages: number of pages from start to pin
455 * @gup_flags: flags modifying pin behaviour
456 * @pages: array that receives pointers to the pages pinned.
457 * Should be at least nr_pages long. Or NULL, if caller
458 * only intends to ensure the pages are faulted in.
459 * @vmas: array of pointers to vmas corresponding to each page.
460 * Or NULL if the caller does not require them.
461 * @nonblocking: whether waiting for disk IO or mmap_sem contention
462 *
463 * Returns number of pages pinned. This may be fewer than the number
464 * requested. If nr_pages is 0 or negative, returns 0. If no pages
465 * were pinned, returns -errno. Each page returned must be released
466 * with a put_page() call when it is finished with. vmas will only
467 * remain valid while mmap_sem is held.
468 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700469 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700470 *
471 * __get_user_pages walks a process's page tables and takes a reference to
472 * each struct page that each user address corresponds to at a given
473 * instant. That is, it takes the page that would be accessed if a user
474 * thread accesses the given user virtual address at that instant.
475 *
476 * This does not guarantee that the page exists in the user mappings when
477 * __get_user_pages returns, and there may even be a completely different
478 * page there in some cases (eg. if mmapped pagecache has been invalidated
479 * and subsequently re faulted). However it does guarantee that the page
480 * won't be freed completely. And mostly callers simply care that the page
481 * contains data that was valid *at some point in time*. Typically, an IO
482 * or similar operation cannot guarantee anything stronger anyway because
483 * locks can't be held over the syscall boundary.
484 *
485 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
486 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
487 * appropriate) must be called after the page is finished with, and
488 * before put_page is called.
489 *
490 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
491 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700492 * *@nonblocking will be set to 0. Further, if @gup_flags does not
493 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
494 * this case.
495 *
496 * A caller using such a combination of @nonblocking and @gup_flags
497 * must therefore hold the mmap_sem for reading only, and recognize
498 * when it's been released. Otherwise, it must be held for either
499 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700500 *
501 * In most cases, get_user_pages or get_user_pages_fast should be used
502 * instead of __get_user_pages. __get_user_pages should be used only if
503 * you need some special @gup_flags.
504 */
505long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
506 unsigned long start, unsigned long nr_pages,
507 unsigned int gup_flags, struct page **pages,
508 struct vm_area_struct **vmas, int *nonblocking)
509{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700510 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700511 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700512 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700513
514 if (!nr_pages)
515 return 0;
516
517 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
518
519 /*
520 * If FOLL_FORCE is set then do not force a full fault as the hinting
521 * fault information is unrelated to the reference behaviour of a task
522 * using the address space
523 */
524 if (!(gup_flags & FOLL_FORCE))
525 gup_flags |= FOLL_NUMA;
526
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700527 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700528 struct page *page;
529 unsigned int foll_flags = gup_flags;
530 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700531
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700532 /* first iteration or cross vma bound */
533 if (!vma || start >= vma->vm_end) {
534 vma = find_extend_vma(mm, start);
535 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700536 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700537 ret = get_gate_page(mm, start & PAGE_MASK,
538 gup_flags, &vma,
539 pages ? &pages[i] : NULL);
540 if (ret)
541 return i ? : ret;
542 page_mask = 0;
543 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700544 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700545
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700546 if (!vma || check_vma_flags(vma, gup_flags))
547 return i ? : -EFAULT;
548 if (is_vm_hugetlb_page(vma)) {
549 i = follow_hugetlb_page(mm, vma, pages, vmas,
550 &start, &nr_pages, i,
551 gup_flags);
552 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700553 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700554 }
555retry:
556 /*
557 * If we have a pending SIGKILL, don't keep faulting pages and
558 * potentially allocating memory.
559 */
560 if (unlikely(fatal_signal_pending(current)))
561 return i ? i : -ERESTARTSYS;
562 cond_resched();
563 page = follow_page_mask(vma, start, foll_flags, &page_mask);
564 if (!page) {
565 int ret;
566 ret = faultin_page(tsk, vma, start, &foll_flags,
567 nonblocking);
568 switch (ret) {
569 case 0:
570 goto retry;
571 case -EFAULT:
572 case -ENOMEM:
573 case -EHWPOISON:
574 return i ? i : ret;
575 case -EBUSY:
576 return i;
577 case -ENOENT:
578 goto next_page;
579 }
580 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700581 } else if (PTR_ERR(page) == -EEXIST) {
582 /*
583 * Proper page table entry exists, but no corresponding
584 * struct page.
585 */
586 goto next_page;
587 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700588 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700589 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700590 if (pages) {
591 pages[i] = page;
592 flush_anon_page(vma, page, start);
593 flush_dcache_page(page);
594 page_mask = 0;
595 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700596next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700597 if (vmas) {
598 vmas[i] = vma;
599 page_mask = 0;
600 }
601 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
602 if (page_increm > nr_pages)
603 page_increm = nr_pages;
604 i += page_increm;
605 start += page_increm * PAGE_SIZE;
606 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700607 } while (nr_pages);
608 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700609}
610EXPORT_SYMBOL(__get_user_pages);
611
612/*
613 * fixup_user_fault() - manually resolve a user page fault
614 * @tsk: the task_struct to use for page fault accounting, or
615 * NULL if faults are not to be recorded.
616 * @mm: mm_struct of target mm
617 * @address: user address
618 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800619 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
620 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700621 *
622 * This is meant to be called in the specific scenario where for locking reasons
623 * we try to access user memory in atomic context (within a pagefault_disable()
624 * section), this returns -EFAULT, and we want to resolve the user fault before
625 * trying again.
626 *
627 * Typically this is meant to be used by the futex code.
628 *
629 * The main difference with get_user_pages() is that this function will
630 * unconditionally call handle_mm_fault() which will in turn perform all the
631 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800632 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700633 *
634 * This is important for some architectures where those bits also gate the
635 * access permission to the page because they are maintained in software. On
636 * such architectures, gup() will not be enough to make a subsequent access
637 * succeed.
638 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800639 * This function will not return with an unlocked mmap_sem. So it has not the
640 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700641 */
642int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800643 unsigned long address, unsigned int fault_flags,
644 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700645{
646 struct vm_area_struct *vma;
647 vm_flags_t vm_flags;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800648 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700649
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800650 if (unlocked)
651 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
652
653retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700654 vma = find_extend_vma(mm, address);
655 if (!vma || address < vma->vm_start)
656 return -EFAULT;
657
658 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
659 if (!(vm_flags & vma->vm_flags))
660 return -EFAULT;
661
662 ret = handle_mm_fault(mm, vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800663 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700664 if (ret & VM_FAULT_ERROR) {
665 if (ret & VM_FAULT_OOM)
666 return -ENOMEM;
667 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
668 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800669 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700670 return -EFAULT;
671 BUG();
672 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800673
674 if (ret & VM_FAULT_RETRY) {
675 down_read(&mm->mmap_sem);
676 if (!(fault_flags & FAULT_FLAG_TRIED)) {
677 *unlocked = true;
678 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
679 fault_flags |= FAULT_FLAG_TRIED;
680 goto retry;
681 }
682 }
683
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700684 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800685 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700686 tsk->maj_flt++;
687 else
688 tsk->min_flt++;
689 }
690 return 0;
691}
692
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800693static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
694 struct mm_struct *mm,
695 unsigned long start,
696 unsigned long nr_pages,
697 int write, int force,
698 struct page **pages,
699 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800700 int *locked, bool notify_drop,
701 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800702{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800703 long ret, pages_done;
704 bool lock_dropped;
705
706 if (locked) {
707 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
708 BUG_ON(vmas);
709 /* check caller initialized locked */
710 BUG_ON(*locked != 1);
711 }
712
713 if (pages)
714 flags |= FOLL_GET;
715 if (write)
716 flags |= FOLL_WRITE;
717 if (force)
718 flags |= FOLL_FORCE;
719
720 pages_done = 0;
721 lock_dropped = false;
722 for (;;) {
723 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
724 vmas, locked);
725 if (!locked)
726 /* VM_FAULT_RETRY couldn't trigger, bypass */
727 return ret;
728
729 /* VM_FAULT_RETRY cannot return errors */
730 if (!*locked) {
731 BUG_ON(ret < 0);
732 BUG_ON(ret >= nr_pages);
733 }
734
735 if (!pages)
736 /* If it's a prefault don't insist harder */
737 return ret;
738
739 if (ret > 0) {
740 nr_pages -= ret;
741 pages_done += ret;
742 if (!nr_pages)
743 break;
744 }
745 if (*locked) {
746 /* VM_FAULT_RETRY didn't trigger */
747 if (!pages_done)
748 pages_done = ret;
749 break;
750 }
751 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
752 pages += ret;
753 start += ret << PAGE_SHIFT;
754
755 /*
756 * Repeat on the address that fired VM_FAULT_RETRY
757 * without FAULT_FLAG_ALLOW_RETRY but with
758 * FAULT_FLAG_TRIED.
759 */
760 *locked = 1;
761 lock_dropped = true;
762 down_read(&mm->mmap_sem);
763 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
764 pages, NULL, NULL);
765 if (ret != 1) {
766 BUG_ON(ret > 1);
767 if (!pages_done)
768 pages_done = ret;
769 break;
770 }
771 nr_pages--;
772 pages_done++;
773 if (!nr_pages)
774 break;
775 pages++;
776 start += PAGE_SIZE;
777 }
778 if (notify_drop && lock_dropped && *locked) {
779 /*
780 * We must let the caller know we temporarily dropped the lock
781 * and so the critical section protected by it was lost.
782 */
783 up_read(&mm->mmap_sem);
784 *locked = 0;
785 }
786 return pages_done;
787}
788
789/*
790 * We can leverage the VM_FAULT_RETRY functionality in the page fault
791 * paths better by using either get_user_pages_locked() or
792 * get_user_pages_unlocked().
793 *
794 * get_user_pages_locked() is suitable to replace the form:
795 *
796 * down_read(&mm->mmap_sem);
797 * do_something()
798 * get_user_pages(tsk, mm, ..., pages, NULL);
799 * up_read(&mm->mmap_sem);
800 *
801 * to:
802 *
803 * int locked = 1;
804 * down_read(&mm->mmap_sem);
805 * do_something()
806 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
807 * if (locked)
808 * up_read(&mm->mmap_sem);
809 */
810long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
811 unsigned long start, unsigned long nr_pages,
812 int write, int force, struct page **pages,
813 int *locked)
814{
815 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800816 pages, NULL, locked, true, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800817}
818EXPORT_SYMBOL(get_user_pages_locked);
819
820/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800821 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
822 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
823 *
824 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
825 * caller if required (just like with __get_user_pages). "FOLL_GET",
826 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
827 * according to the parameters "pages", "write", "force"
828 * respectively.
829 */
830__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
831 unsigned long start, unsigned long nr_pages,
832 int write, int force, struct page **pages,
833 unsigned int gup_flags)
834{
835 long ret;
836 int locked = 1;
837 down_read(&mm->mmap_sem);
838 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
839 pages, NULL, &locked, false, gup_flags);
840 if (locked)
841 up_read(&mm->mmap_sem);
842 return ret;
843}
844EXPORT_SYMBOL(__get_user_pages_unlocked);
845
846/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800847 * get_user_pages_unlocked() is suitable to replace the form:
848 *
849 * down_read(&mm->mmap_sem);
850 * get_user_pages(tsk, mm, ..., pages, NULL);
851 * up_read(&mm->mmap_sem);
852 *
853 * with:
854 *
855 * get_user_pages_unlocked(tsk, mm, ..., pages);
856 *
857 * It is functionally equivalent to get_user_pages_fast so
858 * get_user_pages_fast should be used instead, if the two parameters
859 * "tsk" and "mm" are respectively equal to current and current->mm,
860 * or if "force" shall be set to 1 (get_user_pages_fast misses the
861 * "force" parameter).
862 */
863long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
864 unsigned long start, unsigned long nr_pages,
865 int write, int force, struct page **pages)
866{
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800867 return __get_user_pages_unlocked(tsk, mm, start, nr_pages, write,
868 force, pages, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800869}
870EXPORT_SYMBOL(get_user_pages_unlocked);
871
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700872/*
873 * get_user_pages() - pin user pages in memory
874 * @tsk: the task_struct to use for page fault accounting, or
875 * NULL if faults are not to be recorded.
876 * @mm: mm_struct of target mm
877 * @start: starting user address
878 * @nr_pages: number of pages from start to pin
879 * @write: whether pages will be written to by the caller
880 * @force: whether to force access even when user mapping is currently
881 * protected (but never forces write access to shared mapping).
882 * @pages: array that receives pointers to the pages pinned.
883 * Should be at least nr_pages long. Or NULL, if caller
884 * only intends to ensure the pages are faulted in.
885 * @vmas: array of pointers to vmas corresponding to each page.
886 * Or NULL if the caller does not require them.
887 *
888 * Returns number of pages pinned. This may be fewer than the number
889 * requested. If nr_pages is 0 or negative, returns 0. If no pages
890 * were pinned, returns -errno. Each page returned must be released
891 * with a put_page() call when it is finished with. vmas will only
892 * remain valid while mmap_sem is held.
893 *
894 * Must be called with mmap_sem held for read or write.
895 *
896 * get_user_pages walks a process's page tables and takes a reference to
897 * each struct page that each user address corresponds to at a given
898 * instant. That is, it takes the page that would be accessed if a user
899 * thread accesses the given user virtual address at that instant.
900 *
901 * This does not guarantee that the page exists in the user mappings when
902 * get_user_pages returns, and there may even be a completely different
903 * page there in some cases (eg. if mmapped pagecache has been invalidated
904 * and subsequently re faulted). However it does guarantee that the page
905 * won't be freed completely. And mostly callers simply care that the page
906 * contains data that was valid *at some point in time*. Typically, an IO
907 * or similar operation cannot guarantee anything stronger anyway because
908 * locks can't be held over the syscall boundary.
909 *
910 * If write=0, the page must not be written to. If the page is written to,
911 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
912 * after the page is finished with, and before put_page is called.
913 *
914 * get_user_pages is typically used for fewer-copy IO operations, to get a
915 * handle on the memory by some means other than accesses via the user virtual
916 * addresses. The pages may be submitted for DMA to devices or accessed via
917 * their kernel linear mapping (via the kmap APIs). Care should be taken to
918 * use the correct cache flushing APIs.
919 *
920 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800921 *
922 * get_user_pages should be phased out in favor of
923 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
924 * should use get_user_pages because it cannot pass
925 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700926 */
927long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
928 unsigned long start, unsigned long nr_pages, int write,
929 int force, struct page **pages, struct vm_area_struct **vmas)
930{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800931 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800932 pages, vmas, NULL, false, FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700933}
934EXPORT_SYMBOL(get_user_pages);
935
936/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700937 * populate_vma_page_range() - populate a range of pages in the vma.
938 * @vma: target vma
939 * @start: start address
940 * @end: end address
941 * @nonblocking:
942 *
943 * This takes care of mlocking the pages too if VM_LOCKED is set.
944 *
945 * return 0 on success, negative error code on error.
946 *
947 * vma->vm_mm->mmap_sem must be held.
948 *
949 * If @nonblocking is NULL, it may be held for read or write and will
950 * be unperturbed.
951 *
952 * If @nonblocking is non-NULL, it must held for read only and may be
953 * released. If it's released, *@nonblocking will be set to 0.
954 */
955long populate_vma_page_range(struct vm_area_struct *vma,
956 unsigned long start, unsigned long end, int *nonblocking)
957{
958 struct mm_struct *mm = vma->vm_mm;
959 unsigned long nr_pages = (end - start) / PAGE_SIZE;
960 int gup_flags;
961
962 VM_BUG_ON(start & ~PAGE_MASK);
963 VM_BUG_ON(end & ~PAGE_MASK);
964 VM_BUG_ON_VMA(start < vma->vm_start, vma);
965 VM_BUG_ON_VMA(end > vma->vm_end, vma);
966 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
967
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800968 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
969 if (vma->vm_flags & VM_LOCKONFAULT)
970 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700971 /*
972 * We want to touch writable mappings with a write fault in order
973 * to break COW, except for shared mappings because these don't COW
974 * and we would not want to dirty them for nothing.
975 */
976 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
977 gup_flags |= FOLL_WRITE;
978
979 /*
980 * We want mlock to succeed for regions that have any permissions
981 * other than PROT_NONE.
982 */
983 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
984 gup_flags |= FOLL_FORCE;
985
986 /*
987 * We made sure addr is within a VMA, so the following will
988 * not result in a stack expansion that recurses back here.
989 */
990 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
991 NULL, NULL, nonblocking);
992}
993
994/*
995 * __mm_populate - populate and/or mlock pages within a range of address space.
996 *
997 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
998 * flags. VMAs must be already marked with the desired vm_flags, and
999 * mmap_sem must not be held.
1000 */
1001int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1002{
1003 struct mm_struct *mm = current->mm;
1004 unsigned long end, nstart, nend;
1005 struct vm_area_struct *vma = NULL;
1006 int locked = 0;
1007 long ret = 0;
1008
1009 VM_BUG_ON(start & ~PAGE_MASK);
1010 VM_BUG_ON(len != PAGE_ALIGN(len));
1011 end = start + len;
1012
1013 for (nstart = start; nstart < end; nstart = nend) {
1014 /*
1015 * We want to fault in pages for [nstart; end) address range.
1016 * Find first corresponding VMA.
1017 */
1018 if (!locked) {
1019 locked = 1;
1020 down_read(&mm->mmap_sem);
1021 vma = find_vma(mm, nstart);
1022 } else if (nstart >= vma->vm_end)
1023 vma = vma->vm_next;
1024 if (!vma || vma->vm_start >= end)
1025 break;
1026 /*
1027 * Set [nstart; nend) to intersection of desired address
1028 * range with the first VMA. Also, skip undesirable VMA types.
1029 */
1030 nend = min(end, vma->vm_end);
1031 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1032 continue;
1033 if (nstart < vma->vm_start)
1034 nstart = vma->vm_start;
1035 /*
1036 * Now fault in a range of pages. populate_vma_page_range()
1037 * double checks the vma flags, so that it won't mlock pages
1038 * if the vma was already munlocked.
1039 */
1040 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1041 if (ret < 0) {
1042 if (ignore_errors) {
1043 ret = 0;
1044 continue; /* continue at next VMA */
1045 }
1046 break;
1047 }
1048 nend = nstart + ret * PAGE_SIZE;
1049 ret = 0;
1050 }
1051 if (locked)
1052 up_read(&mm->mmap_sem);
1053 return ret; /* 0 or negative error code */
1054}
1055
1056/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001057 * get_dump_page() - pin user page in memory while writing it to core dump
1058 * @addr: user address
1059 *
1060 * Returns struct page pointer of user page pinned for dump,
1061 * to be freed afterwards by page_cache_release() or put_page().
1062 *
1063 * Returns NULL on any kind of failure - a hole must then be inserted into
1064 * the corefile, to preserve alignment with its headers; and also returns
1065 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1066 * allowing a hole to be left in the corefile to save diskspace.
1067 *
1068 * Called without mmap_sem, but after all other threads have been killed.
1069 */
1070#ifdef CONFIG_ELF_CORE
1071struct page *get_dump_page(unsigned long addr)
1072{
1073 struct vm_area_struct *vma;
1074 struct page *page;
1075
1076 if (__get_user_pages(current, current->mm, addr, 1,
1077 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1078 NULL) < 1)
1079 return NULL;
1080 flush_cache_page(vma, addr, page_to_pfn(page));
1081 return page;
1082}
1083#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001084
1085/*
1086 * Generic RCU Fast GUP
1087 *
1088 * get_user_pages_fast attempts to pin user pages by walking the page
1089 * tables directly and avoids taking locks. Thus the walker needs to be
1090 * protected from page table pages being freed from under it, and should
1091 * block any THP splits.
1092 *
1093 * One way to achieve this is to have the walker disable interrupts, and
1094 * rely on IPIs from the TLB flushing code blocking before the page table
1095 * pages are freed. This is unsuitable for architectures that do not need
1096 * to broadcast an IPI when invalidating TLBs.
1097 *
1098 * Another way to achieve this is to batch up page table containing pages
1099 * belonging to more than one mm_user, then rcu_sched a callback to free those
1100 * pages. Disabling interrupts will allow the fast_gup walker to both block
1101 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1102 * (which is a relatively rare event). The code below adopts this strategy.
1103 *
1104 * Before activating this code, please be aware that the following assumptions
1105 * are currently made:
1106 *
1107 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1108 * pages containing page tables.
1109 *
Steve Capper2667f502014-10-09 15:29:14 -07001110 * *) ptes can be read atomically by the architecture.
1111 *
1112 * *) access_ok is sufficient to validate userspace address ranges.
1113 *
1114 * The last two assumptions can be relaxed by the addition of helper functions.
1115 *
1116 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1117 */
1118#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1119
1120#ifdef __HAVE_ARCH_PTE_SPECIAL
1121static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1122 int write, struct page **pages, int *nr)
1123{
1124 pte_t *ptep, *ptem;
1125 int ret = 0;
1126
1127 ptem = ptep = pte_offset_map(&pmd, addr);
1128 do {
1129 /*
1130 * In the line below we are assuming that the pte can be read
1131 * atomically. If this is not the case for your architecture,
1132 * please wrap this in a helper function!
1133 *
1134 * for an example see gup_get_pte in arch/x86/mm/gup.c
1135 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001136 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001137 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001138
1139 /*
1140 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001141 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001142 */
1143 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001144 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001145 goto pte_unmap;
1146
1147 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1148 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001149 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001150
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001151 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001152 goto pte_unmap;
1153
1154 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001155 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001156 goto pte_unmap;
1157 }
1158
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001159 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001160 pages[*nr] = page;
1161 (*nr)++;
1162
1163 } while (ptep++, addr += PAGE_SIZE, addr != end);
1164
1165 ret = 1;
1166
1167pte_unmap:
1168 pte_unmap(ptem);
1169 return ret;
1170}
1171#else
1172
1173/*
1174 * If we can't determine whether or not a pte is special, then fail immediately
1175 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1176 * to be special.
1177 *
1178 * For a futex to be placed on a THP tail page, get_futex_key requires a
1179 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1180 * useful to have gup_huge_pmd even if we can't operate on ptes.
1181 */
1182static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1183 int write, struct page **pages, int *nr)
1184{
1185 return 0;
1186}
1187#endif /* __HAVE_ARCH_PTE_SPECIAL */
1188
1189static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1190 unsigned long end, int write, struct page **pages, int *nr)
1191{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001192 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001193 int refs;
1194
1195 if (write && !pmd_write(orig))
1196 return 0;
1197
1198 refs = 0;
1199 head = pmd_page(orig);
1200 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001201 do {
1202 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1203 pages[*nr] = page;
1204 (*nr)++;
1205 page++;
1206 refs++;
1207 } while (addr += PAGE_SIZE, addr != end);
1208
1209 if (!page_cache_add_speculative(head, refs)) {
1210 *nr -= refs;
1211 return 0;
1212 }
1213
1214 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1215 *nr -= refs;
1216 while (refs--)
1217 put_page(head);
1218 return 0;
1219 }
1220
Steve Capper2667f502014-10-09 15:29:14 -07001221 return 1;
1222}
1223
1224static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1225 unsigned long end, int write, struct page **pages, int *nr)
1226{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001227 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001228 int refs;
1229
1230 if (write && !pud_write(orig))
1231 return 0;
1232
1233 refs = 0;
1234 head = pud_page(orig);
1235 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001236 do {
1237 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1238 pages[*nr] = page;
1239 (*nr)++;
1240 page++;
1241 refs++;
1242 } while (addr += PAGE_SIZE, addr != end);
1243
1244 if (!page_cache_add_speculative(head, refs)) {
1245 *nr -= refs;
1246 return 0;
1247 }
1248
1249 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1250 *nr -= refs;
1251 while (refs--)
1252 put_page(head);
1253 return 0;
1254 }
1255
Steve Capper2667f502014-10-09 15:29:14 -07001256 return 1;
1257}
1258
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301259static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1260 unsigned long end, int write,
1261 struct page **pages, int *nr)
1262{
1263 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001264 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301265
1266 if (write && !pgd_write(orig))
1267 return 0;
1268
1269 refs = 0;
1270 head = pgd_page(orig);
1271 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301272 do {
1273 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1274 pages[*nr] = page;
1275 (*nr)++;
1276 page++;
1277 refs++;
1278 } while (addr += PAGE_SIZE, addr != end);
1279
1280 if (!page_cache_add_speculative(head, refs)) {
1281 *nr -= refs;
1282 return 0;
1283 }
1284
1285 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1286 *nr -= refs;
1287 while (refs--)
1288 put_page(head);
1289 return 0;
1290 }
1291
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301292 return 1;
1293}
1294
Steve Capper2667f502014-10-09 15:29:14 -07001295static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1296 int write, struct page **pages, int *nr)
1297{
1298 unsigned long next;
1299 pmd_t *pmdp;
1300
1301 pmdp = pmd_offset(&pud, addr);
1302 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001303 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001304
1305 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001306 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001307 return 0;
1308
1309 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1310 /*
1311 * NUMA hinting faults need to be handled in the GUP
1312 * slowpath for accounting purposes and so that they
1313 * can be serialised against THP migration.
1314 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001315 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001316 return 0;
1317
1318 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1319 pages, nr))
1320 return 0;
1321
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301322 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1323 /*
1324 * architecture have different format for hugetlbfs
1325 * pmd format and THP pmd format
1326 */
1327 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1328 PMD_SHIFT, next, write, pages, nr))
1329 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001330 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1331 return 0;
1332 } while (pmdp++, addr = next, addr != end);
1333
1334 return 1;
1335}
1336
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301337static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1338 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001339{
1340 unsigned long next;
1341 pud_t *pudp;
1342
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301343 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001344 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001345 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001346
1347 next = pud_addr_end(addr, end);
1348 if (pud_none(pud))
1349 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301350 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001351 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301352 pages, nr))
1353 return 0;
1354 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1355 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1356 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001357 return 0;
1358 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1359 return 0;
1360 } while (pudp++, addr = next, addr != end);
1361
1362 return 1;
1363}
1364
1365/*
1366 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1367 * the regular GUP. It will only return non-negative values.
1368 */
1369int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1370 struct page **pages)
1371{
1372 struct mm_struct *mm = current->mm;
1373 unsigned long addr, len, end;
1374 unsigned long next, flags;
1375 pgd_t *pgdp;
1376 int nr = 0;
1377
1378 start &= PAGE_MASK;
1379 addr = start;
1380 len = (unsigned long) nr_pages << PAGE_SHIFT;
1381 end = start + len;
1382
1383 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1384 start, len)))
1385 return 0;
1386
1387 /*
1388 * Disable interrupts. We use the nested form as we can already have
1389 * interrupts disabled by get_futex_key.
1390 *
1391 * With interrupts disabled, we block page table pages from being
1392 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1393 * for more details.
1394 *
1395 * We do not adopt an rcu_read_lock(.) here as we also want to
1396 * block IPIs that come from THPs splitting.
1397 */
1398
1399 local_irq_save(flags);
1400 pgdp = pgd_offset(mm, addr);
1401 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001402 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301403
Steve Capper2667f502014-10-09 15:29:14 -07001404 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301405 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001406 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301407 if (unlikely(pgd_huge(pgd))) {
1408 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1409 pages, &nr))
1410 break;
1411 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1412 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1413 PGDIR_SHIFT, next, write, pages, &nr))
1414 break;
1415 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001416 break;
1417 } while (pgdp++, addr = next, addr != end);
1418 local_irq_restore(flags);
1419
1420 return nr;
1421}
1422
1423/**
1424 * get_user_pages_fast() - pin user pages in memory
1425 * @start: starting user address
1426 * @nr_pages: number of pages from start to pin
1427 * @write: whether pages will be written to
1428 * @pages: array that receives pointers to the pages pinned.
1429 * Should be at least nr_pages long.
1430 *
1431 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1432 * If not successful, it will fall back to taking the lock and
1433 * calling get_user_pages().
1434 *
1435 * Returns number of pages pinned. This may be fewer than the number
1436 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1437 * were pinned, returns -errno.
1438 */
1439int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1440 struct page **pages)
1441{
1442 struct mm_struct *mm = current->mm;
1443 int nr, ret;
1444
1445 start &= PAGE_MASK;
1446 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1447 ret = nr;
1448
1449 if (nr < nr_pages) {
1450 /* Try to get the remaining pages with get_user_pages */
1451 start += nr << PAGE_SHIFT;
1452 pages += nr;
1453
Andrea Arcangelia7b78072015-02-11 15:27:23 -08001454 ret = get_user_pages_unlocked(current, mm, start,
1455 nr_pages - nr, write, 0, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001456
1457 /* Have to be a bit careful with return values */
1458 if (nr > 0) {
1459 if (ret < 0)
1460 ret = nr;
1461 else
1462 ret += nr;
1463 }
1464 }
1465
1466 return ret;
1467}
1468
1469#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */