blob: c057784c844456f237adc9065bb4c9a3c230fdc6 [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
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070063static struct page *follow_page_pte(struct vm_area_struct *vma,
64 unsigned long address, pmd_t *pmd, unsigned int flags)
65{
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070066 struct mm_struct *mm = vma->vm_mm;
Dan Williams3565fce2016-01-15 16:56:55 -080067 struct dev_pagemap *pgmap = NULL;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070068 struct page *page;
69 spinlock_t *ptl;
70 pte_t *ptep, pte;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070071
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070072retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070073 if (unlikely(pmd_bad(*pmd)))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070074 return no_page_table(vma, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070075
76 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070077 pte = *ptep;
78 if (!pte_present(pte)) {
79 swp_entry_t entry;
80 /*
81 * KSM's break_ksm() relies upon recognizing a ksm page
82 * even while it is being migrated, so for that case we
83 * need migration_entry_wait().
84 */
85 if (likely(!(flags & FOLL_MIGRATION)))
86 goto no_page;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -080087 if (pte_none(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070088 goto no_page;
89 entry = pte_to_swp_entry(pte);
90 if (!is_migration_entry(entry))
91 goto no_page;
92 pte_unmap_unlock(ptep, ptl);
93 migration_entry_wait(mm, pmd, address);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070094 goto retry;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070095 }
Mel Gorman8a0516e2015-02-12 14:58:22 -080096 if ((flags & FOLL_NUMA) && pte_protnone(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070097 goto no_page;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070098 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
99 pte_unmap_unlock(ptep, ptl);
100 return NULL;
101 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700102
103 page = vm_normal_page(vma, address, pte);
Dan Williams3565fce2016-01-15 16:56:55 -0800104 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
105 /*
106 * Only return device mapping pages in the FOLL_GET case since
107 * they are only valid while holding the pgmap reference.
108 */
109 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
110 if (pgmap)
111 page = pte_page(pte);
112 else
113 goto no_page;
114 } else if (unlikely(!page)) {
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700115 if (flags & FOLL_DUMP) {
116 /* Avoid special (like zero) pages in core dumps */
117 page = ERR_PTR(-EFAULT);
118 goto out;
119 }
120
121 if (is_zero_pfn(pte_pfn(pte))) {
122 page = pte_page(pte);
123 } else {
124 int ret;
125
126 ret = follow_pfn_pte(vma, address, ptep, flags);
127 page = ERR_PTR(ret);
128 goto out;
129 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700130 }
131
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800132 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
133 int ret;
134 get_page(page);
135 pte_unmap_unlock(ptep, ptl);
136 lock_page(page);
137 ret = split_huge_page(page);
138 unlock_page(page);
139 put_page(page);
140 if (ret)
141 return ERR_PTR(ret);
142 goto retry;
143 }
144
Dan Williams3565fce2016-01-15 16:56:55 -0800145 if (flags & FOLL_GET) {
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800146 get_page(page);
Dan Williams3565fce2016-01-15 16:56:55 -0800147
148 /* drop the pgmap reference now that we hold the page */
149 if (pgmap) {
150 put_dev_pagemap(pgmap);
151 pgmap = NULL;
152 }
153 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700154 if (flags & FOLL_TOUCH) {
155 if ((flags & FOLL_WRITE) &&
156 !pte_dirty(pte) && !PageDirty(page))
157 set_page_dirty(page);
158 /*
159 * pte_mkyoung() would be more correct here, but atomic care
160 * is needed to avoid losing the dirty bit: it is easier to use
161 * mark_page_accessed().
162 */
163 mark_page_accessed(page);
164 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800165 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -0800166 /* Do not mlock pte-mapped THP */
167 if (PageTransCompound(page))
168 goto out;
169
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700170 /*
171 * The preliminary mapping check is mainly to avoid the
172 * pointless overhead of lock_page on the ZERO_PAGE
173 * which might bounce very badly if there is contention.
174 *
175 * If the page is already locked, we don't need to
176 * handle it now - vmscan will handle it later if and
177 * when it attempts to reclaim the page.
178 */
179 if (page->mapping && trylock_page(page)) {
180 lru_add_drain(); /* push cached pages to LRU */
181 /*
182 * Because we lock page here, and migration is
183 * blocked by the pte's page reference, and we
184 * know the page is still mapped, we don't even
185 * need to check for file-cache page truncation.
186 */
187 mlock_vma_page(page);
188 unlock_page(page);
189 }
190 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700191out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700192 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700193 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700194no_page:
195 pte_unmap_unlock(ptep, ptl);
196 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700197 return NULL;
198 return no_page_table(vma, flags);
199}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700200
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700201/**
202 * follow_page_mask - look up a page descriptor from a user-virtual address
203 * @vma: vm_area_struct mapping @address
204 * @address: virtual address to look up
205 * @flags: flags modifying lookup behaviour
206 * @page_mask: on output, *page_mask is set according to the size of the page
207 *
208 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
209 *
210 * Returns the mapped (struct page *), %NULL if no mapping exists, or
211 * an error pointer if there is a mapping to something not represented
212 * by a page descriptor (see also vm_normal_page()).
213 */
214struct page *follow_page_mask(struct vm_area_struct *vma,
215 unsigned long address, unsigned int flags,
216 unsigned int *page_mask)
217{
218 pgd_t *pgd;
219 pud_t *pud;
220 pmd_t *pmd;
221 spinlock_t *ptl;
222 struct page *page;
223 struct mm_struct *mm = vma->vm_mm;
224
225 *page_mask = 0;
226
227 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
228 if (!IS_ERR(page)) {
229 BUG_ON(flags & FOLL_GET);
230 return page;
231 }
232
233 pgd = pgd_offset(mm, address);
234 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
235 return no_page_table(vma, flags);
236
237 pud = pud_offset(pgd, address);
238 if (pud_none(*pud))
239 return no_page_table(vma, flags);
240 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800241 page = follow_huge_pud(mm, address, pud, flags);
242 if (page)
243 return page;
244 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700245 }
246 if (unlikely(pud_bad(*pud)))
247 return no_page_table(vma, flags);
248
249 pmd = pmd_offset(pud, address);
250 if (pmd_none(*pmd))
251 return no_page_table(vma, flags);
252 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800253 page = follow_huge_pmd(mm, address, pmd, flags);
254 if (page)
255 return page;
256 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700257 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800258 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700259 return no_page_table(vma, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800260 if (pmd_devmap(*pmd)) {
261 ptl = pmd_lock(mm, pmd);
262 page = follow_devmap_pmd(vma, address, pmd, flags);
263 spin_unlock(ptl);
264 if (page)
265 return page;
266 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800267 if (likely(!pmd_trans_huge(*pmd)))
268 return follow_page_pte(vma, address, pmd, flags);
269
270 ptl = pmd_lock(mm, pmd);
271 if (unlikely(!pmd_trans_huge(*pmd))) {
272 spin_unlock(ptl);
273 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700274 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800275 if (flags & FOLL_SPLIT) {
276 int ret;
277 page = pmd_page(*pmd);
278 if (is_huge_zero_page(page)) {
279 spin_unlock(ptl);
280 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800281 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800282 } else {
283 get_page(page);
284 spin_unlock(ptl);
285 lock_page(page);
286 ret = split_huge_page(page);
287 unlock_page(page);
288 put_page(page);
289 }
290
291 return ret ? ERR_PTR(ret) :
292 follow_page_pte(vma, address, pmd, flags);
293 }
294
295 page = follow_trans_huge_pmd(vma, address, pmd, flags);
296 spin_unlock(ptl);
297 *page_mask = HPAGE_PMD_NR - 1;
298 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700299}
300
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700301static int get_gate_page(struct mm_struct *mm, unsigned long address,
302 unsigned int gup_flags, struct vm_area_struct **vma,
303 struct page **page)
304{
305 pgd_t *pgd;
306 pud_t *pud;
307 pmd_t *pmd;
308 pte_t *pte;
309 int ret = -EFAULT;
310
311 /* user gate pages are read-only */
312 if (gup_flags & FOLL_WRITE)
313 return -EFAULT;
314 if (address > TASK_SIZE)
315 pgd = pgd_offset_k(address);
316 else
317 pgd = pgd_offset_gate(mm, address);
318 BUG_ON(pgd_none(*pgd));
319 pud = pud_offset(pgd, address);
320 BUG_ON(pud_none(*pud));
321 pmd = pmd_offset(pud, address);
322 if (pmd_none(*pmd))
323 return -EFAULT;
324 VM_BUG_ON(pmd_trans_huge(*pmd));
325 pte = pte_offset_map(pmd, address);
326 if (pte_none(*pte))
327 goto unmap;
328 *vma = get_gate_vma(mm);
329 if (!page)
330 goto out;
331 *page = vm_normal_page(*vma, address, *pte);
332 if (!*page) {
333 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
334 goto unmap;
335 *page = pte_page(*pte);
336 }
337 get_page(*page);
338out:
339 ret = 0;
340unmap:
341 pte_unmap(pte);
342 return ret;
343}
344
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700345/*
346 * mmap_sem must be held on entry. If @nonblocking != NULL and
347 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
348 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
349 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700350static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
351 unsigned long address, unsigned int *flags, int *nonblocking)
352{
353 struct mm_struct *mm = vma->vm_mm;
354 unsigned int fault_flags = 0;
355 int ret;
356
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800357 /* mlock all present pages, but do not fault in new pages */
358 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
359 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700360 /* For mm_populate(), just skip the stack guard page. */
361 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700362 (stack_guard_page_start(vma, address) ||
363 stack_guard_page_end(vma, address + PAGE_SIZE)))
364 return -ENOENT;
365 if (*flags & FOLL_WRITE)
366 fault_flags |= FAULT_FLAG_WRITE;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800367 if (*flags & FOLL_REMOTE)
368 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700369 if (nonblocking)
370 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
371 if (*flags & FOLL_NOWAIT)
372 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700373 if (*flags & FOLL_TRIED) {
374 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
375 fault_flags |= FAULT_FLAG_TRIED;
376 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700377
378 ret = handle_mm_fault(mm, vma, address, fault_flags);
379 if (ret & VM_FAULT_ERROR) {
380 if (ret & VM_FAULT_OOM)
381 return -ENOMEM;
382 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
383 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800384 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700385 return -EFAULT;
386 BUG();
387 }
388
389 if (tsk) {
390 if (ret & VM_FAULT_MAJOR)
391 tsk->maj_flt++;
392 else
393 tsk->min_flt++;
394 }
395
396 if (ret & VM_FAULT_RETRY) {
397 if (nonblocking)
398 *nonblocking = 0;
399 return -EBUSY;
400 }
401
402 /*
403 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
404 * necessary, even if maybe_mkwrite decided not to set pte_write. We
405 * can thus safely do subsequent page lookups as if they were reads.
406 * But only do so when looping for pte_write is futile: in some cases
407 * userspace may also be wanting to write to the gotten user page,
408 * which a read fault here might prevent (a readonly page might get
409 * reCOWed by userspace write).
410 */
411 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
412 *flags &= ~FOLL_WRITE;
413 return 0;
414}
415
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700416static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
417{
418 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800419 int write = (gup_flags & FOLL_WRITE);
420 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700421
422 if (vm_flags & (VM_IO | VM_PFNMAP))
423 return -EFAULT;
424
Dave Hansen1b2ee122016-02-12 13:02:21 -0800425 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700426 if (!(vm_flags & VM_WRITE)) {
427 if (!(gup_flags & FOLL_FORCE))
428 return -EFAULT;
429 /*
430 * We used to let the write,force case do COW in a
431 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
432 * set a breakpoint in a read-only mapping of an
433 * executable, without corrupting the file (yet only
434 * when that file had been opened for writing!).
435 * Anon pages in shared mappings are surprising: now
436 * just reject it.
437 */
Hugh Dickins46435362016-01-30 18:03:16 -0800438 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700439 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700440 }
441 } else if (!(vm_flags & VM_READ)) {
442 if (!(gup_flags & FOLL_FORCE))
443 return -EFAULT;
444 /*
445 * Is there actually any vma we can reach here which does not
446 * have VM_MAYREAD set?
447 */
448 if (!(vm_flags & VM_MAYREAD))
449 return -EFAULT;
450 }
Dave Hansend61172b2016-02-12 13:02:24 -0800451 /*
452 * gups are always data accesses, not instruction
453 * fetches, so execute=false here
454 */
455 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800456 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700457 return 0;
458}
459
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700460/**
461 * __get_user_pages() - pin user pages in memory
462 * @tsk: task_struct of target task
463 * @mm: mm_struct of target mm
464 * @start: starting user address
465 * @nr_pages: number of pages from start to pin
466 * @gup_flags: flags modifying pin behaviour
467 * @pages: array that receives pointers to the pages pinned.
468 * Should be at least nr_pages long. Or NULL, if caller
469 * only intends to ensure the pages are faulted in.
470 * @vmas: array of pointers to vmas corresponding to each page.
471 * Or NULL if the caller does not require them.
472 * @nonblocking: whether waiting for disk IO or mmap_sem contention
473 *
474 * Returns number of pages pinned. This may be fewer than the number
475 * requested. If nr_pages is 0 or negative, returns 0. If no pages
476 * were pinned, returns -errno. Each page returned must be released
477 * with a put_page() call when it is finished with. vmas will only
478 * remain valid while mmap_sem is held.
479 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700480 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700481 *
482 * __get_user_pages walks a process's page tables and takes a reference to
483 * each struct page that each user address corresponds to at a given
484 * instant. That is, it takes the page that would be accessed if a user
485 * thread accesses the given user virtual address at that instant.
486 *
487 * This does not guarantee that the page exists in the user mappings when
488 * __get_user_pages returns, and there may even be a completely different
489 * page there in some cases (eg. if mmapped pagecache has been invalidated
490 * and subsequently re faulted). However it does guarantee that the page
491 * won't be freed completely. And mostly callers simply care that the page
492 * contains data that was valid *at some point in time*. Typically, an IO
493 * or similar operation cannot guarantee anything stronger anyway because
494 * locks can't be held over the syscall boundary.
495 *
496 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
497 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
498 * appropriate) must be called after the page is finished with, and
499 * before put_page is called.
500 *
501 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
502 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700503 * *@nonblocking will be set to 0. Further, if @gup_flags does not
504 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
505 * this case.
506 *
507 * A caller using such a combination of @nonblocking and @gup_flags
508 * must therefore hold the mmap_sem for reading only, and recognize
509 * when it's been released. Otherwise, it must be held for either
510 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700511 *
512 * In most cases, get_user_pages or get_user_pages_fast should be used
513 * instead of __get_user_pages. __get_user_pages should be used only if
514 * you need some special @gup_flags.
515 */
516long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
517 unsigned long start, unsigned long nr_pages,
518 unsigned int gup_flags, struct page **pages,
519 struct vm_area_struct **vmas, int *nonblocking)
520{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700521 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700522 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700523 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700524
525 if (!nr_pages)
526 return 0;
527
528 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
529
530 /*
531 * If FOLL_FORCE is set then do not force a full fault as the hinting
532 * fault information is unrelated to the reference behaviour of a task
533 * using the address space
534 */
535 if (!(gup_flags & FOLL_FORCE))
536 gup_flags |= FOLL_NUMA;
537
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700538 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700539 struct page *page;
540 unsigned int foll_flags = gup_flags;
541 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700542
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700543 /* first iteration or cross vma bound */
544 if (!vma || start >= vma->vm_end) {
545 vma = find_extend_vma(mm, start);
546 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700547 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700548 ret = get_gate_page(mm, start & PAGE_MASK,
549 gup_flags, &vma,
550 pages ? &pages[i] : NULL);
551 if (ret)
552 return i ? : ret;
553 page_mask = 0;
554 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700555 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700556
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700557 if (!vma || check_vma_flags(vma, gup_flags))
558 return i ? : -EFAULT;
559 if (is_vm_hugetlb_page(vma)) {
560 i = follow_hugetlb_page(mm, vma, pages, vmas,
561 &start, &nr_pages, i,
562 gup_flags);
563 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700564 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700565 }
566retry:
567 /*
568 * If we have a pending SIGKILL, don't keep faulting pages and
569 * potentially allocating memory.
570 */
571 if (unlikely(fatal_signal_pending(current)))
572 return i ? i : -ERESTARTSYS;
573 cond_resched();
574 page = follow_page_mask(vma, start, foll_flags, &page_mask);
575 if (!page) {
576 int ret;
577 ret = faultin_page(tsk, vma, start, &foll_flags,
578 nonblocking);
579 switch (ret) {
580 case 0:
581 goto retry;
582 case -EFAULT:
583 case -ENOMEM:
584 case -EHWPOISON:
585 return i ? i : ret;
586 case -EBUSY:
587 return i;
588 case -ENOENT:
589 goto next_page;
590 }
591 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700592 } else if (PTR_ERR(page) == -EEXIST) {
593 /*
594 * Proper page table entry exists, but no corresponding
595 * struct page.
596 */
597 goto next_page;
598 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700599 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700600 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700601 if (pages) {
602 pages[i] = page;
603 flush_anon_page(vma, page, start);
604 flush_dcache_page(page);
605 page_mask = 0;
606 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700607next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700608 if (vmas) {
609 vmas[i] = vma;
610 page_mask = 0;
611 }
612 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
613 if (page_increm > nr_pages)
614 page_increm = nr_pages;
615 i += page_increm;
616 start += page_increm * PAGE_SIZE;
617 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700618 } while (nr_pages);
619 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700620}
621EXPORT_SYMBOL(__get_user_pages);
622
Dave Hansend4925e02016-02-12 13:02:16 -0800623bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
624{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800625 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
626 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800627 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800628
629 if (!(vm_flags & vma->vm_flags))
630 return false;
631
Dave Hansen33a709b2016-02-12 13:02:19 -0800632 /*
633 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800634 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800635 *
636 * gup always represents data access, not instruction
637 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800638 */
Dave Hansend61172b2016-02-12 13:02:24 -0800639 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800640 return false;
641
Dave Hansend4925e02016-02-12 13:02:16 -0800642 return true;
643}
644
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700645/*
646 * fixup_user_fault() - manually resolve a user page fault
647 * @tsk: the task_struct to use for page fault accounting, or
648 * NULL if faults are not to be recorded.
649 * @mm: mm_struct of target mm
650 * @address: user address
651 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800652 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
653 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700654 *
655 * This is meant to be called in the specific scenario where for locking reasons
656 * we try to access user memory in atomic context (within a pagefault_disable()
657 * section), this returns -EFAULT, and we want to resolve the user fault before
658 * trying again.
659 *
660 * Typically this is meant to be used by the futex code.
661 *
662 * The main difference with get_user_pages() is that this function will
663 * unconditionally call handle_mm_fault() which will in turn perform all the
664 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800665 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700666 *
667 * This is important for some architectures where those bits also gate the
668 * access permission to the page because they are maintained in software. On
669 * such architectures, gup() will not be enough to make a subsequent access
670 * succeed.
671 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800672 * This function will not return with an unlocked mmap_sem. So it has not the
673 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700674 */
675int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800676 unsigned long address, unsigned int fault_flags,
677 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700678{
679 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800680 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700681
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800682 if (unlocked)
683 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
684
685retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700686 vma = find_extend_vma(mm, address);
687 if (!vma || address < vma->vm_start)
688 return -EFAULT;
689
Dave Hansend4925e02016-02-12 13:02:16 -0800690 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700691 return -EFAULT;
692
693 ret = handle_mm_fault(mm, vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800694 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700695 if (ret & VM_FAULT_ERROR) {
696 if (ret & VM_FAULT_OOM)
697 return -ENOMEM;
698 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
699 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800700 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700701 return -EFAULT;
702 BUG();
703 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800704
705 if (ret & VM_FAULT_RETRY) {
706 down_read(&mm->mmap_sem);
707 if (!(fault_flags & FAULT_FLAG_TRIED)) {
708 *unlocked = true;
709 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
710 fault_flags |= FAULT_FLAG_TRIED;
711 goto retry;
712 }
713 }
714
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700715 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800716 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700717 tsk->maj_flt++;
718 else
719 tsk->min_flt++;
720 }
721 return 0;
722}
723
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800724static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
725 struct mm_struct *mm,
726 unsigned long start,
727 unsigned long nr_pages,
728 int write, int force,
729 struct page **pages,
730 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800731 int *locked, bool notify_drop,
732 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800733{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800734 long ret, pages_done;
735 bool lock_dropped;
736
737 if (locked) {
738 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
739 BUG_ON(vmas);
740 /* check caller initialized locked */
741 BUG_ON(*locked != 1);
742 }
743
744 if (pages)
745 flags |= FOLL_GET;
746 if (write)
747 flags |= FOLL_WRITE;
748 if (force)
749 flags |= FOLL_FORCE;
750
751 pages_done = 0;
752 lock_dropped = false;
753 for (;;) {
754 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
755 vmas, locked);
756 if (!locked)
757 /* VM_FAULT_RETRY couldn't trigger, bypass */
758 return ret;
759
760 /* VM_FAULT_RETRY cannot return errors */
761 if (!*locked) {
762 BUG_ON(ret < 0);
763 BUG_ON(ret >= nr_pages);
764 }
765
766 if (!pages)
767 /* If it's a prefault don't insist harder */
768 return ret;
769
770 if (ret > 0) {
771 nr_pages -= ret;
772 pages_done += ret;
773 if (!nr_pages)
774 break;
775 }
776 if (*locked) {
777 /* VM_FAULT_RETRY didn't trigger */
778 if (!pages_done)
779 pages_done = ret;
780 break;
781 }
782 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
783 pages += ret;
784 start += ret << PAGE_SHIFT;
785
786 /*
787 * Repeat on the address that fired VM_FAULT_RETRY
788 * without FAULT_FLAG_ALLOW_RETRY but with
789 * FAULT_FLAG_TRIED.
790 */
791 *locked = 1;
792 lock_dropped = true;
793 down_read(&mm->mmap_sem);
794 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
795 pages, NULL, NULL);
796 if (ret != 1) {
797 BUG_ON(ret > 1);
798 if (!pages_done)
799 pages_done = ret;
800 break;
801 }
802 nr_pages--;
803 pages_done++;
804 if (!nr_pages)
805 break;
806 pages++;
807 start += PAGE_SIZE;
808 }
809 if (notify_drop && lock_dropped && *locked) {
810 /*
811 * We must let the caller know we temporarily dropped the lock
812 * and so the critical section protected by it was lost.
813 */
814 up_read(&mm->mmap_sem);
815 *locked = 0;
816 }
817 return pages_done;
818}
819
820/*
821 * We can leverage the VM_FAULT_RETRY functionality in the page fault
822 * paths better by using either get_user_pages_locked() or
823 * get_user_pages_unlocked().
824 *
825 * get_user_pages_locked() is suitable to replace the form:
826 *
827 * down_read(&mm->mmap_sem);
828 * do_something()
829 * get_user_pages(tsk, mm, ..., pages, NULL);
830 * up_read(&mm->mmap_sem);
831 *
832 * to:
833 *
834 * int locked = 1;
835 * down_read(&mm->mmap_sem);
836 * do_something()
837 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
838 * if (locked)
839 * up_read(&mm->mmap_sem);
840 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200841long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800842 int write, int force, struct page **pages,
843 int *locked)
844{
Dave Hansencde70142016-02-12 13:01:55 -0800845 return __get_user_pages_locked(current, current->mm, start, nr_pages,
846 write, force, pages, NULL, locked, true,
847 FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800848}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200849EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800850
851/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800852 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
853 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
854 *
855 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
856 * caller if required (just like with __get_user_pages). "FOLL_GET",
857 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
858 * according to the parameters "pages", "write", "force"
859 * respectively.
860 */
861__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
862 unsigned long start, unsigned long nr_pages,
863 int write, int force, struct page **pages,
864 unsigned int gup_flags)
865{
866 long ret;
867 int locked = 1;
868 down_read(&mm->mmap_sem);
869 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
870 pages, NULL, &locked, false, gup_flags);
871 if (locked)
872 up_read(&mm->mmap_sem);
873 return ret;
874}
875EXPORT_SYMBOL(__get_user_pages_unlocked);
876
877/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800878 * get_user_pages_unlocked() is suitable to replace the form:
879 *
880 * down_read(&mm->mmap_sem);
881 * get_user_pages(tsk, mm, ..., pages, NULL);
882 * up_read(&mm->mmap_sem);
883 *
884 * with:
885 *
886 * get_user_pages_unlocked(tsk, mm, ..., pages);
887 *
888 * It is functionally equivalent to get_user_pages_fast so
889 * get_user_pages_fast should be used instead, if the two parameters
890 * "tsk" and "mm" are respectively equal to current and current->mm,
891 * or if "force" shall be set to 1 (get_user_pages_fast misses the
892 * "force" parameter).
893 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200894long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800895 int write, int force, struct page **pages)
896{
Dave Hansencde70142016-02-12 13:01:55 -0800897 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
898 write, force, pages, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800899}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200900EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800901
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700902/*
Dave Hansen1e987792016-02-12 13:01:54 -0800903 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700904 * @tsk: the task_struct to use for page fault accounting, or
905 * NULL if faults are not to be recorded.
906 * @mm: mm_struct of target mm
907 * @start: starting user address
908 * @nr_pages: number of pages from start to pin
909 * @write: whether pages will be written to by the caller
910 * @force: whether to force access even when user mapping is currently
911 * protected (but never forces write access to shared mapping).
912 * @pages: array that receives pointers to the pages pinned.
913 * Should be at least nr_pages long. Or NULL, if caller
914 * only intends to ensure the pages are faulted in.
915 * @vmas: array of pointers to vmas corresponding to each page.
916 * Or NULL if the caller does not require them.
917 *
918 * Returns number of pages pinned. This may be fewer than the number
919 * requested. If nr_pages is 0 or negative, returns 0. If no pages
920 * were pinned, returns -errno. Each page returned must be released
921 * with a put_page() call when it is finished with. vmas will only
922 * remain valid while mmap_sem is held.
923 *
924 * Must be called with mmap_sem held for read or write.
925 *
926 * get_user_pages walks a process's page tables and takes a reference to
927 * each struct page that each user address corresponds to at a given
928 * instant. That is, it takes the page that would be accessed if a user
929 * thread accesses the given user virtual address at that instant.
930 *
931 * This does not guarantee that the page exists in the user mappings when
932 * get_user_pages returns, and there may even be a completely different
933 * page there in some cases (eg. if mmapped pagecache has been invalidated
934 * and subsequently re faulted). However it does guarantee that the page
935 * won't be freed completely. And mostly callers simply care that the page
936 * contains data that was valid *at some point in time*. Typically, an IO
937 * or similar operation cannot guarantee anything stronger anyway because
938 * locks can't be held over the syscall boundary.
939 *
940 * If write=0, the page must not be written to. If the page is written to,
941 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
942 * after the page is finished with, and before put_page is called.
943 *
944 * get_user_pages is typically used for fewer-copy IO operations, to get a
945 * handle on the memory by some means other than accesses via the user virtual
946 * addresses. The pages may be submitted for DMA to devices or accessed via
947 * their kernel linear mapping (via the kmap APIs). Care should be taken to
948 * use the correct cache flushing APIs.
949 *
950 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800951 *
952 * get_user_pages should be phased out in favor of
953 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
954 * should use get_user_pages because it cannot pass
955 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700956 */
Dave Hansen1e987792016-02-12 13:01:54 -0800957long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
958 unsigned long start, unsigned long nr_pages,
959 int write, int force, struct page **pages,
960 struct vm_area_struct **vmas)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700961{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800962 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Dave Hansen1e987792016-02-12 13:01:54 -0800963 pages, vmas, NULL, false,
964 FOLL_TOUCH | FOLL_REMOTE);
965}
966EXPORT_SYMBOL(get_user_pages_remote);
967
968/*
Dave Hansend4edcf02016-02-12 13:01:56 -0800969 * This is the same as get_user_pages_remote(), just with a
970 * less-flexible calling convention where we assume that the task
971 * and mm being operated on are the current task's. We also
972 * obviously don't pass FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -0800973 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200974long get_user_pages(unsigned long start, unsigned long nr_pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800975 int write, int force, struct page **pages,
976 struct vm_area_struct **vmas)
977{
Dave Hansencde70142016-02-12 13:01:55 -0800978 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800979 write, force, pages, vmas, NULL, false,
980 FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700981}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200982EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700983
984/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700985 * populate_vma_page_range() - populate a range of pages in the vma.
986 * @vma: target vma
987 * @start: start address
988 * @end: end address
989 * @nonblocking:
990 *
991 * This takes care of mlocking the pages too if VM_LOCKED is set.
992 *
993 * return 0 on success, negative error code on error.
994 *
995 * vma->vm_mm->mmap_sem must be held.
996 *
997 * If @nonblocking is NULL, it may be held for read or write and will
998 * be unperturbed.
999 *
1000 * If @nonblocking is non-NULL, it must held for read only and may be
1001 * released. If it's released, *@nonblocking will be set to 0.
1002 */
1003long populate_vma_page_range(struct vm_area_struct *vma,
1004 unsigned long start, unsigned long end, int *nonblocking)
1005{
1006 struct mm_struct *mm = vma->vm_mm;
1007 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1008 int gup_flags;
1009
1010 VM_BUG_ON(start & ~PAGE_MASK);
1011 VM_BUG_ON(end & ~PAGE_MASK);
1012 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1013 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1014 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1015
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001016 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1017 if (vma->vm_flags & VM_LOCKONFAULT)
1018 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001019 /*
1020 * We want to touch writable mappings with a write fault in order
1021 * to break COW, except for shared mappings because these don't COW
1022 * and we would not want to dirty them for nothing.
1023 */
1024 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1025 gup_flags |= FOLL_WRITE;
1026
1027 /*
1028 * We want mlock to succeed for regions that have any permissions
1029 * other than PROT_NONE.
1030 */
1031 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1032 gup_flags |= FOLL_FORCE;
1033
1034 /*
1035 * We made sure addr is within a VMA, so the following will
1036 * not result in a stack expansion that recurses back here.
1037 */
1038 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1039 NULL, NULL, nonblocking);
1040}
1041
1042/*
1043 * __mm_populate - populate and/or mlock pages within a range of address space.
1044 *
1045 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1046 * flags. VMAs must be already marked with the desired vm_flags, and
1047 * mmap_sem must not be held.
1048 */
1049int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1050{
1051 struct mm_struct *mm = current->mm;
1052 unsigned long end, nstart, nend;
1053 struct vm_area_struct *vma = NULL;
1054 int locked = 0;
1055 long ret = 0;
1056
1057 VM_BUG_ON(start & ~PAGE_MASK);
1058 VM_BUG_ON(len != PAGE_ALIGN(len));
1059 end = start + len;
1060
1061 for (nstart = start; nstart < end; nstart = nend) {
1062 /*
1063 * We want to fault in pages for [nstart; end) address range.
1064 * Find first corresponding VMA.
1065 */
1066 if (!locked) {
1067 locked = 1;
1068 down_read(&mm->mmap_sem);
1069 vma = find_vma(mm, nstart);
1070 } else if (nstart >= vma->vm_end)
1071 vma = vma->vm_next;
1072 if (!vma || vma->vm_start >= end)
1073 break;
1074 /*
1075 * Set [nstart; nend) to intersection of desired address
1076 * range with the first VMA. Also, skip undesirable VMA types.
1077 */
1078 nend = min(end, vma->vm_end);
1079 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1080 continue;
1081 if (nstart < vma->vm_start)
1082 nstart = vma->vm_start;
1083 /*
1084 * Now fault in a range of pages. populate_vma_page_range()
1085 * double checks the vma flags, so that it won't mlock pages
1086 * if the vma was already munlocked.
1087 */
1088 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1089 if (ret < 0) {
1090 if (ignore_errors) {
1091 ret = 0;
1092 continue; /* continue at next VMA */
1093 }
1094 break;
1095 }
1096 nend = nstart + ret * PAGE_SIZE;
1097 ret = 0;
1098 }
1099 if (locked)
1100 up_read(&mm->mmap_sem);
1101 return ret; /* 0 or negative error code */
1102}
1103
1104/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001105 * get_dump_page() - pin user page in memory while writing it to core dump
1106 * @addr: user address
1107 *
1108 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001109 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001110 *
1111 * Returns NULL on any kind of failure - a hole must then be inserted into
1112 * the corefile, to preserve alignment with its headers; and also returns
1113 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1114 * allowing a hole to be left in the corefile to save diskspace.
1115 *
1116 * Called without mmap_sem, but after all other threads have been killed.
1117 */
1118#ifdef CONFIG_ELF_CORE
1119struct page *get_dump_page(unsigned long addr)
1120{
1121 struct vm_area_struct *vma;
1122 struct page *page;
1123
1124 if (__get_user_pages(current, current->mm, addr, 1,
1125 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1126 NULL) < 1)
1127 return NULL;
1128 flush_cache_page(vma, addr, page_to_pfn(page));
1129 return page;
1130}
1131#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001132
1133/*
1134 * Generic RCU Fast GUP
1135 *
1136 * get_user_pages_fast attempts to pin user pages by walking the page
1137 * tables directly and avoids taking locks. Thus the walker needs to be
1138 * protected from page table pages being freed from under it, and should
1139 * block any THP splits.
1140 *
1141 * One way to achieve this is to have the walker disable interrupts, and
1142 * rely on IPIs from the TLB flushing code blocking before the page table
1143 * pages are freed. This is unsuitable for architectures that do not need
1144 * to broadcast an IPI when invalidating TLBs.
1145 *
1146 * Another way to achieve this is to batch up page table containing pages
1147 * belonging to more than one mm_user, then rcu_sched a callback to free those
1148 * pages. Disabling interrupts will allow the fast_gup walker to both block
1149 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1150 * (which is a relatively rare event). The code below adopts this strategy.
1151 *
1152 * Before activating this code, please be aware that the following assumptions
1153 * are currently made:
1154 *
1155 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1156 * pages containing page tables.
1157 *
Steve Capper2667f502014-10-09 15:29:14 -07001158 * *) ptes can be read atomically by the architecture.
1159 *
1160 * *) access_ok is sufficient to validate userspace address ranges.
1161 *
1162 * The last two assumptions can be relaxed by the addition of helper functions.
1163 *
1164 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1165 */
1166#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1167
1168#ifdef __HAVE_ARCH_PTE_SPECIAL
1169static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1170 int write, struct page **pages, int *nr)
1171{
1172 pte_t *ptep, *ptem;
1173 int ret = 0;
1174
1175 ptem = ptep = pte_offset_map(&pmd, addr);
1176 do {
1177 /*
1178 * In the line below we are assuming that the pte can be read
1179 * atomically. If this is not the case for your architecture,
1180 * please wrap this in a helper function!
1181 *
1182 * for an example see gup_get_pte in arch/x86/mm/gup.c
1183 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001184 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001185 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001186
1187 /*
1188 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001189 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001190 */
1191 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001192 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001193 goto pte_unmap;
1194
Dave Hansen33a709b2016-02-12 13:02:19 -08001195 if (!arch_pte_access_permitted(pte, write))
1196 goto pte_unmap;
1197
Steve Capper2667f502014-10-09 15:29:14 -07001198 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1199 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001200 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001201
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001202 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001203 goto pte_unmap;
1204
1205 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001206 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001207 goto pte_unmap;
1208 }
1209
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001210 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001211 pages[*nr] = page;
1212 (*nr)++;
1213
1214 } while (ptep++, addr += PAGE_SIZE, addr != end);
1215
1216 ret = 1;
1217
1218pte_unmap:
1219 pte_unmap(ptem);
1220 return ret;
1221}
1222#else
1223
1224/*
1225 * If we can't determine whether or not a pte is special, then fail immediately
1226 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1227 * to be special.
1228 *
1229 * For a futex to be placed on a THP tail page, get_futex_key requires a
1230 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1231 * useful to have gup_huge_pmd even if we can't operate on ptes.
1232 */
1233static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1234 int write, struct page **pages, int *nr)
1235{
1236 return 0;
1237}
1238#endif /* __HAVE_ARCH_PTE_SPECIAL */
1239
1240static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1241 unsigned long end, int write, struct page **pages, int *nr)
1242{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001243 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001244 int refs;
1245
1246 if (write && !pmd_write(orig))
1247 return 0;
1248
1249 refs = 0;
1250 head = pmd_page(orig);
1251 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001252 do {
1253 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1254 pages[*nr] = page;
1255 (*nr)++;
1256 page++;
1257 refs++;
1258 } while (addr += PAGE_SIZE, addr != end);
1259
1260 if (!page_cache_add_speculative(head, refs)) {
1261 *nr -= refs;
1262 return 0;
1263 }
1264
1265 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1266 *nr -= refs;
1267 while (refs--)
1268 put_page(head);
1269 return 0;
1270 }
1271
Steve Capper2667f502014-10-09 15:29:14 -07001272 return 1;
1273}
1274
1275static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1276 unsigned long end, int write, struct page **pages, int *nr)
1277{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001278 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001279 int refs;
1280
1281 if (write && !pud_write(orig))
1282 return 0;
1283
1284 refs = 0;
1285 head = pud_page(orig);
1286 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001287 do {
1288 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1289 pages[*nr] = page;
1290 (*nr)++;
1291 page++;
1292 refs++;
1293 } while (addr += PAGE_SIZE, addr != end);
1294
1295 if (!page_cache_add_speculative(head, refs)) {
1296 *nr -= refs;
1297 return 0;
1298 }
1299
1300 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1301 *nr -= refs;
1302 while (refs--)
1303 put_page(head);
1304 return 0;
1305 }
1306
Steve Capper2667f502014-10-09 15:29:14 -07001307 return 1;
1308}
1309
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301310static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1311 unsigned long end, int write,
1312 struct page **pages, int *nr)
1313{
1314 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001315 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301316
1317 if (write && !pgd_write(orig))
1318 return 0;
1319
1320 refs = 0;
1321 head = pgd_page(orig);
1322 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301323 do {
1324 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1325 pages[*nr] = page;
1326 (*nr)++;
1327 page++;
1328 refs++;
1329 } while (addr += PAGE_SIZE, addr != end);
1330
1331 if (!page_cache_add_speculative(head, refs)) {
1332 *nr -= refs;
1333 return 0;
1334 }
1335
1336 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1337 *nr -= refs;
1338 while (refs--)
1339 put_page(head);
1340 return 0;
1341 }
1342
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301343 return 1;
1344}
1345
Steve Capper2667f502014-10-09 15:29:14 -07001346static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1347 int write, struct page **pages, int *nr)
1348{
1349 unsigned long next;
1350 pmd_t *pmdp;
1351
1352 pmdp = pmd_offset(&pud, addr);
1353 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001354 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001355
1356 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001357 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001358 return 0;
1359
1360 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1361 /*
1362 * NUMA hinting faults need to be handled in the GUP
1363 * slowpath for accounting purposes and so that they
1364 * can be serialised against THP migration.
1365 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001366 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001367 return 0;
1368
1369 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1370 pages, nr))
1371 return 0;
1372
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301373 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1374 /*
1375 * architecture have different format for hugetlbfs
1376 * pmd format and THP pmd format
1377 */
1378 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1379 PMD_SHIFT, next, write, pages, nr))
1380 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001381 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1382 return 0;
1383 } while (pmdp++, addr = next, addr != end);
1384
1385 return 1;
1386}
1387
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301388static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1389 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001390{
1391 unsigned long next;
1392 pud_t *pudp;
1393
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301394 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001395 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001396 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001397
1398 next = pud_addr_end(addr, end);
1399 if (pud_none(pud))
1400 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301401 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001402 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301403 pages, nr))
1404 return 0;
1405 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1406 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1407 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001408 return 0;
1409 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1410 return 0;
1411 } while (pudp++, addr = next, addr != end);
1412
1413 return 1;
1414}
1415
1416/*
1417 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1418 * the regular GUP. It will only return non-negative values.
1419 */
1420int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1421 struct page **pages)
1422{
1423 struct mm_struct *mm = current->mm;
1424 unsigned long addr, len, end;
1425 unsigned long next, flags;
1426 pgd_t *pgdp;
1427 int nr = 0;
1428
1429 start &= PAGE_MASK;
1430 addr = start;
1431 len = (unsigned long) nr_pages << PAGE_SHIFT;
1432 end = start + len;
1433
1434 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1435 start, len)))
1436 return 0;
1437
1438 /*
1439 * Disable interrupts. We use the nested form as we can already have
1440 * interrupts disabled by get_futex_key.
1441 *
1442 * With interrupts disabled, we block page table pages from being
1443 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1444 * for more details.
1445 *
1446 * We do not adopt an rcu_read_lock(.) here as we also want to
1447 * block IPIs that come from THPs splitting.
1448 */
1449
1450 local_irq_save(flags);
1451 pgdp = pgd_offset(mm, addr);
1452 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001453 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301454
Steve Capper2667f502014-10-09 15:29:14 -07001455 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301456 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001457 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301458 if (unlikely(pgd_huge(pgd))) {
1459 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1460 pages, &nr))
1461 break;
1462 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1463 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1464 PGDIR_SHIFT, next, write, pages, &nr))
1465 break;
1466 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001467 break;
1468 } while (pgdp++, addr = next, addr != end);
1469 local_irq_restore(flags);
1470
1471 return nr;
1472}
1473
1474/**
1475 * get_user_pages_fast() - pin user pages in memory
1476 * @start: starting user address
1477 * @nr_pages: number of pages from start to pin
1478 * @write: whether pages will be written to
1479 * @pages: array that receives pointers to the pages pinned.
1480 * Should be at least nr_pages long.
1481 *
1482 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1483 * If not successful, it will fall back to taking the lock and
1484 * calling get_user_pages().
1485 *
1486 * Returns number of pages pinned. This may be fewer than the number
1487 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1488 * were pinned, returns -errno.
1489 */
1490int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1491 struct page **pages)
1492{
Steve Capper2667f502014-10-09 15:29:14 -07001493 int nr, ret;
1494
1495 start &= PAGE_MASK;
1496 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1497 ret = nr;
1498
1499 if (nr < nr_pages) {
1500 /* Try to get the remaining pages with get_user_pages */
1501 start += nr << PAGE_SHIFT;
1502 pages += nr;
1503
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001504 ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001505
1506 /* Have to be a bit careful with return values */
1507 if (nr > 0) {
1508 if (ret < 0)
1509 ret = nr;
1510 else
1511 ret += nr;
1512 }
1513 }
1514
1515 return ret;
1516}
1517
1518#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */