blob: 9671e29f8ffd0e59ba5384a39e95b2834ffab133 [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);
Naoya Horiguchi337d9ab2016-07-26 15:24:03 -0700282 if (pmd_trans_unstable(pmd))
283 ret = -EBUSY;
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800284 } else {
285 get_page(page);
286 spin_unlock(ptl);
287 lock_page(page);
288 ret = split_huge_page(page);
289 unlock_page(page);
290 put_page(page);
291 }
292
293 return ret ? ERR_PTR(ret) :
294 follow_page_pte(vma, address, pmd, flags);
295 }
296
297 page = follow_trans_huge_pmd(vma, address, pmd, flags);
298 spin_unlock(ptl);
299 *page_mask = HPAGE_PMD_NR - 1;
300 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700301}
302
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700303static int get_gate_page(struct mm_struct *mm, unsigned long address,
304 unsigned int gup_flags, struct vm_area_struct **vma,
305 struct page **page)
306{
307 pgd_t *pgd;
308 pud_t *pud;
309 pmd_t *pmd;
310 pte_t *pte;
311 int ret = -EFAULT;
312
313 /* user gate pages are read-only */
314 if (gup_flags & FOLL_WRITE)
315 return -EFAULT;
316 if (address > TASK_SIZE)
317 pgd = pgd_offset_k(address);
318 else
319 pgd = pgd_offset_gate(mm, address);
320 BUG_ON(pgd_none(*pgd));
321 pud = pud_offset(pgd, address);
322 BUG_ON(pud_none(*pud));
323 pmd = pmd_offset(pud, address);
324 if (pmd_none(*pmd))
325 return -EFAULT;
326 VM_BUG_ON(pmd_trans_huge(*pmd));
327 pte = pte_offset_map(pmd, address);
328 if (pte_none(*pte))
329 goto unmap;
330 *vma = get_gate_vma(mm);
331 if (!page)
332 goto out;
333 *page = vm_normal_page(*vma, address, *pte);
334 if (!*page) {
335 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
336 goto unmap;
337 *page = pte_page(*pte);
338 }
339 get_page(*page);
340out:
341 ret = 0;
342unmap:
343 pte_unmap(pte);
344 return ret;
345}
346
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700347/*
348 * mmap_sem must be held on entry. If @nonblocking != NULL and
349 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
350 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
351 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700352static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
353 unsigned long address, unsigned int *flags, int *nonblocking)
354{
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700355 unsigned int fault_flags = 0;
356 int ret;
357
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800358 /* mlock all present pages, but do not fault in new pages */
359 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
360 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700361 /* For mm_populate(), just skip the stack guard page. */
362 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700363 (stack_guard_page_start(vma, address) ||
364 stack_guard_page_end(vma, address + PAGE_SIZE)))
365 return -ENOENT;
366 if (*flags & FOLL_WRITE)
367 fault_flags |= FAULT_FLAG_WRITE;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800368 if (*flags & FOLL_REMOTE)
369 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700370 if (nonblocking)
371 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
372 if (*flags & FOLL_NOWAIT)
373 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700374 if (*flags & FOLL_TRIED) {
375 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
376 fault_flags |= FAULT_FLAG_TRIED;
377 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700378
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700379 ret = handle_mm_fault(vma, address, fault_flags);
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700380 if (ret & VM_FAULT_ERROR) {
381 if (ret & VM_FAULT_OOM)
382 return -ENOMEM;
383 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
384 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800385 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700386 return -EFAULT;
387 BUG();
388 }
389
390 if (tsk) {
391 if (ret & VM_FAULT_MAJOR)
392 tsk->maj_flt++;
393 else
394 tsk->min_flt++;
395 }
396
397 if (ret & VM_FAULT_RETRY) {
398 if (nonblocking)
399 *nonblocking = 0;
400 return -EBUSY;
401 }
402
403 /*
404 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
405 * necessary, even if maybe_mkwrite decided not to set pte_write. We
406 * can thus safely do subsequent page lookups as if they were reads.
407 * But only do so when looping for pte_write is futile: in some cases
408 * userspace may also be wanting to write to the gotten user page,
409 * which a read fault here might prevent (a readonly page might get
410 * reCOWed by userspace write).
411 */
412 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
413 *flags &= ~FOLL_WRITE;
414 return 0;
415}
416
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700417static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
418{
419 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800420 int write = (gup_flags & FOLL_WRITE);
421 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700422
423 if (vm_flags & (VM_IO | VM_PFNMAP))
424 return -EFAULT;
425
Dave Hansen1b2ee122016-02-12 13:02:21 -0800426 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700427 if (!(vm_flags & VM_WRITE)) {
428 if (!(gup_flags & FOLL_FORCE))
429 return -EFAULT;
430 /*
431 * We used to let the write,force case do COW in a
432 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
433 * set a breakpoint in a read-only mapping of an
434 * executable, without corrupting the file (yet only
435 * when that file had been opened for writing!).
436 * Anon pages in shared mappings are surprising: now
437 * just reject it.
438 */
Hugh Dickins46435362016-01-30 18:03:16 -0800439 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700440 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700441 }
442 } else if (!(vm_flags & VM_READ)) {
443 if (!(gup_flags & FOLL_FORCE))
444 return -EFAULT;
445 /*
446 * Is there actually any vma we can reach here which does not
447 * have VM_MAYREAD set?
448 */
449 if (!(vm_flags & VM_MAYREAD))
450 return -EFAULT;
451 }
Dave Hansend61172b2016-02-12 13:02:24 -0800452 /*
453 * gups are always data accesses, not instruction
454 * fetches, so execute=false here
455 */
456 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800457 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700458 return 0;
459}
460
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700461/**
462 * __get_user_pages() - pin user pages in memory
463 * @tsk: task_struct of target task
464 * @mm: mm_struct of target mm
465 * @start: starting user address
466 * @nr_pages: number of pages from start to pin
467 * @gup_flags: flags modifying pin behaviour
468 * @pages: array that receives pointers to the pages pinned.
469 * Should be at least nr_pages long. Or NULL, if caller
470 * only intends to ensure the pages are faulted in.
471 * @vmas: array of pointers to vmas corresponding to each page.
472 * Or NULL if the caller does not require them.
473 * @nonblocking: whether waiting for disk IO or mmap_sem contention
474 *
475 * Returns number of pages pinned. This may be fewer than the number
476 * requested. If nr_pages is 0 or negative, returns 0. If no pages
477 * were pinned, returns -errno. Each page returned must be released
478 * with a put_page() call when it is finished with. vmas will only
479 * remain valid while mmap_sem is held.
480 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700481 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700482 *
483 * __get_user_pages walks a process's page tables and takes a reference to
484 * each struct page that each user address corresponds to at a given
485 * instant. That is, it takes the page that would be accessed if a user
486 * thread accesses the given user virtual address at that instant.
487 *
488 * This does not guarantee that the page exists in the user mappings when
489 * __get_user_pages returns, and there may even be a completely different
490 * page there in some cases (eg. if mmapped pagecache has been invalidated
491 * and subsequently re faulted). However it does guarantee that the page
492 * won't be freed completely. And mostly callers simply care that the page
493 * contains data that was valid *at some point in time*. Typically, an IO
494 * or similar operation cannot guarantee anything stronger anyway because
495 * locks can't be held over the syscall boundary.
496 *
497 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
498 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
499 * appropriate) must be called after the page is finished with, and
500 * before put_page is called.
501 *
502 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
503 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700504 * *@nonblocking will be set to 0. Further, if @gup_flags does not
505 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
506 * this case.
507 *
508 * A caller using such a combination of @nonblocking and @gup_flags
509 * must therefore hold the mmap_sem for reading only, and recognize
510 * when it's been released. Otherwise, it must be held for either
511 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700512 *
513 * In most cases, get_user_pages or get_user_pages_fast should be used
514 * instead of __get_user_pages. __get_user_pages should be used only if
515 * you need some special @gup_flags.
516 */
517long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
518 unsigned long start, unsigned long nr_pages,
519 unsigned int gup_flags, struct page **pages,
520 struct vm_area_struct **vmas, int *nonblocking)
521{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700522 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700523 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700524 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700525
526 if (!nr_pages)
527 return 0;
528
529 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
530
531 /*
532 * If FOLL_FORCE is set then do not force a full fault as the hinting
533 * fault information is unrelated to the reference behaviour of a task
534 * using the address space
535 */
536 if (!(gup_flags & FOLL_FORCE))
537 gup_flags |= FOLL_NUMA;
538
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700539 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700540 struct page *page;
541 unsigned int foll_flags = gup_flags;
542 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700543
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700544 /* first iteration or cross vma bound */
545 if (!vma || start >= vma->vm_end) {
546 vma = find_extend_vma(mm, start);
547 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700548 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700549 ret = get_gate_page(mm, start & PAGE_MASK,
550 gup_flags, &vma,
551 pages ? &pages[i] : NULL);
552 if (ret)
553 return i ? : ret;
554 page_mask = 0;
555 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700556 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700557
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700558 if (!vma || check_vma_flags(vma, gup_flags))
559 return i ? : -EFAULT;
560 if (is_vm_hugetlb_page(vma)) {
561 i = follow_hugetlb_page(mm, vma, pages, vmas,
562 &start, &nr_pages, i,
563 gup_flags);
564 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700565 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700566 }
567retry:
568 /*
569 * If we have a pending SIGKILL, don't keep faulting pages and
570 * potentially allocating memory.
571 */
572 if (unlikely(fatal_signal_pending(current)))
573 return i ? i : -ERESTARTSYS;
574 cond_resched();
575 page = follow_page_mask(vma, start, foll_flags, &page_mask);
576 if (!page) {
577 int ret;
578 ret = faultin_page(tsk, vma, start, &foll_flags,
579 nonblocking);
580 switch (ret) {
581 case 0:
582 goto retry;
583 case -EFAULT:
584 case -ENOMEM:
585 case -EHWPOISON:
586 return i ? i : ret;
587 case -EBUSY:
588 return i;
589 case -ENOENT:
590 goto next_page;
591 }
592 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700593 } else if (PTR_ERR(page) == -EEXIST) {
594 /*
595 * Proper page table entry exists, but no corresponding
596 * struct page.
597 */
598 goto next_page;
599 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700600 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700601 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700602 if (pages) {
603 pages[i] = page;
604 flush_anon_page(vma, page, start);
605 flush_dcache_page(page);
606 page_mask = 0;
607 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700608next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700609 if (vmas) {
610 vmas[i] = vma;
611 page_mask = 0;
612 }
613 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
614 if (page_increm > nr_pages)
615 page_increm = nr_pages;
616 i += page_increm;
617 start += page_increm * PAGE_SIZE;
618 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700619 } while (nr_pages);
620 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700621}
622EXPORT_SYMBOL(__get_user_pages);
623
Dave Hansend4925e02016-02-12 13:02:16 -0800624bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
625{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800626 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
627 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800628 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800629
630 if (!(vm_flags & vma->vm_flags))
631 return false;
632
Dave Hansen33a709b2016-02-12 13:02:19 -0800633 /*
634 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800635 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800636 *
637 * gup always represents data access, not instruction
638 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800639 */
Dave Hansend61172b2016-02-12 13:02:24 -0800640 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800641 return false;
642
Dave Hansend4925e02016-02-12 13:02:16 -0800643 return true;
644}
645
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700646/*
647 * fixup_user_fault() - manually resolve a user page fault
648 * @tsk: the task_struct to use for page fault accounting, or
649 * NULL if faults are not to be recorded.
650 * @mm: mm_struct of target mm
651 * @address: user address
652 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800653 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
654 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700655 *
656 * This is meant to be called in the specific scenario where for locking reasons
657 * we try to access user memory in atomic context (within a pagefault_disable()
658 * section), this returns -EFAULT, and we want to resolve the user fault before
659 * trying again.
660 *
661 * Typically this is meant to be used by the futex code.
662 *
663 * The main difference with get_user_pages() is that this function will
664 * unconditionally call handle_mm_fault() which will in turn perform all the
665 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800666 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700667 *
668 * This is important for some architectures where those bits also gate the
669 * access permission to the page because they are maintained in software. On
670 * such architectures, gup() will not be enough to make a subsequent access
671 * succeed.
672 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800673 * This function will not return with an unlocked mmap_sem. So it has not the
674 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700675 */
676int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800677 unsigned long address, unsigned int fault_flags,
678 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700679{
680 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800681 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700682
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800683 if (unlocked)
684 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
685
686retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700687 vma = find_extend_vma(mm, address);
688 if (!vma || address < vma->vm_start)
689 return -EFAULT;
690
Dave Hansend4925e02016-02-12 13:02:16 -0800691 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700692 return -EFAULT;
693
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700694 ret = handle_mm_fault(vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800695 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700696 if (ret & VM_FAULT_ERROR) {
697 if (ret & VM_FAULT_OOM)
698 return -ENOMEM;
699 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
700 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800701 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700702 return -EFAULT;
703 BUG();
704 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800705
706 if (ret & VM_FAULT_RETRY) {
707 down_read(&mm->mmap_sem);
708 if (!(fault_flags & FAULT_FLAG_TRIED)) {
709 *unlocked = true;
710 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
711 fault_flags |= FAULT_FLAG_TRIED;
712 goto retry;
713 }
714 }
715
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700716 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800717 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700718 tsk->maj_flt++;
719 else
720 tsk->min_flt++;
721 }
722 return 0;
723}
724
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800725static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
726 struct mm_struct *mm,
727 unsigned long start,
728 unsigned long nr_pages,
729 int write, int force,
730 struct page **pages,
731 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800732 int *locked, bool notify_drop,
733 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800734{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800735 long ret, pages_done;
736 bool lock_dropped;
737
738 if (locked) {
739 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
740 BUG_ON(vmas);
741 /* check caller initialized locked */
742 BUG_ON(*locked != 1);
743 }
744
745 if (pages)
746 flags |= FOLL_GET;
747 if (write)
748 flags |= FOLL_WRITE;
749 if (force)
750 flags |= FOLL_FORCE;
751
752 pages_done = 0;
753 lock_dropped = false;
754 for (;;) {
755 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
756 vmas, locked);
757 if (!locked)
758 /* VM_FAULT_RETRY couldn't trigger, bypass */
759 return ret;
760
761 /* VM_FAULT_RETRY cannot return errors */
762 if (!*locked) {
763 BUG_ON(ret < 0);
764 BUG_ON(ret >= nr_pages);
765 }
766
767 if (!pages)
768 /* If it's a prefault don't insist harder */
769 return ret;
770
771 if (ret > 0) {
772 nr_pages -= ret;
773 pages_done += ret;
774 if (!nr_pages)
775 break;
776 }
777 if (*locked) {
778 /* VM_FAULT_RETRY didn't trigger */
779 if (!pages_done)
780 pages_done = ret;
781 break;
782 }
783 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
784 pages += ret;
785 start += ret << PAGE_SHIFT;
786
787 /*
788 * Repeat on the address that fired VM_FAULT_RETRY
789 * without FAULT_FLAG_ALLOW_RETRY but with
790 * FAULT_FLAG_TRIED.
791 */
792 *locked = 1;
793 lock_dropped = true;
794 down_read(&mm->mmap_sem);
795 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
796 pages, NULL, NULL);
797 if (ret != 1) {
798 BUG_ON(ret > 1);
799 if (!pages_done)
800 pages_done = ret;
801 break;
802 }
803 nr_pages--;
804 pages_done++;
805 if (!nr_pages)
806 break;
807 pages++;
808 start += PAGE_SIZE;
809 }
810 if (notify_drop && lock_dropped && *locked) {
811 /*
812 * We must let the caller know we temporarily dropped the lock
813 * and so the critical section protected by it was lost.
814 */
815 up_read(&mm->mmap_sem);
816 *locked = 0;
817 }
818 return pages_done;
819}
820
821/*
822 * We can leverage the VM_FAULT_RETRY functionality in the page fault
823 * paths better by using either get_user_pages_locked() or
824 * get_user_pages_unlocked().
825 *
826 * get_user_pages_locked() is suitable to replace the form:
827 *
828 * down_read(&mm->mmap_sem);
829 * do_something()
830 * get_user_pages(tsk, mm, ..., pages, NULL);
831 * up_read(&mm->mmap_sem);
832 *
833 * to:
834 *
835 * int locked = 1;
836 * down_read(&mm->mmap_sem);
837 * do_something()
838 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
839 * if (locked)
840 * up_read(&mm->mmap_sem);
841 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200842long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800843 int write, int force, struct page **pages,
844 int *locked)
845{
Dave Hansencde70142016-02-12 13:01:55 -0800846 return __get_user_pages_locked(current, current->mm, start, nr_pages,
847 write, force, pages, NULL, locked, true,
848 FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800849}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200850EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800851
852/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800853 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
854 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
855 *
856 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
857 * caller if required (just like with __get_user_pages). "FOLL_GET",
858 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
859 * according to the parameters "pages", "write", "force"
860 * respectively.
861 */
862__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
863 unsigned long start, unsigned long nr_pages,
864 int write, int force, struct page **pages,
865 unsigned int gup_flags)
866{
867 long ret;
868 int locked = 1;
869 down_read(&mm->mmap_sem);
870 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
871 pages, NULL, &locked, false, gup_flags);
872 if (locked)
873 up_read(&mm->mmap_sem);
874 return ret;
875}
876EXPORT_SYMBOL(__get_user_pages_unlocked);
877
878/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800879 * get_user_pages_unlocked() is suitable to replace the form:
880 *
881 * down_read(&mm->mmap_sem);
882 * get_user_pages(tsk, mm, ..., pages, NULL);
883 * up_read(&mm->mmap_sem);
884 *
885 * with:
886 *
887 * get_user_pages_unlocked(tsk, mm, ..., pages);
888 *
889 * It is functionally equivalent to get_user_pages_fast so
890 * get_user_pages_fast should be used instead, if the two parameters
891 * "tsk" and "mm" are respectively equal to current and current->mm,
892 * or if "force" shall be set to 1 (get_user_pages_fast misses the
893 * "force" parameter).
894 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200895long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800896 int write, int force, struct page **pages)
897{
Dave Hansencde70142016-02-12 13:01:55 -0800898 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
899 write, force, pages, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800900}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200901EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800902
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700903/*
Dave Hansen1e987792016-02-12 13:01:54 -0800904 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700905 * @tsk: the task_struct to use for page fault accounting, or
906 * NULL if faults are not to be recorded.
907 * @mm: mm_struct of target mm
908 * @start: starting user address
909 * @nr_pages: number of pages from start to pin
910 * @write: whether pages will be written to by the caller
911 * @force: whether to force access even when user mapping is currently
912 * protected (but never forces write access to shared mapping).
913 * @pages: array that receives pointers to the pages pinned.
914 * Should be at least nr_pages long. Or NULL, if caller
915 * only intends to ensure the pages are faulted in.
916 * @vmas: array of pointers to vmas corresponding to each page.
917 * Or NULL if the caller does not require them.
918 *
919 * Returns number of pages pinned. This may be fewer than the number
920 * requested. If nr_pages is 0 or negative, returns 0. If no pages
921 * were pinned, returns -errno. Each page returned must be released
922 * with a put_page() call when it is finished with. vmas will only
923 * remain valid while mmap_sem is held.
924 *
925 * Must be called with mmap_sem held for read or write.
926 *
927 * get_user_pages walks a process's page tables and takes a reference to
928 * each struct page that each user address corresponds to at a given
929 * instant. That is, it takes the page that would be accessed if a user
930 * thread accesses the given user virtual address at that instant.
931 *
932 * This does not guarantee that the page exists in the user mappings when
933 * get_user_pages returns, and there may even be a completely different
934 * page there in some cases (eg. if mmapped pagecache has been invalidated
935 * and subsequently re faulted). However it does guarantee that the page
936 * won't be freed completely. And mostly callers simply care that the page
937 * contains data that was valid *at some point in time*. Typically, an IO
938 * or similar operation cannot guarantee anything stronger anyway because
939 * locks can't be held over the syscall boundary.
940 *
941 * If write=0, the page must not be written to. If the page is written to,
942 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
943 * after the page is finished with, and before put_page is called.
944 *
945 * get_user_pages is typically used for fewer-copy IO operations, to get a
946 * handle on the memory by some means other than accesses via the user virtual
947 * addresses. The pages may be submitted for DMA to devices or accessed via
948 * their kernel linear mapping (via the kmap APIs). Care should be taken to
949 * use the correct cache flushing APIs.
950 *
951 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800952 *
953 * get_user_pages should be phased out in favor of
954 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
955 * should use get_user_pages because it cannot pass
956 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700957 */
Dave Hansen1e987792016-02-12 13:01:54 -0800958long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
959 unsigned long start, unsigned long nr_pages,
960 int write, int force, struct page **pages,
961 struct vm_area_struct **vmas)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700962{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800963 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Dave Hansen1e987792016-02-12 13:01:54 -0800964 pages, vmas, NULL, false,
965 FOLL_TOUCH | FOLL_REMOTE);
966}
967EXPORT_SYMBOL(get_user_pages_remote);
968
969/*
Dave Hansend4edcf02016-02-12 13:01:56 -0800970 * This is the same as get_user_pages_remote(), just with a
971 * less-flexible calling convention where we assume that the task
972 * and mm being operated on are the current task's. We also
973 * obviously don't pass FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -0800974 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200975long get_user_pages(unsigned long start, unsigned long nr_pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800976 int write, int force, struct page **pages,
977 struct vm_area_struct **vmas)
978{
Dave Hansencde70142016-02-12 13:01:55 -0800979 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800980 write, force, pages, vmas, NULL, false,
981 FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700982}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200983EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700984
985/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700986 * populate_vma_page_range() - populate a range of pages in the vma.
987 * @vma: target vma
988 * @start: start address
989 * @end: end address
990 * @nonblocking:
991 *
992 * This takes care of mlocking the pages too if VM_LOCKED is set.
993 *
994 * return 0 on success, negative error code on error.
995 *
996 * vma->vm_mm->mmap_sem must be held.
997 *
998 * If @nonblocking is NULL, it may be held for read or write and will
999 * be unperturbed.
1000 *
1001 * If @nonblocking is non-NULL, it must held for read only and may be
1002 * released. If it's released, *@nonblocking will be set to 0.
1003 */
1004long populate_vma_page_range(struct vm_area_struct *vma,
1005 unsigned long start, unsigned long end, int *nonblocking)
1006{
1007 struct mm_struct *mm = vma->vm_mm;
1008 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1009 int gup_flags;
1010
1011 VM_BUG_ON(start & ~PAGE_MASK);
1012 VM_BUG_ON(end & ~PAGE_MASK);
1013 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1014 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1015 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1016
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001017 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1018 if (vma->vm_flags & VM_LOCKONFAULT)
1019 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001020 /*
1021 * We want to touch writable mappings with a write fault in order
1022 * to break COW, except for shared mappings because these don't COW
1023 * and we would not want to dirty them for nothing.
1024 */
1025 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1026 gup_flags |= FOLL_WRITE;
1027
1028 /*
1029 * We want mlock to succeed for regions that have any permissions
1030 * other than PROT_NONE.
1031 */
1032 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1033 gup_flags |= FOLL_FORCE;
1034
1035 /*
1036 * We made sure addr is within a VMA, so the following will
1037 * not result in a stack expansion that recurses back here.
1038 */
1039 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1040 NULL, NULL, nonblocking);
1041}
1042
1043/*
1044 * __mm_populate - populate and/or mlock pages within a range of address space.
1045 *
1046 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1047 * flags. VMAs must be already marked with the desired vm_flags, and
1048 * mmap_sem must not be held.
1049 */
1050int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1051{
1052 struct mm_struct *mm = current->mm;
1053 unsigned long end, nstart, nend;
1054 struct vm_area_struct *vma = NULL;
1055 int locked = 0;
1056 long ret = 0;
1057
1058 VM_BUG_ON(start & ~PAGE_MASK);
1059 VM_BUG_ON(len != PAGE_ALIGN(len));
1060 end = start + len;
1061
1062 for (nstart = start; nstart < end; nstart = nend) {
1063 /*
1064 * We want to fault in pages for [nstart; end) address range.
1065 * Find first corresponding VMA.
1066 */
1067 if (!locked) {
1068 locked = 1;
1069 down_read(&mm->mmap_sem);
1070 vma = find_vma(mm, nstart);
1071 } else if (nstart >= vma->vm_end)
1072 vma = vma->vm_next;
1073 if (!vma || vma->vm_start >= end)
1074 break;
1075 /*
1076 * Set [nstart; nend) to intersection of desired address
1077 * range with the first VMA. Also, skip undesirable VMA types.
1078 */
1079 nend = min(end, vma->vm_end);
1080 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1081 continue;
1082 if (nstart < vma->vm_start)
1083 nstart = vma->vm_start;
1084 /*
1085 * Now fault in a range of pages. populate_vma_page_range()
1086 * double checks the vma flags, so that it won't mlock pages
1087 * if the vma was already munlocked.
1088 */
1089 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1090 if (ret < 0) {
1091 if (ignore_errors) {
1092 ret = 0;
1093 continue; /* continue at next VMA */
1094 }
1095 break;
1096 }
1097 nend = nstart + ret * PAGE_SIZE;
1098 ret = 0;
1099 }
1100 if (locked)
1101 up_read(&mm->mmap_sem);
1102 return ret; /* 0 or negative error code */
1103}
1104
1105/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001106 * get_dump_page() - pin user page in memory while writing it to core dump
1107 * @addr: user address
1108 *
1109 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001110 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001111 *
1112 * Returns NULL on any kind of failure - a hole must then be inserted into
1113 * the corefile, to preserve alignment with its headers; and also returns
1114 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1115 * allowing a hole to be left in the corefile to save diskspace.
1116 *
1117 * Called without mmap_sem, but after all other threads have been killed.
1118 */
1119#ifdef CONFIG_ELF_CORE
1120struct page *get_dump_page(unsigned long addr)
1121{
1122 struct vm_area_struct *vma;
1123 struct page *page;
1124
1125 if (__get_user_pages(current, current->mm, addr, 1,
1126 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1127 NULL) < 1)
1128 return NULL;
1129 flush_cache_page(vma, addr, page_to_pfn(page));
1130 return page;
1131}
1132#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001133
1134/*
1135 * Generic RCU Fast GUP
1136 *
1137 * get_user_pages_fast attempts to pin user pages by walking the page
1138 * tables directly and avoids taking locks. Thus the walker needs to be
1139 * protected from page table pages being freed from under it, and should
1140 * block any THP splits.
1141 *
1142 * One way to achieve this is to have the walker disable interrupts, and
1143 * rely on IPIs from the TLB flushing code blocking before the page table
1144 * pages are freed. This is unsuitable for architectures that do not need
1145 * to broadcast an IPI when invalidating TLBs.
1146 *
1147 * Another way to achieve this is to batch up page table containing pages
1148 * belonging to more than one mm_user, then rcu_sched a callback to free those
1149 * pages. Disabling interrupts will allow the fast_gup walker to both block
1150 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1151 * (which is a relatively rare event). The code below adopts this strategy.
1152 *
1153 * Before activating this code, please be aware that the following assumptions
1154 * are currently made:
1155 *
1156 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1157 * pages containing page tables.
1158 *
Steve Capper2667f502014-10-09 15:29:14 -07001159 * *) ptes can be read atomically by the architecture.
1160 *
1161 * *) access_ok is sufficient to validate userspace address ranges.
1162 *
1163 * The last two assumptions can be relaxed by the addition of helper functions.
1164 *
1165 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1166 */
1167#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1168
1169#ifdef __HAVE_ARCH_PTE_SPECIAL
1170static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1171 int write, struct page **pages, int *nr)
1172{
1173 pte_t *ptep, *ptem;
1174 int ret = 0;
1175
1176 ptem = ptep = pte_offset_map(&pmd, addr);
1177 do {
1178 /*
1179 * In the line below we are assuming that the pte can be read
1180 * atomically. If this is not the case for your architecture,
1181 * please wrap this in a helper function!
1182 *
1183 * for an example see gup_get_pte in arch/x86/mm/gup.c
1184 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001185 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001186 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001187
1188 /*
1189 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001190 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001191 */
1192 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001193 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001194 goto pte_unmap;
1195
Dave Hansen33a709b2016-02-12 13:02:19 -08001196 if (!arch_pte_access_permitted(pte, write))
1197 goto pte_unmap;
1198
Steve Capper2667f502014-10-09 15:29:14 -07001199 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1200 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001201 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001202
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001203 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001204 goto pte_unmap;
1205
1206 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001207 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001208 goto pte_unmap;
1209 }
1210
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001211 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001212 pages[*nr] = page;
1213 (*nr)++;
1214
1215 } while (ptep++, addr += PAGE_SIZE, addr != end);
1216
1217 ret = 1;
1218
1219pte_unmap:
1220 pte_unmap(ptem);
1221 return ret;
1222}
1223#else
1224
1225/*
1226 * If we can't determine whether or not a pte is special, then fail immediately
1227 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1228 * to be special.
1229 *
1230 * For a futex to be placed on a THP tail page, get_futex_key requires a
1231 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1232 * useful to have gup_huge_pmd even if we can't operate on ptes.
1233 */
1234static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1235 int write, struct page **pages, int *nr)
1236{
1237 return 0;
1238}
1239#endif /* __HAVE_ARCH_PTE_SPECIAL */
1240
1241static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1242 unsigned long end, int write, struct page **pages, int *nr)
1243{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001244 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001245 int refs;
1246
1247 if (write && !pmd_write(orig))
1248 return 0;
1249
1250 refs = 0;
1251 head = pmd_page(orig);
1252 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001253 do {
1254 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1255 pages[*nr] = page;
1256 (*nr)++;
1257 page++;
1258 refs++;
1259 } while (addr += PAGE_SIZE, addr != end);
1260
1261 if (!page_cache_add_speculative(head, refs)) {
1262 *nr -= refs;
1263 return 0;
1264 }
1265
1266 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1267 *nr -= refs;
1268 while (refs--)
1269 put_page(head);
1270 return 0;
1271 }
1272
Steve Capper2667f502014-10-09 15:29:14 -07001273 return 1;
1274}
1275
1276static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1277 unsigned long end, int write, struct page **pages, int *nr)
1278{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001279 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001280 int refs;
1281
1282 if (write && !pud_write(orig))
1283 return 0;
1284
1285 refs = 0;
1286 head = pud_page(orig);
1287 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001288 do {
1289 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1290 pages[*nr] = page;
1291 (*nr)++;
1292 page++;
1293 refs++;
1294 } while (addr += PAGE_SIZE, addr != end);
1295
1296 if (!page_cache_add_speculative(head, refs)) {
1297 *nr -= refs;
1298 return 0;
1299 }
1300
1301 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1302 *nr -= refs;
1303 while (refs--)
1304 put_page(head);
1305 return 0;
1306 }
1307
Steve Capper2667f502014-10-09 15:29:14 -07001308 return 1;
1309}
1310
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301311static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1312 unsigned long end, int write,
1313 struct page **pages, int *nr)
1314{
1315 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001316 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301317
1318 if (write && !pgd_write(orig))
1319 return 0;
1320
1321 refs = 0;
1322 head = pgd_page(orig);
1323 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301324 do {
1325 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1326 pages[*nr] = page;
1327 (*nr)++;
1328 page++;
1329 refs++;
1330 } while (addr += PAGE_SIZE, addr != end);
1331
1332 if (!page_cache_add_speculative(head, refs)) {
1333 *nr -= refs;
1334 return 0;
1335 }
1336
1337 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1338 *nr -= refs;
1339 while (refs--)
1340 put_page(head);
1341 return 0;
1342 }
1343
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301344 return 1;
1345}
1346
Steve Capper2667f502014-10-09 15:29:14 -07001347static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1348 int write, struct page **pages, int *nr)
1349{
1350 unsigned long next;
1351 pmd_t *pmdp;
1352
1353 pmdp = pmd_offset(&pud, addr);
1354 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001355 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001356
1357 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001358 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001359 return 0;
1360
1361 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1362 /*
1363 * NUMA hinting faults need to be handled in the GUP
1364 * slowpath for accounting purposes and so that they
1365 * can be serialised against THP migration.
1366 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001367 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001368 return 0;
1369
1370 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1371 pages, nr))
1372 return 0;
1373
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301374 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1375 /*
1376 * architecture have different format for hugetlbfs
1377 * pmd format and THP pmd format
1378 */
1379 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1380 PMD_SHIFT, next, write, pages, nr))
1381 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001382 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1383 return 0;
1384 } while (pmdp++, addr = next, addr != end);
1385
1386 return 1;
1387}
1388
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301389static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1390 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001391{
1392 unsigned long next;
1393 pud_t *pudp;
1394
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301395 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001396 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001397 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001398
1399 next = pud_addr_end(addr, end);
1400 if (pud_none(pud))
1401 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301402 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001403 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301404 pages, nr))
1405 return 0;
1406 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1407 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1408 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001409 return 0;
1410 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1411 return 0;
1412 } while (pudp++, addr = next, addr != end);
1413
1414 return 1;
1415}
1416
1417/*
1418 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1419 * the regular GUP. It will only return non-negative values.
1420 */
1421int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1422 struct page **pages)
1423{
1424 struct mm_struct *mm = current->mm;
1425 unsigned long addr, len, end;
1426 unsigned long next, flags;
1427 pgd_t *pgdp;
1428 int nr = 0;
1429
1430 start &= PAGE_MASK;
1431 addr = start;
1432 len = (unsigned long) nr_pages << PAGE_SHIFT;
1433 end = start + len;
1434
1435 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1436 start, len)))
1437 return 0;
1438
1439 /*
1440 * Disable interrupts. We use the nested form as we can already have
1441 * interrupts disabled by get_futex_key.
1442 *
1443 * With interrupts disabled, we block page table pages from being
1444 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1445 * for more details.
1446 *
1447 * We do not adopt an rcu_read_lock(.) here as we also want to
1448 * block IPIs that come from THPs splitting.
1449 */
1450
1451 local_irq_save(flags);
1452 pgdp = pgd_offset(mm, addr);
1453 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001454 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301455
Steve Capper2667f502014-10-09 15:29:14 -07001456 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301457 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001458 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301459 if (unlikely(pgd_huge(pgd))) {
1460 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1461 pages, &nr))
1462 break;
1463 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1464 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1465 PGDIR_SHIFT, next, write, pages, &nr))
1466 break;
1467 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001468 break;
1469 } while (pgdp++, addr = next, addr != end);
1470 local_irq_restore(flags);
1471
1472 return nr;
1473}
1474
1475/**
1476 * get_user_pages_fast() - pin user pages in memory
1477 * @start: starting user address
1478 * @nr_pages: number of pages from start to pin
1479 * @write: whether pages will be written to
1480 * @pages: array that receives pointers to the pages pinned.
1481 * Should be at least nr_pages long.
1482 *
1483 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1484 * If not successful, it will fall back to taking the lock and
1485 * calling get_user_pages().
1486 *
1487 * Returns number of pages pinned. This may be fewer than the number
1488 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1489 * were pinned, returns -errno.
1490 */
1491int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1492 struct page **pages)
1493{
Steve Capper2667f502014-10-09 15:29:14 -07001494 int nr, ret;
1495
1496 start &= PAGE_MASK;
1497 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1498 ret = nr;
1499
1500 if (nr < nr_pages) {
1501 /* Try to get the remaining pages with get_user_pages */
1502 start += nr << PAGE_SHIFT;
1503 pages += nr;
1504
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001505 ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001506
1507 /* Have to be a bit careful with return values */
1508 if (nr > 0) {
1509 if (ret < 0)
1510 ret = nr;
1511 else
1512 ret += nr;
1513 }
1514 }
1515
1516 return ret;
1517}
1518
1519#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */