blob: dee142e100f4cbf988a1bf7a8fe8d02501d724d6 [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{
355 struct mm_struct *mm = vma->vm_mm;
356 unsigned int fault_flags = 0;
357 int ret;
358
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800359 /* mlock all present pages, but do not fault in new pages */
360 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
361 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700362 /* For mm_populate(), just skip the stack guard page. */
363 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700364 (stack_guard_page_start(vma, address) ||
365 stack_guard_page_end(vma, address + PAGE_SIZE)))
366 return -ENOENT;
367 if (*flags & FOLL_WRITE)
368 fault_flags |= FAULT_FLAG_WRITE;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800369 if (*flags & FOLL_REMOTE)
370 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700371 if (nonblocking)
372 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
373 if (*flags & FOLL_NOWAIT)
374 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700375 if (*flags & FOLL_TRIED) {
376 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
377 fault_flags |= FAULT_FLAG_TRIED;
378 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700379
380 ret = handle_mm_fault(mm, vma, address, fault_flags);
381 if (ret & VM_FAULT_ERROR) {
382 if (ret & VM_FAULT_OOM)
383 return -ENOMEM;
384 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
385 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800386 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700387 return -EFAULT;
388 BUG();
389 }
390
391 if (tsk) {
392 if (ret & VM_FAULT_MAJOR)
393 tsk->maj_flt++;
394 else
395 tsk->min_flt++;
396 }
397
398 if (ret & VM_FAULT_RETRY) {
399 if (nonblocking)
400 *nonblocking = 0;
401 return -EBUSY;
402 }
403
404 /*
405 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
406 * necessary, even if maybe_mkwrite decided not to set pte_write. We
407 * can thus safely do subsequent page lookups as if they were reads.
408 * But only do so when looping for pte_write is futile: in some cases
409 * userspace may also be wanting to write to the gotten user page,
410 * which a read fault here might prevent (a readonly page might get
411 * reCOWed by userspace write).
412 */
413 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
414 *flags &= ~FOLL_WRITE;
415 return 0;
416}
417
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700418static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
419{
420 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800421 int write = (gup_flags & FOLL_WRITE);
422 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700423
424 if (vm_flags & (VM_IO | VM_PFNMAP))
425 return -EFAULT;
426
Dave Hansen1b2ee122016-02-12 13:02:21 -0800427 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700428 if (!(vm_flags & VM_WRITE)) {
429 if (!(gup_flags & FOLL_FORCE))
430 return -EFAULT;
431 /*
432 * We used to let the write,force case do COW in a
433 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
434 * set a breakpoint in a read-only mapping of an
435 * executable, without corrupting the file (yet only
436 * when that file had been opened for writing!).
437 * Anon pages in shared mappings are surprising: now
438 * just reject it.
439 */
Hugh Dickins46435362016-01-30 18:03:16 -0800440 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700441 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700442 }
443 } else if (!(vm_flags & VM_READ)) {
444 if (!(gup_flags & FOLL_FORCE))
445 return -EFAULT;
446 /*
447 * Is there actually any vma we can reach here which does not
448 * have VM_MAYREAD set?
449 */
450 if (!(vm_flags & VM_MAYREAD))
451 return -EFAULT;
452 }
Dave Hansend61172b2016-02-12 13:02:24 -0800453 /*
454 * gups are always data accesses, not instruction
455 * fetches, so execute=false here
456 */
457 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800458 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700459 return 0;
460}
461
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700462/**
463 * __get_user_pages() - pin user pages in memory
464 * @tsk: task_struct of target task
465 * @mm: mm_struct of target mm
466 * @start: starting user address
467 * @nr_pages: number of pages from start to pin
468 * @gup_flags: flags modifying pin behaviour
469 * @pages: array that receives pointers to the pages pinned.
470 * Should be at least nr_pages long. Or NULL, if caller
471 * only intends to ensure the pages are faulted in.
472 * @vmas: array of pointers to vmas corresponding to each page.
473 * Or NULL if the caller does not require them.
474 * @nonblocking: whether waiting for disk IO or mmap_sem contention
475 *
476 * Returns number of pages pinned. This may be fewer than the number
477 * requested. If nr_pages is 0 or negative, returns 0. If no pages
478 * were pinned, returns -errno. Each page returned must be released
479 * with a put_page() call when it is finished with. vmas will only
480 * remain valid while mmap_sem is held.
481 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700482 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700483 *
484 * __get_user_pages walks a process's page tables and takes a reference to
485 * each struct page that each user address corresponds to at a given
486 * instant. That is, it takes the page that would be accessed if a user
487 * thread accesses the given user virtual address at that instant.
488 *
489 * This does not guarantee that the page exists in the user mappings when
490 * __get_user_pages returns, and there may even be a completely different
491 * page there in some cases (eg. if mmapped pagecache has been invalidated
492 * and subsequently re faulted). However it does guarantee that the page
493 * won't be freed completely. And mostly callers simply care that the page
494 * contains data that was valid *at some point in time*. Typically, an IO
495 * or similar operation cannot guarantee anything stronger anyway because
496 * locks can't be held over the syscall boundary.
497 *
498 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
499 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
500 * appropriate) must be called after the page is finished with, and
501 * before put_page is called.
502 *
503 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
504 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700505 * *@nonblocking will be set to 0. Further, if @gup_flags does not
506 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
507 * this case.
508 *
509 * A caller using such a combination of @nonblocking and @gup_flags
510 * must therefore hold the mmap_sem for reading only, and recognize
511 * when it's been released. Otherwise, it must be held for either
512 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700513 *
514 * In most cases, get_user_pages or get_user_pages_fast should be used
515 * instead of __get_user_pages. __get_user_pages should be used only if
516 * you need some special @gup_flags.
517 */
518long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
519 unsigned long start, unsigned long nr_pages,
520 unsigned int gup_flags, struct page **pages,
521 struct vm_area_struct **vmas, int *nonblocking)
522{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700523 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700524 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700525 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700526
527 if (!nr_pages)
528 return 0;
529
530 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
531
532 /*
533 * If FOLL_FORCE is set then do not force a full fault as the hinting
534 * fault information is unrelated to the reference behaviour of a task
535 * using the address space
536 */
537 if (!(gup_flags & FOLL_FORCE))
538 gup_flags |= FOLL_NUMA;
539
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700540 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700541 struct page *page;
542 unsigned int foll_flags = gup_flags;
543 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700544
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700545 /* first iteration or cross vma bound */
546 if (!vma || start >= vma->vm_end) {
547 vma = find_extend_vma(mm, start);
548 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700549 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700550 ret = get_gate_page(mm, start & PAGE_MASK,
551 gup_flags, &vma,
552 pages ? &pages[i] : NULL);
553 if (ret)
554 return i ? : ret;
555 page_mask = 0;
556 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700557 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700558
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700559 if (!vma || check_vma_flags(vma, gup_flags))
560 return i ? : -EFAULT;
561 if (is_vm_hugetlb_page(vma)) {
562 i = follow_hugetlb_page(mm, vma, pages, vmas,
563 &start, &nr_pages, i,
564 gup_flags);
565 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700566 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700567 }
568retry:
569 /*
570 * If we have a pending SIGKILL, don't keep faulting pages and
571 * potentially allocating memory.
572 */
573 if (unlikely(fatal_signal_pending(current)))
574 return i ? i : -ERESTARTSYS;
575 cond_resched();
576 page = follow_page_mask(vma, start, foll_flags, &page_mask);
577 if (!page) {
578 int ret;
579 ret = faultin_page(tsk, vma, start, &foll_flags,
580 nonblocking);
581 switch (ret) {
582 case 0:
583 goto retry;
584 case -EFAULT:
585 case -ENOMEM:
586 case -EHWPOISON:
587 return i ? i : ret;
588 case -EBUSY:
589 return i;
590 case -ENOENT:
591 goto next_page;
592 }
593 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700594 } else if (PTR_ERR(page) == -EEXIST) {
595 /*
596 * Proper page table entry exists, but no corresponding
597 * struct page.
598 */
599 goto next_page;
600 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700601 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700602 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700603 if (pages) {
604 pages[i] = page;
605 flush_anon_page(vma, page, start);
606 flush_dcache_page(page);
607 page_mask = 0;
608 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700609next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700610 if (vmas) {
611 vmas[i] = vma;
612 page_mask = 0;
613 }
614 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
615 if (page_increm > nr_pages)
616 page_increm = nr_pages;
617 i += page_increm;
618 start += page_increm * PAGE_SIZE;
619 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700620 } while (nr_pages);
621 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700622}
623EXPORT_SYMBOL(__get_user_pages);
624
Dave Hansend4925e02016-02-12 13:02:16 -0800625bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
626{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800627 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
628 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800629 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800630
631 if (!(vm_flags & vma->vm_flags))
632 return false;
633
Dave Hansen33a709b2016-02-12 13:02:19 -0800634 /*
635 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800636 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800637 *
638 * gup always represents data access, not instruction
639 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800640 */
Dave Hansend61172b2016-02-12 13:02:24 -0800641 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800642 return false;
643
Dave Hansend4925e02016-02-12 13:02:16 -0800644 return true;
645}
646
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700647/*
648 * fixup_user_fault() - manually resolve a user page fault
649 * @tsk: the task_struct to use for page fault accounting, or
650 * NULL if faults are not to be recorded.
651 * @mm: mm_struct of target mm
652 * @address: user address
653 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800654 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
655 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700656 *
657 * This is meant to be called in the specific scenario where for locking reasons
658 * we try to access user memory in atomic context (within a pagefault_disable()
659 * section), this returns -EFAULT, and we want to resolve the user fault before
660 * trying again.
661 *
662 * Typically this is meant to be used by the futex code.
663 *
664 * The main difference with get_user_pages() is that this function will
665 * unconditionally call handle_mm_fault() which will in turn perform all the
666 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800667 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700668 *
669 * This is important for some architectures where those bits also gate the
670 * access permission to the page because they are maintained in software. On
671 * such architectures, gup() will not be enough to make a subsequent access
672 * succeed.
673 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800674 * This function will not return with an unlocked mmap_sem. So it has not the
675 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700676 */
677int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800678 unsigned long address, unsigned int fault_flags,
679 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700680{
681 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800682 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700683
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800684 if (unlocked)
685 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
686
687retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700688 vma = find_extend_vma(mm, address);
689 if (!vma || address < vma->vm_start)
690 return -EFAULT;
691
Dave Hansend4925e02016-02-12 13:02:16 -0800692 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700693 return -EFAULT;
694
695 ret = handle_mm_fault(mm, vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800696 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700697 if (ret & VM_FAULT_ERROR) {
698 if (ret & VM_FAULT_OOM)
699 return -ENOMEM;
700 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
701 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800702 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700703 return -EFAULT;
704 BUG();
705 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800706
707 if (ret & VM_FAULT_RETRY) {
708 down_read(&mm->mmap_sem);
709 if (!(fault_flags & FAULT_FLAG_TRIED)) {
710 *unlocked = true;
711 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
712 fault_flags |= FAULT_FLAG_TRIED;
713 goto retry;
714 }
715 }
716
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700717 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800718 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700719 tsk->maj_flt++;
720 else
721 tsk->min_flt++;
722 }
723 return 0;
724}
725
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800726static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
727 struct mm_struct *mm,
728 unsigned long start,
729 unsigned long nr_pages,
730 int write, int force,
731 struct page **pages,
732 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800733 int *locked, bool notify_drop,
734 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800735{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800736 long ret, pages_done;
737 bool lock_dropped;
738
739 if (locked) {
740 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
741 BUG_ON(vmas);
742 /* check caller initialized locked */
743 BUG_ON(*locked != 1);
744 }
745
746 if (pages)
747 flags |= FOLL_GET;
748 if (write)
749 flags |= FOLL_WRITE;
750 if (force)
751 flags |= FOLL_FORCE;
752
753 pages_done = 0;
754 lock_dropped = false;
755 for (;;) {
756 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
757 vmas, locked);
758 if (!locked)
759 /* VM_FAULT_RETRY couldn't trigger, bypass */
760 return ret;
761
762 /* VM_FAULT_RETRY cannot return errors */
763 if (!*locked) {
764 BUG_ON(ret < 0);
765 BUG_ON(ret >= nr_pages);
766 }
767
768 if (!pages)
769 /* If it's a prefault don't insist harder */
770 return ret;
771
772 if (ret > 0) {
773 nr_pages -= ret;
774 pages_done += ret;
775 if (!nr_pages)
776 break;
777 }
778 if (*locked) {
779 /* VM_FAULT_RETRY didn't trigger */
780 if (!pages_done)
781 pages_done = ret;
782 break;
783 }
784 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
785 pages += ret;
786 start += ret << PAGE_SHIFT;
787
788 /*
789 * Repeat on the address that fired VM_FAULT_RETRY
790 * without FAULT_FLAG_ALLOW_RETRY but with
791 * FAULT_FLAG_TRIED.
792 */
793 *locked = 1;
794 lock_dropped = true;
795 down_read(&mm->mmap_sem);
796 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
797 pages, NULL, NULL);
798 if (ret != 1) {
799 BUG_ON(ret > 1);
800 if (!pages_done)
801 pages_done = ret;
802 break;
803 }
804 nr_pages--;
805 pages_done++;
806 if (!nr_pages)
807 break;
808 pages++;
809 start += PAGE_SIZE;
810 }
811 if (notify_drop && lock_dropped && *locked) {
812 /*
813 * We must let the caller know we temporarily dropped the lock
814 * and so the critical section protected by it was lost.
815 */
816 up_read(&mm->mmap_sem);
817 *locked = 0;
818 }
819 return pages_done;
820}
821
822/*
823 * We can leverage the VM_FAULT_RETRY functionality in the page fault
824 * paths better by using either get_user_pages_locked() or
825 * get_user_pages_unlocked().
826 *
827 * get_user_pages_locked() is suitable to replace the form:
828 *
829 * down_read(&mm->mmap_sem);
830 * do_something()
831 * get_user_pages(tsk, mm, ..., pages, NULL);
832 * up_read(&mm->mmap_sem);
833 *
834 * to:
835 *
836 * int locked = 1;
837 * down_read(&mm->mmap_sem);
838 * do_something()
839 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
840 * if (locked)
841 * up_read(&mm->mmap_sem);
842 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200843long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800844 int write, int force, struct page **pages,
845 int *locked)
846{
Dave Hansencde70142016-02-12 13:01:55 -0800847 return __get_user_pages_locked(current, current->mm, start, nr_pages,
848 write, force, pages, NULL, locked, true,
849 FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800850}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200851EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800852
853/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800854 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
855 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
856 *
857 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
858 * caller if required (just like with __get_user_pages). "FOLL_GET",
859 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
860 * according to the parameters "pages", "write", "force"
861 * respectively.
862 */
863__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
864 unsigned long start, unsigned long nr_pages,
865 int write, int force, struct page **pages,
866 unsigned int gup_flags)
867{
868 long ret;
869 int locked = 1;
870 down_read(&mm->mmap_sem);
871 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
872 pages, NULL, &locked, false, gup_flags);
873 if (locked)
874 up_read(&mm->mmap_sem);
875 return ret;
876}
877EXPORT_SYMBOL(__get_user_pages_unlocked);
878
879/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800880 * get_user_pages_unlocked() is suitable to replace the form:
881 *
882 * down_read(&mm->mmap_sem);
883 * get_user_pages(tsk, mm, ..., pages, NULL);
884 * up_read(&mm->mmap_sem);
885 *
886 * with:
887 *
888 * get_user_pages_unlocked(tsk, mm, ..., pages);
889 *
890 * It is functionally equivalent to get_user_pages_fast so
891 * get_user_pages_fast should be used instead, if the two parameters
892 * "tsk" and "mm" are respectively equal to current and current->mm,
893 * or if "force" shall be set to 1 (get_user_pages_fast misses the
894 * "force" parameter).
895 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200896long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800897 int write, int force, struct page **pages)
898{
Dave Hansencde70142016-02-12 13:01:55 -0800899 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
900 write, force, pages, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800901}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200902EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800903
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700904/*
Dave Hansen1e987792016-02-12 13:01:54 -0800905 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700906 * @tsk: the task_struct to use for page fault accounting, or
907 * NULL if faults are not to be recorded.
908 * @mm: mm_struct of target mm
909 * @start: starting user address
910 * @nr_pages: number of pages from start to pin
911 * @write: whether pages will be written to by the caller
912 * @force: whether to force access even when user mapping is currently
913 * protected (but never forces write access to shared mapping).
914 * @pages: array that receives pointers to the pages pinned.
915 * Should be at least nr_pages long. Or NULL, if caller
916 * only intends to ensure the pages are faulted in.
917 * @vmas: array of pointers to vmas corresponding to each page.
918 * Or NULL if the caller does not require them.
919 *
920 * Returns number of pages pinned. This may be fewer than the number
921 * requested. If nr_pages is 0 or negative, returns 0. If no pages
922 * were pinned, returns -errno. Each page returned must be released
923 * with a put_page() call when it is finished with. vmas will only
924 * remain valid while mmap_sem is held.
925 *
926 * Must be called with mmap_sem held for read or write.
927 *
928 * get_user_pages walks a process's page tables and takes a reference to
929 * each struct page that each user address corresponds to at a given
930 * instant. That is, it takes the page that would be accessed if a user
931 * thread accesses the given user virtual address at that instant.
932 *
933 * This does not guarantee that the page exists in the user mappings when
934 * get_user_pages returns, and there may even be a completely different
935 * page there in some cases (eg. if mmapped pagecache has been invalidated
936 * and subsequently re faulted). However it does guarantee that the page
937 * won't be freed completely. And mostly callers simply care that the page
938 * contains data that was valid *at some point in time*. Typically, an IO
939 * or similar operation cannot guarantee anything stronger anyway because
940 * locks can't be held over the syscall boundary.
941 *
942 * If write=0, the page must not be written to. If the page is written to,
943 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
944 * after the page is finished with, and before put_page is called.
945 *
946 * get_user_pages is typically used for fewer-copy IO operations, to get a
947 * handle on the memory by some means other than accesses via the user virtual
948 * addresses. The pages may be submitted for DMA to devices or accessed via
949 * their kernel linear mapping (via the kmap APIs). Care should be taken to
950 * use the correct cache flushing APIs.
951 *
952 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800953 *
954 * get_user_pages should be phased out in favor of
955 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
956 * should use get_user_pages because it cannot pass
957 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700958 */
Dave Hansen1e987792016-02-12 13:01:54 -0800959long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
960 unsigned long start, unsigned long nr_pages,
961 int write, int force, struct page **pages,
962 struct vm_area_struct **vmas)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700963{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800964 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Dave Hansen1e987792016-02-12 13:01:54 -0800965 pages, vmas, NULL, false,
966 FOLL_TOUCH | FOLL_REMOTE);
967}
968EXPORT_SYMBOL(get_user_pages_remote);
969
970/*
Dave Hansend4edcf02016-02-12 13:01:56 -0800971 * This is the same as get_user_pages_remote(), just with a
972 * less-flexible calling convention where we assume that the task
973 * and mm being operated on are the current task's. We also
974 * obviously don't pass FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -0800975 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200976long get_user_pages(unsigned long start, unsigned long nr_pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800977 int write, int force, struct page **pages,
978 struct vm_area_struct **vmas)
979{
Dave Hansencde70142016-02-12 13:01:55 -0800980 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800981 write, force, pages, vmas, NULL, false,
982 FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700983}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200984EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700985
986/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700987 * populate_vma_page_range() - populate a range of pages in the vma.
988 * @vma: target vma
989 * @start: start address
990 * @end: end address
991 * @nonblocking:
992 *
993 * This takes care of mlocking the pages too if VM_LOCKED is set.
994 *
995 * return 0 on success, negative error code on error.
996 *
997 * vma->vm_mm->mmap_sem must be held.
998 *
999 * If @nonblocking is NULL, it may be held for read or write and will
1000 * be unperturbed.
1001 *
1002 * If @nonblocking is non-NULL, it must held for read only and may be
1003 * released. If it's released, *@nonblocking will be set to 0.
1004 */
1005long populate_vma_page_range(struct vm_area_struct *vma,
1006 unsigned long start, unsigned long end, int *nonblocking)
1007{
1008 struct mm_struct *mm = vma->vm_mm;
1009 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1010 int gup_flags;
1011
1012 VM_BUG_ON(start & ~PAGE_MASK);
1013 VM_BUG_ON(end & ~PAGE_MASK);
1014 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1015 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1016 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1017
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001018 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1019 if (vma->vm_flags & VM_LOCKONFAULT)
1020 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001021 /*
1022 * We want to touch writable mappings with a write fault in order
1023 * to break COW, except for shared mappings because these don't COW
1024 * and we would not want to dirty them for nothing.
1025 */
1026 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1027 gup_flags |= FOLL_WRITE;
1028
1029 /*
1030 * We want mlock to succeed for regions that have any permissions
1031 * other than PROT_NONE.
1032 */
1033 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1034 gup_flags |= FOLL_FORCE;
1035
1036 /*
1037 * We made sure addr is within a VMA, so the following will
1038 * not result in a stack expansion that recurses back here.
1039 */
1040 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1041 NULL, NULL, nonblocking);
1042}
1043
1044/*
1045 * __mm_populate - populate and/or mlock pages within a range of address space.
1046 *
1047 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1048 * flags. VMAs must be already marked with the desired vm_flags, and
1049 * mmap_sem must not be held.
1050 */
1051int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1052{
1053 struct mm_struct *mm = current->mm;
1054 unsigned long end, nstart, nend;
1055 struct vm_area_struct *vma = NULL;
1056 int locked = 0;
1057 long ret = 0;
1058
1059 VM_BUG_ON(start & ~PAGE_MASK);
1060 VM_BUG_ON(len != PAGE_ALIGN(len));
1061 end = start + len;
1062
1063 for (nstart = start; nstart < end; nstart = nend) {
1064 /*
1065 * We want to fault in pages for [nstart; end) address range.
1066 * Find first corresponding VMA.
1067 */
1068 if (!locked) {
1069 locked = 1;
1070 down_read(&mm->mmap_sem);
1071 vma = find_vma(mm, nstart);
1072 } else if (nstart >= vma->vm_end)
1073 vma = vma->vm_next;
1074 if (!vma || vma->vm_start >= end)
1075 break;
1076 /*
1077 * Set [nstart; nend) to intersection of desired address
1078 * range with the first VMA. Also, skip undesirable VMA types.
1079 */
1080 nend = min(end, vma->vm_end);
1081 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1082 continue;
1083 if (nstart < vma->vm_start)
1084 nstart = vma->vm_start;
1085 /*
1086 * Now fault in a range of pages. populate_vma_page_range()
1087 * double checks the vma flags, so that it won't mlock pages
1088 * if the vma was already munlocked.
1089 */
1090 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1091 if (ret < 0) {
1092 if (ignore_errors) {
1093 ret = 0;
1094 continue; /* continue at next VMA */
1095 }
1096 break;
1097 }
1098 nend = nstart + ret * PAGE_SIZE;
1099 ret = 0;
1100 }
1101 if (locked)
1102 up_read(&mm->mmap_sem);
1103 return ret; /* 0 or negative error code */
1104}
1105
1106/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001107 * get_dump_page() - pin user page in memory while writing it to core dump
1108 * @addr: user address
1109 *
1110 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001111 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001112 *
1113 * Returns NULL on any kind of failure - a hole must then be inserted into
1114 * the corefile, to preserve alignment with its headers; and also returns
1115 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1116 * allowing a hole to be left in the corefile to save diskspace.
1117 *
1118 * Called without mmap_sem, but after all other threads have been killed.
1119 */
1120#ifdef CONFIG_ELF_CORE
1121struct page *get_dump_page(unsigned long addr)
1122{
1123 struct vm_area_struct *vma;
1124 struct page *page;
1125
1126 if (__get_user_pages(current, current->mm, addr, 1,
1127 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1128 NULL) < 1)
1129 return NULL;
1130 flush_cache_page(vma, addr, page_to_pfn(page));
1131 return page;
1132}
1133#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001134
1135/*
1136 * Generic RCU Fast GUP
1137 *
1138 * get_user_pages_fast attempts to pin user pages by walking the page
1139 * tables directly and avoids taking locks. Thus the walker needs to be
1140 * protected from page table pages being freed from under it, and should
1141 * block any THP splits.
1142 *
1143 * One way to achieve this is to have the walker disable interrupts, and
1144 * rely on IPIs from the TLB flushing code blocking before the page table
1145 * pages are freed. This is unsuitable for architectures that do not need
1146 * to broadcast an IPI when invalidating TLBs.
1147 *
1148 * Another way to achieve this is to batch up page table containing pages
1149 * belonging to more than one mm_user, then rcu_sched a callback to free those
1150 * pages. Disabling interrupts will allow the fast_gup walker to both block
1151 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1152 * (which is a relatively rare event). The code below adopts this strategy.
1153 *
1154 * Before activating this code, please be aware that the following assumptions
1155 * are currently made:
1156 *
1157 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1158 * pages containing page tables.
1159 *
Steve Capper2667f502014-10-09 15:29:14 -07001160 * *) ptes can be read atomically by the architecture.
1161 *
1162 * *) access_ok is sufficient to validate userspace address ranges.
1163 *
1164 * The last two assumptions can be relaxed by the addition of helper functions.
1165 *
1166 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1167 */
1168#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1169
1170#ifdef __HAVE_ARCH_PTE_SPECIAL
1171static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1172 int write, struct page **pages, int *nr)
1173{
1174 pte_t *ptep, *ptem;
1175 int ret = 0;
1176
1177 ptem = ptep = pte_offset_map(&pmd, addr);
1178 do {
1179 /*
1180 * In the line below we are assuming that the pte can be read
1181 * atomically. If this is not the case for your architecture,
1182 * please wrap this in a helper function!
1183 *
1184 * for an example see gup_get_pte in arch/x86/mm/gup.c
1185 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001186 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001187 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001188
1189 /*
1190 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001191 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001192 */
1193 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001194 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001195 goto pte_unmap;
1196
Dave Hansen33a709b2016-02-12 13:02:19 -08001197 if (!arch_pte_access_permitted(pte, write))
1198 goto pte_unmap;
1199
Steve Capper2667f502014-10-09 15:29:14 -07001200 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1201 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001202 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001203
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001204 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001205 goto pte_unmap;
1206
1207 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001208 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001209 goto pte_unmap;
1210 }
1211
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001212 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001213 pages[*nr] = page;
1214 (*nr)++;
1215
1216 } while (ptep++, addr += PAGE_SIZE, addr != end);
1217
1218 ret = 1;
1219
1220pte_unmap:
1221 pte_unmap(ptem);
1222 return ret;
1223}
1224#else
1225
1226/*
1227 * If we can't determine whether or not a pte is special, then fail immediately
1228 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1229 * to be special.
1230 *
1231 * For a futex to be placed on a THP tail page, get_futex_key requires a
1232 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1233 * useful to have gup_huge_pmd even if we can't operate on ptes.
1234 */
1235static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1236 int write, struct page **pages, int *nr)
1237{
1238 return 0;
1239}
1240#endif /* __HAVE_ARCH_PTE_SPECIAL */
1241
1242static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1243 unsigned long end, int write, struct page **pages, int *nr)
1244{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001245 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001246 int refs;
1247
1248 if (write && !pmd_write(orig))
1249 return 0;
1250
1251 refs = 0;
1252 head = pmd_page(orig);
1253 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001254 do {
1255 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1256 pages[*nr] = page;
1257 (*nr)++;
1258 page++;
1259 refs++;
1260 } while (addr += PAGE_SIZE, addr != end);
1261
1262 if (!page_cache_add_speculative(head, refs)) {
1263 *nr -= refs;
1264 return 0;
1265 }
1266
1267 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1268 *nr -= refs;
1269 while (refs--)
1270 put_page(head);
1271 return 0;
1272 }
1273
Steve Capper2667f502014-10-09 15:29:14 -07001274 return 1;
1275}
1276
1277static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1278 unsigned long end, int write, struct page **pages, int *nr)
1279{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001280 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001281 int refs;
1282
1283 if (write && !pud_write(orig))
1284 return 0;
1285
1286 refs = 0;
1287 head = pud_page(orig);
1288 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001289 do {
1290 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1291 pages[*nr] = page;
1292 (*nr)++;
1293 page++;
1294 refs++;
1295 } while (addr += PAGE_SIZE, addr != end);
1296
1297 if (!page_cache_add_speculative(head, refs)) {
1298 *nr -= refs;
1299 return 0;
1300 }
1301
1302 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1303 *nr -= refs;
1304 while (refs--)
1305 put_page(head);
1306 return 0;
1307 }
1308
Steve Capper2667f502014-10-09 15:29:14 -07001309 return 1;
1310}
1311
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301312static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1313 unsigned long end, int write,
1314 struct page **pages, int *nr)
1315{
1316 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001317 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301318
1319 if (write && !pgd_write(orig))
1320 return 0;
1321
1322 refs = 0;
1323 head = pgd_page(orig);
1324 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301325 do {
1326 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1327 pages[*nr] = page;
1328 (*nr)++;
1329 page++;
1330 refs++;
1331 } while (addr += PAGE_SIZE, addr != end);
1332
1333 if (!page_cache_add_speculative(head, refs)) {
1334 *nr -= refs;
1335 return 0;
1336 }
1337
1338 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1339 *nr -= refs;
1340 while (refs--)
1341 put_page(head);
1342 return 0;
1343 }
1344
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301345 return 1;
1346}
1347
Steve Capper2667f502014-10-09 15:29:14 -07001348static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1349 int write, struct page **pages, int *nr)
1350{
1351 unsigned long next;
1352 pmd_t *pmdp;
1353
1354 pmdp = pmd_offset(&pud, addr);
1355 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001356 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001357
1358 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001359 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001360 return 0;
1361
1362 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1363 /*
1364 * NUMA hinting faults need to be handled in the GUP
1365 * slowpath for accounting purposes and so that they
1366 * can be serialised against THP migration.
1367 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001368 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001369 return 0;
1370
1371 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1372 pages, nr))
1373 return 0;
1374
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301375 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1376 /*
1377 * architecture have different format for hugetlbfs
1378 * pmd format and THP pmd format
1379 */
1380 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1381 PMD_SHIFT, next, write, pages, nr))
1382 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001383 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1384 return 0;
1385 } while (pmdp++, addr = next, addr != end);
1386
1387 return 1;
1388}
1389
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301390static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1391 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001392{
1393 unsigned long next;
1394 pud_t *pudp;
1395
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301396 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001397 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001398 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001399
1400 next = pud_addr_end(addr, end);
1401 if (pud_none(pud))
1402 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301403 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001404 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301405 pages, nr))
1406 return 0;
1407 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1408 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1409 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001410 return 0;
1411 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1412 return 0;
1413 } while (pudp++, addr = next, addr != end);
1414
1415 return 1;
1416}
1417
1418/*
1419 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1420 * the regular GUP. It will only return non-negative values.
1421 */
1422int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1423 struct page **pages)
1424{
1425 struct mm_struct *mm = current->mm;
1426 unsigned long addr, len, end;
1427 unsigned long next, flags;
1428 pgd_t *pgdp;
1429 int nr = 0;
1430
1431 start &= PAGE_MASK;
1432 addr = start;
1433 len = (unsigned long) nr_pages << PAGE_SHIFT;
1434 end = start + len;
1435
1436 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1437 start, len)))
1438 return 0;
1439
1440 /*
1441 * Disable interrupts. We use the nested form as we can already have
1442 * interrupts disabled by get_futex_key.
1443 *
1444 * With interrupts disabled, we block page table pages from being
1445 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1446 * for more details.
1447 *
1448 * We do not adopt an rcu_read_lock(.) here as we also want to
1449 * block IPIs that come from THPs splitting.
1450 */
1451
1452 local_irq_save(flags);
1453 pgdp = pgd_offset(mm, addr);
1454 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001455 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301456
Steve Capper2667f502014-10-09 15:29:14 -07001457 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301458 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001459 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301460 if (unlikely(pgd_huge(pgd))) {
1461 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1462 pages, &nr))
1463 break;
1464 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1465 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1466 PGDIR_SHIFT, next, write, pages, &nr))
1467 break;
1468 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001469 break;
1470 } while (pgdp++, addr = next, addr != end);
1471 local_irq_restore(flags);
1472
1473 return nr;
1474}
1475
1476/**
1477 * get_user_pages_fast() - pin user pages in memory
1478 * @start: starting user address
1479 * @nr_pages: number of pages from start to pin
1480 * @write: whether pages will be written to
1481 * @pages: array that receives pointers to the pages pinned.
1482 * Should be at least nr_pages long.
1483 *
1484 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1485 * If not successful, it will fall back to taking the lock and
1486 * calling get_user_pages().
1487 *
1488 * Returns number of pages pinned. This may be fewer than the number
1489 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1490 * were pinned, returns -errno.
1491 */
1492int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1493 struct page **pages)
1494{
Steve Capper2667f502014-10-09 15:29:14 -07001495 int nr, ret;
1496
1497 start &= PAGE_MASK;
1498 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1499 ret = nr;
1500
1501 if (nr < nr_pages) {
1502 /* Try to get the remaining pages with get_user_pages */
1503 start += nr << PAGE_SHIFT;
1504 pages += nr;
1505
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001506 ret = get_user_pages_unlocked(start, nr_pages - nr, write, 0, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001507
1508 /* Have to be a bit careful with return values */
1509 if (nr > 0) {
1510 if (ret < 0)
1511 ret = nr;
1512 else
1513 ret += nr;
1514 }
1515 }
1516
1517 return ret;
1518}
1519
1520#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */