blob: b64a36175884e07604b0e216bc2d545a2892dcb7 [file] [log] [blame]
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001#include <linux/kernel.h>
2#include <linux/errno.h>
3#include <linux/err.h>
4#include <linux/spinlock.h>
5
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07006#include <linux/mm.h>
Dan Williams3565fce2016-01-15 16:56:55 -08007#include <linux/memremap.h>
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07008#include <linux/pagemap.h>
9#include <linux/rmap.h>
10#include <linux/swap.h>
11#include <linux/swapops.h>
12
Steve Capper2667f502014-10-09 15:29:14 -070013#include <linux/sched.h>
14#include <linux/rwsem.h>
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +053015#include <linux/hugetlb.h>
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070016
Steve Capper2667f502014-10-09 15:29:14 -070017#include <asm/pgtable.h>
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070018#include <asm/tlbflush.h>
Steve Capper2667f502014-10-09 15:29:14 -070019
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070020#include "internal.h"
21
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070022static struct page *no_page_table(struct vm_area_struct *vma,
23 unsigned int flags)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070024{
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070025 /*
26 * When core dumping an enormous anonymous area that nobody
27 * has touched so far, we don't want to allocate unnecessary pages or
28 * page tables. Return error instead of NULL to skip handle_mm_fault,
29 * then get_dump_page() will return NULL to leave a hole in the dump.
30 * But we can only make this optimization where a hole would surely
31 * be zero-filled if handle_mm_fault() actually did handle it.
32 */
33 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
34 return ERR_PTR(-EFAULT);
35 return NULL;
36}
37
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070038static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
39 pte_t *pte, unsigned int flags)
40{
41 /* No page to get reference */
42 if (flags & FOLL_GET)
43 return -EFAULT;
44
45 if (flags & FOLL_TOUCH) {
46 pte_t entry = *pte;
47
48 if (flags & FOLL_WRITE)
49 entry = pte_mkdirty(entry);
50 entry = pte_mkyoung(entry);
51
52 if (!pte_same(*pte, entry)) {
53 set_pte_at(vma->vm_mm, address, pte, entry);
54 update_mmu_cache(vma, address, pte);
55 }
56 }
57
58 /* Proper page table entry exists, but no corresponding struct page */
59 return -EEXIST;
60}
61
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070062static struct page *follow_page_pte(struct vm_area_struct *vma,
63 unsigned long address, pmd_t *pmd, unsigned int flags)
64{
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070065 struct mm_struct *mm = vma->vm_mm;
Dan Williams3565fce2016-01-15 16:56:55 -080066 struct dev_pagemap *pgmap = NULL;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070067 struct page *page;
68 spinlock_t *ptl;
69 pte_t *ptep, pte;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070070
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070071retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070072 if (unlikely(pmd_bad(*pmd)))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070073 return no_page_table(vma, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070074
75 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070076 pte = *ptep;
77 if (!pte_present(pte)) {
78 swp_entry_t entry;
79 /*
80 * KSM's break_ksm() relies upon recognizing a ksm page
81 * even while it is being migrated, so for that case we
82 * need migration_entry_wait().
83 */
84 if (likely(!(flags & FOLL_MIGRATION)))
85 goto no_page;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -080086 if (pte_none(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070087 goto no_page;
88 entry = pte_to_swp_entry(pte);
89 if (!is_migration_entry(entry))
90 goto no_page;
91 pte_unmap_unlock(ptep, ptl);
92 migration_entry_wait(mm, pmd, address);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070093 goto retry;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070094 }
Mel Gorman8a0516e2015-02-12 14:58:22 -080095 if ((flags & FOLL_NUMA) && pte_protnone(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070096 goto no_page;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070097 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
98 pte_unmap_unlock(ptep, ptl);
99 return NULL;
100 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700101
102 page = vm_normal_page(vma, address, pte);
Dan Williams3565fce2016-01-15 16:56:55 -0800103 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
104 /*
105 * Only return device mapping pages in the FOLL_GET case since
106 * they are only valid while holding the pgmap reference.
107 */
108 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
109 if (pgmap)
110 page = pte_page(pte);
111 else
112 goto no_page;
113 } else if (unlikely(!page)) {
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700114 if (flags & FOLL_DUMP) {
115 /* Avoid special (like zero) pages in core dumps */
116 page = ERR_PTR(-EFAULT);
117 goto out;
118 }
119
120 if (is_zero_pfn(pte_pfn(pte))) {
121 page = pte_page(pte);
122 } else {
123 int ret;
124
125 ret = follow_pfn_pte(vma, address, ptep, flags);
126 page = ERR_PTR(ret);
127 goto out;
128 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700129 }
130
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800131 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
132 int ret;
133 get_page(page);
134 pte_unmap_unlock(ptep, ptl);
135 lock_page(page);
136 ret = split_huge_page(page);
137 unlock_page(page);
138 put_page(page);
139 if (ret)
140 return ERR_PTR(ret);
141 goto retry;
142 }
143
Dan Williams3565fce2016-01-15 16:56:55 -0800144 if (flags & FOLL_GET) {
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800145 get_page(page);
Dan Williams3565fce2016-01-15 16:56:55 -0800146
147 /* drop the pgmap reference now that we hold the page */
148 if (pgmap) {
149 put_dev_pagemap(pgmap);
150 pgmap = NULL;
151 }
152 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700153 if (flags & FOLL_TOUCH) {
154 if ((flags & FOLL_WRITE) &&
155 !pte_dirty(pte) && !PageDirty(page))
156 set_page_dirty(page);
157 /*
158 * pte_mkyoung() would be more correct here, but atomic care
159 * is needed to avoid losing the dirty bit: it is easier to use
160 * mark_page_accessed().
161 */
162 mark_page_accessed(page);
163 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800164 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -0800165 /* Do not mlock pte-mapped THP */
166 if (PageTransCompound(page))
167 goto out;
168
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700169 /*
170 * The preliminary mapping check is mainly to avoid the
171 * pointless overhead of lock_page on the ZERO_PAGE
172 * which might bounce very badly if there is contention.
173 *
174 * If the page is already locked, we don't need to
175 * handle it now - vmscan will handle it later if and
176 * when it attempts to reclaim the page.
177 */
178 if (page->mapping && trylock_page(page)) {
179 lru_add_drain(); /* push cached pages to LRU */
180 /*
181 * Because we lock page here, and migration is
182 * blocked by the pte's page reference, and we
183 * know the page is still mapped, we don't even
184 * need to check for file-cache page truncation.
185 */
186 mlock_vma_page(page);
187 unlock_page(page);
188 }
189 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700190out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700191 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700192 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700193no_page:
194 pte_unmap_unlock(ptep, ptl);
195 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700196 return NULL;
197 return no_page_table(vma, flags);
198}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700199
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700200/**
201 * follow_page_mask - look up a page descriptor from a user-virtual address
202 * @vma: vm_area_struct mapping @address
203 * @address: virtual address to look up
204 * @flags: flags modifying lookup behaviour
205 * @page_mask: on output, *page_mask is set according to the size of the page
206 *
207 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
208 *
209 * Returns the mapped (struct page *), %NULL if no mapping exists, or
210 * an error pointer if there is a mapping to something not represented
211 * by a page descriptor (see also vm_normal_page()).
212 */
213struct page *follow_page_mask(struct vm_area_struct *vma,
214 unsigned long address, unsigned int flags,
215 unsigned int *page_mask)
216{
217 pgd_t *pgd;
218 pud_t *pud;
219 pmd_t *pmd;
220 spinlock_t *ptl;
221 struct page *page;
222 struct mm_struct *mm = vma->vm_mm;
223
224 *page_mask = 0;
225
226 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
227 if (!IS_ERR(page)) {
228 BUG_ON(flags & FOLL_GET);
229 return page;
230 }
231
232 pgd = pgd_offset(mm, address);
233 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
234 return no_page_table(vma, flags);
235
236 pud = pud_offset(pgd, address);
237 if (pud_none(*pud))
238 return no_page_table(vma, flags);
239 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800240 page = follow_huge_pud(mm, address, pud, flags);
241 if (page)
242 return page;
243 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700244 }
245 if (unlikely(pud_bad(*pud)))
246 return no_page_table(vma, flags);
247
248 pmd = pmd_offset(pud, address);
249 if (pmd_none(*pmd))
250 return no_page_table(vma, flags);
251 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800252 page = follow_huge_pmd(mm, address, pmd, flags);
253 if (page)
254 return page;
255 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700256 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800257 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700258 return no_page_table(vma, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800259 if (pmd_devmap(*pmd)) {
260 ptl = pmd_lock(mm, pmd);
261 page = follow_devmap_pmd(vma, address, pmd, flags);
262 spin_unlock(ptl);
263 if (page)
264 return page;
265 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800266 if (likely(!pmd_trans_huge(*pmd)))
267 return follow_page_pte(vma, address, pmd, flags);
268
269 ptl = pmd_lock(mm, pmd);
270 if (unlikely(!pmd_trans_huge(*pmd))) {
271 spin_unlock(ptl);
272 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700273 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800274 if (flags & FOLL_SPLIT) {
275 int ret;
276 page = pmd_page(*pmd);
277 if (is_huge_zero_page(page)) {
278 spin_unlock(ptl);
279 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800280 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800281 } else {
282 get_page(page);
283 spin_unlock(ptl);
284 lock_page(page);
285 ret = split_huge_page(page);
286 unlock_page(page);
287 put_page(page);
288 }
289
290 return ret ? ERR_PTR(ret) :
291 follow_page_pte(vma, address, pmd, flags);
292 }
293
294 page = follow_trans_huge_pmd(vma, address, pmd, flags);
295 spin_unlock(ptl);
296 *page_mask = HPAGE_PMD_NR - 1;
297 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700298}
299
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700300static int get_gate_page(struct mm_struct *mm, unsigned long address,
301 unsigned int gup_flags, struct vm_area_struct **vma,
302 struct page **page)
303{
304 pgd_t *pgd;
305 pud_t *pud;
306 pmd_t *pmd;
307 pte_t *pte;
308 int ret = -EFAULT;
309
310 /* user gate pages are read-only */
311 if (gup_flags & FOLL_WRITE)
312 return -EFAULT;
313 if (address > TASK_SIZE)
314 pgd = pgd_offset_k(address);
315 else
316 pgd = pgd_offset_gate(mm, address);
317 BUG_ON(pgd_none(*pgd));
318 pud = pud_offset(pgd, address);
319 BUG_ON(pud_none(*pud));
320 pmd = pmd_offset(pud, address);
321 if (pmd_none(*pmd))
322 return -EFAULT;
323 VM_BUG_ON(pmd_trans_huge(*pmd));
324 pte = pte_offset_map(pmd, address);
325 if (pte_none(*pte))
326 goto unmap;
327 *vma = get_gate_vma(mm);
328 if (!page)
329 goto out;
330 *page = vm_normal_page(*vma, address, *pte);
331 if (!*page) {
332 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
333 goto unmap;
334 *page = pte_page(*pte);
335 }
336 get_page(*page);
337out:
338 ret = 0;
339unmap:
340 pte_unmap(pte);
341 return ret;
342}
343
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700344/*
345 * mmap_sem must be held on entry. If @nonblocking != NULL and
346 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
347 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
348 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700349static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
350 unsigned long address, unsigned int *flags, int *nonblocking)
351{
352 struct mm_struct *mm = vma->vm_mm;
353 unsigned int fault_flags = 0;
354 int ret;
355
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800356 /* mlock all present pages, but do not fault in new pages */
357 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
358 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700359 /* For mm_populate(), just skip the stack guard page. */
360 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700361 (stack_guard_page_start(vma, address) ||
362 stack_guard_page_end(vma, address + PAGE_SIZE)))
363 return -ENOENT;
364 if (*flags & FOLL_WRITE)
365 fault_flags |= FAULT_FLAG_WRITE;
366 if (nonblocking)
367 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
368 if (*flags & FOLL_NOWAIT)
369 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700370 if (*flags & FOLL_TRIED) {
371 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
372 fault_flags |= FAULT_FLAG_TRIED;
373 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700374
375 ret = handle_mm_fault(mm, vma, address, fault_flags);
376 if (ret & VM_FAULT_ERROR) {
377 if (ret & VM_FAULT_OOM)
378 return -ENOMEM;
379 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
380 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800381 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700382 return -EFAULT;
383 BUG();
384 }
385
386 if (tsk) {
387 if (ret & VM_FAULT_MAJOR)
388 tsk->maj_flt++;
389 else
390 tsk->min_flt++;
391 }
392
393 if (ret & VM_FAULT_RETRY) {
394 if (nonblocking)
395 *nonblocking = 0;
396 return -EBUSY;
397 }
398
399 /*
400 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
401 * necessary, even if maybe_mkwrite decided not to set pte_write. We
402 * can thus safely do subsequent page lookups as if they were reads.
403 * But only do so when looping for pte_write is futile: in some cases
404 * userspace may also be wanting to write to the gotten user page,
405 * which a read fault here might prevent (a readonly page might get
406 * reCOWed by userspace write).
407 */
408 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
409 *flags &= ~FOLL_WRITE;
410 return 0;
411}
412
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700413static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
414{
415 vm_flags_t vm_flags = vma->vm_flags;
416
417 if (vm_flags & (VM_IO | VM_PFNMAP))
418 return -EFAULT;
419
420 if (gup_flags & FOLL_WRITE) {
421 if (!(vm_flags & VM_WRITE)) {
422 if (!(gup_flags & FOLL_FORCE))
423 return -EFAULT;
424 /*
425 * We used to let the write,force case do COW in a
426 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
427 * set a breakpoint in a read-only mapping of an
428 * executable, without corrupting the file (yet only
429 * when that file had been opened for writing!).
430 * Anon pages in shared mappings are surprising: now
431 * just reject it.
432 */
433 if (!is_cow_mapping(vm_flags)) {
434 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
435 return -EFAULT;
436 }
437 }
438 } else if (!(vm_flags & VM_READ)) {
439 if (!(gup_flags & FOLL_FORCE))
440 return -EFAULT;
441 /*
442 * Is there actually any vma we can reach here which does not
443 * have VM_MAYREAD set?
444 */
445 if (!(vm_flags & VM_MAYREAD))
446 return -EFAULT;
447 }
448 return 0;
449}
450
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700451/**
452 * __get_user_pages() - pin user pages in memory
453 * @tsk: task_struct of target task
454 * @mm: mm_struct of target mm
455 * @start: starting user address
456 * @nr_pages: number of pages from start to pin
457 * @gup_flags: flags modifying pin behaviour
458 * @pages: array that receives pointers to the pages pinned.
459 * Should be at least nr_pages long. Or NULL, if caller
460 * only intends to ensure the pages are faulted in.
461 * @vmas: array of pointers to vmas corresponding to each page.
462 * Or NULL if the caller does not require them.
463 * @nonblocking: whether waiting for disk IO or mmap_sem contention
464 *
465 * Returns number of pages pinned. This may be fewer than the number
466 * requested. If nr_pages is 0 or negative, returns 0. If no pages
467 * were pinned, returns -errno. Each page returned must be released
468 * with a put_page() call when it is finished with. vmas will only
469 * remain valid while mmap_sem is held.
470 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700471 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700472 *
473 * __get_user_pages walks a process's page tables and takes a reference to
474 * each struct page that each user address corresponds to at a given
475 * instant. That is, it takes the page that would be accessed if a user
476 * thread accesses the given user virtual address at that instant.
477 *
478 * This does not guarantee that the page exists in the user mappings when
479 * __get_user_pages returns, and there may even be a completely different
480 * page there in some cases (eg. if mmapped pagecache has been invalidated
481 * and subsequently re faulted). However it does guarantee that the page
482 * won't be freed completely. And mostly callers simply care that the page
483 * contains data that was valid *at some point in time*. Typically, an IO
484 * or similar operation cannot guarantee anything stronger anyway because
485 * locks can't be held over the syscall boundary.
486 *
487 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
488 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
489 * appropriate) must be called after the page is finished with, and
490 * before put_page is called.
491 *
492 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
493 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700494 * *@nonblocking will be set to 0. Further, if @gup_flags does not
495 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
496 * this case.
497 *
498 * A caller using such a combination of @nonblocking and @gup_flags
499 * must therefore hold the mmap_sem for reading only, and recognize
500 * when it's been released. Otherwise, it must be held for either
501 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700502 *
503 * In most cases, get_user_pages or get_user_pages_fast should be used
504 * instead of __get_user_pages. __get_user_pages should be used only if
505 * you need some special @gup_flags.
506 */
507long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
508 unsigned long start, unsigned long nr_pages,
509 unsigned int gup_flags, struct page **pages,
510 struct vm_area_struct **vmas, int *nonblocking)
511{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700512 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700513 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700514 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700515
516 if (!nr_pages)
517 return 0;
518
519 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
520
521 /*
522 * If FOLL_FORCE is set then do not force a full fault as the hinting
523 * fault information is unrelated to the reference behaviour of a task
524 * using the address space
525 */
526 if (!(gup_flags & FOLL_FORCE))
527 gup_flags |= FOLL_NUMA;
528
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700529 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700530 struct page *page;
531 unsigned int foll_flags = gup_flags;
532 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700533
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700534 /* first iteration or cross vma bound */
535 if (!vma || start >= vma->vm_end) {
536 vma = find_extend_vma(mm, start);
537 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700538 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700539 ret = get_gate_page(mm, start & PAGE_MASK,
540 gup_flags, &vma,
541 pages ? &pages[i] : NULL);
542 if (ret)
543 return i ? : ret;
544 page_mask = 0;
545 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700546 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700547
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700548 if (!vma || check_vma_flags(vma, gup_flags))
549 return i ? : -EFAULT;
550 if (is_vm_hugetlb_page(vma)) {
551 i = follow_hugetlb_page(mm, vma, pages, vmas,
552 &start, &nr_pages, i,
553 gup_flags);
554 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700555 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700556 }
557retry:
558 /*
559 * If we have a pending SIGKILL, don't keep faulting pages and
560 * potentially allocating memory.
561 */
562 if (unlikely(fatal_signal_pending(current)))
563 return i ? i : -ERESTARTSYS;
564 cond_resched();
565 page = follow_page_mask(vma, start, foll_flags, &page_mask);
566 if (!page) {
567 int ret;
568 ret = faultin_page(tsk, vma, start, &foll_flags,
569 nonblocking);
570 switch (ret) {
571 case 0:
572 goto retry;
573 case -EFAULT:
574 case -ENOMEM:
575 case -EHWPOISON:
576 return i ? i : ret;
577 case -EBUSY:
578 return i;
579 case -ENOENT:
580 goto next_page;
581 }
582 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700583 } else if (PTR_ERR(page) == -EEXIST) {
584 /*
585 * Proper page table entry exists, but no corresponding
586 * struct page.
587 */
588 goto next_page;
589 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700590 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700591 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700592 if (pages) {
593 pages[i] = page;
594 flush_anon_page(vma, page, start);
595 flush_dcache_page(page);
596 page_mask = 0;
597 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700598next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700599 if (vmas) {
600 vmas[i] = vma;
601 page_mask = 0;
602 }
603 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
604 if (page_increm > nr_pages)
605 page_increm = nr_pages;
606 i += page_increm;
607 start += page_increm * PAGE_SIZE;
608 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700609 } while (nr_pages);
610 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700611}
612EXPORT_SYMBOL(__get_user_pages);
613
614/*
615 * fixup_user_fault() - manually resolve a user page fault
616 * @tsk: the task_struct to use for page fault accounting, or
617 * NULL if faults are not to be recorded.
618 * @mm: mm_struct of target mm
619 * @address: user address
620 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800621 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
622 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700623 *
624 * This is meant to be called in the specific scenario where for locking reasons
625 * we try to access user memory in atomic context (within a pagefault_disable()
626 * section), this returns -EFAULT, and we want to resolve the user fault before
627 * trying again.
628 *
629 * Typically this is meant to be used by the futex code.
630 *
631 * The main difference with get_user_pages() is that this function will
632 * unconditionally call handle_mm_fault() which will in turn perform all the
633 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800634 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700635 *
636 * This is important for some architectures where those bits also gate the
637 * access permission to the page because they are maintained in software. On
638 * such architectures, gup() will not be enough to make a subsequent access
639 * succeed.
640 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800641 * This function will not return with an unlocked mmap_sem. So it has not the
642 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700643 */
644int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800645 unsigned long address, unsigned int fault_flags,
646 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700647{
648 struct vm_area_struct *vma;
649 vm_flags_t vm_flags;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800650 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700651
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800652 if (unlocked)
653 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
654
655retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700656 vma = find_extend_vma(mm, address);
657 if (!vma || address < vma->vm_start)
658 return -EFAULT;
659
660 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
661 if (!(vm_flags & vma->vm_flags))
662 return -EFAULT;
663
664 ret = handle_mm_fault(mm, vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800665 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700666 if (ret & VM_FAULT_ERROR) {
667 if (ret & VM_FAULT_OOM)
668 return -ENOMEM;
669 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
670 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800671 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700672 return -EFAULT;
673 BUG();
674 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800675
676 if (ret & VM_FAULT_RETRY) {
677 down_read(&mm->mmap_sem);
678 if (!(fault_flags & FAULT_FLAG_TRIED)) {
679 *unlocked = true;
680 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
681 fault_flags |= FAULT_FLAG_TRIED;
682 goto retry;
683 }
684 }
685
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700686 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800687 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700688 tsk->maj_flt++;
689 else
690 tsk->min_flt++;
691 }
692 return 0;
693}
694
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800695static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
696 struct mm_struct *mm,
697 unsigned long start,
698 unsigned long nr_pages,
699 int write, int force,
700 struct page **pages,
701 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800702 int *locked, bool notify_drop,
703 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800704{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800705 long ret, pages_done;
706 bool lock_dropped;
707
708 if (locked) {
709 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
710 BUG_ON(vmas);
711 /* check caller initialized locked */
712 BUG_ON(*locked != 1);
713 }
714
715 if (pages)
716 flags |= FOLL_GET;
717 if (write)
718 flags |= FOLL_WRITE;
719 if (force)
720 flags |= FOLL_FORCE;
721
722 pages_done = 0;
723 lock_dropped = false;
724 for (;;) {
725 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
726 vmas, locked);
727 if (!locked)
728 /* VM_FAULT_RETRY couldn't trigger, bypass */
729 return ret;
730
731 /* VM_FAULT_RETRY cannot return errors */
732 if (!*locked) {
733 BUG_ON(ret < 0);
734 BUG_ON(ret >= nr_pages);
735 }
736
737 if (!pages)
738 /* If it's a prefault don't insist harder */
739 return ret;
740
741 if (ret > 0) {
742 nr_pages -= ret;
743 pages_done += ret;
744 if (!nr_pages)
745 break;
746 }
747 if (*locked) {
748 /* VM_FAULT_RETRY didn't trigger */
749 if (!pages_done)
750 pages_done = ret;
751 break;
752 }
753 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
754 pages += ret;
755 start += ret << PAGE_SHIFT;
756
757 /*
758 * Repeat on the address that fired VM_FAULT_RETRY
759 * without FAULT_FLAG_ALLOW_RETRY but with
760 * FAULT_FLAG_TRIED.
761 */
762 *locked = 1;
763 lock_dropped = true;
764 down_read(&mm->mmap_sem);
765 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
766 pages, NULL, NULL);
767 if (ret != 1) {
768 BUG_ON(ret > 1);
769 if (!pages_done)
770 pages_done = ret;
771 break;
772 }
773 nr_pages--;
774 pages_done++;
775 if (!nr_pages)
776 break;
777 pages++;
778 start += PAGE_SIZE;
779 }
780 if (notify_drop && lock_dropped && *locked) {
781 /*
782 * We must let the caller know we temporarily dropped the lock
783 * and so the critical section protected by it was lost.
784 */
785 up_read(&mm->mmap_sem);
786 *locked = 0;
787 }
788 return pages_done;
789}
790
791/*
792 * We can leverage the VM_FAULT_RETRY functionality in the page fault
793 * paths better by using either get_user_pages_locked() or
794 * get_user_pages_unlocked().
795 *
796 * get_user_pages_locked() is suitable to replace the form:
797 *
798 * down_read(&mm->mmap_sem);
799 * do_something()
800 * get_user_pages(tsk, mm, ..., pages, NULL);
801 * up_read(&mm->mmap_sem);
802 *
803 * to:
804 *
805 * int locked = 1;
806 * down_read(&mm->mmap_sem);
807 * do_something()
808 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
809 * if (locked)
810 * up_read(&mm->mmap_sem);
811 */
812long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
813 unsigned long start, unsigned long nr_pages,
814 int write, int force, struct page **pages,
815 int *locked)
816{
817 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800818 pages, NULL, locked, true, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800819}
820EXPORT_SYMBOL(get_user_pages_locked);
821
822/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800823 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
824 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
825 *
826 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
827 * caller if required (just like with __get_user_pages). "FOLL_GET",
828 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
829 * according to the parameters "pages", "write", "force"
830 * respectively.
831 */
832__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
833 unsigned long start, unsigned long nr_pages,
834 int write, int force, struct page **pages,
835 unsigned int gup_flags)
836{
837 long ret;
838 int locked = 1;
839 down_read(&mm->mmap_sem);
840 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
841 pages, NULL, &locked, false, gup_flags);
842 if (locked)
843 up_read(&mm->mmap_sem);
844 return ret;
845}
846EXPORT_SYMBOL(__get_user_pages_unlocked);
847
848/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800849 * get_user_pages_unlocked() is suitable to replace the form:
850 *
851 * down_read(&mm->mmap_sem);
852 * get_user_pages(tsk, mm, ..., pages, NULL);
853 * up_read(&mm->mmap_sem);
854 *
855 * with:
856 *
857 * get_user_pages_unlocked(tsk, mm, ..., pages);
858 *
859 * It is functionally equivalent to get_user_pages_fast so
860 * get_user_pages_fast should be used instead, if the two parameters
861 * "tsk" and "mm" are respectively equal to current and current->mm,
862 * or if "force" shall be set to 1 (get_user_pages_fast misses the
863 * "force" parameter).
864 */
865long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
866 unsigned long start, unsigned long nr_pages,
867 int write, int force, struct page **pages)
868{
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800869 return __get_user_pages_unlocked(tsk, mm, start, nr_pages, write,
870 force, pages, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800871}
872EXPORT_SYMBOL(get_user_pages_unlocked);
873
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700874/*
875 * get_user_pages() - pin user pages in memory
876 * @tsk: the task_struct to use for page fault accounting, or
877 * NULL if faults are not to be recorded.
878 * @mm: mm_struct of target mm
879 * @start: starting user address
880 * @nr_pages: number of pages from start to pin
881 * @write: whether pages will be written to by the caller
882 * @force: whether to force access even when user mapping is currently
883 * protected (but never forces write access to shared mapping).
884 * @pages: array that receives pointers to the pages pinned.
885 * Should be at least nr_pages long. Or NULL, if caller
886 * only intends to ensure the pages are faulted in.
887 * @vmas: array of pointers to vmas corresponding to each page.
888 * Or NULL if the caller does not require them.
889 *
890 * Returns number of pages pinned. This may be fewer than the number
891 * requested. If nr_pages is 0 or negative, returns 0. If no pages
892 * were pinned, returns -errno. Each page returned must be released
893 * with a put_page() call when it is finished with. vmas will only
894 * remain valid while mmap_sem is held.
895 *
896 * Must be called with mmap_sem held for read or write.
897 *
898 * get_user_pages walks a process's page tables and takes a reference to
899 * each struct page that each user address corresponds to at a given
900 * instant. That is, it takes the page that would be accessed if a user
901 * thread accesses the given user virtual address at that instant.
902 *
903 * This does not guarantee that the page exists in the user mappings when
904 * get_user_pages returns, and there may even be a completely different
905 * page there in some cases (eg. if mmapped pagecache has been invalidated
906 * and subsequently re faulted). However it does guarantee that the page
907 * won't be freed completely. And mostly callers simply care that the page
908 * contains data that was valid *at some point in time*. Typically, an IO
909 * or similar operation cannot guarantee anything stronger anyway because
910 * locks can't be held over the syscall boundary.
911 *
912 * If write=0, the page must not be written to. If the page is written to,
913 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
914 * after the page is finished with, and before put_page is called.
915 *
916 * get_user_pages is typically used for fewer-copy IO operations, to get a
917 * handle on the memory by some means other than accesses via the user virtual
918 * addresses. The pages may be submitted for DMA to devices or accessed via
919 * their kernel linear mapping (via the kmap APIs). Care should be taken to
920 * use the correct cache flushing APIs.
921 *
922 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800923 *
924 * get_user_pages should be phased out in favor of
925 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
926 * should use get_user_pages because it cannot pass
927 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700928 */
929long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
930 unsigned long start, unsigned long nr_pages, int write,
931 int force, struct page **pages, struct vm_area_struct **vmas)
932{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800933 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800934 pages, vmas, NULL, false, FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700935}
936EXPORT_SYMBOL(get_user_pages);
937
938/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700939 * populate_vma_page_range() - populate a range of pages in the vma.
940 * @vma: target vma
941 * @start: start address
942 * @end: end address
943 * @nonblocking:
944 *
945 * This takes care of mlocking the pages too if VM_LOCKED is set.
946 *
947 * return 0 on success, negative error code on error.
948 *
949 * vma->vm_mm->mmap_sem must be held.
950 *
951 * If @nonblocking is NULL, it may be held for read or write and will
952 * be unperturbed.
953 *
954 * If @nonblocking is non-NULL, it must held for read only and may be
955 * released. If it's released, *@nonblocking will be set to 0.
956 */
957long populate_vma_page_range(struct vm_area_struct *vma,
958 unsigned long start, unsigned long end, int *nonblocking)
959{
960 struct mm_struct *mm = vma->vm_mm;
961 unsigned long nr_pages = (end - start) / PAGE_SIZE;
962 int gup_flags;
963
964 VM_BUG_ON(start & ~PAGE_MASK);
965 VM_BUG_ON(end & ~PAGE_MASK);
966 VM_BUG_ON_VMA(start < vma->vm_start, vma);
967 VM_BUG_ON_VMA(end > vma->vm_end, vma);
968 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
969
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800970 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
971 if (vma->vm_flags & VM_LOCKONFAULT)
972 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700973 /*
974 * We want to touch writable mappings with a write fault in order
975 * to break COW, except for shared mappings because these don't COW
976 * and we would not want to dirty them for nothing.
977 */
978 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
979 gup_flags |= FOLL_WRITE;
980
981 /*
982 * We want mlock to succeed for regions that have any permissions
983 * other than PROT_NONE.
984 */
985 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
986 gup_flags |= FOLL_FORCE;
987
988 /*
989 * We made sure addr is within a VMA, so the following will
990 * not result in a stack expansion that recurses back here.
991 */
992 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
993 NULL, NULL, nonblocking);
994}
995
996/*
997 * __mm_populate - populate and/or mlock pages within a range of address space.
998 *
999 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1000 * flags. VMAs must be already marked with the desired vm_flags, and
1001 * mmap_sem must not be held.
1002 */
1003int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1004{
1005 struct mm_struct *mm = current->mm;
1006 unsigned long end, nstart, nend;
1007 struct vm_area_struct *vma = NULL;
1008 int locked = 0;
1009 long ret = 0;
1010
1011 VM_BUG_ON(start & ~PAGE_MASK);
1012 VM_BUG_ON(len != PAGE_ALIGN(len));
1013 end = start + len;
1014
1015 for (nstart = start; nstart < end; nstart = nend) {
1016 /*
1017 * We want to fault in pages for [nstart; end) address range.
1018 * Find first corresponding VMA.
1019 */
1020 if (!locked) {
1021 locked = 1;
1022 down_read(&mm->mmap_sem);
1023 vma = find_vma(mm, nstart);
1024 } else if (nstart >= vma->vm_end)
1025 vma = vma->vm_next;
1026 if (!vma || vma->vm_start >= end)
1027 break;
1028 /*
1029 * Set [nstart; nend) to intersection of desired address
1030 * range with the first VMA. Also, skip undesirable VMA types.
1031 */
1032 nend = min(end, vma->vm_end);
1033 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1034 continue;
1035 if (nstart < vma->vm_start)
1036 nstart = vma->vm_start;
1037 /*
1038 * Now fault in a range of pages. populate_vma_page_range()
1039 * double checks the vma flags, so that it won't mlock pages
1040 * if the vma was already munlocked.
1041 */
1042 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1043 if (ret < 0) {
1044 if (ignore_errors) {
1045 ret = 0;
1046 continue; /* continue at next VMA */
1047 }
1048 break;
1049 }
1050 nend = nstart + ret * PAGE_SIZE;
1051 ret = 0;
1052 }
1053 if (locked)
1054 up_read(&mm->mmap_sem);
1055 return ret; /* 0 or negative error code */
1056}
1057
1058/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001059 * get_dump_page() - pin user page in memory while writing it to core dump
1060 * @addr: user address
1061 *
1062 * Returns struct page pointer of user page pinned for dump,
1063 * to be freed afterwards by page_cache_release() or put_page().
1064 *
1065 * Returns NULL on any kind of failure - a hole must then be inserted into
1066 * the corefile, to preserve alignment with its headers; and also returns
1067 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1068 * allowing a hole to be left in the corefile to save diskspace.
1069 *
1070 * Called without mmap_sem, but after all other threads have been killed.
1071 */
1072#ifdef CONFIG_ELF_CORE
1073struct page *get_dump_page(unsigned long addr)
1074{
1075 struct vm_area_struct *vma;
1076 struct page *page;
1077
1078 if (__get_user_pages(current, current->mm, addr, 1,
1079 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1080 NULL) < 1)
1081 return NULL;
1082 flush_cache_page(vma, addr, page_to_pfn(page));
1083 return page;
1084}
1085#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001086
1087/*
1088 * Generic RCU Fast GUP
1089 *
1090 * get_user_pages_fast attempts to pin user pages by walking the page
1091 * tables directly and avoids taking locks. Thus the walker needs to be
1092 * protected from page table pages being freed from under it, and should
1093 * block any THP splits.
1094 *
1095 * One way to achieve this is to have the walker disable interrupts, and
1096 * rely on IPIs from the TLB flushing code blocking before the page table
1097 * pages are freed. This is unsuitable for architectures that do not need
1098 * to broadcast an IPI when invalidating TLBs.
1099 *
1100 * Another way to achieve this is to batch up page table containing pages
1101 * belonging to more than one mm_user, then rcu_sched a callback to free those
1102 * pages. Disabling interrupts will allow the fast_gup walker to both block
1103 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1104 * (which is a relatively rare event). The code below adopts this strategy.
1105 *
1106 * Before activating this code, please be aware that the following assumptions
1107 * are currently made:
1108 *
1109 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1110 * pages containing page tables.
1111 *
Steve Capper2667f502014-10-09 15:29:14 -07001112 * *) ptes can be read atomically by the architecture.
1113 *
1114 * *) access_ok is sufficient to validate userspace address ranges.
1115 *
1116 * The last two assumptions can be relaxed by the addition of helper functions.
1117 *
1118 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1119 */
1120#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1121
1122#ifdef __HAVE_ARCH_PTE_SPECIAL
1123static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1124 int write, struct page **pages, int *nr)
1125{
1126 pte_t *ptep, *ptem;
1127 int ret = 0;
1128
1129 ptem = ptep = pte_offset_map(&pmd, addr);
1130 do {
1131 /*
1132 * In the line below we are assuming that the pte can be read
1133 * atomically. If this is not the case for your architecture,
1134 * please wrap this in a helper function!
1135 *
1136 * for an example see gup_get_pte in arch/x86/mm/gup.c
1137 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001138 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001139 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001140
1141 /*
1142 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001143 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001144 */
1145 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001146 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001147 goto pte_unmap;
1148
1149 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1150 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001151 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001152
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001153 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001154 goto pte_unmap;
1155
1156 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001157 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001158 goto pte_unmap;
1159 }
1160
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001161 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001162 pages[*nr] = page;
1163 (*nr)++;
1164
1165 } while (ptep++, addr += PAGE_SIZE, addr != end);
1166
1167 ret = 1;
1168
1169pte_unmap:
1170 pte_unmap(ptem);
1171 return ret;
1172}
1173#else
1174
1175/*
1176 * If we can't determine whether or not a pte is special, then fail immediately
1177 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1178 * to be special.
1179 *
1180 * For a futex to be placed on a THP tail page, get_futex_key requires a
1181 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1182 * useful to have gup_huge_pmd even if we can't operate on ptes.
1183 */
1184static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1185 int write, struct page **pages, int *nr)
1186{
1187 return 0;
1188}
1189#endif /* __HAVE_ARCH_PTE_SPECIAL */
1190
1191static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1192 unsigned long end, int write, struct page **pages, int *nr)
1193{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001194 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001195 int refs;
1196
1197 if (write && !pmd_write(orig))
1198 return 0;
1199
1200 refs = 0;
1201 head = pmd_page(orig);
1202 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001203 do {
1204 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1205 pages[*nr] = page;
1206 (*nr)++;
1207 page++;
1208 refs++;
1209 } while (addr += PAGE_SIZE, addr != end);
1210
1211 if (!page_cache_add_speculative(head, refs)) {
1212 *nr -= refs;
1213 return 0;
1214 }
1215
1216 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1217 *nr -= refs;
1218 while (refs--)
1219 put_page(head);
1220 return 0;
1221 }
1222
Steve Capper2667f502014-10-09 15:29:14 -07001223 return 1;
1224}
1225
1226static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1227 unsigned long end, int write, struct page **pages, int *nr)
1228{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001229 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001230 int refs;
1231
1232 if (write && !pud_write(orig))
1233 return 0;
1234
1235 refs = 0;
1236 head = pud_page(orig);
1237 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001238 do {
1239 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1240 pages[*nr] = page;
1241 (*nr)++;
1242 page++;
1243 refs++;
1244 } while (addr += PAGE_SIZE, addr != end);
1245
1246 if (!page_cache_add_speculative(head, refs)) {
1247 *nr -= refs;
1248 return 0;
1249 }
1250
1251 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1252 *nr -= refs;
1253 while (refs--)
1254 put_page(head);
1255 return 0;
1256 }
1257
Steve Capper2667f502014-10-09 15:29:14 -07001258 return 1;
1259}
1260
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301261static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1262 unsigned long end, int write,
1263 struct page **pages, int *nr)
1264{
1265 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001266 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301267
1268 if (write && !pgd_write(orig))
1269 return 0;
1270
1271 refs = 0;
1272 head = pgd_page(orig);
1273 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301274 do {
1275 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1276 pages[*nr] = page;
1277 (*nr)++;
1278 page++;
1279 refs++;
1280 } while (addr += PAGE_SIZE, addr != end);
1281
1282 if (!page_cache_add_speculative(head, refs)) {
1283 *nr -= refs;
1284 return 0;
1285 }
1286
1287 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1288 *nr -= refs;
1289 while (refs--)
1290 put_page(head);
1291 return 0;
1292 }
1293
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301294 return 1;
1295}
1296
Steve Capper2667f502014-10-09 15:29:14 -07001297static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1298 int write, struct page **pages, int *nr)
1299{
1300 unsigned long next;
1301 pmd_t *pmdp;
1302
1303 pmdp = pmd_offset(&pud, addr);
1304 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001305 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001306
1307 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001308 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001309 return 0;
1310
1311 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1312 /*
1313 * NUMA hinting faults need to be handled in the GUP
1314 * slowpath for accounting purposes and so that they
1315 * can be serialised against THP migration.
1316 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001317 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001318 return 0;
1319
1320 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1321 pages, nr))
1322 return 0;
1323
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301324 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1325 /*
1326 * architecture have different format for hugetlbfs
1327 * pmd format and THP pmd format
1328 */
1329 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1330 PMD_SHIFT, next, write, pages, nr))
1331 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001332 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1333 return 0;
1334 } while (pmdp++, addr = next, addr != end);
1335
1336 return 1;
1337}
1338
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301339static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1340 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001341{
1342 unsigned long next;
1343 pud_t *pudp;
1344
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301345 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001346 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001347 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001348
1349 next = pud_addr_end(addr, end);
1350 if (pud_none(pud))
1351 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301352 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001353 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301354 pages, nr))
1355 return 0;
1356 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1357 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1358 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001359 return 0;
1360 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1361 return 0;
1362 } while (pudp++, addr = next, addr != end);
1363
1364 return 1;
1365}
1366
1367/*
1368 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1369 * the regular GUP. It will only return non-negative values.
1370 */
1371int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1372 struct page **pages)
1373{
1374 struct mm_struct *mm = current->mm;
1375 unsigned long addr, len, end;
1376 unsigned long next, flags;
1377 pgd_t *pgdp;
1378 int nr = 0;
1379
1380 start &= PAGE_MASK;
1381 addr = start;
1382 len = (unsigned long) nr_pages << PAGE_SHIFT;
1383 end = start + len;
1384
1385 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1386 start, len)))
1387 return 0;
1388
1389 /*
1390 * Disable interrupts. We use the nested form as we can already have
1391 * interrupts disabled by get_futex_key.
1392 *
1393 * With interrupts disabled, we block page table pages from being
1394 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1395 * for more details.
1396 *
1397 * We do not adopt an rcu_read_lock(.) here as we also want to
1398 * block IPIs that come from THPs splitting.
1399 */
1400
1401 local_irq_save(flags);
1402 pgdp = pgd_offset(mm, addr);
1403 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001404 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301405
Steve Capper2667f502014-10-09 15:29:14 -07001406 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301407 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001408 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301409 if (unlikely(pgd_huge(pgd))) {
1410 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1411 pages, &nr))
1412 break;
1413 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1414 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1415 PGDIR_SHIFT, next, write, pages, &nr))
1416 break;
1417 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001418 break;
1419 } while (pgdp++, addr = next, addr != end);
1420 local_irq_restore(flags);
1421
1422 return nr;
1423}
1424
1425/**
1426 * get_user_pages_fast() - pin user pages in memory
1427 * @start: starting user address
1428 * @nr_pages: number of pages from start to pin
1429 * @write: whether pages will be written to
1430 * @pages: array that receives pointers to the pages pinned.
1431 * Should be at least nr_pages long.
1432 *
1433 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1434 * If not successful, it will fall back to taking the lock and
1435 * calling get_user_pages().
1436 *
1437 * Returns number of pages pinned. This may be fewer than the number
1438 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1439 * were pinned, returns -errno.
1440 */
1441int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1442 struct page **pages)
1443{
1444 struct mm_struct *mm = current->mm;
1445 int nr, ret;
1446
1447 start &= PAGE_MASK;
1448 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1449 ret = nr;
1450
1451 if (nr < nr_pages) {
1452 /* Try to get the remaining pages with get_user_pages */
1453 start += nr << PAGE_SHIFT;
1454 pages += nr;
1455
Andrea Arcangelia7b78072015-02-11 15:27:23 -08001456 ret = get_user_pages_unlocked(current, mm, start,
1457 nr_pages - nr, write, 0, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001458
1459 /* Have to be a bit careful with return values */
1460 if (nr > 0) {
1461 if (ret < 0)
1462 ret = nr;
1463 else
1464 ret += nr;
1465 }
1466 }
1467
1468 return ret;
1469}
1470
1471#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */