blob: 6c3b4e822946f533529852e8bf06abc007918e62 [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
Linus Torvalds19be0ea2016-10-13 13:07:36 -070063/*
64 * FOLL_FORCE can write to even unwritable pte's, but only
65 * after we've gone through a COW cycle and they are dirty.
66 */
67static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
68{
69 return pte_write(pte) ||
70 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
71}
72
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070073static struct page *follow_page_pte(struct vm_area_struct *vma,
74 unsigned long address, pmd_t *pmd, unsigned int flags)
75{
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070076 struct mm_struct *mm = vma->vm_mm;
Dan Williams3565fce2016-01-15 16:56:55 -080077 struct dev_pagemap *pgmap = NULL;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070078 struct page *page;
79 spinlock_t *ptl;
80 pte_t *ptep, pte;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070081
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070082retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070083 if (unlikely(pmd_bad(*pmd)))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070084 return no_page_table(vma, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070085
86 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070087 pte = *ptep;
88 if (!pte_present(pte)) {
89 swp_entry_t entry;
90 /*
91 * KSM's break_ksm() relies upon recognizing a ksm page
92 * even while it is being migrated, so for that case we
93 * need migration_entry_wait().
94 */
95 if (likely(!(flags & FOLL_MIGRATION)))
96 goto no_page;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -080097 if (pte_none(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070098 goto no_page;
99 entry = pte_to_swp_entry(pte);
100 if (!is_migration_entry(entry))
101 goto no_page;
102 pte_unmap_unlock(ptep, ptl);
103 migration_entry_wait(mm, pmd, address);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700104 goto retry;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700105 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800106 if ((flags & FOLL_NUMA) && pte_protnone(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700107 goto no_page;
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700108 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700109 pte_unmap_unlock(ptep, ptl);
110 return NULL;
111 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700112
113 page = vm_normal_page(vma, address, pte);
Dan Williams3565fce2016-01-15 16:56:55 -0800114 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
115 /*
116 * Only return device mapping pages in the FOLL_GET case since
117 * they are only valid while holding the pgmap reference.
118 */
119 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
120 if (pgmap)
121 page = pte_page(pte);
122 else
123 goto no_page;
124 } else if (unlikely(!page)) {
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700125 if (flags & FOLL_DUMP) {
126 /* Avoid special (like zero) pages in core dumps */
127 page = ERR_PTR(-EFAULT);
128 goto out;
129 }
130
131 if (is_zero_pfn(pte_pfn(pte))) {
132 page = pte_page(pte);
133 } else {
134 int ret;
135
136 ret = follow_pfn_pte(vma, address, ptep, flags);
137 page = ERR_PTR(ret);
138 goto out;
139 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700140 }
141
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800142 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
143 int ret;
144 get_page(page);
145 pte_unmap_unlock(ptep, ptl);
146 lock_page(page);
147 ret = split_huge_page(page);
148 unlock_page(page);
149 put_page(page);
150 if (ret)
151 return ERR_PTR(ret);
152 goto retry;
153 }
154
Dan Williams3565fce2016-01-15 16:56:55 -0800155 if (flags & FOLL_GET) {
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800156 get_page(page);
Dan Williams3565fce2016-01-15 16:56:55 -0800157
158 /* drop the pgmap reference now that we hold the page */
159 if (pgmap) {
160 put_dev_pagemap(pgmap);
161 pgmap = NULL;
162 }
163 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700164 if (flags & FOLL_TOUCH) {
165 if ((flags & FOLL_WRITE) &&
166 !pte_dirty(pte) && !PageDirty(page))
167 set_page_dirty(page);
168 /*
169 * pte_mkyoung() would be more correct here, but atomic care
170 * is needed to avoid losing the dirty bit: it is easier to use
171 * mark_page_accessed().
172 */
173 mark_page_accessed(page);
174 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800175 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -0800176 /* Do not mlock pte-mapped THP */
177 if (PageTransCompound(page))
178 goto out;
179
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700180 /*
181 * The preliminary mapping check is mainly to avoid the
182 * pointless overhead of lock_page on the ZERO_PAGE
183 * which might bounce very badly if there is contention.
184 *
185 * If the page is already locked, we don't need to
186 * handle it now - vmscan will handle it later if and
187 * when it attempts to reclaim the page.
188 */
189 if (page->mapping && trylock_page(page)) {
190 lru_add_drain(); /* push cached pages to LRU */
191 /*
192 * Because we lock page here, and migration is
193 * blocked by the pte's page reference, and we
194 * know the page is still mapped, we don't even
195 * need to check for file-cache page truncation.
196 */
197 mlock_vma_page(page);
198 unlock_page(page);
199 }
200 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700201out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700202 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700203 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700204no_page:
205 pte_unmap_unlock(ptep, ptl);
206 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700207 return NULL;
208 return no_page_table(vma, flags);
209}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700210
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700211/**
212 * follow_page_mask - look up a page descriptor from a user-virtual address
213 * @vma: vm_area_struct mapping @address
214 * @address: virtual address to look up
215 * @flags: flags modifying lookup behaviour
216 * @page_mask: on output, *page_mask is set according to the size of the page
217 *
218 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
219 *
220 * Returns the mapped (struct page *), %NULL if no mapping exists, or
221 * an error pointer if there is a mapping to something not represented
222 * by a page descriptor (see also vm_normal_page()).
223 */
224struct page *follow_page_mask(struct vm_area_struct *vma,
225 unsigned long address, unsigned int flags,
226 unsigned int *page_mask)
227{
228 pgd_t *pgd;
229 pud_t *pud;
230 pmd_t *pmd;
231 spinlock_t *ptl;
232 struct page *page;
233 struct mm_struct *mm = vma->vm_mm;
234
235 *page_mask = 0;
236
237 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
238 if (!IS_ERR(page)) {
239 BUG_ON(flags & FOLL_GET);
240 return page;
241 }
242
243 pgd = pgd_offset(mm, address);
244 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
245 return no_page_table(vma, flags);
246
247 pud = pud_offset(pgd, address);
248 if (pud_none(*pud))
249 return no_page_table(vma, flags);
250 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800251 page = follow_huge_pud(mm, address, pud, flags);
252 if (page)
253 return page;
254 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700255 }
256 if (unlikely(pud_bad(*pud)))
257 return no_page_table(vma, flags);
258
259 pmd = pmd_offset(pud, address);
260 if (pmd_none(*pmd))
261 return no_page_table(vma, flags);
262 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800263 page = follow_huge_pmd(mm, address, pmd, flags);
264 if (page)
265 return page;
266 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700267 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800268 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700269 return no_page_table(vma, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800270 if (pmd_devmap(*pmd)) {
271 ptl = pmd_lock(mm, pmd);
272 page = follow_devmap_pmd(vma, address, pmd, flags);
273 spin_unlock(ptl);
274 if (page)
275 return page;
276 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800277 if (likely(!pmd_trans_huge(*pmd)))
278 return follow_page_pte(vma, address, pmd, flags);
279
280 ptl = pmd_lock(mm, pmd);
281 if (unlikely(!pmd_trans_huge(*pmd))) {
282 spin_unlock(ptl);
283 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700284 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800285 if (flags & FOLL_SPLIT) {
286 int ret;
287 page = pmd_page(*pmd);
288 if (is_huge_zero_page(page)) {
289 spin_unlock(ptl);
290 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800291 split_huge_pmd(vma, pmd, address);
Naoya Horiguchi337d9ab2016-07-26 15:24:03 -0700292 if (pmd_trans_unstable(pmd))
293 ret = -EBUSY;
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800294 } else {
295 get_page(page);
296 spin_unlock(ptl);
297 lock_page(page);
298 ret = split_huge_page(page);
299 unlock_page(page);
300 put_page(page);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -0700301 if (pmd_none(*pmd))
302 return no_page_table(vma, flags);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800303 }
304
305 return ret ? ERR_PTR(ret) :
306 follow_page_pte(vma, address, pmd, flags);
307 }
308
309 page = follow_trans_huge_pmd(vma, address, pmd, flags);
310 spin_unlock(ptl);
311 *page_mask = HPAGE_PMD_NR - 1;
312 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700313}
314
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700315static int get_gate_page(struct mm_struct *mm, unsigned long address,
316 unsigned int gup_flags, struct vm_area_struct **vma,
317 struct page **page)
318{
319 pgd_t *pgd;
320 pud_t *pud;
321 pmd_t *pmd;
322 pte_t *pte;
323 int ret = -EFAULT;
324
325 /* user gate pages are read-only */
326 if (gup_flags & FOLL_WRITE)
327 return -EFAULT;
328 if (address > TASK_SIZE)
329 pgd = pgd_offset_k(address);
330 else
331 pgd = pgd_offset_gate(mm, address);
332 BUG_ON(pgd_none(*pgd));
333 pud = pud_offset(pgd, address);
334 BUG_ON(pud_none(*pud));
335 pmd = pmd_offset(pud, address);
336 if (pmd_none(*pmd))
337 return -EFAULT;
338 VM_BUG_ON(pmd_trans_huge(*pmd));
339 pte = pte_offset_map(pmd, address);
340 if (pte_none(*pte))
341 goto unmap;
342 *vma = get_gate_vma(mm);
343 if (!page)
344 goto out;
345 *page = vm_normal_page(*vma, address, *pte);
346 if (!*page) {
347 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
348 goto unmap;
349 *page = pte_page(*pte);
350 }
351 get_page(*page);
352out:
353 ret = 0;
354unmap:
355 pte_unmap(pte);
356 return ret;
357}
358
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700359/*
360 * mmap_sem must be held on entry. If @nonblocking != NULL and
361 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
362 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
363 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700364static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
365 unsigned long address, unsigned int *flags, int *nonblocking)
366{
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700367 unsigned int fault_flags = 0;
368 int ret;
369
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800370 /* mlock all present pages, but do not fault in new pages */
371 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
372 return -ENOENT;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700373 if (*flags & FOLL_WRITE)
374 fault_flags |= FAULT_FLAG_WRITE;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800375 if (*flags & FOLL_REMOTE)
376 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700377 if (nonblocking)
378 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
379 if (*flags & FOLL_NOWAIT)
380 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700381 if (*flags & FOLL_TRIED) {
382 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
383 fault_flags |= FAULT_FLAG_TRIED;
384 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700385
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700386 ret = handle_mm_fault(vma, address, fault_flags);
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700387 if (ret & VM_FAULT_ERROR) {
388 if (ret & VM_FAULT_OOM)
389 return -ENOMEM;
390 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
391 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800392 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700393 return -EFAULT;
394 BUG();
395 }
396
397 if (tsk) {
398 if (ret & VM_FAULT_MAJOR)
399 tsk->maj_flt++;
400 else
401 tsk->min_flt++;
402 }
403
404 if (ret & VM_FAULT_RETRY) {
405 if (nonblocking)
406 *nonblocking = 0;
407 return -EBUSY;
408 }
409
410 /*
411 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
412 * necessary, even if maybe_mkwrite decided not to set pte_write. We
413 * can thus safely do subsequent page lookups as if they were reads.
414 * But only do so when looping for pte_write is futile: in some cases
415 * userspace may also be wanting to write to the gotten user page,
416 * which a read fault here might prevent (a readonly page might get
417 * reCOWed by userspace write).
418 */
419 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700420 *flags |= FOLL_COW;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700421 return 0;
422}
423
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700424static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
425{
426 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800427 int write = (gup_flags & FOLL_WRITE);
428 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700429
430 if (vm_flags & (VM_IO | VM_PFNMAP))
431 return -EFAULT;
432
Dave Hansen1b2ee122016-02-12 13:02:21 -0800433 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700434 if (!(vm_flags & VM_WRITE)) {
435 if (!(gup_flags & FOLL_FORCE))
436 return -EFAULT;
437 /*
438 * We used to let the write,force case do COW in a
439 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
440 * set a breakpoint in a read-only mapping of an
441 * executable, without corrupting the file (yet only
442 * when that file had been opened for writing!).
443 * Anon pages in shared mappings are surprising: now
444 * just reject it.
445 */
Hugh Dickins46435362016-01-30 18:03:16 -0800446 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700447 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700448 }
449 } else if (!(vm_flags & VM_READ)) {
450 if (!(gup_flags & FOLL_FORCE))
451 return -EFAULT;
452 /*
453 * Is there actually any vma we can reach here which does not
454 * have VM_MAYREAD set?
455 */
456 if (!(vm_flags & VM_MAYREAD))
457 return -EFAULT;
458 }
Dave Hansend61172b2016-02-12 13:02:24 -0800459 /*
460 * gups are always data accesses, not instruction
461 * fetches, so execute=false here
462 */
463 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800464 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700465 return 0;
466}
467
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700468/**
469 * __get_user_pages() - pin user pages in memory
470 * @tsk: task_struct of target task
471 * @mm: mm_struct of target mm
472 * @start: starting user address
473 * @nr_pages: number of pages from start to pin
474 * @gup_flags: flags modifying pin behaviour
475 * @pages: array that receives pointers to the pages pinned.
476 * Should be at least nr_pages long. Or NULL, if caller
477 * only intends to ensure the pages are faulted in.
478 * @vmas: array of pointers to vmas corresponding to each page.
479 * Or NULL if the caller does not require them.
480 * @nonblocking: whether waiting for disk IO or mmap_sem contention
481 *
482 * Returns number of pages pinned. This may be fewer than the number
483 * requested. If nr_pages is 0 or negative, returns 0. If no pages
484 * were pinned, returns -errno. Each page returned must be released
485 * with a put_page() call when it is finished with. vmas will only
486 * remain valid while mmap_sem is held.
487 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700488 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700489 *
490 * __get_user_pages walks a process's page tables and takes a reference to
491 * each struct page that each user address corresponds to at a given
492 * instant. That is, it takes the page that would be accessed if a user
493 * thread accesses the given user virtual address at that instant.
494 *
495 * This does not guarantee that the page exists in the user mappings when
496 * __get_user_pages returns, and there may even be a completely different
497 * page there in some cases (eg. if mmapped pagecache has been invalidated
498 * and subsequently re faulted). However it does guarantee that the page
499 * won't be freed completely. And mostly callers simply care that the page
500 * contains data that was valid *at some point in time*. Typically, an IO
501 * or similar operation cannot guarantee anything stronger anyway because
502 * locks can't be held over the syscall boundary.
503 *
504 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
505 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
506 * appropriate) must be called after the page is finished with, and
507 * before put_page is called.
508 *
509 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
510 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700511 * *@nonblocking will be set to 0. Further, if @gup_flags does not
512 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
513 * this case.
514 *
515 * A caller using such a combination of @nonblocking and @gup_flags
516 * must therefore hold the mmap_sem for reading only, and recognize
517 * when it's been released. Otherwise, it must be held for either
518 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700519 *
520 * In most cases, get_user_pages or get_user_pages_fast should be used
521 * instead of __get_user_pages. __get_user_pages should be used only if
522 * you need some special @gup_flags.
523 */
Lorenzo Stoakes0d731752016-10-24 10:57:25 +0100524static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700525 unsigned long start, unsigned long nr_pages,
526 unsigned int gup_flags, struct page **pages,
527 struct vm_area_struct **vmas, int *nonblocking)
528{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700529 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700530 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700531 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700532
533 if (!nr_pages)
534 return 0;
535
536 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
537
538 /*
539 * If FOLL_FORCE is set then do not force a full fault as the hinting
540 * fault information is unrelated to the reference behaviour of a task
541 * using the address space
542 */
543 if (!(gup_flags & FOLL_FORCE))
544 gup_flags |= FOLL_NUMA;
545
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700546 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700547 struct page *page;
548 unsigned int foll_flags = gup_flags;
549 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700550
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700551 /* first iteration or cross vma bound */
552 if (!vma || start >= vma->vm_end) {
553 vma = find_extend_vma(mm, start);
554 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700555 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700556 ret = get_gate_page(mm, start & PAGE_MASK,
557 gup_flags, &vma,
558 pages ? &pages[i] : NULL);
559 if (ret)
560 return i ? : ret;
561 page_mask = 0;
562 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700563 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700564
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700565 if (!vma || check_vma_flags(vma, gup_flags))
566 return i ? : -EFAULT;
567 if (is_vm_hugetlb_page(vma)) {
568 i = follow_hugetlb_page(mm, vma, pages, vmas,
569 &start, &nr_pages, i,
570 gup_flags);
571 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700572 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700573 }
574retry:
575 /*
576 * If we have a pending SIGKILL, don't keep faulting pages and
577 * potentially allocating memory.
578 */
579 if (unlikely(fatal_signal_pending(current)))
580 return i ? i : -ERESTARTSYS;
581 cond_resched();
582 page = follow_page_mask(vma, start, foll_flags, &page_mask);
583 if (!page) {
584 int ret;
585 ret = faultin_page(tsk, vma, start, &foll_flags,
586 nonblocking);
587 switch (ret) {
588 case 0:
589 goto retry;
590 case -EFAULT:
591 case -ENOMEM:
592 case -EHWPOISON:
593 return i ? i : ret;
594 case -EBUSY:
595 return i;
596 case -ENOENT:
597 goto next_page;
598 }
599 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700600 } else if (PTR_ERR(page) == -EEXIST) {
601 /*
602 * Proper page table entry exists, but no corresponding
603 * struct page.
604 */
605 goto next_page;
606 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700607 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700608 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700609 if (pages) {
610 pages[i] = page;
611 flush_anon_page(vma, page, start);
612 flush_dcache_page(page);
613 page_mask = 0;
614 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700615next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700616 if (vmas) {
617 vmas[i] = vma;
618 page_mask = 0;
619 }
620 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
621 if (page_increm > nr_pages)
622 page_increm = nr_pages;
623 i += page_increm;
624 start += page_increm * PAGE_SIZE;
625 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700626 } while (nr_pages);
627 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700628}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700629
Dave Hansend4925e02016-02-12 13:02:16 -0800630bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
631{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800632 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
633 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800634 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800635
636 if (!(vm_flags & vma->vm_flags))
637 return false;
638
Dave Hansen33a709b2016-02-12 13:02:19 -0800639 /*
640 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800641 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800642 *
643 * gup always represents data access, not instruction
644 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800645 */
Dave Hansend61172b2016-02-12 13:02:24 -0800646 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800647 return false;
648
Dave Hansend4925e02016-02-12 13:02:16 -0800649 return true;
650}
651
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700652/*
653 * fixup_user_fault() - manually resolve a user page fault
654 * @tsk: the task_struct to use for page fault accounting, or
655 * NULL if faults are not to be recorded.
656 * @mm: mm_struct of target mm
657 * @address: user address
658 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800659 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
660 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700661 *
662 * This is meant to be called in the specific scenario where for locking reasons
663 * we try to access user memory in atomic context (within a pagefault_disable()
664 * section), this returns -EFAULT, and we want to resolve the user fault before
665 * trying again.
666 *
667 * Typically this is meant to be used by the futex code.
668 *
669 * The main difference with get_user_pages() is that this function will
670 * unconditionally call handle_mm_fault() which will in turn perform all the
671 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800672 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700673 *
674 * This is important for some architectures where those bits also gate the
675 * access permission to the page because they are maintained in software. On
676 * such architectures, gup() will not be enough to make a subsequent access
677 * succeed.
678 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800679 * This function will not return with an unlocked mmap_sem. So it has not the
680 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700681 */
682int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800683 unsigned long address, unsigned int fault_flags,
684 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700685{
686 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800687 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700688
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800689 if (unlocked)
690 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
691
692retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700693 vma = find_extend_vma(mm, address);
694 if (!vma || address < vma->vm_start)
695 return -EFAULT;
696
Dave Hansend4925e02016-02-12 13:02:16 -0800697 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700698 return -EFAULT;
699
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700700 ret = handle_mm_fault(vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800701 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700702 if (ret & VM_FAULT_ERROR) {
703 if (ret & VM_FAULT_OOM)
704 return -ENOMEM;
705 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
706 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800707 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700708 return -EFAULT;
709 BUG();
710 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800711
712 if (ret & VM_FAULT_RETRY) {
713 down_read(&mm->mmap_sem);
714 if (!(fault_flags & FAULT_FLAG_TRIED)) {
715 *unlocked = true;
716 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
717 fault_flags |= FAULT_FLAG_TRIED;
718 goto retry;
719 }
720 }
721
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700722 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800723 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700724 tsk->maj_flt++;
725 else
726 tsk->min_flt++;
727 }
728 return 0;
729}
Paolo Bonziniadd6a0c2016-06-07 17:51:18 +0200730EXPORT_SYMBOL_GPL(fixup_user_fault);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700731
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800732static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
733 struct mm_struct *mm,
734 unsigned long start,
735 unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800736 struct page **pages,
737 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800738 int *locked, bool notify_drop,
739 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800740{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800741 long ret, pages_done;
742 bool lock_dropped;
743
744 if (locked) {
745 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
746 BUG_ON(vmas);
747 /* check caller initialized locked */
748 BUG_ON(*locked != 1);
749 }
750
751 if (pages)
752 flags |= FOLL_GET;
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800753
754 pages_done = 0;
755 lock_dropped = false;
756 for (;;) {
757 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
758 vmas, locked);
759 if (!locked)
760 /* VM_FAULT_RETRY couldn't trigger, bypass */
761 return ret;
762
763 /* VM_FAULT_RETRY cannot return errors */
764 if (!*locked) {
765 BUG_ON(ret < 0);
766 BUG_ON(ret >= nr_pages);
767 }
768
769 if (!pages)
770 /* If it's a prefault don't insist harder */
771 return ret;
772
773 if (ret > 0) {
774 nr_pages -= ret;
775 pages_done += ret;
776 if (!nr_pages)
777 break;
778 }
779 if (*locked) {
780 /* VM_FAULT_RETRY didn't trigger */
781 if (!pages_done)
782 pages_done = ret;
783 break;
784 }
785 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
786 pages += ret;
787 start += ret << PAGE_SHIFT;
788
789 /*
790 * Repeat on the address that fired VM_FAULT_RETRY
791 * without FAULT_FLAG_ALLOW_RETRY but with
792 * FAULT_FLAG_TRIED.
793 */
794 *locked = 1;
795 lock_dropped = true;
796 down_read(&mm->mmap_sem);
797 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
798 pages, NULL, NULL);
799 if (ret != 1) {
800 BUG_ON(ret > 1);
801 if (!pages_done)
802 pages_done = ret;
803 break;
804 }
805 nr_pages--;
806 pages_done++;
807 if (!nr_pages)
808 break;
809 pages++;
810 start += PAGE_SIZE;
811 }
812 if (notify_drop && lock_dropped && *locked) {
813 /*
814 * We must let the caller know we temporarily dropped the lock
815 * and so the critical section protected by it was lost.
816 */
817 up_read(&mm->mmap_sem);
818 *locked = 0;
819 }
820 return pages_done;
821}
822
823/*
824 * We can leverage the VM_FAULT_RETRY functionality in the page fault
825 * paths better by using either get_user_pages_locked() or
826 * get_user_pages_unlocked().
827 *
828 * get_user_pages_locked() is suitable to replace the form:
829 *
830 * down_read(&mm->mmap_sem);
831 * do_something()
832 * get_user_pages(tsk, mm, ..., pages, NULL);
833 * up_read(&mm->mmap_sem);
834 *
835 * to:
836 *
837 * int locked = 1;
838 * down_read(&mm->mmap_sem);
839 * do_something()
840 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
841 * if (locked)
842 * up_read(&mm->mmap_sem);
843 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200844long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100845 unsigned int gup_flags, struct page **pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800846 int *locked)
847{
Dave Hansencde70142016-02-12 13:01:55 -0800848 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100849 pages, NULL, locked, true,
850 gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800851}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200852EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800853
854/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800855 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
856 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
857 *
858 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
859 * caller if required (just like with __get_user_pages). "FOLL_GET",
860 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
861 * according to the parameters "pages", "write", "force"
862 * respectively.
863 */
864__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
865 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesd4944b02016-10-13 01:20:12 +0100866 struct page **pages, unsigned int gup_flags)
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800867{
868 long ret;
869 int locked = 1;
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100870
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800871 down_read(&mm->mmap_sem);
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100872 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
873 &locked, false, gup_flags);
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800874 if (locked)
875 up_read(&mm->mmap_sem);
876 return ret;
877}
878EXPORT_SYMBOL(__get_user_pages_unlocked);
879
880/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800881 * get_user_pages_unlocked() is suitable to replace the form:
882 *
883 * down_read(&mm->mmap_sem);
884 * get_user_pages(tsk, mm, ..., pages, NULL);
885 * up_read(&mm->mmap_sem);
886 *
887 * with:
888 *
889 * get_user_pages_unlocked(tsk, mm, ..., pages);
890 *
891 * It is functionally equivalent to get_user_pages_fast so
892 * get_user_pages_fast should be used instead, if the two parameters
893 * "tsk" and "mm" are respectively equal to current and current->mm,
894 * or if "force" shall be set to 1 (get_user_pages_fast misses the
895 * "force" parameter).
896 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200897long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100898 struct page **pages, unsigned int gup_flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800899{
Dave Hansencde70142016-02-12 13:01:55 -0800900 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100901 pages, gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800902}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200903EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800904
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700905/*
Dave Hansen1e987792016-02-12 13:01:54 -0800906 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700907 * @tsk: the task_struct to use for page fault accounting, or
908 * NULL if faults are not to be recorded.
909 * @mm: mm_struct of target mm
910 * @start: starting user address
911 * @nr_pages: number of pages from start to pin
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100912 * @gup_flags: flags modifying lookup behaviour
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700913 * @pages: array that receives pointers to the pages pinned.
914 * Should be at least nr_pages long. Or NULL, if caller
915 * only intends to ensure the pages are faulted in.
916 * @vmas: array of pointers to vmas corresponding to each page.
917 * Or NULL if the caller does not require them.
918 *
919 * Returns number of pages pinned. This may be fewer than the number
920 * requested. If nr_pages is 0 or negative, returns 0. If no pages
921 * were pinned, returns -errno. Each page returned must be released
922 * with a put_page() call when it is finished with. vmas will only
923 * remain valid while mmap_sem is held.
924 *
925 * Must be called with mmap_sem held for read or write.
926 *
927 * get_user_pages walks a process's page tables and takes a reference to
928 * each struct page that each user address corresponds to at a given
929 * instant. That is, it takes the page that would be accessed if a user
930 * thread accesses the given user virtual address at that instant.
931 *
932 * This does not guarantee that the page exists in the user mappings when
933 * get_user_pages returns, and there may even be a completely different
934 * page there in some cases (eg. if mmapped pagecache has been invalidated
935 * and subsequently re faulted). However it does guarantee that the page
936 * won't be freed completely. And mostly callers simply care that the page
937 * contains data that was valid *at some point in time*. Typically, an IO
938 * or similar operation cannot guarantee anything stronger anyway because
939 * locks can't be held over the syscall boundary.
940 *
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100941 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
942 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
943 * be called after the page is finished with, and before put_page is called.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700944 *
945 * get_user_pages is typically used for fewer-copy IO operations, to get a
946 * handle on the memory by some means other than accesses via the user virtual
947 * addresses. The pages may be submitted for DMA to devices or accessed via
948 * their kernel linear mapping (via the kmap APIs). Care should be taken to
949 * use the correct cache flushing APIs.
950 *
951 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800952 *
953 * get_user_pages should be phased out in favor of
954 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
955 * should use get_user_pages because it cannot pass
956 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700957 */
Dave Hansen1e987792016-02-12 13:01:54 -0800958long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
959 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100960 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800961 struct vm_area_struct **vmas)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700962{
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100963 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100964 NULL, false,
965 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
Dave Hansen1e987792016-02-12 13:01:54 -0800966}
967EXPORT_SYMBOL(get_user_pages_remote);
968
969/*
Dave Hansend4edcf02016-02-12 13:01:56 -0800970 * This is the same as get_user_pages_remote(), just with a
971 * less-flexible calling convention where we assume that the task
972 * and mm being operated on are the current task's. We also
973 * obviously don't pass FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -0800974 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200975long get_user_pages(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +0100976 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800977 struct vm_area_struct **vmas)
978{
Dave Hansencde70142016-02-12 13:01:55 -0800979 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +0100980 pages, vmas, NULL, false,
981 gup_flags | FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700982}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200983EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700984
Dan Williamsb29ea3c2018-02-23 14:05:49 -0800985#ifdef CONFIG_FS_DAX
986/*
987 * This is the same as get_user_pages() in that it assumes we are
988 * operating on the current task's mm, but it goes further to validate
989 * that the vmas associated with the address range are suitable for
990 * longterm elevated page reference counts. For example, filesystem-dax
991 * mappings are subject to the lifetime enforced by the filesystem and
992 * we need guarantees that longterm users like RDMA and V4L2 only
993 * establish mappings that have a kernel enforced revocation mechanism.
994 *
995 * "longterm" == userspace controlled elevated page count lifetime.
996 * Contrast this to iov_iter_get_pages() usages which are transient.
997 */
998long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
999 unsigned int gup_flags, struct page **pages,
1000 struct vm_area_struct **vmas_arg)
1001{
1002 struct vm_area_struct **vmas = vmas_arg;
1003 struct vm_area_struct *vma_prev = NULL;
1004 long rc, i;
1005
1006 if (!pages)
1007 return -EINVAL;
1008
1009 if (!vmas) {
1010 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1011 GFP_KERNEL);
1012 if (!vmas)
1013 return -ENOMEM;
1014 }
1015
1016 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1017
1018 for (i = 0; i < rc; i++) {
1019 struct vm_area_struct *vma = vmas[i];
1020
1021 if (vma == vma_prev)
1022 continue;
1023
1024 vma_prev = vma;
1025
1026 if (vma_is_fsdax(vma))
1027 break;
1028 }
1029
1030 /*
1031 * Either get_user_pages() failed, or the vma validation
1032 * succeeded, in either case we don't need to put_page() before
1033 * returning.
1034 */
1035 if (i >= rc)
1036 goto out;
1037
1038 for (i = 0; i < rc; i++)
1039 put_page(pages[i]);
1040 rc = -EOPNOTSUPP;
1041out:
1042 if (vmas != vmas_arg)
1043 kfree(vmas);
1044 return rc;
1045}
1046EXPORT_SYMBOL(get_user_pages_longterm);
1047#endif /* CONFIG_FS_DAX */
1048
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001049/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001050 * populate_vma_page_range() - populate a range of pages in the vma.
1051 * @vma: target vma
1052 * @start: start address
1053 * @end: end address
1054 * @nonblocking:
1055 *
1056 * This takes care of mlocking the pages too if VM_LOCKED is set.
1057 *
1058 * return 0 on success, negative error code on error.
1059 *
1060 * vma->vm_mm->mmap_sem must be held.
1061 *
1062 * If @nonblocking is NULL, it may be held for read or write and will
1063 * be unperturbed.
1064 *
1065 * If @nonblocking is non-NULL, it must held for read only and may be
1066 * released. If it's released, *@nonblocking will be set to 0.
1067 */
1068long populate_vma_page_range(struct vm_area_struct *vma,
1069 unsigned long start, unsigned long end, int *nonblocking)
1070{
1071 struct mm_struct *mm = vma->vm_mm;
1072 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1073 int gup_flags;
1074
1075 VM_BUG_ON(start & ~PAGE_MASK);
1076 VM_BUG_ON(end & ~PAGE_MASK);
1077 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1078 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1079 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1080
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001081 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1082 if (vma->vm_flags & VM_LOCKONFAULT)
1083 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001084 /*
1085 * We want to touch writable mappings with a write fault in order
1086 * to break COW, except for shared mappings because these don't COW
1087 * and we would not want to dirty them for nothing.
1088 */
1089 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1090 gup_flags |= FOLL_WRITE;
1091
1092 /*
1093 * We want mlock to succeed for regions that have any permissions
1094 * other than PROT_NONE.
1095 */
1096 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1097 gup_flags |= FOLL_FORCE;
1098
1099 /*
1100 * We made sure addr is within a VMA, so the following will
1101 * not result in a stack expansion that recurses back here.
1102 */
1103 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1104 NULL, NULL, nonblocking);
1105}
1106
1107/*
1108 * __mm_populate - populate and/or mlock pages within a range of address space.
1109 *
1110 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1111 * flags. VMAs must be already marked with the desired vm_flags, and
1112 * mmap_sem must not be held.
1113 */
1114int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1115{
1116 struct mm_struct *mm = current->mm;
1117 unsigned long end, nstart, nend;
1118 struct vm_area_struct *vma = NULL;
1119 int locked = 0;
1120 long ret = 0;
1121
1122 VM_BUG_ON(start & ~PAGE_MASK);
1123 VM_BUG_ON(len != PAGE_ALIGN(len));
1124 end = start + len;
1125
1126 for (nstart = start; nstart < end; nstart = nend) {
1127 /*
1128 * We want to fault in pages for [nstart; end) address range.
1129 * Find first corresponding VMA.
1130 */
1131 if (!locked) {
1132 locked = 1;
1133 down_read(&mm->mmap_sem);
1134 vma = find_vma(mm, nstart);
1135 } else if (nstart >= vma->vm_end)
1136 vma = vma->vm_next;
1137 if (!vma || vma->vm_start >= end)
1138 break;
1139 /*
1140 * Set [nstart; nend) to intersection of desired address
1141 * range with the first VMA. Also, skip undesirable VMA types.
1142 */
1143 nend = min(end, vma->vm_end);
1144 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1145 continue;
1146 if (nstart < vma->vm_start)
1147 nstart = vma->vm_start;
1148 /*
1149 * Now fault in a range of pages. populate_vma_page_range()
1150 * double checks the vma flags, so that it won't mlock pages
1151 * if the vma was already munlocked.
1152 */
1153 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1154 if (ret < 0) {
1155 if (ignore_errors) {
1156 ret = 0;
1157 continue; /* continue at next VMA */
1158 }
1159 break;
1160 }
1161 nend = nstart + ret * PAGE_SIZE;
1162 ret = 0;
1163 }
1164 if (locked)
1165 up_read(&mm->mmap_sem);
1166 return ret; /* 0 or negative error code */
1167}
1168
1169/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001170 * get_dump_page() - pin user page in memory while writing it to core dump
1171 * @addr: user address
1172 *
1173 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001174 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001175 *
1176 * Returns NULL on any kind of failure - a hole must then be inserted into
1177 * the corefile, to preserve alignment with its headers; and also returns
1178 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1179 * allowing a hole to be left in the corefile to save diskspace.
1180 *
1181 * Called without mmap_sem, but after all other threads have been killed.
1182 */
1183#ifdef CONFIG_ELF_CORE
1184struct page *get_dump_page(unsigned long addr)
1185{
1186 struct vm_area_struct *vma;
1187 struct page *page;
1188
1189 if (__get_user_pages(current, current->mm, addr, 1,
1190 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1191 NULL) < 1)
1192 return NULL;
1193 flush_cache_page(vma, addr, page_to_pfn(page));
1194 return page;
1195}
1196#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001197
1198/*
1199 * Generic RCU Fast GUP
1200 *
1201 * get_user_pages_fast attempts to pin user pages by walking the page
1202 * tables directly and avoids taking locks. Thus the walker needs to be
1203 * protected from page table pages being freed from under it, and should
1204 * block any THP splits.
1205 *
1206 * One way to achieve this is to have the walker disable interrupts, and
1207 * rely on IPIs from the TLB flushing code blocking before the page table
1208 * pages are freed. This is unsuitable for architectures that do not need
1209 * to broadcast an IPI when invalidating TLBs.
1210 *
1211 * Another way to achieve this is to batch up page table containing pages
1212 * belonging to more than one mm_user, then rcu_sched a callback to free those
1213 * pages. Disabling interrupts will allow the fast_gup walker to both block
1214 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1215 * (which is a relatively rare event). The code below adopts this strategy.
1216 *
1217 * Before activating this code, please be aware that the following assumptions
1218 * are currently made:
1219 *
1220 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1221 * pages containing page tables.
1222 *
Steve Capper2667f502014-10-09 15:29:14 -07001223 * *) ptes can be read atomically by the architecture.
1224 *
1225 * *) access_ok is sufficient to validate userspace address ranges.
1226 *
1227 * The last two assumptions can be relaxed by the addition of helper functions.
1228 *
1229 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1230 */
1231#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1232
1233#ifdef __HAVE_ARCH_PTE_SPECIAL
1234static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1235 int write, struct page **pages, int *nr)
1236{
1237 pte_t *ptep, *ptem;
1238 int ret = 0;
1239
1240 ptem = ptep = pte_offset_map(&pmd, addr);
1241 do {
1242 /*
1243 * In the line below we are assuming that the pte can be read
1244 * atomically. If this is not the case for your architecture,
1245 * please wrap this in a helper function!
1246 *
1247 * for an example see gup_get_pte in arch/x86/mm/gup.c
1248 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001249 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001250 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001251
1252 /*
1253 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001254 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001255 */
1256 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001257 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001258 goto pte_unmap;
1259
Dave Hansen33a709b2016-02-12 13:02:19 -08001260 if (!arch_pte_access_permitted(pte, write))
1261 goto pte_unmap;
1262
Steve Capper2667f502014-10-09 15:29:14 -07001263 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1264 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001265 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001266
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001267 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001268 goto pte_unmap;
1269
1270 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001271 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001272 goto pte_unmap;
1273 }
1274
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001275 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001276 pages[*nr] = page;
1277 (*nr)++;
1278
1279 } while (ptep++, addr += PAGE_SIZE, addr != end);
1280
1281 ret = 1;
1282
1283pte_unmap:
1284 pte_unmap(ptem);
1285 return ret;
1286}
1287#else
1288
1289/*
1290 * If we can't determine whether or not a pte is special, then fail immediately
1291 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1292 * to be special.
1293 *
1294 * For a futex to be placed on a THP tail page, get_futex_key requires a
1295 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1296 * useful to have gup_huge_pmd even if we can't operate on ptes.
1297 */
1298static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1299 int write, struct page **pages, int *nr)
1300{
1301 return 0;
1302}
1303#endif /* __HAVE_ARCH_PTE_SPECIAL */
1304
1305static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1306 unsigned long end, int write, struct page **pages, int *nr)
1307{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001308 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001309 int refs;
1310
1311 if (write && !pmd_write(orig))
1312 return 0;
1313
1314 refs = 0;
1315 head = pmd_page(orig);
1316 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001317 do {
1318 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1319 pages[*nr] = page;
1320 (*nr)++;
1321 page++;
1322 refs++;
1323 } while (addr += PAGE_SIZE, addr != end);
1324
1325 if (!page_cache_add_speculative(head, refs)) {
1326 *nr -= refs;
1327 return 0;
1328 }
1329
1330 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1331 *nr -= refs;
1332 while (refs--)
1333 put_page(head);
1334 return 0;
1335 }
1336
Steve Capper2667f502014-10-09 15:29:14 -07001337 return 1;
1338}
1339
1340static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1341 unsigned long end, int write, struct page **pages, int *nr)
1342{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001343 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001344 int refs;
1345
1346 if (write && !pud_write(orig))
1347 return 0;
1348
1349 refs = 0;
1350 head = pud_page(orig);
1351 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001352 do {
1353 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1354 pages[*nr] = page;
1355 (*nr)++;
1356 page++;
1357 refs++;
1358 } while (addr += PAGE_SIZE, addr != end);
1359
1360 if (!page_cache_add_speculative(head, refs)) {
1361 *nr -= refs;
1362 return 0;
1363 }
1364
1365 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1366 *nr -= refs;
1367 while (refs--)
1368 put_page(head);
1369 return 0;
1370 }
1371
Steve Capper2667f502014-10-09 15:29:14 -07001372 return 1;
1373}
1374
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301375static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1376 unsigned long end, int write,
1377 struct page **pages, int *nr)
1378{
1379 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001380 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301381
1382 if (write && !pgd_write(orig))
1383 return 0;
1384
1385 refs = 0;
1386 head = pgd_page(orig);
1387 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301388 do {
1389 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1390 pages[*nr] = page;
1391 (*nr)++;
1392 page++;
1393 refs++;
1394 } while (addr += PAGE_SIZE, addr != end);
1395
1396 if (!page_cache_add_speculative(head, refs)) {
1397 *nr -= refs;
1398 return 0;
1399 }
1400
1401 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1402 *nr -= refs;
1403 while (refs--)
1404 put_page(head);
1405 return 0;
1406 }
1407
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301408 return 1;
1409}
1410
Steve Capper2667f502014-10-09 15:29:14 -07001411static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1412 int write, struct page **pages, int *nr)
1413{
1414 unsigned long next;
1415 pmd_t *pmdp;
1416
1417 pmdp = pmd_offset(&pud, addr);
1418 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001419 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001420
1421 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001422 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001423 return 0;
1424
1425 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1426 /*
1427 * NUMA hinting faults need to be handled in the GUP
1428 * slowpath for accounting purposes and so that they
1429 * can be serialised against THP migration.
1430 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001431 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001432 return 0;
1433
1434 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1435 pages, nr))
1436 return 0;
1437
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301438 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1439 /*
1440 * architecture have different format for hugetlbfs
1441 * pmd format and THP pmd format
1442 */
1443 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1444 PMD_SHIFT, next, write, pages, nr))
1445 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001446 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1447 return 0;
1448 } while (pmdp++, addr = next, addr != end);
1449
1450 return 1;
1451}
1452
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301453static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1454 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001455{
1456 unsigned long next;
1457 pud_t *pudp;
1458
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301459 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001460 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001461 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001462
1463 next = pud_addr_end(addr, end);
1464 if (pud_none(pud))
1465 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301466 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001467 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301468 pages, nr))
1469 return 0;
1470 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1471 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1472 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001473 return 0;
1474 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1475 return 0;
1476 } while (pudp++, addr = next, addr != end);
1477
1478 return 1;
1479}
1480
1481/*
1482 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1483 * the regular GUP. It will only return non-negative values.
1484 */
1485int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1486 struct page **pages)
1487{
1488 struct mm_struct *mm = current->mm;
1489 unsigned long addr, len, end;
1490 unsigned long next, flags;
1491 pgd_t *pgdp;
1492 int nr = 0;
1493
1494 start &= PAGE_MASK;
1495 addr = start;
1496 len = (unsigned long) nr_pages << PAGE_SHIFT;
1497 end = start + len;
1498
1499 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1500 start, len)))
1501 return 0;
1502
1503 /*
1504 * Disable interrupts. We use the nested form as we can already have
1505 * interrupts disabled by get_futex_key.
1506 *
1507 * With interrupts disabled, we block page table pages from being
1508 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1509 * for more details.
1510 *
1511 * We do not adopt an rcu_read_lock(.) here as we also want to
1512 * block IPIs that come from THPs splitting.
1513 */
1514
1515 local_irq_save(flags);
1516 pgdp = pgd_offset(mm, addr);
1517 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001518 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301519
Steve Capper2667f502014-10-09 15:29:14 -07001520 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301521 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001522 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301523 if (unlikely(pgd_huge(pgd))) {
1524 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1525 pages, &nr))
1526 break;
1527 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1528 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1529 PGDIR_SHIFT, next, write, pages, &nr))
1530 break;
1531 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001532 break;
1533 } while (pgdp++, addr = next, addr != end);
1534 local_irq_restore(flags);
1535
1536 return nr;
1537}
1538
1539/**
1540 * get_user_pages_fast() - pin user pages in memory
1541 * @start: starting user address
1542 * @nr_pages: number of pages from start to pin
1543 * @write: whether pages will be written to
1544 * @pages: array that receives pointers to the pages pinned.
1545 * Should be at least nr_pages long.
1546 *
1547 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1548 * If not successful, it will fall back to taking the lock and
1549 * calling get_user_pages().
1550 *
1551 * Returns number of pages pinned. This may be fewer than the number
1552 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1553 * were pinned, returns -errno.
1554 */
1555int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1556 struct page **pages)
1557{
Steve Capper2667f502014-10-09 15:29:14 -07001558 int nr, ret;
1559
1560 start &= PAGE_MASK;
1561 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1562 ret = nr;
1563
1564 if (nr < nr_pages) {
1565 /* Try to get the remaining pages with get_user_pages */
1566 start += nr << PAGE_SHIFT;
1567 pages += nr;
1568
Lorenzo Stoakesc1641542016-10-13 01:20:13 +01001569 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1570 write ? FOLL_WRITE : 0);
Steve Capper2667f502014-10-09 15:29:14 -07001571
1572 /* Have to be a bit careful with return values */
1573 if (nr > 0) {
1574 if (ret < 0)
1575 ret = nr;
1576 else
1577 ret += nr;
1578 }
1579 }
1580
1581 return ret;
1582}
1583
1584#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */