blob: 0b80bf3878dcf5c3e4b35b8177d8b0ca75f6dfed [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/*
Linus Torvalds0c296402022-01-24 14:41:50 +010064 * FOLL_FORCE or a forced COW break can write even to unwritable pte's,
65 * but only after we've gone through a COW cycle and they are dirty.
Linus Torvalds19be0ea2016-10-13 13:07:36 -070066 */
67static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
68{
Linus Torvalds0c296402022-01-24 14:41:50 +010069 return pte_write(pte) || ((flags & FOLL_COW) && pte_dirty(pte));
70}
71
72/*
73 * A (separate) COW fault might break the page the other way and
74 * get_user_pages() would return the page from what is now the wrong
75 * VM. So we need to force a COW break at GUP time even for reads.
76 */
77static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned int flags)
78{
79 return is_cow_mapping(vma->vm_flags) && (flags & FOLL_GET);
Linus Torvalds19be0ea2016-10-13 13:07:36 -070080}
81
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070082static struct page *follow_page_pte(struct vm_area_struct *vma,
83 unsigned long address, pmd_t *pmd, unsigned int flags)
84{
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070085 struct mm_struct *mm = vma->vm_mm;
Dan Williams3565fce2016-01-15 16:56:55 -080086 struct dev_pagemap *pgmap = NULL;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070087 struct page *page;
88 spinlock_t *ptl;
89 pte_t *ptep, pte;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070090
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070091retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070092 if (unlikely(pmd_bad(*pmd)))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070093 return no_page_table(vma, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070094
95 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070096 pte = *ptep;
97 if (!pte_present(pte)) {
98 swp_entry_t entry;
99 /*
100 * KSM's break_ksm() relies upon recognizing a ksm page
101 * even while it is being migrated, so for that case we
102 * need migration_entry_wait().
103 */
104 if (likely(!(flags & FOLL_MIGRATION)))
105 goto no_page;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800106 if (pte_none(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700107 goto no_page;
108 entry = pte_to_swp_entry(pte);
109 if (!is_migration_entry(entry))
110 goto no_page;
111 pte_unmap_unlock(ptep, ptl);
112 migration_entry_wait(mm, pmd, address);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700113 goto retry;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700114 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800115 if ((flags & FOLL_NUMA) && pte_protnone(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700116 goto no_page;
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700117 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700118 pte_unmap_unlock(ptep, ptl);
119 return NULL;
120 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700121
122 page = vm_normal_page(vma, address, pte);
Dan Williams3565fce2016-01-15 16:56:55 -0800123 if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
124 /*
125 * Only return device mapping pages in the FOLL_GET case since
126 * they are only valid while holding the pgmap reference.
127 */
128 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
129 if (pgmap)
130 page = pte_page(pte);
131 else
132 goto no_page;
133 } else if (unlikely(!page)) {
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700134 if (flags & FOLL_DUMP) {
135 /* Avoid special (like zero) pages in core dumps */
136 page = ERR_PTR(-EFAULT);
137 goto out;
138 }
139
140 if (is_zero_pfn(pte_pfn(pte))) {
141 page = pte_page(pte);
142 } else {
143 int ret;
144
145 ret = follow_pfn_pte(vma, address, ptep, flags);
146 page = ERR_PTR(ret);
147 goto out;
148 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700149 }
150
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800151 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
152 int ret;
153 get_page(page);
154 pte_unmap_unlock(ptep, ptl);
155 lock_page(page);
156 ret = split_huge_page(page);
157 unlock_page(page);
158 put_page(page);
159 if (ret)
160 return ERR_PTR(ret);
161 goto retry;
162 }
163
Dan Williams3565fce2016-01-15 16:56:55 -0800164 if (flags & FOLL_GET) {
Linus Torvalds2ed768c2019-04-11 10:49:19 -0700165 if (unlikely(!try_get_page(page))) {
166 page = ERR_PTR(-ENOMEM);
167 goto out;
168 }
Dan Williams3565fce2016-01-15 16:56:55 -0800169
170 /* drop the pgmap reference now that we hold the page */
171 if (pgmap) {
172 put_dev_pagemap(pgmap);
173 pgmap = NULL;
174 }
175 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700176 if (flags & FOLL_TOUCH) {
177 if ((flags & FOLL_WRITE) &&
178 !pte_dirty(pte) && !PageDirty(page))
179 set_page_dirty(page);
180 /*
181 * pte_mkyoung() would be more correct here, but atomic care
182 * is needed to avoid losing the dirty bit: it is easier to use
183 * mark_page_accessed().
184 */
185 mark_page_accessed(page);
186 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800187 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemove90309c2016-01-15 16:54:33 -0800188 /* Do not mlock pte-mapped THP */
189 if (PageTransCompound(page))
190 goto out;
191
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700192 /*
193 * The preliminary mapping check is mainly to avoid the
194 * pointless overhead of lock_page on the ZERO_PAGE
195 * which might bounce very badly if there is contention.
196 *
197 * If the page is already locked, we don't need to
198 * handle it now - vmscan will handle it later if and
199 * when it attempts to reclaim the page.
200 */
201 if (page->mapping && trylock_page(page)) {
202 lru_add_drain(); /* push cached pages to LRU */
203 /*
204 * Because we lock page here, and migration is
205 * blocked by the pte's page reference, and we
206 * know the page is still mapped, we don't even
207 * need to check for file-cache page truncation.
208 */
209 mlock_vma_page(page);
210 unlock_page(page);
211 }
212 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700213out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700214 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700215 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700216no_page:
217 pte_unmap_unlock(ptep, ptl);
218 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700219 return NULL;
220 return no_page_table(vma, flags);
221}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700222
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700223/**
224 * follow_page_mask - look up a page descriptor from a user-virtual address
225 * @vma: vm_area_struct mapping @address
226 * @address: virtual address to look up
227 * @flags: flags modifying lookup behaviour
228 * @page_mask: on output, *page_mask is set according to the size of the page
229 *
230 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
231 *
232 * Returns the mapped (struct page *), %NULL if no mapping exists, or
233 * an error pointer if there is a mapping to something not represented
234 * by a page descriptor (see also vm_normal_page()).
235 */
236struct page *follow_page_mask(struct vm_area_struct *vma,
237 unsigned long address, unsigned int flags,
238 unsigned int *page_mask)
239{
240 pgd_t *pgd;
241 pud_t *pud;
242 pmd_t *pmd;
243 spinlock_t *ptl;
244 struct page *page;
245 struct mm_struct *mm = vma->vm_mm;
246
247 *page_mask = 0;
248
249 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
250 if (!IS_ERR(page)) {
251 BUG_ON(flags & FOLL_GET);
252 return page;
253 }
254
255 pgd = pgd_offset(mm, address);
256 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
257 return no_page_table(vma, flags);
258
259 pud = pud_offset(pgd, address);
260 if (pud_none(*pud))
261 return no_page_table(vma, flags);
262 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800263 page = follow_huge_pud(mm, address, pud, flags);
264 if (page)
265 return page;
266 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700267 }
268 if (unlikely(pud_bad(*pud)))
269 return no_page_table(vma, flags);
270
271 pmd = pmd_offset(pud, address);
272 if (pmd_none(*pmd))
273 return no_page_table(vma, flags);
274 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800275 page = follow_huge_pmd(mm, address, pmd, flags);
276 if (page)
277 return page;
278 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700279 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800280 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700281 return no_page_table(vma, flags);
Dan Williams3565fce2016-01-15 16:56:55 -0800282 if (pmd_devmap(*pmd)) {
283 ptl = pmd_lock(mm, pmd);
284 page = follow_devmap_pmd(vma, address, pmd, flags);
285 spin_unlock(ptl);
286 if (page)
287 return page;
288 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800289 if (likely(!pmd_trans_huge(*pmd)))
290 return follow_page_pte(vma, address, pmd, flags);
291
292 ptl = pmd_lock(mm, pmd);
293 if (unlikely(!pmd_trans_huge(*pmd))) {
294 spin_unlock(ptl);
295 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700296 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800297 if (flags & FOLL_SPLIT) {
298 int ret;
299 page = pmd_page(*pmd);
300 if (is_huge_zero_page(page)) {
301 spin_unlock(ptl);
302 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800303 split_huge_pmd(vma, pmd, address);
Naoya Horiguchi337d9ab2016-07-26 15:24:03 -0700304 if (pmd_trans_unstable(pmd))
305 ret = -EBUSY;
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800306 } else {
Linus Torvalds2ed768c2019-04-11 10:49:19 -0700307 if (unlikely(!try_get_page(page))) {
308 spin_unlock(ptl);
309 return ERR_PTR(-ENOMEM);
310 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800311 spin_unlock(ptl);
312 lock_page(page);
313 ret = split_huge_page(page);
314 unlock_page(page);
315 put_page(page);
Kirill A. Shutemovbaa355f2016-07-26 15:25:51 -0700316 if (pmd_none(*pmd))
317 return no_page_table(vma, flags);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800318 }
319
320 return ret ? ERR_PTR(ret) :
321 follow_page_pte(vma, address, pmd, flags);
322 }
323
324 page = follow_trans_huge_pmd(vma, address, pmd, flags);
325 spin_unlock(ptl);
326 *page_mask = HPAGE_PMD_NR - 1;
327 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700328}
329
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700330static int get_gate_page(struct mm_struct *mm, unsigned long address,
331 unsigned int gup_flags, struct vm_area_struct **vma,
332 struct page **page)
333{
334 pgd_t *pgd;
335 pud_t *pud;
336 pmd_t *pmd;
337 pte_t *pte;
338 int ret = -EFAULT;
339
340 /* user gate pages are read-only */
341 if (gup_flags & FOLL_WRITE)
342 return -EFAULT;
343 if (address > TASK_SIZE)
344 pgd = pgd_offset_k(address);
345 else
346 pgd = pgd_offset_gate(mm, address);
347 BUG_ON(pgd_none(*pgd));
348 pud = pud_offset(pgd, address);
349 BUG_ON(pud_none(*pud));
350 pmd = pmd_offset(pud, address);
351 if (pmd_none(*pmd))
352 return -EFAULT;
353 VM_BUG_ON(pmd_trans_huge(*pmd));
354 pte = pte_offset_map(pmd, address);
355 if (pte_none(*pte))
356 goto unmap;
357 *vma = get_gate_vma(mm);
358 if (!page)
359 goto out;
360 *page = vm_normal_page(*vma, address, *pte);
361 if (!*page) {
362 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
363 goto unmap;
364 *page = pte_page(*pte);
365 }
Linus Torvalds2ed768c2019-04-11 10:49:19 -0700366 if (unlikely(!try_get_page(*page))) {
367 ret = -ENOMEM;
368 goto unmap;
369 }
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700370out:
371 ret = 0;
372unmap:
373 pte_unmap(pte);
374 return ret;
375}
376
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700377/*
378 * mmap_sem must be held on entry. If @nonblocking != NULL and
379 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
380 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
381 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700382static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
383 unsigned long address, unsigned int *flags, int *nonblocking)
384{
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700385 unsigned int fault_flags = 0;
386 int ret;
387
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800388 /* mlock all present pages, but do not fault in new pages */
389 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
390 return -ENOENT;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700391 if (*flags & FOLL_WRITE)
392 fault_flags |= FAULT_FLAG_WRITE;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800393 if (*flags & FOLL_REMOTE)
394 fault_flags |= FAULT_FLAG_REMOTE;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700395 if (nonblocking)
396 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
397 if (*flags & FOLL_NOWAIT)
398 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700399 if (*flags & FOLL_TRIED) {
400 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
401 fault_flags |= FAULT_FLAG_TRIED;
402 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700403
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700404 ret = handle_mm_fault(vma, address, fault_flags);
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700405 if (ret & VM_FAULT_ERROR) {
406 if (ret & VM_FAULT_OOM)
407 return -ENOMEM;
408 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
409 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800410 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700411 return -EFAULT;
412 BUG();
413 }
414
415 if (tsk) {
416 if (ret & VM_FAULT_MAJOR)
417 tsk->maj_flt++;
418 else
419 tsk->min_flt++;
420 }
421
422 if (ret & VM_FAULT_RETRY) {
423 if (nonblocking)
424 *nonblocking = 0;
425 return -EBUSY;
426 }
427
428 /*
429 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
430 * necessary, even if maybe_mkwrite decided not to set pte_write. We
431 * can thus safely do subsequent page lookups as if they were reads.
432 * But only do so when looping for pte_write is futile: in some cases
433 * userspace may also be wanting to write to the gotten user page,
434 * which a read fault here might prevent (a readonly page might get
435 * reCOWed by userspace write).
436 */
437 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
Linus Torvalds19be0ea2016-10-13 13:07:36 -0700438 *flags |= FOLL_COW;
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700439 return 0;
440}
441
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700442static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
443{
444 vm_flags_t vm_flags = vma->vm_flags;
Dave Hansen1b2ee122016-02-12 13:02:21 -0800445 int write = (gup_flags & FOLL_WRITE);
446 int foreign = (gup_flags & FOLL_REMOTE);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700447
448 if (vm_flags & (VM_IO | VM_PFNMAP))
449 return -EFAULT;
450
Willy Tarreau6f1abf82018-05-11 08:11:44 +0200451 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
452 return -EFAULT;
453
Dave Hansen1b2ee122016-02-12 13:02:21 -0800454 if (write) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700455 if (!(vm_flags & VM_WRITE)) {
456 if (!(gup_flags & FOLL_FORCE))
457 return -EFAULT;
458 /*
459 * We used to let the write,force case do COW in a
460 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
461 * set a breakpoint in a read-only mapping of an
462 * executable, without corrupting the file (yet only
463 * when that file had been opened for writing!).
464 * Anon pages in shared mappings are surprising: now
465 * just reject it.
466 */
Hugh Dickins46435362016-01-30 18:03:16 -0800467 if (!is_cow_mapping(vm_flags))
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700468 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700469 }
470 } else if (!(vm_flags & VM_READ)) {
471 if (!(gup_flags & FOLL_FORCE))
472 return -EFAULT;
473 /*
474 * Is there actually any vma we can reach here which does not
475 * have VM_MAYREAD set?
476 */
477 if (!(vm_flags & VM_MAYREAD))
478 return -EFAULT;
479 }
Dave Hansend61172b2016-02-12 13:02:24 -0800480 /*
481 * gups are always data accesses, not instruction
482 * fetches, so execute=false here
483 */
484 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800485 return -EFAULT;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700486 return 0;
487}
488
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700489/**
490 * __get_user_pages() - pin user pages in memory
491 * @tsk: task_struct of target task
492 * @mm: mm_struct of target mm
493 * @start: starting user address
494 * @nr_pages: number of pages from start to pin
495 * @gup_flags: flags modifying pin behaviour
496 * @pages: array that receives pointers to the pages pinned.
497 * Should be at least nr_pages long. Or NULL, if caller
498 * only intends to ensure the pages are faulted in.
499 * @vmas: array of pointers to vmas corresponding to each page.
500 * Or NULL if the caller does not require them.
501 * @nonblocking: whether waiting for disk IO or mmap_sem contention
502 *
503 * Returns number of pages pinned. This may be fewer than the number
504 * requested. If nr_pages is 0 or negative, returns 0. If no pages
505 * were pinned, returns -errno. Each page returned must be released
506 * with a put_page() call when it is finished with. vmas will only
507 * remain valid while mmap_sem is held.
508 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700509 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700510 *
511 * __get_user_pages walks a process's page tables and takes a reference to
512 * each struct page that each user address corresponds to at a given
513 * instant. That is, it takes the page that would be accessed if a user
514 * thread accesses the given user virtual address at that instant.
515 *
516 * This does not guarantee that the page exists in the user mappings when
517 * __get_user_pages returns, and there may even be a completely different
518 * page there in some cases (eg. if mmapped pagecache has been invalidated
519 * and subsequently re faulted). However it does guarantee that the page
520 * won't be freed completely. And mostly callers simply care that the page
521 * contains data that was valid *at some point in time*. Typically, an IO
522 * or similar operation cannot guarantee anything stronger anyway because
523 * locks can't be held over the syscall boundary.
524 *
525 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
526 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
527 * appropriate) must be called after the page is finished with, and
528 * before put_page is called.
529 *
530 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
531 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700532 * *@nonblocking will be set to 0. Further, if @gup_flags does not
533 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
534 * this case.
535 *
536 * A caller using such a combination of @nonblocking and @gup_flags
537 * must therefore hold the mmap_sem for reading only, and recognize
538 * when it's been released. Otherwise, it must be held for either
539 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700540 *
541 * In most cases, get_user_pages or get_user_pages_fast should be used
542 * instead of __get_user_pages. __get_user_pages should be used only if
543 * you need some special @gup_flags.
544 */
Lorenzo Stoakes0d731752016-10-24 10:57:25 +0100545static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700546 unsigned long start, unsigned long nr_pages,
547 unsigned int gup_flags, struct page **pages,
548 struct vm_area_struct **vmas, int *nonblocking)
549{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700550 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700551 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700552 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700553
554 if (!nr_pages)
555 return 0;
556
557 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
558
559 /*
560 * If FOLL_FORCE is set then do not force a full fault as the hinting
561 * fault information is unrelated to the reference behaviour of a task
562 * using the address space
563 */
564 if (!(gup_flags & FOLL_FORCE))
565 gup_flags |= FOLL_NUMA;
566
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700567 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700568 struct page *page;
569 unsigned int foll_flags = gup_flags;
570 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700571
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700572 /* first iteration or cross vma bound */
573 if (!vma || start >= vma->vm_end) {
574 vma = find_extend_vma(mm, start);
575 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700576 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700577 ret = get_gate_page(mm, start & PAGE_MASK,
578 gup_flags, &vma,
579 pages ? &pages[i] : NULL);
580 if (ret)
581 return i ? : ret;
582 page_mask = 0;
583 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700584 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700585
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700586 if (!vma || check_vma_flags(vma, gup_flags))
587 return i ? : -EFAULT;
588 if (is_vm_hugetlb_page(vma)) {
Linus Torvalds0c296402022-01-24 14:41:50 +0100589 if (should_force_cow_break(vma, foll_flags))
590 foll_flags |= FOLL_WRITE;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700591 i = follow_hugetlb_page(mm, vma, pages, vmas,
592 &start, &nr_pages, i,
Linus Torvalds0c296402022-01-24 14:41:50 +0100593 foll_flags);
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700594 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700595 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700596 }
Linus Torvalds0c296402022-01-24 14:41:50 +0100597
598 if (should_force_cow_break(vma, foll_flags))
599 foll_flags |= FOLL_WRITE;
600
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700601retry:
602 /*
603 * If we have a pending SIGKILL, don't keep faulting pages and
604 * potentially allocating memory.
605 */
606 if (unlikely(fatal_signal_pending(current)))
607 return i ? i : -ERESTARTSYS;
608 cond_resched();
609 page = follow_page_mask(vma, start, foll_flags, &page_mask);
610 if (!page) {
611 int ret;
612 ret = faultin_page(tsk, vma, start, &foll_flags,
613 nonblocking);
614 switch (ret) {
615 case 0:
616 goto retry;
617 case -EFAULT:
618 case -ENOMEM:
619 case -EHWPOISON:
620 return i ? i : ret;
621 case -EBUSY:
622 return i;
623 case -ENOENT:
624 goto next_page;
625 }
626 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700627 } else if (PTR_ERR(page) == -EEXIST) {
628 /*
629 * Proper page table entry exists, but no corresponding
630 * struct page.
631 */
632 goto next_page;
633 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700634 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700635 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700636 if (pages) {
637 pages[i] = page;
638 flush_anon_page(vma, page, start);
639 flush_dcache_page(page);
640 page_mask = 0;
641 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700642next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700643 if (vmas) {
644 vmas[i] = vma;
645 page_mask = 0;
646 }
647 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
648 if (page_increm > nr_pages)
649 page_increm = nr_pages;
650 i += page_increm;
651 start += page_increm * PAGE_SIZE;
652 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700653 } while (nr_pages);
654 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700655}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700656
Dave Hansend4925e02016-02-12 13:02:16 -0800657bool vma_permits_fault(struct vm_area_struct *vma, unsigned int fault_flags)
658{
Dave Hansen1b2ee122016-02-12 13:02:21 -0800659 bool write = !!(fault_flags & FAULT_FLAG_WRITE);
660 bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
Dave Hansen33a709b2016-02-12 13:02:19 -0800661 vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
Dave Hansend4925e02016-02-12 13:02:16 -0800662
663 if (!(vm_flags & vma->vm_flags))
664 return false;
665
Dave Hansen33a709b2016-02-12 13:02:19 -0800666 /*
667 * The architecture might have a hardware protection
Dave Hansen1b2ee122016-02-12 13:02:21 -0800668 * mechanism other than read/write that can deny access.
Dave Hansend61172b2016-02-12 13:02:24 -0800669 *
670 * gup always represents data access, not instruction
671 * fetches, so execute=false here:
Dave Hansen33a709b2016-02-12 13:02:19 -0800672 */
Dave Hansend61172b2016-02-12 13:02:24 -0800673 if (!arch_vma_access_permitted(vma, write, false, foreign))
Dave Hansen33a709b2016-02-12 13:02:19 -0800674 return false;
675
Dave Hansend4925e02016-02-12 13:02:16 -0800676 return true;
677}
678
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700679/*
680 * fixup_user_fault() - manually resolve a user page fault
681 * @tsk: the task_struct to use for page fault accounting, or
682 * NULL if faults are not to be recorded.
683 * @mm: mm_struct of target mm
684 * @address: user address
685 * @fault_flags:flags to pass down to handle_mm_fault()
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800686 * @unlocked: did we unlock the mmap_sem while retrying, maybe NULL if caller
687 * does not allow retry
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700688 *
689 * This is meant to be called in the specific scenario where for locking reasons
690 * we try to access user memory in atomic context (within a pagefault_disable()
691 * section), this returns -EFAULT, and we want to resolve the user fault before
692 * trying again.
693 *
694 * Typically this is meant to be used by the futex code.
695 *
696 * The main difference with get_user_pages() is that this function will
697 * unconditionally call handle_mm_fault() which will in turn perform all the
698 * necessary SW fixup of the dirty and young bits in the PTE, while
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800699 * get_user_pages() only guarantees to update these in the struct page.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700700 *
701 * This is important for some architectures where those bits also gate the
702 * access permission to the page because they are maintained in software. On
703 * such architectures, gup() will not be enough to make a subsequent access
704 * succeed.
705 *
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800706 * This function will not return with an unlocked mmap_sem. So it has not the
707 * same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700708 */
709int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800710 unsigned long address, unsigned int fault_flags,
711 bool *unlocked)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700712{
713 struct vm_area_struct *vma;
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800714 int ret, major = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700715
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800716 if (unlocked)
717 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
718
719retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700720 vma = find_extend_vma(mm, address);
721 if (!vma || address < vma->vm_start)
722 return -EFAULT;
723
Dave Hansend4925e02016-02-12 13:02:16 -0800724 if (!vma_permits_fault(vma, fault_flags))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700725 return -EFAULT;
726
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -0700727 ret = handle_mm_fault(vma, address, fault_flags);
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800728 major |= ret & VM_FAULT_MAJOR;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700729 if (ret & VM_FAULT_ERROR) {
730 if (ret & VM_FAULT_OOM)
731 return -ENOMEM;
732 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
733 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800734 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700735 return -EFAULT;
736 BUG();
737 }
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800738
739 if (ret & VM_FAULT_RETRY) {
740 down_read(&mm->mmap_sem);
741 if (!(fault_flags & FAULT_FLAG_TRIED)) {
742 *unlocked = true;
743 fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
744 fault_flags |= FAULT_FLAG_TRIED;
745 goto retry;
746 }
747 }
748
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700749 if (tsk) {
Dominik Dingel4a9e1cd2016-01-15 16:57:04 -0800750 if (major)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700751 tsk->maj_flt++;
752 else
753 tsk->min_flt++;
754 }
755 return 0;
756}
Paolo Bonziniadd6a0c2016-06-07 17:51:18 +0200757EXPORT_SYMBOL_GPL(fixup_user_fault);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700758
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800759static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
760 struct mm_struct *mm,
761 unsigned long start,
762 unsigned long nr_pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800763 struct page **pages,
764 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800765 int *locked, bool notify_drop,
766 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800767{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800768 long ret, pages_done;
769 bool lock_dropped;
770
771 if (locked) {
772 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
773 BUG_ON(vmas);
774 /* check caller initialized locked */
775 BUG_ON(*locked != 1);
776 }
777
778 if (pages)
779 flags |= FOLL_GET;
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800780
781 pages_done = 0;
782 lock_dropped = false;
783 for (;;) {
784 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
785 vmas, locked);
786 if (!locked)
787 /* VM_FAULT_RETRY couldn't trigger, bypass */
788 return ret;
789
790 /* VM_FAULT_RETRY cannot return errors */
791 if (!*locked) {
792 BUG_ON(ret < 0);
793 BUG_ON(ret >= nr_pages);
794 }
795
796 if (!pages)
797 /* If it's a prefault don't insist harder */
798 return ret;
799
800 if (ret > 0) {
801 nr_pages -= ret;
802 pages_done += ret;
803 if (!nr_pages)
804 break;
805 }
806 if (*locked) {
807 /* VM_FAULT_RETRY didn't trigger */
808 if (!pages_done)
809 pages_done = ret;
810 break;
811 }
812 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
813 pages += ret;
814 start += ret << PAGE_SHIFT;
815
816 /*
817 * Repeat on the address that fired VM_FAULT_RETRY
818 * without FAULT_FLAG_ALLOW_RETRY but with
819 * FAULT_FLAG_TRIED.
820 */
821 *locked = 1;
822 lock_dropped = true;
823 down_read(&mm->mmap_sem);
824 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
825 pages, NULL, NULL);
826 if (ret != 1) {
827 BUG_ON(ret > 1);
828 if (!pages_done)
829 pages_done = ret;
830 break;
831 }
832 nr_pages--;
833 pages_done++;
834 if (!nr_pages)
835 break;
836 pages++;
837 start += PAGE_SIZE;
838 }
839 if (notify_drop && lock_dropped && *locked) {
840 /*
841 * We must let the caller know we temporarily dropped the lock
842 * and so the critical section protected by it was lost.
843 */
844 up_read(&mm->mmap_sem);
845 *locked = 0;
846 }
847 return pages_done;
848}
849
850/*
851 * We can leverage the VM_FAULT_RETRY functionality in the page fault
852 * paths better by using either get_user_pages_locked() or
853 * get_user_pages_unlocked().
854 *
855 * get_user_pages_locked() is suitable to replace the form:
856 *
857 * down_read(&mm->mmap_sem);
858 * do_something()
859 * get_user_pages(tsk, mm, ..., pages, NULL);
860 * up_read(&mm->mmap_sem);
861 *
862 * to:
863 *
864 * int locked = 1;
865 * down_read(&mm->mmap_sem);
866 * do_something()
867 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
868 * if (locked)
869 * up_read(&mm->mmap_sem);
870 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200871long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100872 unsigned int gup_flags, struct page **pages,
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800873 int *locked)
874{
Dave Hansencde70142016-02-12 13:01:55 -0800875 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes3b913172016-10-13 01:20:14 +0100876 pages, NULL, locked, true,
877 gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800878}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200879EXPORT_SYMBOL(get_user_pages_locked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800880
881/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800882 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
883 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
884 *
885 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
886 * caller if required (just like with __get_user_pages). "FOLL_GET",
887 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
888 * according to the parameters "pages", "write", "force"
889 * respectively.
890 */
891__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
892 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesd4944b02016-10-13 01:20:12 +0100893 struct page **pages, unsigned int gup_flags)
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800894{
895 long ret;
896 int locked = 1;
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100897
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800898 down_read(&mm->mmap_sem);
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100899 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
900 &locked, false, gup_flags);
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800901 if (locked)
902 up_read(&mm->mmap_sem);
903 return ret;
904}
905EXPORT_SYMBOL(__get_user_pages_unlocked);
906
907/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800908 * get_user_pages_unlocked() is suitable to replace the form:
909 *
910 * down_read(&mm->mmap_sem);
911 * get_user_pages(tsk, mm, ..., pages, NULL);
912 * up_read(&mm->mmap_sem);
913 *
914 * with:
915 *
916 * get_user_pages_unlocked(tsk, mm, ..., pages);
917 *
918 * It is functionally equivalent to get_user_pages_fast so
919 * get_user_pages_fast should be used instead, if the two parameters
920 * "tsk" and "mm" are respectively equal to current and current->mm,
921 * or if "force" shall be set to 1 (get_user_pages_fast misses the
922 * "force" parameter).
923 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200924long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100925 struct page **pages, unsigned int gup_flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800926{
Dave Hansencde70142016-02-12 13:01:55 -0800927 return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
Lorenzo Stoakesc1641542016-10-13 01:20:13 +0100928 pages, gup_flags | FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800929}
Ingo Molnarc12d2da2016-04-04 10:24:58 +0200930EXPORT_SYMBOL(get_user_pages_unlocked);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800931
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700932/*
Dave Hansen1e987792016-02-12 13:01:54 -0800933 * get_user_pages_remote() - pin user pages in memory
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700934 * @tsk: the task_struct to use for page fault accounting, or
935 * NULL if faults are not to be recorded.
936 * @mm: mm_struct of target mm
937 * @start: starting user address
938 * @nr_pages: number of pages from start to pin
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100939 * @gup_flags: flags modifying lookup behaviour
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700940 * @pages: array that receives pointers to the pages pinned.
941 * Should be at least nr_pages long. Or NULL, if caller
942 * only intends to ensure the pages are faulted in.
943 * @vmas: array of pointers to vmas corresponding to each page.
944 * Or NULL if the caller does not require them.
945 *
946 * Returns number of pages pinned. This may be fewer than the number
947 * requested. If nr_pages is 0 or negative, returns 0. If no pages
948 * were pinned, returns -errno. Each page returned must be released
949 * with a put_page() call when it is finished with. vmas will only
950 * remain valid while mmap_sem is held.
951 *
952 * Must be called with mmap_sem held for read or write.
953 *
954 * get_user_pages walks a process's page tables and takes a reference to
955 * each struct page that each user address corresponds to at a given
956 * instant. That is, it takes the page that would be accessed if a user
957 * thread accesses the given user virtual address at that instant.
958 *
959 * This does not guarantee that the page exists in the user mappings when
960 * get_user_pages returns, and there may even be a completely different
961 * page there in some cases (eg. if mmapped pagecache has been invalidated
962 * and subsequently re faulted). However it does guarantee that the page
963 * won't be freed completely. And mostly callers simply care that the page
964 * contains data that was valid *at some point in time*. Typically, an IO
965 * or similar operation cannot guarantee anything stronger anyway because
966 * locks can't be held over the syscall boundary.
967 *
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100968 * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
969 * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
970 * be called after the page is finished with, and before put_page is called.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700971 *
972 * get_user_pages is typically used for fewer-copy IO operations, to get a
973 * handle on the memory by some means other than accesses via the user virtual
974 * addresses. The pages may be submitted for DMA to devices or accessed via
975 * their kernel linear mapping (via the kmap APIs). Care should be taken to
976 * use the correct cache flushing APIs.
977 *
978 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800979 *
980 * get_user_pages should be phased out in favor of
981 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
982 * should use get_user_pages because it cannot pass
983 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700984 */
Dave Hansen1e987792016-02-12 13:01:54 -0800985long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
986 unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100987 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -0800988 struct vm_area_struct **vmas)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700989{
Lorenzo Stoakes859110d2016-10-13 01:20:11 +0100990 return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
Lorenzo Stoakes9beae1e2016-10-13 01:20:17 +0100991 NULL, false,
992 gup_flags | FOLL_TOUCH | FOLL_REMOTE);
Dave Hansen1e987792016-02-12 13:01:54 -0800993}
994EXPORT_SYMBOL(get_user_pages_remote);
995
996/*
Dave Hansend4edcf02016-02-12 13:01:56 -0800997 * This is the same as get_user_pages_remote(), just with a
998 * less-flexible calling convention where we assume that the task
999 * and mm being operated on are the current task's. We also
1000 * obviously don't pass FOLL_REMOTE in here.
Dave Hansen1e987792016-02-12 13:01:54 -08001001 */
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001002long get_user_pages(unsigned long start, unsigned long nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +01001003 unsigned int gup_flags, struct page **pages,
Dave Hansen1e987792016-02-12 13:01:54 -08001004 struct vm_area_struct **vmas)
1005{
Dave Hansencde70142016-02-12 13:01:55 -08001006 return __get_user_pages_locked(current, current->mm, start, nr_pages,
Lorenzo Stoakes768ae302016-10-13 01:20:16 +01001007 pages, vmas, NULL, false,
1008 gup_flags | FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001009}
Ingo Molnarc12d2da2016-04-04 10:24:58 +02001010EXPORT_SYMBOL(get_user_pages);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001011
Dan Williamsb29ea3c2018-02-23 14:05:49 -08001012#ifdef CONFIG_FS_DAX
1013/*
1014 * This is the same as get_user_pages() in that it assumes we are
1015 * operating on the current task's mm, but it goes further to validate
1016 * that the vmas associated with the address range are suitable for
1017 * longterm elevated page reference counts. For example, filesystem-dax
1018 * mappings are subject to the lifetime enforced by the filesystem and
1019 * we need guarantees that longterm users like RDMA and V4L2 only
1020 * establish mappings that have a kernel enforced revocation mechanism.
1021 *
1022 * "longterm" == userspace controlled elevated page count lifetime.
1023 * Contrast this to iov_iter_get_pages() usages which are transient.
1024 */
1025long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1026 unsigned int gup_flags, struct page **pages,
1027 struct vm_area_struct **vmas_arg)
1028{
1029 struct vm_area_struct **vmas = vmas_arg;
1030 struct vm_area_struct *vma_prev = NULL;
1031 long rc, i;
1032
1033 if (!pages)
1034 return -EINVAL;
1035
1036 if (!vmas) {
1037 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1038 GFP_KERNEL);
1039 if (!vmas)
1040 return -ENOMEM;
1041 }
1042
1043 rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1044
1045 for (i = 0; i < rc; i++) {
1046 struct vm_area_struct *vma = vmas[i];
1047
1048 if (vma == vma_prev)
1049 continue;
1050
1051 vma_prev = vma;
1052
1053 if (vma_is_fsdax(vma))
1054 break;
1055 }
1056
1057 /*
1058 * Either get_user_pages() failed, or the vma validation
1059 * succeeded, in either case we don't need to put_page() before
1060 * returning.
1061 */
1062 if (i >= rc)
1063 goto out;
1064
1065 for (i = 0; i < rc; i++)
1066 put_page(pages[i]);
1067 rc = -EOPNOTSUPP;
1068out:
1069 if (vmas != vmas_arg)
1070 kfree(vmas);
1071 return rc;
1072}
1073EXPORT_SYMBOL(get_user_pages_longterm);
1074#endif /* CONFIG_FS_DAX */
1075
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001076/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001077 * populate_vma_page_range() - populate a range of pages in the vma.
1078 * @vma: target vma
1079 * @start: start address
1080 * @end: end address
1081 * @nonblocking:
1082 *
1083 * This takes care of mlocking the pages too if VM_LOCKED is set.
1084 *
1085 * return 0 on success, negative error code on error.
1086 *
1087 * vma->vm_mm->mmap_sem must be held.
1088 *
1089 * If @nonblocking is NULL, it may be held for read or write and will
1090 * be unperturbed.
1091 *
1092 * If @nonblocking is non-NULL, it must held for read only and may be
1093 * released. If it's released, *@nonblocking will be set to 0.
1094 */
1095long populate_vma_page_range(struct vm_area_struct *vma,
1096 unsigned long start, unsigned long end, int *nonblocking)
1097{
1098 struct mm_struct *mm = vma->vm_mm;
1099 unsigned long nr_pages = (end - start) / PAGE_SIZE;
1100 int gup_flags;
1101
1102 VM_BUG_ON(start & ~PAGE_MASK);
1103 VM_BUG_ON(end & ~PAGE_MASK);
1104 VM_BUG_ON_VMA(start < vma->vm_start, vma);
1105 VM_BUG_ON_VMA(end > vma->vm_end, vma);
1106 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1107
Eric B Munsonde60f5f2015-11-05 18:51:36 -08001108 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1109 if (vma->vm_flags & VM_LOCKONFAULT)
1110 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001111 /*
1112 * We want to touch writable mappings with a write fault in order
1113 * to break COW, except for shared mappings because these don't COW
1114 * and we would not want to dirty them for nothing.
1115 */
1116 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1117 gup_flags |= FOLL_WRITE;
1118
1119 /*
1120 * We want mlock to succeed for regions that have any permissions
1121 * other than PROT_NONE.
1122 */
1123 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1124 gup_flags |= FOLL_FORCE;
1125
1126 /*
1127 * We made sure addr is within a VMA, so the following will
1128 * not result in a stack expansion that recurses back here.
1129 */
1130 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1131 NULL, NULL, nonblocking);
1132}
1133
1134/*
1135 * __mm_populate - populate and/or mlock pages within a range of address space.
1136 *
1137 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1138 * flags. VMAs must be already marked with the desired vm_flags, and
1139 * mmap_sem must not be held.
1140 */
1141int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1142{
1143 struct mm_struct *mm = current->mm;
1144 unsigned long end, nstart, nend;
1145 struct vm_area_struct *vma = NULL;
1146 int locked = 0;
1147 long ret = 0;
1148
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -07001149 end = start + len;
1150
1151 for (nstart = start; nstart < end; nstart = nend) {
1152 /*
1153 * We want to fault in pages for [nstart; end) address range.
1154 * Find first corresponding VMA.
1155 */
1156 if (!locked) {
1157 locked = 1;
1158 down_read(&mm->mmap_sem);
1159 vma = find_vma(mm, nstart);
1160 } else if (nstart >= vma->vm_end)
1161 vma = vma->vm_next;
1162 if (!vma || vma->vm_start >= end)
1163 break;
1164 /*
1165 * Set [nstart; nend) to intersection of desired address
1166 * range with the first VMA. Also, skip undesirable VMA types.
1167 */
1168 nend = min(end, vma->vm_end);
1169 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1170 continue;
1171 if (nstart < vma->vm_start)
1172 nstart = vma->vm_start;
1173 /*
1174 * Now fault in a range of pages. populate_vma_page_range()
1175 * double checks the vma flags, so that it won't mlock pages
1176 * if the vma was already munlocked.
1177 */
1178 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1179 if (ret < 0) {
1180 if (ignore_errors) {
1181 ret = 0;
1182 continue; /* continue at next VMA */
1183 }
1184 break;
1185 }
1186 nend = nstart + ret * PAGE_SIZE;
1187 ret = 0;
1188 }
1189 if (locked)
1190 up_read(&mm->mmap_sem);
1191 return ret; /* 0 or negative error code */
1192}
1193
1194/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001195 * get_dump_page() - pin user page in memory while writing it to core dump
1196 * @addr: user address
1197 *
1198 * Returns struct page pointer of user page pinned for dump,
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +03001199 * to be freed afterwards by put_page().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001200 *
1201 * Returns NULL on any kind of failure - a hole must then be inserted into
1202 * the corefile, to preserve alignment with its headers; and also returns
1203 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1204 * allowing a hole to be left in the corefile to save diskspace.
1205 *
1206 * Called without mmap_sem, but after all other threads have been killed.
1207 */
1208#ifdef CONFIG_ELF_CORE
1209struct page *get_dump_page(unsigned long addr)
1210{
1211 struct vm_area_struct *vma;
1212 struct page *page;
1213
1214 if (__get_user_pages(current, current->mm, addr, 1,
1215 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1216 NULL) < 1)
1217 return NULL;
1218 flush_cache_page(vma, addr, page_to_pfn(page));
1219 return page;
1220}
1221#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001222
1223/*
1224 * Generic RCU Fast GUP
1225 *
1226 * get_user_pages_fast attempts to pin user pages by walking the page
1227 * tables directly and avoids taking locks. Thus the walker needs to be
1228 * protected from page table pages being freed from under it, and should
1229 * block any THP splits.
1230 *
1231 * One way to achieve this is to have the walker disable interrupts, and
1232 * rely on IPIs from the TLB flushing code blocking before the page table
1233 * pages are freed. This is unsuitable for architectures that do not need
1234 * to broadcast an IPI when invalidating TLBs.
1235 *
1236 * Another way to achieve this is to batch up page table containing pages
1237 * belonging to more than one mm_user, then rcu_sched a callback to free those
1238 * pages. Disabling interrupts will allow the fast_gup walker to both block
1239 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1240 * (which is a relatively rare event). The code below adopts this strategy.
1241 *
1242 * Before activating this code, please be aware that the following assumptions
1243 * are currently made:
1244 *
1245 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1246 * pages containing page tables.
1247 *
Steve Capper2667f502014-10-09 15:29:14 -07001248 * *) ptes can be read atomically by the architecture.
1249 *
1250 * *) access_ok is sufficient to validate userspace address ranges.
1251 *
1252 * The last two assumptions can be relaxed by the addition of helper functions.
1253 *
1254 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1255 */
1256#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1257
Linus Torvalds2ed768c2019-04-11 10:49:19 -07001258/*
1259 * Return the compund head page with ref appropriately incremented,
1260 * or NULL if that failed.
1261 */
1262static inline struct page *try_get_compound_head(struct page *page, int refs)
1263{
1264 struct page *head = compound_head(page);
1265 if (WARN_ON_ONCE(page_ref_count(head) < 0))
1266 return NULL;
1267 if (unlikely(!page_cache_add_speculative(head, refs)))
1268 return NULL;
1269 return head;
1270}
1271
Steve Capper2667f502014-10-09 15:29:14 -07001272#ifdef __HAVE_ARCH_PTE_SPECIAL
1273static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1274 int write, struct page **pages, int *nr)
1275{
1276 pte_t *ptep, *ptem;
1277 int ret = 0;
1278
1279 ptem = ptep = pte_offset_map(&pmd, addr);
1280 do {
1281 /*
1282 * In the line below we are assuming that the pte can be read
1283 * atomically. If this is not the case for your architecture,
1284 * please wrap this in a helper function!
1285 *
1286 * for an example see gup_get_pte in arch/x86/mm/gup.c
1287 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001288 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001289 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001290
1291 /*
1292 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001293 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001294 */
1295 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001296 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001297 goto pte_unmap;
1298
Dave Hansen33a709b2016-02-12 13:02:19 -08001299 if (!arch_pte_access_permitted(pte, write))
1300 goto pte_unmap;
1301
Steve Capper2667f502014-10-09 15:29:14 -07001302 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1303 page = pte_page(pte);
1304
Linus Torvalds2ed768c2019-04-11 10:49:19 -07001305 head = try_get_compound_head(page, 1);
1306 if (!head)
Steve Capper2667f502014-10-09 15:29:14 -07001307 goto pte_unmap;
1308
1309 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001310 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001311 goto pte_unmap;
1312 }
1313
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001314 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001315 pages[*nr] = page;
1316 (*nr)++;
1317
1318 } while (ptep++, addr += PAGE_SIZE, addr != end);
1319
1320 ret = 1;
1321
1322pte_unmap:
1323 pte_unmap(ptem);
1324 return ret;
1325}
1326#else
1327
1328/*
1329 * If we can't determine whether or not a pte is special, then fail immediately
1330 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1331 * to be special.
1332 *
1333 * For a futex to be placed on a THP tail page, get_futex_key requires a
1334 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1335 * useful to have gup_huge_pmd even if we can't operate on ptes.
1336 */
1337static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1338 int write, struct page **pages, int *nr)
1339{
1340 return 0;
1341}
1342#endif /* __HAVE_ARCH_PTE_SPECIAL */
1343
1344static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1345 unsigned long end, int write, struct page **pages, int *nr)
1346{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001347 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001348 int refs;
1349
1350 if (write && !pmd_write(orig))
1351 return 0;
1352
1353 refs = 0;
Punit Agrawal26c02ad2017-07-06 15:39:39 -07001354 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001355 do {
Steve Capper2667f502014-10-09 15:29:14 -07001356 pages[*nr] = page;
1357 (*nr)++;
1358 page++;
1359 refs++;
1360 } while (addr += PAGE_SIZE, addr != end);
1361
Linus Torvalds2ed768c2019-04-11 10:49:19 -07001362 head = try_get_compound_head(pmd_page(orig), refs);
1363 if (!head) {
Steve Capper2667f502014-10-09 15:29:14 -07001364 *nr -= refs;
1365 return 0;
1366 }
1367
1368 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1369 *nr -= refs;
1370 while (refs--)
1371 put_page(head);
1372 return 0;
1373 }
1374
Steve Capper2667f502014-10-09 15:29:14 -07001375 return 1;
1376}
1377
1378static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1379 unsigned long end, int write, struct page **pages, int *nr)
1380{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001381 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001382 int refs;
1383
1384 if (write && !pud_write(orig))
1385 return 0;
1386
1387 refs = 0;
Punit Agrawal26c02ad2017-07-06 15:39:39 -07001388 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001389 do {
Steve Capper2667f502014-10-09 15:29:14 -07001390 pages[*nr] = page;
1391 (*nr)++;
1392 page++;
1393 refs++;
1394 } while (addr += PAGE_SIZE, addr != end);
1395
Linus Torvalds2ed768c2019-04-11 10:49:19 -07001396 head = try_get_compound_head(pud_page(orig), refs);
1397 if (!head) {
Steve Capper2667f502014-10-09 15:29:14 -07001398 *nr -= refs;
1399 return 0;
1400 }
1401
1402 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1403 *nr -= refs;
1404 while (refs--)
1405 put_page(head);
1406 return 0;
1407 }
1408
Steve Capper2667f502014-10-09 15:29:14 -07001409 return 1;
1410}
1411
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301412static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1413 unsigned long end, int write,
1414 struct page **pages, int *nr)
1415{
1416 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001417 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301418
1419 if (write && !pgd_write(orig))
1420 return 0;
1421
1422 refs = 0;
Punit Agrawal26c02ad2017-07-06 15:39:39 -07001423 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301424 do {
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301425 pages[*nr] = page;
1426 (*nr)++;
1427 page++;
1428 refs++;
1429 } while (addr += PAGE_SIZE, addr != end);
1430
Linus Torvalds2ed768c2019-04-11 10:49:19 -07001431 head = try_get_compound_head(pgd_page(orig), refs);
1432 if (!head) {
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301433 *nr -= refs;
1434 return 0;
1435 }
1436
1437 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1438 *nr -= refs;
1439 while (refs--)
1440 put_page(head);
1441 return 0;
1442 }
1443
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301444 return 1;
1445}
1446
Steve Capper2667f502014-10-09 15:29:14 -07001447static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1448 int write, struct page **pages, int *nr)
1449{
1450 unsigned long next;
1451 pmd_t *pmdp;
1452
1453 pmdp = pmd_offset(&pud, addr);
1454 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001455 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001456
1457 next = pmd_addr_end(addr, end);
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -08001458 if (pmd_none(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001459 return 0;
1460
Yu Zhaoc133d8e2019-02-12 15:35:58 -08001461 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
1462 pmd_devmap(pmd))) {
Steve Capper2667f502014-10-09 15:29:14 -07001463 /*
1464 * NUMA hinting faults need to be handled in the GUP
1465 * slowpath for accounting purposes and so that they
1466 * can be serialised against THP migration.
1467 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001468 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001469 return 0;
1470
1471 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1472 pages, nr))
1473 return 0;
1474
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301475 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1476 /*
1477 * architecture have different format for hugetlbfs
1478 * pmd format and THP pmd format
1479 */
1480 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1481 PMD_SHIFT, next, write, pages, nr))
1482 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001483 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1484 return 0;
1485 } while (pmdp++, addr = next, addr != end);
1486
1487 return 1;
1488}
1489
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301490static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1491 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001492{
1493 unsigned long next;
1494 pud_t *pudp;
1495
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301496 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001497 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001498 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001499
1500 next = pud_addr_end(addr, end);
1501 if (pud_none(pud))
1502 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301503 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001504 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301505 pages, nr))
1506 return 0;
1507 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1508 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1509 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001510 return 0;
1511 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1512 return 0;
1513 } while (pudp++, addr = next, addr != end);
1514
1515 return 1;
1516}
1517
1518/*
1519 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1520 * the regular GUP. It will only return non-negative values.
Linus Torvalds0c296402022-01-24 14:41:50 +01001521 *
1522 * Careful, careful! COW breaking can go either way, so a non-write
1523 * access can get ambiguous page results. If you call this function without
1524 * 'write' set, you'd better be sure that you're ok with that ambiguity.
Steve Capper2667f502014-10-09 15:29:14 -07001525 */
1526int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1527 struct page **pages)
1528{
1529 struct mm_struct *mm = current->mm;
1530 unsigned long addr, len, end;
1531 unsigned long next, flags;
1532 pgd_t *pgdp;
1533 int nr = 0;
1534
1535 start &= PAGE_MASK;
1536 addr = start;
1537 len = (unsigned long) nr_pages << PAGE_SHIFT;
1538 end = start + len;
1539
1540 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1541 start, len)))
1542 return 0;
1543
1544 /*
1545 * Disable interrupts. We use the nested form as we can already have
1546 * interrupts disabled by get_futex_key.
1547 *
1548 * With interrupts disabled, we block page table pages from being
1549 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1550 * for more details.
1551 *
1552 * We do not adopt an rcu_read_lock(.) here as we also want to
1553 * block IPIs that come from THPs splitting.
Linus Torvalds0c296402022-01-24 14:41:50 +01001554 *
1555 * NOTE! We allow read-only gup_fast() here, but you'd better be
1556 * careful about possible COW pages. You'll get _a_ COW page, but
1557 * not necessarily the one you intended to get depending on what
1558 * COW event happens after this. COW may break the page copy in a
1559 * random direction.
Steve Capper2667f502014-10-09 15:29:14 -07001560 */
1561
1562 local_irq_save(flags);
1563 pgdp = pgd_offset(mm, addr);
1564 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001565 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301566
Steve Capper2667f502014-10-09 15:29:14 -07001567 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301568 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001569 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301570 if (unlikely(pgd_huge(pgd))) {
Ben Hutchings6fbb8382022-01-24 14:38:28 +01001571 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301572 pages, &nr))
1573 break;
1574 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1575 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
Ben Hutchings6fbb8382022-01-24 14:38:28 +01001576 PGDIR_SHIFT, next, write, pages, &nr))
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301577 break;
Ben Hutchings6fbb8382022-01-24 14:38:28 +01001578 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001579 break;
1580 } while (pgdp++, addr = next, addr != end);
1581 local_irq_restore(flags);
1582
1583 return nr;
1584}
1585
1586/**
1587 * get_user_pages_fast() - pin user pages in memory
1588 * @start: starting user address
1589 * @nr_pages: number of pages from start to pin
1590 * @write: whether pages will be written to
1591 * @pages: array that receives pointers to the pages pinned.
1592 * Should be at least nr_pages long.
1593 *
1594 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1595 * If not successful, it will fall back to taking the lock and
1596 * calling get_user_pages().
1597 *
1598 * Returns number of pages pinned. This may be fewer than the number
1599 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1600 * were pinned, returns -errno.
1601 */
1602int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1603 struct page **pages)
1604{
Steve Capper2667f502014-10-09 15:29:14 -07001605 int nr, ret;
1606
1607 start &= PAGE_MASK;
Linus Torvalds0c296402022-01-24 14:41:50 +01001608 /*
1609 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
1610 * because get_user_pages() may need to cause an early COW in
1611 * order to avoid confusing the normal COW routines. So only
1612 * targets that are already writable are safe to do by just
1613 * looking at the page tables.
1614 */
1615 nr = __get_user_pages_fast(start, nr_pages, 1, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001616 ret = nr;
1617
1618 if (nr < nr_pages) {
1619 /* Try to get the remaining pages with get_user_pages */
1620 start += nr << PAGE_SHIFT;
1621 pages += nr;
1622
Lorenzo Stoakesc1641542016-10-13 01:20:13 +01001623 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1624 write ? FOLL_WRITE : 0);
Steve Capper2667f502014-10-09 15:29:14 -07001625
1626 /* Have to be a bit careful with return values */
1627 if (nr > 0) {
1628 if (ret < 0)
1629 ret = nr;
1630 else
1631 ret += nr;
1632 }
1633 }
1634
1635 return ret;
1636}
1637
1638#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */