blob: 7017abea9fd6b9cb081246eb430d3b43c02d39bc [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>
7#include <linux/pagemap.h>
8#include <linux/rmap.h>
9#include <linux/swap.h>
10#include <linux/swapops.h>
11
Steve Capper2667f502014-10-09 15:29:14 -070012#include <linux/sched.h>
13#include <linux/rwsem.h>
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +053014#include <linux/hugetlb.h>
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070015
Steve Capper2667f502014-10-09 15:29:14 -070016#include <asm/pgtable.h>
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070017#include <asm/tlbflush.h>
Steve Capper2667f502014-10-09 15:29:14 -070018
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070019#include "internal.h"
20
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070021static struct page *no_page_table(struct vm_area_struct *vma,
22 unsigned int flags)
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070023{
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070024 /*
25 * When core dumping an enormous anonymous area that nobody
26 * has touched so far, we don't want to allocate unnecessary pages or
27 * page tables. Return error instead of NULL to skip handle_mm_fault,
28 * then get_dump_page() will return NULL to leave a hole in the dump.
29 * But we can only make this optimization where a hole would surely
30 * be zero-filled if handle_mm_fault() actually did handle it.
31 */
32 if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
33 return ERR_PTR(-EFAULT);
34 return NULL;
35}
36
Kirill A. Shutemov1027e442015-09-04 15:47:55 -070037static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
38 pte_t *pte, unsigned int flags)
39{
40 /* No page to get reference */
41 if (flags & FOLL_GET)
42 return -EFAULT;
43
44 if (flags & FOLL_TOUCH) {
45 pte_t entry = *pte;
46
47 if (flags & FOLL_WRITE)
48 entry = pte_mkdirty(entry);
49 entry = pte_mkyoung(entry);
50
51 if (!pte_same(*pte, entry)) {
52 set_pte_at(vma->vm_mm, address, pte, entry);
53 update_mmu_cache(vma, address, pte);
54 }
55 }
56
57 /* Proper page table entry exists, but no corresponding struct page */
58 return -EEXIST;
59}
60
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070061static struct page *follow_page_pte(struct vm_area_struct *vma,
62 unsigned long address, pmd_t *pmd, unsigned int flags)
63{
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070064 struct mm_struct *mm = vma->vm_mm;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070065 struct page *page;
66 spinlock_t *ptl;
67 pte_t *ptep, pte;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070068
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070069retry:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070070 if (unlikely(pmd_bad(*pmd)))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070071 return no_page_table(vma, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070072
73 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070074 pte = *ptep;
75 if (!pte_present(pte)) {
76 swp_entry_t entry;
77 /*
78 * KSM's break_ksm() relies upon recognizing a ksm page
79 * even while it is being migrated, so for that case we
80 * need migration_entry_wait().
81 */
82 if (likely(!(flags & FOLL_MIGRATION)))
83 goto no_page;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -080084 if (pte_none(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070085 goto no_page;
86 entry = pte_to_swp_entry(pte);
87 if (!is_migration_entry(entry))
88 goto no_page;
89 pte_unmap_unlock(ptep, ptl);
90 migration_entry_wait(mm, pmd, address);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070091 goto retry;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070092 }
Mel Gorman8a0516e2015-02-12 14:58:22 -080093 if ((flags & FOLL_NUMA) && pte_protnone(pte))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070094 goto no_page;
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -070095 if ((flags & FOLL_WRITE) && !pte_write(pte)) {
96 pte_unmap_unlock(ptep, ptl);
97 return NULL;
98 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -070099
100 page = vm_normal_page(vma, address, pte);
101 if (unlikely(!page)) {
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700102 if (flags & FOLL_DUMP) {
103 /* Avoid special (like zero) pages in core dumps */
104 page = ERR_PTR(-EFAULT);
105 goto out;
106 }
107
108 if (is_zero_pfn(pte_pfn(pte))) {
109 page = pte_page(pte);
110 } else {
111 int ret;
112
113 ret = follow_pfn_pte(vma, address, ptep, flags);
114 page = ERR_PTR(ret);
115 goto out;
116 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700117 }
118
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800119 if (flags & FOLL_SPLIT && PageTransCompound(page)) {
120 int ret;
121 get_page(page);
122 pte_unmap_unlock(ptep, ptl);
123 lock_page(page);
124 ret = split_huge_page(page);
125 unlock_page(page);
126 put_page(page);
127 if (ret)
128 return ERR_PTR(ret);
129 goto retry;
130 }
131
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700132 if (flags & FOLL_GET)
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -0800133 get_page(page);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700134 if (flags & FOLL_TOUCH) {
135 if ((flags & FOLL_WRITE) &&
136 !pte_dirty(pte) && !PageDirty(page))
137 set_page_dirty(page);
138 /*
139 * pte_mkyoung() would be more correct here, but atomic care
140 * is needed to avoid losing the dirty bit: it is easier to use
141 * mark_page_accessed().
142 */
143 mark_page_accessed(page);
144 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800145 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700146 /*
147 * The preliminary mapping check is mainly to avoid the
148 * pointless overhead of lock_page on the ZERO_PAGE
149 * which might bounce very badly if there is contention.
150 *
151 * If the page is already locked, we don't need to
152 * handle it now - vmscan will handle it later if and
153 * when it attempts to reclaim the page.
154 */
155 if (page->mapping && trylock_page(page)) {
156 lru_add_drain(); /* push cached pages to LRU */
157 /*
158 * Because we lock page here, and migration is
159 * blocked by the pte's page reference, and we
160 * know the page is still mapped, we don't even
161 * need to check for file-cache page truncation.
162 */
163 mlock_vma_page(page);
164 unlock_page(page);
165 }
166 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700167out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700168 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700169 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700170no_page:
171 pte_unmap_unlock(ptep, ptl);
172 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700173 return NULL;
174 return no_page_table(vma, flags);
175}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700176
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700177/**
178 * follow_page_mask - look up a page descriptor from a user-virtual address
179 * @vma: vm_area_struct mapping @address
180 * @address: virtual address to look up
181 * @flags: flags modifying lookup behaviour
182 * @page_mask: on output, *page_mask is set according to the size of the page
183 *
184 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
185 *
186 * Returns the mapped (struct page *), %NULL if no mapping exists, or
187 * an error pointer if there is a mapping to something not represented
188 * by a page descriptor (see also vm_normal_page()).
189 */
190struct page *follow_page_mask(struct vm_area_struct *vma,
191 unsigned long address, unsigned int flags,
192 unsigned int *page_mask)
193{
194 pgd_t *pgd;
195 pud_t *pud;
196 pmd_t *pmd;
197 spinlock_t *ptl;
198 struct page *page;
199 struct mm_struct *mm = vma->vm_mm;
200
201 *page_mask = 0;
202
203 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
204 if (!IS_ERR(page)) {
205 BUG_ON(flags & FOLL_GET);
206 return page;
207 }
208
209 pgd = pgd_offset(mm, address);
210 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
211 return no_page_table(vma, flags);
212
213 pud = pud_offset(pgd, address);
214 if (pud_none(*pud))
215 return no_page_table(vma, flags);
216 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800217 page = follow_huge_pud(mm, address, pud, flags);
218 if (page)
219 return page;
220 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700221 }
222 if (unlikely(pud_bad(*pud)))
223 return no_page_table(vma, flags);
224
225 pmd = pmd_offset(pud, address);
226 if (pmd_none(*pmd))
227 return no_page_table(vma, flags);
228 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800229 page = follow_huge_pmd(mm, address, pmd, flags);
230 if (page)
231 return page;
232 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700233 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800234 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700235 return no_page_table(vma, flags);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800236 if (likely(!pmd_trans_huge(*pmd)))
237 return follow_page_pte(vma, address, pmd, flags);
238
239 ptl = pmd_lock(mm, pmd);
240 if (unlikely(!pmd_trans_huge(*pmd))) {
241 spin_unlock(ptl);
242 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700243 }
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800244
245 if (unlikely(pmd_trans_splitting(*pmd))) {
246 spin_unlock(ptl);
247 wait_split_huge_page(vma->anon_vma, pmd);
248 return follow_page_pte(vma, address, pmd, flags);
249 }
250
251 if (flags & FOLL_SPLIT) {
252 int ret;
253 page = pmd_page(*pmd);
254 if (is_huge_zero_page(page)) {
255 spin_unlock(ptl);
256 ret = 0;
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -0800257 split_huge_pmd(vma, pmd, address);
Kirill A. Shutemov6742d292016-01-15 16:52:28 -0800258 } else {
259 get_page(page);
260 spin_unlock(ptl);
261 lock_page(page);
262 ret = split_huge_page(page);
263 unlock_page(page);
264 put_page(page);
265 }
266
267 return ret ? ERR_PTR(ret) :
268 follow_page_pte(vma, address, pmd, flags);
269 }
270
271 page = follow_trans_huge_pmd(vma, address, pmd, flags);
272 spin_unlock(ptl);
273 *page_mask = HPAGE_PMD_NR - 1;
274 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700275}
276
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700277static int get_gate_page(struct mm_struct *mm, unsigned long address,
278 unsigned int gup_flags, struct vm_area_struct **vma,
279 struct page **page)
280{
281 pgd_t *pgd;
282 pud_t *pud;
283 pmd_t *pmd;
284 pte_t *pte;
285 int ret = -EFAULT;
286
287 /* user gate pages are read-only */
288 if (gup_flags & FOLL_WRITE)
289 return -EFAULT;
290 if (address > TASK_SIZE)
291 pgd = pgd_offset_k(address);
292 else
293 pgd = pgd_offset_gate(mm, address);
294 BUG_ON(pgd_none(*pgd));
295 pud = pud_offset(pgd, address);
296 BUG_ON(pud_none(*pud));
297 pmd = pmd_offset(pud, address);
298 if (pmd_none(*pmd))
299 return -EFAULT;
300 VM_BUG_ON(pmd_trans_huge(*pmd));
301 pte = pte_offset_map(pmd, address);
302 if (pte_none(*pte))
303 goto unmap;
304 *vma = get_gate_vma(mm);
305 if (!page)
306 goto out;
307 *page = vm_normal_page(*vma, address, *pte);
308 if (!*page) {
309 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
310 goto unmap;
311 *page = pte_page(*pte);
312 }
313 get_page(*page);
314out:
315 ret = 0;
316unmap:
317 pte_unmap(pte);
318 return ret;
319}
320
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700321/*
322 * mmap_sem must be held on entry. If @nonblocking != NULL and
323 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
324 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
325 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700326static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
327 unsigned long address, unsigned int *flags, int *nonblocking)
328{
329 struct mm_struct *mm = vma->vm_mm;
330 unsigned int fault_flags = 0;
331 int ret;
332
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800333 /* mlock all present pages, but do not fault in new pages */
334 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
335 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700336 /* For mm_populate(), just skip the stack guard page. */
337 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700338 (stack_guard_page_start(vma, address) ||
339 stack_guard_page_end(vma, address + PAGE_SIZE)))
340 return -ENOENT;
341 if (*flags & FOLL_WRITE)
342 fault_flags |= FAULT_FLAG_WRITE;
343 if (nonblocking)
344 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
345 if (*flags & FOLL_NOWAIT)
346 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700347 if (*flags & FOLL_TRIED) {
348 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
349 fault_flags |= FAULT_FLAG_TRIED;
350 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700351
352 ret = handle_mm_fault(mm, vma, address, fault_flags);
353 if (ret & VM_FAULT_ERROR) {
354 if (ret & VM_FAULT_OOM)
355 return -ENOMEM;
356 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
357 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800358 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700359 return -EFAULT;
360 BUG();
361 }
362
363 if (tsk) {
364 if (ret & VM_FAULT_MAJOR)
365 tsk->maj_flt++;
366 else
367 tsk->min_flt++;
368 }
369
370 if (ret & VM_FAULT_RETRY) {
371 if (nonblocking)
372 *nonblocking = 0;
373 return -EBUSY;
374 }
375
376 /*
377 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
378 * necessary, even if maybe_mkwrite decided not to set pte_write. We
379 * can thus safely do subsequent page lookups as if they were reads.
380 * But only do so when looping for pte_write is futile: in some cases
381 * userspace may also be wanting to write to the gotten user page,
382 * which a read fault here might prevent (a readonly page might get
383 * reCOWed by userspace write).
384 */
385 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
386 *flags &= ~FOLL_WRITE;
387 return 0;
388}
389
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700390static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
391{
392 vm_flags_t vm_flags = vma->vm_flags;
393
394 if (vm_flags & (VM_IO | VM_PFNMAP))
395 return -EFAULT;
396
397 if (gup_flags & FOLL_WRITE) {
398 if (!(vm_flags & VM_WRITE)) {
399 if (!(gup_flags & FOLL_FORCE))
400 return -EFAULT;
401 /*
402 * We used to let the write,force case do COW in a
403 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
404 * set a breakpoint in a read-only mapping of an
405 * executable, without corrupting the file (yet only
406 * when that file had been opened for writing!).
407 * Anon pages in shared mappings are surprising: now
408 * just reject it.
409 */
410 if (!is_cow_mapping(vm_flags)) {
411 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
412 return -EFAULT;
413 }
414 }
415 } else if (!(vm_flags & VM_READ)) {
416 if (!(gup_flags & FOLL_FORCE))
417 return -EFAULT;
418 /*
419 * Is there actually any vma we can reach here which does not
420 * have VM_MAYREAD set?
421 */
422 if (!(vm_flags & VM_MAYREAD))
423 return -EFAULT;
424 }
425 return 0;
426}
427
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700428/**
429 * __get_user_pages() - pin user pages in memory
430 * @tsk: task_struct of target task
431 * @mm: mm_struct of target mm
432 * @start: starting user address
433 * @nr_pages: number of pages from start to pin
434 * @gup_flags: flags modifying pin behaviour
435 * @pages: array that receives pointers to the pages pinned.
436 * Should be at least nr_pages long. Or NULL, if caller
437 * only intends to ensure the pages are faulted in.
438 * @vmas: array of pointers to vmas corresponding to each page.
439 * Or NULL if the caller does not require them.
440 * @nonblocking: whether waiting for disk IO or mmap_sem contention
441 *
442 * Returns number of pages pinned. This may be fewer than the number
443 * requested. If nr_pages is 0 or negative, returns 0. If no pages
444 * were pinned, returns -errno. Each page returned must be released
445 * with a put_page() call when it is finished with. vmas will only
446 * remain valid while mmap_sem is held.
447 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700448 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700449 *
450 * __get_user_pages walks a process's page tables and takes a reference to
451 * each struct page that each user address corresponds to at a given
452 * instant. That is, it takes the page that would be accessed if a user
453 * thread accesses the given user virtual address at that instant.
454 *
455 * This does not guarantee that the page exists in the user mappings when
456 * __get_user_pages returns, and there may even be a completely different
457 * page there in some cases (eg. if mmapped pagecache has been invalidated
458 * and subsequently re faulted). However it does guarantee that the page
459 * won't be freed completely. And mostly callers simply care that the page
460 * contains data that was valid *at some point in time*. Typically, an IO
461 * or similar operation cannot guarantee anything stronger anyway because
462 * locks can't be held over the syscall boundary.
463 *
464 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
465 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
466 * appropriate) must be called after the page is finished with, and
467 * before put_page is called.
468 *
469 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
470 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700471 * *@nonblocking will be set to 0. Further, if @gup_flags does not
472 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
473 * this case.
474 *
475 * A caller using such a combination of @nonblocking and @gup_flags
476 * must therefore hold the mmap_sem for reading only, and recognize
477 * when it's been released. Otherwise, it must be held for either
478 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700479 *
480 * In most cases, get_user_pages or get_user_pages_fast should be used
481 * instead of __get_user_pages. __get_user_pages should be used only if
482 * you need some special @gup_flags.
483 */
484long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
485 unsigned long start, unsigned long nr_pages,
486 unsigned int gup_flags, struct page **pages,
487 struct vm_area_struct **vmas, int *nonblocking)
488{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700489 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700490 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700491 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700492
493 if (!nr_pages)
494 return 0;
495
496 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
497
498 /*
499 * If FOLL_FORCE is set then do not force a full fault as the hinting
500 * fault information is unrelated to the reference behaviour of a task
501 * using the address space
502 */
503 if (!(gup_flags & FOLL_FORCE))
504 gup_flags |= FOLL_NUMA;
505
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700506 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700507 struct page *page;
508 unsigned int foll_flags = gup_flags;
509 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700510
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700511 /* first iteration or cross vma bound */
512 if (!vma || start >= vma->vm_end) {
513 vma = find_extend_vma(mm, start);
514 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700515 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700516 ret = get_gate_page(mm, start & PAGE_MASK,
517 gup_flags, &vma,
518 pages ? &pages[i] : NULL);
519 if (ret)
520 return i ? : ret;
521 page_mask = 0;
522 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700523 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700524
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700525 if (!vma || check_vma_flags(vma, gup_flags))
526 return i ? : -EFAULT;
527 if (is_vm_hugetlb_page(vma)) {
528 i = follow_hugetlb_page(mm, vma, pages, vmas,
529 &start, &nr_pages, i,
530 gup_flags);
531 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700532 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700533 }
534retry:
535 /*
536 * If we have a pending SIGKILL, don't keep faulting pages and
537 * potentially allocating memory.
538 */
539 if (unlikely(fatal_signal_pending(current)))
540 return i ? i : -ERESTARTSYS;
541 cond_resched();
542 page = follow_page_mask(vma, start, foll_flags, &page_mask);
543 if (!page) {
544 int ret;
545 ret = faultin_page(tsk, vma, start, &foll_flags,
546 nonblocking);
547 switch (ret) {
548 case 0:
549 goto retry;
550 case -EFAULT:
551 case -ENOMEM:
552 case -EHWPOISON:
553 return i ? i : ret;
554 case -EBUSY:
555 return i;
556 case -ENOENT:
557 goto next_page;
558 }
559 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700560 } else if (PTR_ERR(page) == -EEXIST) {
561 /*
562 * Proper page table entry exists, but no corresponding
563 * struct page.
564 */
565 goto next_page;
566 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700567 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700568 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700569 if (pages) {
570 pages[i] = page;
571 flush_anon_page(vma, page, start);
572 flush_dcache_page(page);
573 page_mask = 0;
574 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700575next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700576 if (vmas) {
577 vmas[i] = vma;
578 page_mask = 0;
579 }
580 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
581 if (page_increm > nr_pages)
582 page_increm = nr_pages;
583 i += page_increm;
584 start += page_increm * PAGE_SIZE;
585 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700586 } while (nr_pages);
587 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700588}
589EXPORT_SYMBOL(__get_user_pages);
590
591/*
592 * fixup_user_fault() - manually resolve a user page fault
593 * @tsk: the task_struct to use for page fault accounting, or
594 * NULL if faults are not to be recorded.
595 * @mm: mm_struct of target mm
596 * @address: user address
597 * @fault_flags:flags to pass down to handle_mm_fault()
598 *
599 * This is meant to be called in the specific scenario where for locking reasons
600 * we try to access user memory in atomic context (within a pagefault_disable()
601 * section), this returns -EFAULT, and we want to resolve the user fault before
602 * trying again.
603 *
604 * Typically this is meant to be used by the futex code.
605 *
606 * The main difference with get_user_pages() is that this function will
607 * unconditionally call handle_mm_fault() which will in turn perform all the
608 * necessary SW fixup of the dirty and young bits in the PTE, while
609 * handle_mm_fault() only guarantees to update these in the struct page.
610 *
611 * This is important for some architectures where those bits also gate the
612 * access permission to the page because they are maintained in software. On
613 * such architectures, gup() will not be enough to make a subsequent access
614 * succeed.
615 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700616 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700617 */
618int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
619 unsigned long address, unsigned int fault_flags)
620{
621 struct vm_area_struct *vma;
622 vm_flags_t vm_flags;
623 int ret;
624
625 vma = find_extend_vma(mm, address);
626 if (!vma || address < vma->vm_start)
627 return -EFAULT;
628
629 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
630 if (!(vm_flags & vma->vm_flags))
631 return -EFAULT;
632
633 ret = handle_mm_fault(mm, vma, address, fault_flags);
634 if (ret & VM_FAULT_ERROR) {
635 if (ret & VM_FAULT_OOM)
636 return -ENOMEM;
637 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
638 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800639 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700640 return -EFAULT;
641 BUG();
642 }
643 if (tsk) {
644 if (ret & VM_FAULT_MAJOR)
645 tsk->maj_flt++;
646 else
647 tsk->min_flt++;
648 }
649 return 0;
650}
651
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800652static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
653 struct mm_struct *mm,
654 unsigned long start,
655 unsigned long nr_pages,
656 int write, int force,
657 struct page **pages,
658 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800659 int *locked, bool notify_drop,
660 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800661{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800662 long ret, pages_done;
663 bool lock_dropped;
664
665 if (locked) {
666 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
667 BUG_ON(vmas);
668 /* check caller initialized locked */
669 BUG_ON(*locked != 1);
670 }
671
672 if (pages)
673 flags |= FOLL_GET;
674 if (write)
675 flags |= FOLL_WRITE;
676 if (force)
677 flags |= FOLL_FORCE;
678
679 pages_done = 0;
680 lock_dropped = false;
681 for (;;) {
682 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
683 vmas, locked);
684 if (!locked)
685 /* VM_FAULT_RETRY couldn't trigger, bypass */
686 return ret;
687
688 /* VM_FAULT_RETRY cannot return errors */
689 if (!*locked) {
690 BUG_ON(ret < 0);
691 BUG_ON(ret >= nr_pages);
692 }
693
694 if (!pages)
695 /* If it's a prefault don't insist harder */
696 return ret;
697
698 if (ret > 0) {
699 nr_pages -= ret;
700 pages_done += ret;
701 if (!nr_pages)
702 break;
703 }
704 if (*locked) {
705 /* VM_FAULT_RETRY didn't trigger */
706 if (!pages_done)
707 pages_done = ret;
708 break;
709 }
710 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
711 pages += ret;
712 start += ret << PAGE_SHIFT;
713
714 /*
715 * Repeat on the address that fired VM_FAULT_RETRY
716 * without FAULT_FLAG_ALLOW_RETRY but with
717 * FAULT_FLAG_TRIED.
718 */
719 *locked = 1;
720 lock_dropped = true;
721 down_read(&mm->mmap_sem);
722 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
723 pages, NULL, NULL);
724 if (ret != 1) {
725 BUG_ON(ret > 1);
726 if (!pages_done)
727 pages_done = ret;
728 break;
729 }
730 nr_pages--;
731 pages_done++;
732 if (!nr_pages)
733 break;
734 pages++;
735 start += PAGE_SIZE;
736 }
737 if (notify_drop && lock_dropped && *locked) {
738 /*
739 * We must let the caller know we temporarily dropped the lock
740 * and so the critical section protected by it was lost.
741 */
742 up_read(&mm->mmap_sem);
743 *locked = 0;
744 }
745 return pages_done;
746}
747
748/*
749 * We can leverage the VM_FAULT_RETRY functionality in the page fault
750 * paths better by using either get_user_pages_locked() or
751 * get_user_pages_unlocked().
752 *
753 * get_user_pages_locked() is suitable to replace the form:
754 *
755 * down_read(&mm->mmap_sem);
756 * do_something()
757 * get_user_pages(tsk, mm, ..., pages, NULL);
758 * up_read(&mm->mmap_sem);
759 *
760 * to:
761 *
762 * int locked = 1;
763 * down_read(&mm->mmap_sem);
764 * do_something()
765 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
766 * if (locked)
767 * up_read(&mm->mmap_sem);
768 */
769long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
770 unsigned long start, unsigned long nr_pages,
771 int write, int force, struct page **pages,
772 int *locked)
773{
774 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800775 pages, NULL, locked, true, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800776}
777EXPORT_SYMBOL(get_user_pages_locked);
778
779/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800780 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
781 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
782 *
783 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
784 * caller if required (just like with __get_user_pages). "FOLL_GET",
785 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
786 * according to the parameters "pages", "write", "force"
787 * respectively.
788 */
789__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
790 unsigned long start, unsigned long nr_pages,
791 int write, int force, struct page **pages,
792 unsigned int gup_flags)
793{
794 long ret;
795 int locked = 1;
796 down_read(&mm->mmap_sem);
797 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
798 pages, NULL, &locked, false, gup_flags);
799 if (locked)
800 up_read(&mm->mmap_sem);
801 return ret;
802}
803EXPORT_SYMBOL(__get_user_pages_unlocked);
804
805/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800806 * get_user_pages_unlocked() is suitable to replace the form:
807 *
808 * down_read(&mm->mmap_sem);
809 * get_user_pages(tsk, mm, ..., pages, NULL);
810 * up_read(&mm->mmap_sem);
811 *
812 * with:
813 *
814 * get_user_pages_unlocked(tsk, mm, ..., pages);
815 *
816 * It is functionally equivalent to get_user_pages_fast so
817 * get_user_pages_fast should be used instead, if the two parameters
818 * "tsk" and "mm" are respectively equal to current and current->mm,
819 * or if "force" shall be set to 1 (get_user_pages_fast misses the
820 * "force" parameter).
821 */
822long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
823 unsigned long start, unsigned long nr_pages,
824 int write, int force, struct page **pages)
825{
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800826 return __get_user_pages_unlocked(tsk, mm, start, nr_pages, write,
827 force, pages, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800828}
829EXPORT_SYMBOL(get_user_pages_unlocked);
830
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700831/*
832 * get_user_pages() - pin user pages in memory
833 * @tsk: the task_struct to use for page fault accounting, or
834 * NULL if faults are not to be recorded.
835 * @mm: mm_struct of target mm
836 * @start: starting user address
837 * @nr_pages: number of pages from start to pin
838 * @write: whether pages will be written to by the caller
839 * @force: whether to force access even when user mapping is currently
840 * protected (but never forces write access to shared mapping).
841 * @pages: array that receives pointers to the pages pinned.
842 * Should be at least nr_pages long. Or NULL, if caller
843 * only intends to ensure the pages are faulted in.
844 * @vmas: array of pointers to vmas corresponding to each page.
845 * Or NULL if the caller does not require them.
846 *
847 * Returns number of pages pinned. This may be fewer than the number
848 * requested. If nr_pages is 0 or negative, returns 0. If no pages
849 * were pinned, returns -errno. Each page returned must be released
850 * with a put_page() call when it is finished with. vmas will only
851 * remain valid while mmap_sem is held.
852 *
853 * Must be called with mmap_sem held for read or write.
854 *
855 * get_user_pages walks a process's page tables and takes a reference to
856 * each struct page that each user address corresponds to at a given
857 * instant. That is, it takes the page that would be accessed if a user
858 * thread accesses the given user virtual address at that instant.
859 *
860 * This does not guarantee that the page exists in the user mappings when
861 * get_user_pages returns, and there may even be a completely different
862 * page there in some cases (eg. if mmapped pagecache has been invalidated
863 * and subsequently re faulted). However it does guarantee that the page
864 * won't be freed completely. And mostly callers simply care that the page
865 * contains data that was valid *at some point in time*. Typically, an IO
866 * or similar operation cannot guarantee anything stronger anyway because
867 * locks can't be held over the syscall boundary.
868 *
869 * If write=0, the page must not be written to. If the page is written to,
870 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
871 * after the page is finished with, and before put_page is called.
872 *
873 * get_user_pages is typically used for fewer-copy IO operations, to get a
874 * handle on the memory by some means other than accesses via the user virtual
875 * addresses. The pages may be submitted for DMA to devices or accessed via
876 * their kernel linear mapping (via the kmap APIs). Care should be taken to
877 * use the correct cache flushing APIs.
878 *
879 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800880 *
881 * get_user_pages should be phased out in favor of
882 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
883 * should use get_user_pages because it cannot pass
884 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700885 */
886long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
887 unsigned long start, unsigned long nr_pages, int write,
888 int force, struct page **pages, struct vm_area_struct **vmas)
889{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800890 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800891 pages, vmas, NULL, false, FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700892}
893EXPORT_SYMBOL(get_user_pages);
894
895/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700896 * populate_vma_page_range() - populate a range of pages in the vma.
897 * @vma: target vma
898 * @start: start address
899 * @end: end address
900 * @nonblocking:
901 *
902 * This takes care of mlocking the pages too if VM_LOCKED is set.
903 *
904 * return 0 on success, negative error code on error.
905 *
906 * vma->vm_mm->mmap_sem must be held.
907 *
908 * If @nonblocking is NULL, it may be held for read or write and will
909 * be unperturbed.
910 *
911 * If @nonblocking is non-NULL, it must held for read only and may be
912 * released. If it's released, *@nonblocking will be set to 0.
913 */
914long populate_vma_page_range(struct vm_area_struct *vma,
915 unsigned long start, unsigned long end, int *nonblocking)
916{
917 struct mm_struct *mm = vma->vm_mm;
918 unsigned long nr_pages = (end - start) / PAGE_SIZE;
919 int gup_flags;
920
921 VM_BUG_ON(start & ~PAGE_MASK);
922 VM_BUG_ON(end & ~PAGE_MASK);
923 VM_BUG_ON_VMA(start < vma->vm_start, vma);
924 VM_BUG_ON_VMA(end > vma->vm_end, vma);
925 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
926
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800927 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
928 if (vma->vm_flags & VM_LOCKONFAULT)
929 gup_flags &= ~FOLL_POPULATE;
Kirill A. Shutemov7479df62016-01-15 16:52:35 -0800930 if (vma->vm_flags & VM_LOCKED)
931 gup_flags |= FOLL_SPLIT;
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700932 /*
933 * We want to touch writable mappings with a write fault in order
934 * to break COW, except for shared mappings because these don't COW
935 * and we would not want to dirty them for nothing.
936 */
937 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
938 gup_flags |= FOLL_WRITE;
939
940 /*
941 * We want mlock to succeed for regions that have any permissions
942 * other than PROT_NONE.
943 */
944 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
945 gup_flags |= FOLL_FORCE;
946
947 /*
948 * We made sure addr is within a VMA, so the following will
949 * not result in a stack expansion that recurses back here.
950 */
951 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
952 NULL, NULL, nonblocking);
953}
954
955/*
956 * __mm_populate - populate and/or mlock pages within a range of address space.
957 *
958 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
959 * flags. VMAs must be already marked with the desired vm_flags, and
960 * mmap_sem must not be held.
961 */
962int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
963{
964 struct mm_struct *mm = current->mm;
965 unsigned long end, nstart, nend;
966 struct vm_area_struct *vma = NULL;
967 int locked = 0;
968 long ret = 0;
969
970 VM_BUG_ON(start & ~PAGE_MASK);
971 VM_BUG_ON(len != PAGE_ALIGN(len));
972 end = start + len;
973
974 for (nstart = start; nstart < end; nstart = nend) {
975 /*
976 * We want to fault in pages for [nstart; end) address range.
977 * Find first corresponding VMA.
978 */
979 if (!locked) {
980 locked = 1;
981 down_read(&mm->mmap_sem);
982 vma = find_vma(mm, nstart);
983 } else if (nstart >= vma->vm_end)
984 vma = vma->vm_next;
985 if (!vma || vma->vm_start >= end)
986 break;
987 /*
988 * Set [nstart; nend) to intersection of desired address
989 * range with the first VMA. Also, skip undesirable VMA types.
990 */
991 nend = min(end, vma->vm_end);
992 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
993 continue;
994 if (nstart < vma->vm_start)
995 nstart = vma->vm_start;
996 /*
997 * Now fault in a range of pages. populate_vma_page_range()
998 * double checks the vma flags, so that it won't mlock pages
999 * if the vma was already munlocked.
1000 */
1001 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1002 if (ret < 0) {
1003 if (ignore_errors) {
1004 ret = 0;
1005 continue; /* continue at next VMA */
1006 }
1007 break;
1008 }
1009 nend = nstart + ret * PAGE_SIZE;
1010 ret = 0;
1011 }
1012 if (locked)
1013 up_read(&mm->mmap_sem);
1014 return ret; /* 0 or negative error code */
1015}
1016
1017/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -07001018 * get_dump_page() - pin user page in memory while writing it to core dump
1019 * @addr: user address
1020 *
1021 * Returns struct page pointer of user page pinned for dump,
1022 * to be freed afterwards by page_cache_release() or put_page().
1023 *
1024 * Returns NULL on any kind of failure - a hole must then be inserted into
1025 * the corefile, to preserve alignment with its headers; and also returns
1026 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1027 * allowing a hole to be left in the corefile to save diskspace.
1028 *
1029 * Called without mmap_sem, but after all other threads have been killed.
1030 */
1031#ifdef CONFIG_ELF_CORE
1032struct page *get_dump_page(unsigned long addr)
1033{
1034 struct vm_area_struct *vma;
1035 struct page *page;
1036
1037 if (__get_user_pages(current, current->mm, addr, 1,
1038 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1039 NULL) < 1)
1040 return NULL;
1041 flush_cache_page(vma, addr, page_to_pfn(page));
1042 return page;
1043}
1044#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001045
1046/*
1047 * Generic RCU Fast GUP
1048 *
1049 * get_user_pages_fast attempts to pin user pages by walking the page
1050 * tables directly and avoids taking locks. Thus the walker needs to be
1051 * protected from page table pages being freed from under it, and should
1052 * block any THP splits.
1053 *
1054 * One way to achieve this is to have the walker disable interrupts, and
1055 * rely on IPIs from the TLB flushing code blocking before the page table
1056 * pages are freed. This is unsuitable for architectures that do not need
1057 * to broadcast an IPI when invalidating TLBs.
1058 *
1059 * Another way to achieve this is to batch up page table containing pages
1060 * belonging to more than one mm_user, then rcu_sched a callback to free those
1061 * pages. Disabling interrupts will allow the fast_gup walker to both block
1062 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1063 * (which is a relatively rare event). The code below adopts this strategy.
1064 *
1065 * Before activating this code, please be aware that the following assumptions
1066 * are currently made:
1067 *
1068 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1069 * pages containing page tables.
1070 *
1071 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1072 * pmdp_splitting_flush.
1073 *
1074 * *) ptes can be read atomically by the architecture.
1075 *
1076 * *) access_ok is sufficient to validate userspace address ranges.
1077 *
1078 * The last two assumptions can be relaxed by the addition of helper functions.
1079 *
1080 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1081 */
1082#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1083
1084#ifdef __HAVE_ARCH_PTE_SPECIAL
1085static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1086 int write, struct page **pages, int *nr)
1087{
1088 pte_t *ptep, *ptem;
1089 int ret = 0;
1090
1091 ptem = ptep = pte_offset_map(&pmd, addr);
1092 do {
1093 /*
1094 * In the line below we are assuming that the pte can be read
1095 * atomically. If this is not the case for your architecture,
1096 * please wrap this in a helper function!
1097 *
1098 * for an example see gup_get_pte in arch/x86/mm/gup.c
1099 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001100 pte_t pte = READ_ONCE(*ptep);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001101 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001102
1103 /*
1104 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001105 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001106 */
1107 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001108 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001109 goto pte_unmap;
1110
1111 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1112 page = pte_page(pte);
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001113 head = compound_head(page);
Steve Capper2667f502014-10-09 15:29:14 -07001114
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001115 if (!page_cache_get_speculative(head))
Steve Capper2667f502014-10-09 15:29:14 -07001116 goto pte_unmap;
1117
1118 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001119 put_page(head);
Steve Capper2667f502014-10-09 15:29:14 -07001120 goto pte_unmap;
1121 }
1122
Kirill A. Shutemov7aef4172016-01-15 16:52:32 -08001123 VM_BUG_ON_PAGE(compound_head(page) != head, page);
Steve Capper2667f502014-10-09 15:29:14 -07001124 pages[*nr] = page;
1125 (*nr)++;
1126
1127 } while (ptep++, addr += PAGE_SIZE, addr != end);
1128
1129 ret = 1;
1130
1131pte_unmap:
1132 pte_unmap(ptem);
1133 return ret;
1134}
1135#else
1136
1137/*
1138 * If we can't determine whether or not a pte is special, then fail immediately
1139 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1140 * to be special.
1141 *
1142 * For a futex to be placed on a THP tail page, get_futex_key requires a
1143 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1144 * useful to have gup_huge_pmd even if we can't operate on ptes.
1145 */
1146static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1147 int write, struct page **pages, int *nr)
1148{
1149 return 0;
1150}
1151#endif /* __HAVE_ARCH_PTE_SPECIAL */
1152
1153static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1154 unsigned long end, int write, struct page **pages, int *nr)
1155{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001156 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001157 int refs;
1158
1159 if (write && !pmd_write(orig))
1160 return 0;
1161
1162 refs = 0;
1163 head = pmd_page(orig);
1164 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001165 do {
1166 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1167 pages[*nr] = page;
1168 (*nr)++;
1169 page++;
1170 refs++;
1171 } while (addr += PAGE_SIZE, addr != end);
1172
1173 if (!page_cache_add_speculative(head, refs)) {
1174 *nr -= refs;
1175 return 0;
1176 }
1177
1178 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1179 *nr -= refs;
1180 while (refs--)
1181 put_page(head);
1182 return 0;
1183 }
1184
Steve Capper2667f502014-10-09 15:29:14 -07001185 return 1;
1186}
1187
1188static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1189 unsigned long end, int write, struct page **pages, int *nr)
1190{
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001191 struct page *head, *page;
Steve Capper2667f502014-10-09 15:29:14 -07001192 int refs;
1193
1194 if (write && !pud_write(orig))
1195 return 0;
1196
1197 refs = 0;
1198 head = pud_page(orig);
1199 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper2667f502014-10-09 15:29:14 -07001200 do {
1201 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1202 pages[*nr] = page;
1203 (*nr)++;
1204 page++;
1205 refs++;
1206 } while (addr += PAGE_SIZE, addr != end);
1207
1208 if (!page_cache_add_speculative(head, refs)) {
1209 *nr -= refs;
1210 return 0;
1211 }
1212
1213 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1214 *nr -= refs;
1215 while (refs--)
1216 put_page(head);
1217 return 0;
1218 }
1219
Steve Capper2667f502014-10-09 15:29:14 -07001220 return 1;
1221}
1222
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301223static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1224 unsigned long end, int write,
1225 struct page **pages, int *nr)
1226{
1227 int refs;
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08001228 struct page *head, *page;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301229
1230 if (write && !pgd_write(orig))
1231 return 0;
1232
1233 refs = 0;
1234 head = pgd_page(orig);
1235 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301236 do {
1237 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1238 pages[*nr] = page;
1239 (*nr)++;
1240 page++;
1241 refs++;
1242 } while (addr += PAGE_SIZE, addr != end);
1243
1244 if (!page_cache_add_speculative(head, refs)) {
1245 *nr -= refs;
1246 return 0;
1247 }
1248
1249 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1250 *nr -= refs;
1251 while (refs--)
1252 put_page(head);
1253 return 0;
1254 }
1255
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301256 return 1;
1257}
1258
Steve Capper2667f502014-10-09 15:29:14 -07001259static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1260 int write, struct page **pages, int *nr)
1261{
1262 unsigned long next;
1263 pmd_t *pmdp;
1264
1265 pmdp = pmd_offset(&pud, addr);
1266 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001267 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001268
1269 next = pmd_addr_end(addr, end);
1270 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1271 return 0;
1272
1273 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1274 /*
1275 * NUMA hinting faults need to be handled in the GUP
1276 * slowpath for accounting purposes and so that they
1277 * can be serialised against THP migration.
1278 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001279 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001280 return 0;
1281
1282 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1283 pages, nr))
1284 return 0;
1285
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301286 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1287 /*
1288 * architecture have different format for hugetlbfs
1289 * pmd format and THP pmd format
1290 */
1291 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1292 PMD_SHIFT, next, write, pages, nr))
1293 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001294 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1295 return 0;
1296 } while (pmdp++, addr = next, addr != end);
1297
1298 return 1;
1299}
1300
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301301static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1302 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001303{
1304 unsigned long next;
1305 pud_t *pudp;
1306
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301307 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001308 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001309 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001310
1311 next = pud_addr_end(addr, end);
1312 if (pud_none(pud))
1313 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301314 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001315 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301316 pages, nr))
1317 return 0;
1318 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1319 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1320 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001321 return 0;
1322 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1323 return 0;
1324 } while (pudp++, addr = next, addr != end);
1325
1326 return 1;
1327}
1328
1329/*
1330 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1331 * the regular GUP. It will only return non-negative values.
1332 */
1333int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1334 struct page **pages)
1335{
1336 struct mm_struct *mm = current->mm;
1337 unsigned long addr, len, end;
1338 unsigned long next, flags;
1339 pgd_t *pgdp;
1340 int nr = 0;
1341
1342 start &= PAGE_MASK;
1343 addr = start;
1344 len = (unsigned long) nr_pages << PAGE_SHIFT;
1345 end = start + len;
1346
1347 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1348 start, len)))
1349 return 0;
1350
1351 /*
1352 * Disable interrupts. We use the nested form as we can already have
1353 * interrupts disabled by get_futex_key.
1354 *
1355 * With interrupts disabled, we block page table pages from being
1356 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1357 * for more details.
1358 *
1359 * We do not adopt an rcu_read_lock(.) here as we also want to
1360 * block IPIs that come from THPs splitting.
1361 */
1362
1363 local_irq_save(flags);
1364 pgdp = pgd_offset(mm, addr);
1365 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001366 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301367
Steve Capper2667f502014-10-09 15:29:14 -07001368 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301369 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001370 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301371 if (unlikely(pgd_huge(pgd))) {
1372 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1373 pages, &nr))
1374 break;
1375 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1376 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1377 PGDIR_SHIFT, next, write, pages, &nr))
1378 break;
1379 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001380 break;
1381 } while (pgdp++, addr = next, addr != end);
1382 local_irq_restore(flags);
1383
1384 return nr;
1385}
1386
1387/**
1388 * get_user_pages_fast() - pin user pages in memory
1389 * @start: starting user address
1390 * @nr_pages: number of pages from start to pin
1391 * @write: whether pages will be written to
1392 * @pages: array that receives pointers to the pages pinned.
1393 * Should be at least nr_pages long.
1394 *
1395 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1396 * If not successful, it will fall back to taking the lock and
1397 * calling get_user_pages().
1398 *
1399 * Returns number of pages pinned. This may be fewer than the number
1400 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1401 * were pinned, returns -errno.
1402 */
1403int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1404 struct page **pages)
1405{
1406 struct mm_struct *mm = current->mm;
1407 int nr, ret;
1408
1409 start &= PAGE_MASK;
1410 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1411 ret = nr;
1412
1413 if (nr < nr_pages) {
1414 /* Try to get the remaining pages with get_user_pages */
1415 start += nr << PAGE_SHIFT;
1416 pages += nr;
1417
Andrea Arcangelia7b78072015-02-11 15:27:23 -08001418 ret = get_user_pages_unlocked(current, mm, start,
1419 nr_pages - nr, write, 0, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001420
1421 /* Have to be a bit careful with return values */
1422 if (nr > 0) {
1423 if (ret < 0)
1424 ret = nr;
1425 else
1426 ret += nr;
1427 }
1428 }
1429
1430 return ret;
1431}
1432
1433#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */