blob: deafa2c91b362206b7ef56aec918509a9cc24dd2 [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
119 if (flags & FOLL_GET)
120 get_page_foll(page);
121 if (flags & FOLL_TOUCH) {
122 if ((flags & FOLL_WRITE) &&
123 !pte_dirty(pte) && !PageDirty(page))
124 set_page_dirty(page);
125 /*
126 * pte_mkyoung() would be more correct here, but atomic care
127 * is needed to avoid losing the dirty bit: it is easier to use
128 * mark_page_accessed().
129 */
130 mark_page_accessed(page);
131 }
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800132 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700133 /*
134 * The preliminary mapping check is mainly to avoid the
135 * pointless overhead of lock_page on the ZERO_PAGE
136 * which might bounce very badly if there is contention.
137 *
138 * If the page is already locked, we don't need to
139 * handle it now - vmscan will handle it later if and
140 * when it attempts to reclaim the page.
141 */
142 if (page->mapping && trylock_page(page)) {
143 lru_add_drain(); /* push cached pages to LRU */
144 /*
145 * Because we lock page here, and migration is
146 * blocked by the pte's page reference, and we
147 * know the page is still mapped, we don't even
148 * need to check for file-cache page truncation.
149 */
150 mlock_vma_page(page);
151 unlock_page(page);
152 }
153 }
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700154out:
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700155 pte_unmap_unlock(ptep, ptl);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700156 return page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700157no_page:
158 pte_unmap_unlock(ptep, ptl);
159 if (!pte_none(pte))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700160 return NULL;
161 return no_page_table(vma, flags);
162}
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700163
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700164/**
165 * follow_page_mask - look up a page descriptor from a user-virtual address
166 * @vma: vm_area_struct mapping @address
167 * @address: virtual address to look up
168 * @flags: flags modifying lookup behaviour
169 * @page_mask: on output, *page_mask is set according to the size of the page
170 *
171 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
172 *
173 * Returns the mapped (struct page *), %NULL if no mapping exists, or
174 * an error pointer if there is a mapping to something not represented
175 * by a page descriptor (see also vm_normal_page()).
176 */
177struct page *follow_page_mask(struct vm_area_struct *vma,
178 unsigned long address, unsigned int flags,
179 unsigned int *page_mask)
180{
181 pgd_t *pgd;
182 pud_t *pud;
183 pmd_t *pmd;
184 spinlock_t *ptl;
185 struct page *page;
186 struct mm_struct *mm = vma->vm_mm;
187
188 *page_mask = 0;
189
190 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
191 if (!IS_ERR(page)) {
192 BUG_ON(flags & FOLL_GET);
193 return page;
194 }
195
196 pgd = pgd_offset(mm, address);
197 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
198 return no_page_table(vma, flags);
199
200 pud = pud_offset(pgd, address);
201 if (pud_none(*pud))
202 return no_page_table(vma, flags);
203 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800204 page = follow_huge_pud(mm, address, pud, flags);
205 if (page)
206 return page;
207 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700208 }
209 if (unlikely(pud_bad(*pud)))
210 return no_page_table(vma, flags);
211
212 pmd = pmd_offset(pud, address);
213 if (pmd_none(*pmd))
214 return no_page_table(vma, flags);
215 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -0800216 page = follow_huge_pmd(mm, address, pmd, flags);
217 if (page)
218 return page;
219 return no_page_table(vma, flags);
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700220 }
Mel Gorman8a0516e2015-02-12 14:58:22 -0800221 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
Kirill A. Shutemov69e68b42014-06-04 16:08:11 -0700222 return no_page_table(vma, flags);
223 if (pmd_trans_huge(*pmd)) {
224 if (flags & FOLL_SPLIT) {
225 split_huge_page_pmd(vma, address, pmd);
226 return follow_page_pte(vma, address, pmd, flags);
227 }
228 ptl = pmd_lock(mm, pmd);
229 if (likely(pmd_trans_huge(*pmd))) {
230 if (unlikely(pmd_trans_splitting(*pmd))) {
231 spin_unlock(ptl);
232 wait_split_huge_page(vma->anon_vma, pmd);
233 } else {
234 page = follow_trans_huge_pmd(vma, address,
235 pmd, flags);
236 spin_unlock(ptl);
237 *page_mask = HPAGE_PMD_NR - 1;
238 return page;
239 }
240 } else
241 spin_unlock(ptl);
242 }
243 return follow_page_pte(vma, address, pmd, flags);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700244}
245
Kirill A. Shutemovf2b495c2014-06-04 16:08:11 -0700246static int get_gate_page(struct mm_struct *mm, unsigned long address,
247 unsigned int gup_flags, struct vm_area_struct **vma,
248 struct page **page)
249{
250 pgd_t *pgd;
251 pud_t *pud;
252 pmd_t *pmd;
253 pte_t *pte;
254 int ret = -EFAULT;
255
256 /* user gate pages are read-only */
257 if (gup_flags & FOLL_WRITE)
258 return -EFAULT;
259 if (address > TASK_SIZE)
260 pgd = pgd_offset_k(address);
261 else
262 pgd = pgd_offset_gate(mm, address);
263 BUG_ON(pgd_none(*pgd));
264 pud = pud_offset(pgd, address);
265 BUG_ON(pud_none(*pud));
266 pmd = pmd_offset(pud, address);
267 if (pmd_none(*pmd))
268 return -EFAULT;
269 VM_BUG_ON(pmd_trans_huge(*pmd));
270 pte = pte_offset_map(pmd, address);
271 if (pte_none(*pte))
272 goto unmap;
273 *vma = get_gate_vma(mm);
274 if (!page)
275 goto out;
276 *page = vm_normal_page(*vma, address, *pte);
277 if (!*page) {
278 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
279 goto unmap;
280 *page = pte_page(*pte);
281 }
282 get_page(*page);
283out:
284 ret = 0;
285unmap:
286 pte_unmap(pte);
287 return ret;
288}
289
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700290/*
291 * mmap_sem must be held on entry. If @nonblocking != NULL and
292 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
293 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
294 */
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700295static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
296 unsigned long address, unsigned int *flags, int *nonblocking)
297{
298 struct mm_struct *mm = vma->vm_mm;
299 unsigned int fault_flags = 0;
300 int ret;
301
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800302 /* mlock all present pages, but do not fault in new pages */
303 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
304 return -ENOENT;
Kirill A. Shutemov84d33df2015-04-14 15:44:37 -0700305 /* For mm_populate(), just skip the stack guard page. */
306 if ((*flags & FOLL_POPULATE) &&
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700307 (stack_guard_page_start(vma, address) ||
308 stack_guard_page_end(vma, address + PAGE_SIZE)))
309 return -ENOENT;
310 if (*flags & FOLL_WRITE)
311 fault_flags |= FAULT_FLAG_WRITE;
312 if (nonblocking)
313 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
314 if (*flags & FOLL_NOWAIT)
315 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
Andres Lagar-Cavilla234b2392014-09-17 10:51:48 -0700316 if (*flags & FOLL_TRIED) {
317 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
318 fault_flags |= FAULT_FLAG_TRIED;
319 }
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700320
321 ret = handle_mm_fault(mm, vma, address, fault_flags);
322 if (ret & VM_FAULT_ERROR) {
323 if (ret & VM_FAULT_OOM)
324 return -ENOMEM;
325 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
326 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
Linus Torvalds33692f22015-01-29 10:51:32 -0800327 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov16744482014-06-04 16:08:12 -0700328 return -EFAULT;
329 BUG();
330 }
331
332 if (tsk) {
333 if (ret & VM_FAULT_MAJOR)
334 tsk->maj_flt++;
335 else
336 tsk->min_flt++;
337 }
338
339 if (ret & VM_FAULT_RETRY) {
340 if (nonblocking)
341 *nonblocking = 0;
342 return -EBUSY;
343 }
344
345 /*
346 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
347 * necessary, even if maybe_mkwrite decided not to set pte_write. We
348 * can thus safely do subsequent page lookups as if they were reads.
349 * But only do so when looping for pte_write is futile: in some cases
350 * userspace may also be wanting to write to the gotten user page,
351 * which a read fault here might prevent (a readonly page might get
352 * reCOWed by userspace write).
353 */
354 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
355 *flags &= ~FOLL_WRITE;
356 return 0;
357}
358
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700359static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
360{
361 vm_flags_t vm_flags = vma->vm_flags;
362
363 if (vm_flags & (VM_IO | VM_PFNMAP))
364 return -EFAULT;
365
366 if (gup_flags & FOLL_WRITE) {
367 if (!(vm_flags & VM_WRITE)) {
368 if (!(gup_flags & FOLL_FORCE))
369 return -EFAULT;
370 /*
371 * We used to let the write,force case do COW in a
372 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
373 * set a breakpoint in a read-only mapping of an
374 * executable, without corrupting the file (yet only
375 * when that file had been opened for writing!).
376 * Anon pages in shared mappings are surprising: now
377 * just reject it.
378 */
379 if (!is_cow_mapping(vm_flags)) {
380 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
381 return -EFAULT;
382 }
383 }
384 } else if (!(vm_flags & VM_READ)) {
385 if (!(gup_flags & FOLL_FORCE))
386 return -EFAULT;
387 /*
388 * Is there actually any vma we can reach here which does not
389 * have VM_MAYREAD set?
390 */
391 if (!(vm_flags & VM_MAYREAD))
392 return -EFAULT;
393 }
394 return 0;
395}
396
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700397/**
398 * __get_user_pages() - pin user pages in memory
399 * @tsk: task_struct of target task
400 * @mm: mm_struct of target mm
401 * @start: starting user address
402 * @nr_pages: number of pages from start to pin
403 * @gup_flags: flags modifying pin behaviour
404 * @pages: array that receives pointers to the pages pinned.
405 * Should be at least nr_pages long. Or NULL, if caller
406 * only intends to ensure the pages are faulted in.
407 * @vmas: array of pointers to vmas corresponding to each page.
408 * Or NULL if the caller does not require them.
409 * @nonblocking: whether waiting for disk IO or mmap_sem contention
410 *
411 * Returns number of pages pinned. This may be fewer than the number
412 * requested. If nr_pages is 0 or negative, returns 0. If no pages
413 * were pinned, returns -errno. Each page returned must be released
414 * with a put_page() call when it is finished with. vmas will only
415 * remain valid while mmap_sem is held.
416 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700417 * Must be called with mmap_sem held. It may be released. See below.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700418 *
419 * __get_user_pages walks a process's page tables and takes a reference to
420 * each struct page that each user address corresponds to at a given
421 * instant. That is, it takes the page that would be accessed if a user
422 * thread accesses the given user virtual address at that instant.
423 *
424 * This does not guarantee that the page exists in the user mappings when
425 * __get_user_pages returns, and there may even be a completely different
426 * page there in some cases (eg. if mmapped pagecache has been invalidated
427 * and subsequently re faulted). However it does guarantee that the page
428 * won't be freed completely. And mostly callers simply care that the page
429 * contains data that was valid *at some point in time*. Typically, an IO
430 * or similar operation cannot guarantee anything stronger anyway because
431 * locks can't be held over the syscall boundary.
432 *
433 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
434 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
435 * appropriate) must be called after the page is finished with, and
436 * before put_page is called.
437 *
438 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
439 * or mmap_sem contention, and if waiting is needed to pin all pages,
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700440 * *@nonblocking will be set to 0. Further, if @gup_flags does not
441 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
442 * this case.
443 *
444 * A caller using such a combination of @nonblocking and @gup_flags
445 * must therefore hold the mmap_sem for reading only, and recognize
446 * when it's been released. Otherwise, it must be held for either
447 * reading or writing and will not be released.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700448 *
449 * In most cases, get_user_pages or get_user_pages_fast should be used
450 * instead of __get_user_pages. __get_user_pages should be used only if
451 * you need some special @gup_flags.
452 */
453long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
454 unsigned long start, unsigned long nr_pages,
455 unsigned int gup_flags, struct page **pages,
456 struct vm_area_struct **vmas, int *nonblocking)
457{
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700458 long i = 0;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700459 unsigned int page_mask;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700460 struct vm_area_struct *vma = NULL;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700461
462 if (!nr_pages)
463 return 0;
464
465 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
466
467 /*
468 * If FOLL_FORCE is set then do not force a full fault as the hinting
469 * fault information is unrelated to the reference behaviour of a task
470 * using the address space
471 */
472 if (!(gup_flags & FOLL_FORCE))
473 gup_flags |= FOLL_NUMA;
474
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700475 do {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700476 struct page *page;
477 unsigned int foll_flags = gup_flags;
478 unsigned int page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700479
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700480 /* first iteration or cross vma bound */
481 if (!vma || start >= vma->vm_end) {
482 vma = find_extend_vma(mm, start);
483 if (!vma && in_gate_area(mm, start)) {
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700484 int ret;
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700485 ret = get_gate_page(mm, start & PAGE_MASK,
486 gup_flags, &vma,
487 pages ? &pages[i] : NULL);
488 if (ret)
489 return i ? : ret;
490 page_mask = 0;
491 goto next_page;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700492 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700493
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700494 if (!vma || check_vma_flags(vma, gup_flags))
495 return i ? : -EFAULT;
496 if (is_vm_hugetlb_page(vma)) {
497 i = follow_hugetlb_page(mm, vma, pages, vmas,
498 &start, &nr_pages, i,
499 gup_flags);
500 continue;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700501 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700502 }
503retry:
504 /*
505 * If we have a pending SIGKILL, don't keep faulting pages and
506 * potentially allocating memory.
507 */
508 if (unlikely(fatal_signal_pending(current)))
509 return i ? i : -ERESTARTSYS;
510 cond_resched();
511 page = follow_page_mask(vma, start, foll_flags, &page_mask);
512 if (!page) {
513 int ret;
514 ret = faultin_page(tsk, vma, start, &foll_flags,
515 nonblocking);
516 switch (ret) {
517 case 0:
518 goto retry;
519 case -EFAULT:
520 case -ENOMEM:
521 case -EHWPOISON:
522 return i ? i : ret;
523 case -EBUSY:
524 return i;
525 case -ENOENT:
526 goto next_page;
527 }
528 BUG();
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700529 } else if (PTR_ERR(page) == -EEXIST) {
530 /*
531 * Proper page table entry exists, but no corresponding
532 * struct page.
533 */
534 goto next_page;
535 } else if (IS_ERR(page)) {
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700536 return i ? i : PTR_ERR(page);
Kirill A. Shutemov1027e442015-09-04 15:47:55 -0700537 }
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700538 if (pages) {
539 pages[i] = page;
540 flush_anon_page(vma, page, start);
541 flush_dcache_page(page);
542 page_mask = 0;
543 }
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700544next_page:
Kirill A. Shutemovfa5bb202014-06-04 16:08:13 -0700545 if (vmas) {
546 vmas[i] = vma;
547 page_mask = 0;
548 }
549 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
550 if (page_increm > nr_pages)
551 page_increm = nr_pages;
552 i += page_increm;
553 start += page_increm * PAGE_SIZE;
554 nr_pages -= page_increm;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700555 } while (nr_pages);
556 return i;
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700557}
558EXPORT_SYMBOL(__get_user_pages);
559
560/*
561 * fixup_user_fault() - manually resolve a user page fault
562 * @tsk: the task_struct to use for page fault accounting, or
563 * NULL if faults are not to be recorded.
564 * @mm: mm_struct of target mm
565 * @address: user address
566 * @fault_flags:flags to pass down to handle_mm_fault()
567 *
568 * This is meant to be called in the specific scenario where for locking reasons
569 * we try to access user memory in atomic context (within a pagefault_disable()
570 * section), this returns -EFAULT, and we want to resolve the user fault before
571 * trying again.
572 *
573 * Typically this is meant to be used by the futex code.
574 *
575 * The main difference with get_user_pages() is that this function will
576 * unconditionally call handle_mm_fault() which will in turn perform all the
577 * necessary SW fixup of the dirty and young bits in the PTE, while
578 * handle_mm_fault() only guarantees to update these in the struct page.
579 *
580 * This is important for some architectures where those bits also gate the
581 * access permission to the page because they are maintained in software. On
582 * such architectures, gup() will not be enough to make a subsequent access
583 * succeed.
584 *
Paul Cassella9a95f3c2014-08-06 16:07:24 -0700585 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700586 */
587int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
588 unsigned long address, unsigned int fault_flags)
589{
590 struct vm_area_struct *vma;
591 vm_flags_t vm_flags;
592 int ret;
593
594 vma = find_extend_vma(mm, address);
595 if (!vma || address < vma->vm_start)
596 return -EFAULT;
597
598 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
599 if (!(vm_flags & vma->vm_flags))
600 return -EFAULT;
601
602 ret = handle_mm_fault(mm, vma, address, fault_flags);
603 if (ret & VM_FAULT_ERROR) {
604 if (ret & VM_FAULT_OOM)
605 return -ENOMEM;
606 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
607 return -EHWPOISON;
Linus Torvalds33692f22015-01-29 10:51:32 -0800608 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700609 return -EFAULT;
610 BUG();
611 }
612 if (tsk) {
613 if (ret & VM_FAULT_MAJOR)
614 tsk->maj_flt++;
615 else
616 tsk->min_flt++;
617 }
618 return 0;
619}
620
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800621static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
622 struct mm_struct *mm,
623 unsigned long start,
624 unsigned long nr_pages,
625 int write, int force,
626 struct page **pages,
627 struct vm_area_struct **vmas,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800628 int *locked, bool notify_drop,
629 unsigned int flags)
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800630{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800631 long ret, pages_done;
632 bool lock_dropped;
633
634 if (locked) {
635 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
636 BUG_ON(vmas);
637 /* check caller initialized locked */
638 BUG_ON(*locked != 1);
639 }
640
641 if (pages)
642 flags |= FOLL_GET;
643 if (write)
644 flags |= FOLL_WRITE;
645 if (force)
646 flags |= FOLL_FORCE;
647
648 pages_done = 0;
649 lock_dropped = false;
650 for (;;) {
651 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
652 vmas, locked);
653 if (!locked)
654 /* VM_FAULT_RETRY couldn't trigger, bypass */
655 return ret;
656
657 /* VM_FAULT_RETRY cannot return errors */
658 if (!*locked) {
659 BUG_ON(ret < 0);
660 BUG_ON(ret >= nr_pages);
661 }
662
663 if (!pages)
664 /* If it's a prefault don't insist harder */
665 return ret;
666
667 if (ret > 0) {
668 nr_pages -= ret;
669 pages_done += ret;
670 if (!nr_pages)
671 break;
672 }
673 if (*locked) {
674 /* VM_FAULT_RETRY didn't trigger */
675 if (!pages_done)
676 pages_done = ret;
677 break;
678 }
679 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
680 pages += ret;
681 start += ret << PAGE_SHIFT;
682
683 /*
684 * Repeat on the address that fired VM_FAULT_RETRY
685 * without FAULT_FLAG_ALLOW_RETRY but with
686 * FAULT_FLAG_TRIED.
687 */
688 *locked = 1;
689 lock_dropped = true;
690 down_read(&mm->mmap_sem);
691 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
692 pages, NULL, NULL);
693 if (ret != 1) {
694 BUG_ON(ret > 1);
695 if (!pages_done)
696 pages_done = ret;
697 break;
698 }
699 nr_pages--;
700 pages_done++;
701 if (!nr_pages)
702 break;
703 pages++;
704 start += PAGE_SIZE;
705 }
706 if (notify_drop && lock_dropped && *locked) {
707 /*
708 * We must let the caller know we temporarily dropped the lock
709 * and so the critical section protected by it was lost.
710 */
711 up_read(&mm->mmap_sem);
712 *locked = 0;
713 }
714 return pages_done;
715}
716
717/*
718 * We can leverage the VM_FAULT_RETRY functionality in the page fault
719 * paths better by using either get_user_pages_locked() or
720 * get_user_pages_unlocked().
721 *
722 * get_user_pages_locked() is suitable to replace the form:
723 *
724 * down_read(&mm->mmap_sem);
725 * do_something()
726 * get_user_pages(tsk, mm, ..., pages, NULL);
727 * up_read(&mm->mmap_sem);
728 *
729 * to:
730 *
731 * int locked = 1;
732 * down_read(&mm->mmap_sem);
733 * do_something()
734 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
735 * if (locked)
736 * up_read(&mm->mmap_sem);
737 */
738long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
739 unsigned long start, unsigned long nr_pages,
740 int write, int force, struct page **pages,
741 int *locked)
742{
743 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800744 pages, NULL, locked, true, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800745}
746EXPORT_SYMBOL(get_user_pages_locked);
747
748/*
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800749 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
750 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
751 *
752 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
753 * caller if required (just like with __get_user_pages). "FOLL_GET",
754 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
755 * according to the parameters "pages", "write", "force"
756 * respectively.
757 */
758__always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
759 unsigned long start, unsigned long nr_pages,
760 int write, int force, struct page **pages,
761 unsigned int gup_flags)
762{
763 long ret;
764 int locked = 1;
765 down_read(&mm->mmap_sem);
766 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
767 pages, NULL, &locked, false, gup_flags);
768 if (locked)
769 up_read(&mm->mmap_sem);
770 return ret;
771}
772EXPORT_SYMBOL(__get_user_pages_unlocked);
773
774/*
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800775 * get_user_pages_unlocked() is suitable to replace the form:
776 *
777 * down_read(&mm->mmap_sem);
778 * get_user_pages(tsk, mm, ..., pages, NULL);
779 * up_read(&mm->mmap_sem);
780 *
781 * with:
782 *
783 * get_user_pages_unlocked(tsk, mm, ..., pages);
784 *
785 * It is functionally equivalent to get_user_pages_fast so
786 * get_user_pages_fast should be used instead, if the two parameters
787 * "tsk" and "mm" are respectively equal to current and current->mm,
788 * or if "force" shall be set to 1 (get_user_pages_fast misses the
789 * "force" parameter).
790 */
791long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
792 unsigned long start, unsigned long nr_pages,
793 int write, int force, struct page **pages)
794{
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800795 return __get_user_pages_unlocked(tsk, mm, start, nr_pages, write,
796 force, pages, FOLL_TOUCH);
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800797}
798EXPORT_SYMBOL(get_user_pages_unlocked);
799
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700800/*
801 * get_user_pages() - pin user pages in memory
802 * @tsk: the task_struct to use for page fault accounting, or
803 * NULL if faults are not to be recorded.
804 * @mm: mm_struct of target mm
805 * @start: starting user address
806 * @nr_pages: number of pages from start to pin
807 * @write: whether pages will be written to by the caller
808 * @force: whether to force access even when user mapping is currently
809 * protected (but never forces write access to shared mapping).
810 * @pages: array that receives pointers to the pages pinned.
811 * Should be at least nr_pages long. Or NULL, if caller
812 * only intends to ensure the pages are faulted in.
813 * @vmas: array of pointers to vmas corresponding to each page.
814 * Or NULL if the caller does not require them.
815 *
816 * Returns number of pages pinned. This may be fewer than the number
817 * requested. If nr_pages is 0 or negative, returns 0. If no pages
818 * were pinned, returns -errno. Each page returned must be released
819 * with a put_page() call when it is finished with. vmas will only
820 * remain valid while mmap_sem is held.
821 *
822 * Must be called with mmap_sem held for read or write.
823 *
824 * get_user_pages walks a process's page tables and takes a reference to
825 * each struct page that each user address corresponds to at a given
826 * instant. That is, it takes the page that would be accessed if a user
827 * thread accesses the given user virtual address at that instant.
828 *
829 * This does not guarantee that the page exists in the user mappings when
830 * get_user_pages returns, and there may even be a completely different
831 * page there in some cases (eg. if mmapped pagecache has been invalidated
832 * and subsequently re faulted). However it does guarantee that the page
833 * won't be freed completely. And mostly callers simply care that the page
834 * contains data that was valid *at some point in time*. Typically, an IO
835 * or similar operation cannot guarantee anything stronger anyway because
836 * locks can't be held over the syscall boundary.
837 *
838 * If write=0, the page must not be written to. If the page is written to,
839 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
840 * after the page is finished with, and before put_page is called.
841 *
842 * get_user_pages is typically used for fewer-copy IO operations, to get a
843 * handle on the memory by some means other than accesses via the user virtual
844 * addresses. The pages may be submitted for DMA to devices or accessed via
845 * their kernel linear mapping (via the kmap APIs). Care should be taken to
846 * use the correct cache flushing APIs.
847 *
848 * See also get_user_pages_fast, for performance critical applications.
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800849 *
850 * get_user_pages should be phased out in favor of
851 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
852 * should use get_user_pages because it cannot pass
853 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700854 */
855long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
856 unsigned long start, unsigned long nr_pages, int write,
857 int force, struct page **pages, struct vm_area_struct **vmas)
858{
Andrea Arcangelif0818f42015-02-11 15:27:17 -0800859 return __get_user_pages_locked(tsk, mm, start, nr_pages, write, force,
Andrea Arcangeli0fd71a52015-02-11 15:27:20 -0800860 pages, vmas, NULL, false, FOLL_TOUCH);
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700861}
862EXPORT_SYMBOL(get_user_pages);
863
864/**
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700865 * populate_vma_page_range() - populate a range of pages in the vma.
866 * @vma: target vma
867 * @start: start address
868 * @end: end address
869 * @nonblocking:
870 *
871 * This takes care of mlocking the pages too if VM_LOCKED is set.
872 *
873 * return 0 on success, negative error code on error.
874 *
875 * vma->vm_mm->mmap_sem must be held.
876 *
877 * If @nonblocking is NULL, it may be held for read or write and will
878 * be unperturbed.
879 *
880 * If @nonblocking is non-NULL, it must held for read only and may be
881 * released. If it's released, *@nonblocking will be set to 0.
882 */
883long populate_vma_page_range(struct vm_area_struct *vma,
884 unsigned long start, unsigned long end, int *nonblocking)
885{
886 struct mm_struct *mm = vma->vm_mm;
887 unsigned long nr_pages = (end - start) / PAGE_SIZE;
888 int gup_flags;
889
890 VM_BUG_ON(start & ~PAGE_MASK);
891 VM_BUG_ON(end & ~PAGE_MASK);
892 VM_BUG_ON_VMA(start < vma->vm_start, vma);
893 VM_BUG_ON_VMA(end > vma->vm_end, vma);
894 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
895
Eric B Munsonde60f5f2015-11-05 18:51:36 -0800896 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
897 if (vma->vm_flags & VM_LOCKONFAULT)
898 gup_flags &= ~FOLL_POPULATE;
899
Kirill A. Shutemovacc3c8d2015-04-14 15:44:45 -0700900 /*
901 * We want to touch writable mappings with a write fault in order
902 * to break COW, except for shared mappings because these don't COW
903 * and we would not want to dirty them for nothing.
904 */
905 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
906 gup_flags |= FOLL_WRITE;
907
908 /*
909 * We want mlock to succeed for regions that have any permissions
910 * other than PROT_NONE.
911 */
912 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
913 gup_flags |= FOLL_FORCE;
914
915 /*
916 * We made sure addr is within a VMA, so the following will
917 * not result in a stack expansion that recurses back here.
918 */
919 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
920 NULL, NULL, nonblocking);
921}
922
923/*
924 * __mm_populate - populate and/or mlock pages within a range of address space.
925 *
926 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
927 * flags. VMAs must be already marked with the desired vm_flags, and
928 * mmap_sem must not be held.
929 */
930int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
931{
932 struct mm_struct *mm = current->mm;
933 unsigned long end, nstart, nend;
934 struct vm_area_struct *vma = NULL;
935 int locked = 0;
936 long ret = 0;
937
938 VM_BUG_ON(start & ~PAGE_MASK);
939 VM_BUG_ON(len != PAGE_ALIGN(len));
940 end = start + len;
941
942 for (nstart = start; nstart < end; nstart = nend) {
943 /*
944 * We want to fault in pages for [nstart; end) address range.
945 * Find first corresponding VMA.
946 */
947 if (!locked) {
948 locked = 1;
949 down_read(&mm->mmap_sem);
950 vma = find_vma(mm, nstart);
951 } else if (nstart >= vma->vm_end)
952 vma = vma->vm_next;
953 if (!vma || vma->vm_start >= end)
954 break;
955 /*
956 * Set [nstart; nend) to intersection of desired address
957 * range with the first VMA. Also, skip undesirable VMA types.
958 */
959 nend = min(end, vma->vm_end);
960 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
961 continue;
962 if (nstart < vma->vm_start)
963 nstart = vma->vm_start;
964 /*
965 * Now fault in a range of pages. populate_vma_page_range()
966 * double checks the vma flags, so that it won't mlock pages
967 * if the vma was already munlocked.
968 */
969 ret = populate_vma_page_range(vma, nstart, nend, &locked);
970 if (ret < 0) {
971 if (ignore_errors) {
972 ret = 0;
973 continue; /* continue at next VMA */
974 }
975 break;
976 }
977 nend = nstart + ret * PAGE_SIZE;
978 ret = 0;
979 }
980 if (locked)
981 up_read(&mm->mmap_sem);
982 return ret; /* 0 or negative error code */
983}
984
985/**
Kirill A. Shutemov4bbd4c72014-06-04 16:08:10 -0700986 * get_dump_page() - pin user page in memory while writing it to core dump
987 * @addr: user address
988 *
989 * Returns struct page pointer of user page pinned for dump,
990 * to be freed afterwards by page_cache_release() or put_page().
991 *
992 * Returns NULL on any kind of failure - a hole must then be inserted into
993 * the corefile, to preserve alignment with its headers; and also returns
994 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
995 * allowing a hole to be left in the corefile to save diskspace.
996 *
997 * Called without mmap_sem, but after all other threads have been killed.
998 */
999#ifdef CONFIG_ELF_CORE
1000struct page *get_dump_page(unsigned long addr)
1001{
1002 struct vm_area_struct *vma;
1003 struct page *page;
1004
1005 if (__get_user_pages(current, current->mm, addr, 1,
1006 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1007 NULL) < 1)
1008 return NULL;
1009 flush_cache_page(vma, addr, page_to_pfn(page));
1010 return page;
1011}
1012#endif /* CONFIG_ELF_CORE */
Steve Capper2667f502014-10-09 15:29:14 -07001013
1014/*
1015 * Generic RCU Fast GUP
1016 *
1017 * get_user_pages_fast attempts to pin user pages by walking the page
1018 * tables directly and avoids taking locks. Thus the walker needs to be
1019 * protected from page table pages being freed from under it, and should
1020 * block any THP splits.
1021 *
1022 * One way to achieve this is to have the walker disable interrupts, and
1023 * rely on IPIs from the TLB flushing code blocking before the page table
1024 * pages are freed. This is unsuitable for architectures that do not need
1025 * to broadcast an IPI when invalidating TLBs.
1026 *
1027 * Another way to achieve this is to batch up page table containing pages
1028 * belonging to more than one mm_user, then rcu_sched a callback to free those
1029 * pages. Disabling interrupts will allow the fast_gup walker to both block
1030 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1031 * (which is a relatively rare event). The code below adopts this strategy.
1032 *
1033 * Before activating this code, please be aware that the following assumptions
1034 * are currently made:
1035 *
1036 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1037 * pages containing page tables.
1038 *
1039 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1040 * pmdp_splitting_flush.
1041 *
1042 * *) ptes can be read atomically by the architecture.
1043 *
1044 * *) access_ok is sufficient to validate userspace address ranges.
1045 *
1046 * The last two assumptions can be relaxed by the addition of helper functions.
1047 *
1048 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1049 */
1050#ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1051
1052#ifdef __HAVE_ARCH_PTE_SPECIAL
1053static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1054 int write, struct page **pages, int *nr)
1055{
1056 pte_t *ptep, *ptem;
1057 int ret = 0;
1058
1059 ptem = ptep = pte_offset_map(&pmd, addr);
1060 do {
1061 /*
1062 * In the line below we are assuming that the pte can be read
1063 * atomically. If this is not the case for your architecture,
1064 * please wrap this in a helper function!
1065 *
1066 * for an example see gup_get_pte in arch/x86/mm/gup.c
1067 */
Jason Low9d8c47e2015-04-15 16:14:05 -07001068 pte_t pte = READ_ONCE(*ptep);
Steve Capper2667f502014-10-09 15:29:14 -07001069 struct page *page;
1070
1071 /*
1072 * Similar to the PMD case below, NUMA hinting must take slow
Mel Gorman8a0516e2015-02-12 14:58:22 -08001073 * path using the pte_protnone check.
Steve Capper2667f502014-10-09 15:29:14 -07001074 */
1075 if (!pte_present(pte) || pte_special(pte) ||
Mel Gorman8a0516e2015-02-12 14:58:22 -08001076 pte_protnone(pte) || (write && !pte_write(pte)))
Steve Capper2667f502014-10-09 15:29:14 -07001077 goto pte_unmap;
1078
1079 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1080 page = pte_page(pte);
1081
1082 if (!page_cache_get_speculative(page))
1083 goto pte_unmap;
1084
1085 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1086 put_page(page);
1087 goto pte_unmap;
1088 }
1089
1090 pages[*nr] = page;
1091 (*nr)++;
1092
1093 } while (ptep++, addr += PAGE_SIZE, addr != end);
1094
1095 ret = 1;
1096
1097pte_unmap:
1098 pte_unmap(ptem);
1099 return ret;
1100}
1101#else
1102
1103/*
1104 * If we can't determine whether or not a pte is special, then fail immediately
1105 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1106 * to be special.
1107 *
1108 * For a futex to be placed on a THP tail page, get_futex_key requires a
1109 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1110 * useful to have gup_huge_pmd even if we can't operate on ptes.
1111 */
1112static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1113 int write, struct page **pages, int *nr)
1114{
1115 return 0;
1116}
1117#endif /* __HAVE_ARCH_PTE_SPECIAL */
1118
1119static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1120 unsigned long end, int write, struct page **pages, int *nr)
1121{
1122 struct page *head, *page, *tail;
1123 int refs;
1124
1125 if (write && !pmd_write(orig))
1126 return 0;
1127
1128 refs = 0;
1129 head = pmd_page(orig);
1130 page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1131 tail = page;
1132 do {
1133 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1134 pages[*nr] = page;
1135 (*nr)++;
1136 page++;
1137 refs++;
1138 } while (addr += PAGE_SIZE, addr != end);
1139
1140 if (!page_cache_add_speculative(head, refs)) {
1141 *nr -= refs;
1142 return 0;
1143 }
1144
1145 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1146 *nr -= refs;
1147 while (refs--)
1148 put_page(head);
1149 return 0;
1150 }
1151
1152 /*
1153 * Any tail pages need their mapcount reference taken before we
1154 * return. (This allows the THP code to bump their ref count when
1155 * they are split into base pages).
1156 */
1157 while (refs--) {
1158 if (PageTail(tail))
1159 get_huge_page_tail(tail);
1160 tail++;
1161 }
1162
1163 return 1;
1164}
1165
1166static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1167 unsigned long end, int write, struct page **pages, int *nr)
1168{
1169 struct page *head, *page, *tail;
1170 int refs;
1171
1172 if (write && !pud_write(orig))
1173 return 0;
1174
1175 refs = 0;
1176 head = pud_page(orig);
1177 page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1178 tail = page;
1179 do {
1180 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1181 pages[*nr] = page;
1182 (*nr)++;
1183 page++;
1184 refs++;
1185 } while (addr += PAGE_SIZE, addr != end);
1186
1187 if (!page_cache_add_speculative(head, refs)) {
1188 *nr -= refs;
1189 return 0;
1190 }
1191
1192 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1193 *nr -= refs;
1194 while (refs--)
1195 put_page(head);
1196 return 0;
1197 }
1198
1199 while (refs--) {
1200 if (PageTail(tail))
1201 get_huge_page_tail(tail);
1202 tail++;
1203 }
1204
1205 return 1;
1206}
1207
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301208static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1209 unsigned long end, int write,
1210 struct page **pages, int *nr)
1211{
1212 int refs;
1213 struct page *head, *page, *tail;
1214
1215 if (write && !pgd_write(orig))
1216 return 0;
1217
1218 refs = 0;
1219 head = pgd_page(orig);
1220 page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1221 tail = page;
1222 do {
1223 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1224 pages[*nr] = page;
1225 (*nr)++;
1226 page++;
1227 refs++;
1228 } while (addr += PAGE_SIZE, addr != end);
1229
1230 if (!page_cache_add_speculative(head, refs)) {
1231 *nr -= refs;
1232 return 0;
1233 }
1234
1235 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1236 *nr -= refs;
1237 while (refs--)
1238 put_page(head);
1239 return 0;
1240 }
1241
1242 while (refs--) {
1243 if (PageTail(tail))
1244 get_huge_page_tail(tail);
1245 tail++;
1246 }
1247
1248 return 1;
1249}
1250
Steve Capper2667f502014-10-09 15:29:14 -07001251static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1252 int write, struct page **pages, int *nr)
1253{
1254 unsigned long next;
1255 pmd_t *pmdp;
1256
1257 pmdp = pmd_offset(&pud, addr);
1258 do {
Christian Borntraeger38c5ce92015-01-06 22:54:46 +01001259 pmd_t pmd = READ_ONCE(*pmdp);
Steve Capper2667f502014-10-09 15:29:14 -07001260
1261 next = pmd_addr_end(addr, end);
1262 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1263 return 0;
1264
1265 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1266 /*
1267 * NUMA hinting faults need to be handled in the GUP
1268 * slowpath for accounting purposes and so that they
1269 * can be serialised against THP migration.
1270 */
Mel Gorman8a0516e2015-02-12 14:58:22 -08001271 if (pmd_protnone(pmd))
Steve Capper2667f502014-10-09 15:29:14 -07001272 return 0;
1273
1274 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1275 pages, nr))
1276 return 0;
1277
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301278 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1279 /*
1280 * architecture have different format for hugetlbfs
1281 * pmd format and THP pmd format
1282 */
1283 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1284 PMD_SHIFT, next, write, pages, nr))
1285 return 0;
Steve Capper2667f502014-10-09 15:29:14 -07001286 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1287 return 0;
1288 } while (pmdp++, addr = next, addr != end);
1289
1290 return 1;
1291}
1292
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301293static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1294 int write, struct page **pages, int *nr)
Steve Capper2667f502014-10-09 15:29:14 -07001295{
1296 unsigned long next;
1297 pud_t *pudp;
1298
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301299 pudp = pud_offset(&pgd, addr);
Steve Capper2667f502014-10-09 15:29:14 -07001300 do {
Christian Borntraegere37c6982014-12-07 21:41:33 +01001301 pud_t pud = READ_ONCE(*pudp);
Steve Capper2667f502014-10-09 15:29:14 -07001302
1303 next = pud_addr_end(addr, end);
1304 if (pud_none(pud))
1305 return 0;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301306 if (unlikely(pud_huge(pud))) {
Steve Capper2667f502014-10-09 15:29:14 -07001307 if (!gup_huge_pud(pud, pudp, addr, next, write,
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301308 pages, nr))
1309 return 0;
1310 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1311 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1312 PUD_SHIFT, next, write, pages, nr))
Steve Capper2667f502014-10-09 15:29:14 -07001313 return 0;
1314 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1315 return 0;
1316 } while (pudp++, addr = next, addr != end);
1317
1318 return 1;
1319}
1320
1321/*
1322 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1323 * the regular GUP. It will only return non-negative values.
1324 */
1325int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1326 struct page **pages)
1327{
1328 struct mm_struct *mm = current->mm;
1329 unsigned long addr, len, end;
1330 unsigned long next, flags;
1331 pgd_t *pgdp;
1332 int nr = 0;
1333
1334 start &= PAGE_MASK;
1335 addr = start;
1336 len = (unsigned long) nr_pages << PAGE_SHIFT;
1337 end = start + len;
1338
1339 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1340 start, len)))
1341 return 0;
1342
1343 /*
1344 * Disable interrupts. We use the nested form as we can already have
1345 * interrupts disabled by get_futex_key.
1346 *
1347 * With interrupts disabled, we block page table pages from being
1348 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1349 * for more details.
1350 *
1351 * We do not adopt an rcu_read_lock(.) here as we also want to
1352 * block IPIs that come from THPs splitting.
1353 */
1354
1355 local_irq_save(flags);
1356 pgdp = pgd_offset(mm, addr);
1357 do {
Jason Low9d8c47e2015-04-15 16:14:05 -07001358 pgd_t pgd = READ_ONCE(*pgdp);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301359
Steve Capper2667f502014-10-09 15:29:14 -07001360 next = pgd_addr_end(addr, end);
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301361 if (pgd_none(pgd))
Steve Capper2667f502014-10-09 15:29:14 -07001362 break;
Aneesh Kumar K.Vf30c59e2014-11-05 21:57:40 +05301363 if (unlikely(pgd_huge(pgd))) {
1364 if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1365 pages, &nr))
1366 break;
1367 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1368 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1369 PGDIR_SHIFT, next, write, pages, &nr))
1370 break;
1371 } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
Steve Capper2667f502014-10-09 15:29:14 -07001372 break;
1373 } while (pgdp++, addr = next, addr != end);
1374 local_irq_restore(flags);
1375
1376 return nr;
1377}
1378
1379/**
1380 * get_user_pages_fast() - pin user pages in memory
1381 * @start: starting user address
1382 * @nr_pages: number of pages from start to pin
1383 * @write: whether pages will be written to
1384 * @pages: array that receives pointers to the pages pinned.
1385 * Should be at least nr_pages long.
1386 *
1387 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1388 * If not successful, it will fall back to taking the lock and
1389 * calling get_user_pages().
1390 *
1391 * Returns number of pages pinned. This may be fewer than the number
1392 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1393 * were pinned, returns -errno.
1394 */
1395int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1396 struct page **pages)
1397{
1398 struct mm_struct *mm = current->mm;
1399 int nr, ret;
1400
1401 start &= PAGE_MASK;
1402 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1403 ret = nr;
1404
1405 if (nr < nr_pages) {
1406 /* Try to get the remaining pages with get_user_pages */
1407 start += nr << PAGE_SHIFT;
1408 pages += nr;
1409
Andrea Arcangelia7b78072015-02-11 15:27:23 -08001410 ret = get_user_pages_unlocked(current, mm, start,
1411 nr_pages - nr, write, 0, pages);
Steve Capper2667f502014-10-09 15:29:14 -07001412
1413 /* Have to be a bit careful with return values */
1414 if (nr > 0) {
1415 if (ret < 0)
1416 ret = nr;
1417 else
1418 ret += nr;
1419 }
1420 }
1421
1422 return ret;
1423}
1424
1425#endif /* CONFIG_HAVE_GENERIC_RCU_GUP */