~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/gup.c

Version: ~ [ linux-5.10-rc5 ] ~ [ linux-5.9.10 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.79 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.159 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.208 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.245 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.245 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 #include <linux/kernel.h>
  2 #include <linux/errno.h>
  3 #include <linux/err.h>
  4 #include <linux/spinlock.h>
  5 
  6 #include <linux/mm.h>
  7 #include <linux/memremap.h>
  8 #include <linux/pagemap.h>
  9 #include <linux/rmap.h>
 10 #include <linux/swap.h>
 11 #include <linux/swapops.h>
 12 
 13 #include <linux/sched/signal.h>
 14 #include <linux/rwsem.h>
 15 #include <linux/hugetlb.h>
 16 
 17 #include <asm/mmu_context.h>
 18 #include <asm/pgtable.h>
 19 #include <asm/tlbflush.h>
 20 
 21 #include "internal.h"
 22 
 23 static struct page *no_page_table(struct vm_area_struct *vma,
 24                 unsigned int flags)
 25 {
 26         /*
 27          * When core dumping an enormous anonymous area that nobody
 28          * has touched so far, we don't want to allocate unnecessary pages or
 29          * page tables.  Return error instead of NULL to skip handle_mm_fault,
 30          * then get_dump_page() will return NULL to leave a hole in the dump.
 31          * But we can only make this optimization where a hole would surely
 32          * be zero-filled if handle_mm_fault() actually did handle it.
 33          */
 34         if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
 35                 return ERR_PTR(-EFAULT);
 36         return NULL;
 37 }
 38 
 39 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
 40                 pte_t *pte, unsigned int flags)
 41 {
 42         /* No page to get reference */
 43         if (flags & FOLL_GET)
 44                 return -EFAULT;
 45 
 46         if (flags & FOLL_TOUCH) {
 47                 pte_t entry = *pte;
 48 
 49                 if (flags & FOLL_WRITE)
 50                         entry = pte_mkdirty(entry);
 51                 entry = pte_mkyoung(entry);
 52 
 53                 if (!pte_same(*pte, entry)) {
 54                         set_pte_at(vma->vm_mm, address, pte, entry);
 55                         update_mmu_cache(vma, address, pte);
 56                 }
 57         }
 58 
 59         /* Proper page table entry exists, but no corresponding struct page */
 60         return -EEXIST;
 61 }
 62 
 63 /*
 64  * FOLL_FORCE can write to even unwritable pte's, but only
 65  * after we've gone through a COW cycle and they are dirty.
 66  */
 67 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
 68 {
 69         return pte_write(pte) ||
 70                 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
 71 }
 72 
 73 static struct page *follow_page_pte(struct vm_area_struct *vma,
 74                 unsigned long address, pmd_t *pmd, unsigned int flags)
 75 {
 76         struct mm_struct *mm = vma->vm_mm;
 77         struct dev_pagemap *pgmap = NULL;
 78         struct page *page;
 79         spinlock_t *ptl;
 80         pte_t *ptep, pte;
 81 
 82 retry:
 83         if (unlikely(pmd_bad(*pmd)))
 84                 return no_page_table(vma, flags);
 85 
 86         ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 87         pte = *ptep;
 88         if (!pte_present(pte)) {
 89                 swp_entry_t entry;
 90                 /*
 91                  * KSM's break_ksm() relies upon recognizing a ksm page
 92                  * even while it is being migrated, so for that case we
 93                  * need migration_entry_wait().
 94                  */
 95                 if (likely(!(flags & FOLL_MIGRATION)))
 96                         goto no_page;
 97                 if (pte_none(pte))
 98                         goto no_page;
 99                 entry = pte_to_swp_entry(pte);
100                 if (!is_migration_entry(entry))
101                         goto no_page;
102                 pte_unmap_unlock(ptep, ptl);
103                 migration_entry_wait(mm, pmd, address);
104                 goto retry;
105         }
106         if ((flags & FOLL_NUMA) && pte_protnone(pte))
107                 goto no_page;
108         if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
109                 pte_unmap_unlock(ptep, ptl);
110                 return NULL;
111         }
112 
113         page = vm_normal_page(vma, address, pte);
114         if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
115                 /*
116                  * Only return device mapping pages in the FOLL_GET case since
117                  * they are only valid while holding the pgmap reference.
118                  */
119                 pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
120                 if (pgmap)
121                         page = pte_page(pte);
122                 else
123                         goto no_page;
124         } else if (unlikely(!page)) {
125                 if (flags & FOLL_DUMP) {
126                         /* Avoid special (like zero) pages in core dumps */
127                         page = ERR_PTR(-EFAULT);
128                         goto out;
129                 }
130 
131                 if (is_zero_pfn(pte_pfn(pte))) {
132                         page = pte_page(pte);
133                 } else {
134                         int ret;
135 
136                         ret = follow_pfn_pte(vma, address, ptep, flags);
137                         page = ERR_PTR(ret);
138                         goto out;
139                 }
140         }
141 
142         if (flags & FOLL_SPLIT && PageTransCompound(page)) {
143                 int ret;
144                 get_page(page);
145                 pte_unmap_unlock(ptep, ptl);
146                 lock_page(page);
147                 ret = split_huge_page(page);
148                 unlock_page(page);
149                 put_page(page);
150                 if (ret)
151                         return ERR_PTR(ret);
152                 goto retry;
153         }
154 
155         if (flags & FOLL_GET) {
156                 get_page(page);
157 
158                 /* drop the pgmap reference now that we hold the page */
159                 if (pgmap) {
160                         put_dev_pagemap(pgmap);
161                         pgmap = NULL;
162                 }
163         }
164         if (flags & FOLL_TOUCH) {
165                 if ((flags & FOLL_WRITE) &&
166                     !pte_dirty(pte) && !PageDirty(page))
167                         set_page_dirty(page);
168                 /*
169                  * pte_mkyoung() would be more correct here, but atomic care
170                  * is needed to avoid losing the dirty bit: it is easier to use
171                  * mark_page_accessed().
172                  */
173                 mark_page_accessed(page);
174         }
175         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
176                 /* Do not mlock pte-mapped THP */
177                 if (PageTransCompound(page))
178                         goto out;
179 
180                 /*
181                  * The preliminary mapping check is mainly to avoid the
182                  * pointless overhead of lock_page on the ZERO_PAGE
183                  * which might bounce very badly if there is contention.
184                  *
185                  * If the page is already locked, we don't need to
186                  * handle it now - vmscan will handle it later if and
187                  * when it attempts to reclaim the page.
188                  */
189                 if (page->mapping && trylock_page(page)) {
190                         lru_add_drain();  /* push cached pages to LRU */
191                         /*
192                          * Because we lock page here, and migration is
193                          * blocked by the pte's page reference, and we
194                          * know the page is still mapped, we don't even
195                          * need to check for file-cache page truncation.
196                          */
197                         mlock_vma_page(page);
198                         unlock_page(page);
199                 }
200         }
201 out:
202         pte_unmap_unlock(ptep, ptl);
203         return page;
204 no_page:
205         pte_unmap_unlock(ptep, ptl);
206         if (!pte_none(pte))
207                 return NULL;
208         return no_page_table(vma, flags);
209 }
210 
211 static struct page *follow_pmd_mask(struct vm_area_struct *vma,
212                                     unsigned long address, pud_t *pudp,
213                                     unsigned int flags, unsigned int *page_mask)
214 {
215         pmd_t *pmd;
216         spinlock_t *ptl;
217         struct page *page;
218         struct mm_struct *mm = vma->vm_mm;
219 
220         pmd = pmd_offset(pudp, address);
221         if (pmd_none(*pmd))
222                 return no_page_table(vma, flags);
223         if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
224                 page = follow_huge_pmd(mm, address, pmd, flags);
225                 if (page)
226                         return page;
227                 return no_page_table(vma, flags);
228         }
229         if (is_hugepd(__hugepd(pmd_val(*pmd)))) {
230                 page = follow_huge_pd(vma, address,
231                                       __hugepd(pmd_val(*pmd)), flags,
232                                       PMD_SHIFT);
233                 if (page)
234                         return page;
235                 return no_page_table(vma, flags);
236         }
237 retry:
238         if (!pmd_present(*pmd)) {
239                 if (likely(!(flags & FOLL_MIGRATION)))
240                         return no_page_table(vma, flags);
241                 VM_BUG_ON(thp_migration_supported() &&
242                                   !is_pmd_migration_entry(*pmd));
243                 if (is_pmd_migration_entry(*pmd))
244                         pmd_migration_entry_wait(mm, pmd);
245                 goto retry;
246         }
247         if (pmd_devmap(*pmd)) {
248                 ptl = pmd_lock(mm, pmd);
249                 page = follow_devmap_pmd(vma, address, pmd, flags);
250                 spin_unlock(ptl);
251                 if (page)
252                         return page;
253         }
254         if (likely(!pmd_trans_huge(*pmd)))
255                 return follow_page_pte(vma, address, pmd, flags);
256 
257         if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
258                 return no_page_table(vma, flags);
259 
260 retry_locked:
261         ptl = pmd_lock(mm, pmd);
262         if (unlikely(!pmd_present(*pmd))) {
263                 spin_unlock(ptl);
264                 if (likely(!(flags & FOLL_MIGRATION)))
265                         return no_page_table(vma, flags);
266                 pmd_migration_entry_wait(mm, pmd);
267                 goto retry_locked;
268         }
269         if (unlikely(!pmd_trans_huge(*pmd))) {
270                 spin_unlock(ptl);
271                 return follow_page_pte(vma, address, pmd, flags);
272         }
273         if (flags & FOLL_SPLIT) {
274                 int ret;
275                 page = pmd_page(*pmd);
276                 if (is_huge_zero_page(page)) {
277                         spin_unlock(ptl);
278                         ret = 0;
279                         split_huge_pmd(vma, pmd, address);
280                         if (pmd_trans_unstable(pmd))
281                                 ret = -EBUSY;
282                 } else {
283                         get_page(page);
284                         spin_unlock(ptl);
285                         lock_page(page);
286                         ret = split_huge_page(page);
287                         unlock_page(page);
288                         put_page(page);
289                         if (pmd_none(*pmd))
290                                 return no_page_table(vma, flags);
291                 }
292 
293                 return ret ? ERR_PTR(ret) :
294                         follow_page_pte(vma, address, pmd, flags);
295         }
296         page = follow_trans_huge_pmd(vma, address, pmd, flags);
297         spin_unlock(ptl);
298         *page_mask = HPAGE_PMD_NR - 1;
299         return page;
300 }
301 
302 
303 static struct page *follow_pud_mask(struct vm_area_struct *vma,
304                                     unsigned long address, p4d_t *p4dp,
305                                     unsigned int flags, unsigned int *page_mask)
306 {
307         pud_t *pud;
308         spinlock_t *ptl;
309         struct page *page;
310         struct mm_struct *mm = vma->vm_mm;
311 
312         pud = pud_offset(p4dp, address);
313         if (pud_none(*pud))
314                 return no_page_table(vma, flags);
315         if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
316                 page = follow_huge_pud(mm, address, pud, flags);
317                 if (page)
318                         return page;
319                 return no_page_table(vma, flags);
320         }
321         if (is_hugepd(__hugepd(pud_val(*pud)))) {
322                 page = follow_huge_pd(vma, address,
323                                       __hugepd(pud_val(*pud)), flags,
324                                       PUD_SHIFT);
325                 if (page)
326                         return page;
327                 return no_page_table(vma, flags);
328         }
329         if (pud_devmap(*pud)) {
330                 ptl = pud_lock(mm, pud);
331                 page = follow_devmap_pud(vma, address, pud, flags);
332                 spin_unlock(ptl);
333                 if (page)
334                         return page;
335         }
336         if (unlikely(pud_bad(*pud)))
337                 return no_page_table(vma, flags);
338 
339         return follow_pmd_mask(vma, address, pud, flags, page_mask);
340 }
341 
342 
343 static struct page *follow_p4d_mask(struct vm_area_struct *vma,
344                                     unsigned long address, pgd_t *pgdp,
345                                     unsigned int flags, unsigned int *page_mask)
346 {
347         p4d_t *p4d;
348         struct page *page;
349 
350         p4d = p4d_offset(pgdp, address);
351         if (p4d_none(*p4d))
352                 return no_page_table(vma, flags);
353         BUILD_BUG_ON(p4d_huge(*p4d));
354         if (unlikely(p4d_bad(*p4d)))
355                 return no_page_table(vma, flags);
356 
357         if (is_hugepd(__hugepd(p4d_val(*p4d)))) {
358                 page = follow_huge_pd(vma, address,
359                                       __hugepd(p4d_val(*p4d)), flags,
360                                       P4D_SHIFT);
361                 if (page)
362                         return page;
363                 return no_page_table(vma, flags);
364         }
365         return follow_pud_mask(vma, address, p4d, flags, page_mask);
366 }
367 
368 /**
369  * follow_page_mask - look up a page descriptor from a user-virtual address
370  * @vma: vm_area_struct mapping @address
371  * @address: virtual address to look up
372  * @flags: flags modifying lookup behaviour
373  * @page_mask: on output, *page_mask is set according to the size of the page
374  *
375  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
376  *
377  * Returns the mapped (struct page *), %NULL if no mapping exists, or
378  * an error pointer if there is a mapping to something not represented
379  * by a page descriptor (see also vm_normal_page()).
380  */
381 struct page *follow_page_mask(struct vm_area_struct *vma,
382                               unsigned long address, unsigned int flags,
383                               unsigned int *page_mask)
384 {
385         pgd_t *pgd;
386         struct page *page;
387         struct mm_struct *mm = vma->vm_mm;
388 
389         *page_mask = 0;
390 
391         /* make this handle hugepd */
392         page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
393         if (!IS_ERR(page)) {
394                 BUG_ON(flags & FOLL_GET);
395                 return page;
396         }
397 
398         pgd = pgd_offset(mm, address);
399 
400         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
401                 return no_page_table(vma, flags);
402 
403         if (pgd_huge(*pgd)) {
404                 page = follow_huge_pgd(mm, address, pgd, flags);
405                 if (page)
406                         return page;
407                 return no_page_table(vma, flags);
408         }
409         if (is_hugepd(__hugepd(pgd_val(*pgd)))) {
410                 page = follow_huge_pd(vma, address,
411                                       __hugepd(pgd_val(*pgd)), flags,
412                                       PGDIR_SHIFT);
413                 if (page)
414                         return page;
415                 return no_page_table(vma, flags);
416         }
417 
418         return follow_p4d_mask(vma, address, pgd, flags, page_mask);
419 }
420 
421 static int get_gate_page(struct mm_struct *mm, unsigned long address,
422                 unsigned int gup_flags, struct vm_area_struct **vma,
423                 struct page **page)
424 {
425         pgd_t *pgd;
426         p4d_t *p4d;
427         pud_t *pud;
428         pmd_t *pmd;
429         pte_t *pte;
430         int ret = -EFAULT;
431 
432         /* user gate pages are read-only */
433         if (gup_flags & FOLL_WRITE)
434                 return -EFAULT;
435         if (address > TASK_SIZE)
436                 pgd = pgd_offset_k(address);
437         else
438                 pgd = pgd_offset_gate(mm, address);
439         BUG_ON(pgd_none(*pgd));
440         p4d = p4d_offset(pgd, address);
441         BUG_ON(p4d_none(*p4d));
442         pud = pud_offset(p4d, address);
443         BUG_ON(pud_none(*pud));
444         pmd = pmd_offset(pud, address);
445         if (!pmd_present(*pmd))
446                 return -EFAULT;
447         VM_BUG_ON(pmd_trans_huge(*pmd));
448         pte = pte_offset_map(pmd, address);
449         if (pte_none(*pte))
450                 goto unmap;
451         *vma = get_gate_vma(mm);
452         if (!page)
453                 goto out;
454         *page = vm_normal_page(*vma, address, *pte);
455         if (!*page) {
456                 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
457                         goto unmap;
458                 *page = pte_page(*pte);
459 
460                 /*
461                  * This should never happen (a device public page in the gate
462                  * area).
463                  */
464                 if (is_device_public_page(*page))
465                         goto unmap;
466         }
467         get_page(*page);
468 out:
469         ret = 0;
470 unmap:
471         pte_unmap(pte);
472         return ret;
473 }
474 
475 /*
476  * mmap_sem must be held on entry.  If @nonblocking != NULL and
477  * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
478  * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
479  */
480 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
481                 unsigned long address, unsigned int *flags, int *nonblocking)
482 {
483         unsigned int fault_flags = 0;
484         int ret;
485 
486         /* mlock all present pages, but do not fault in new pages */
487         if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
488                 return -ENOENT;
489         if (*flags & FOLL_WRITE)
490                 fault_flags |= FAULT_FLAG_WRITE;
491         if (*flags & FOLL_REMOTE)
492                 fault_flags |= FAULT_FLAG_REMOTE;
493         if (nonblocking)
494                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
495         if (*flags & FOLL_NOWAIT)
496                 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
497         if (*flags & FOLL_TRIED) {
498                 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
499                 fault_flags |= FAULT_FLAG_TRIED;
500         }
501 
502         ret = handle_mm_fault(vma, address, fault_flags);
503         if (ret & VM_FAULT_ERROR) {
504                 int err = vm_fault_to_errno(ret, *flags);
505 
506                 if (err)
507                         return err;
508                 BUG();
509         }
510 
511         if (tsk) {
512                 if (ret & VM_FAULT_MAJOR)
513                         tsk->maj_flt++;
514                 else
515                         tsk->min_flt++;
516         }
517 
518         if (ret & VM_FAULT_RETRY) {
519                 if (nonblocking && !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
520                         *nonblocking = 0;
521                 return -EBUSY;
522         }
523 
524         /*
525          * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
526          * necessary, even if maybe_mkwrite decided not to set pte_write. We
527          * can thus safely do subsequent page lookups as if they were reads.
528          * But only do so when looping for pte_write is futile: in some cases
529          * userspace may also be wanting to write to the gotten user page,
530          * which a read fault here might prevent (a readonly page might get
531          * reCOWed by userspace write).
532          */
533         if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
534                 *flags |= FOLL_COW;
535         return 0;
536 }
537 
538 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
539 {
540         vm_flags_t vm_flags = vma->vm_flags;
541         int write = (gup_flags & FOLL_WRITE);
542         int foreign = (gup_flags & FOLL_REMOTE);
543 
544         if (vm_flags & (VM_IO | VM_PFNMAP))
545                 return -EFAULT;
546 
547         if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
548                 return -EFAULT;
549 
550         if (write) {
551                 if (!(vm_flags & VM_WRITE)) {
552                         if (!(gup_flags & FOLL_FORCE))
553                                 return -EFAULT;
554                         /*
555                          * We used to let the write,force case do COW in a
556                          * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
557                          * set a breakpoint in a read-only mapping of an
558                          * executable, without corrupting the file (yet only
559                          * when that file had been opened for writing!).
560                          * Anon pages in shared mappings are surprising: now
561                          * just reject it.
562                          */
563                         if (!is_cow_mapping(vm_flags))
564                                 return -EFAULT;
565                 }
566         } else if (!(vm_flags & VM_READ)) {
567                 if (!(gup_flags & FOLL_FORCE))
568                         return -EFAULT;
569                 /*
570                  * Is there actually any vma we can reach here which does not
571                  * have VM_MAYREAD set?
572                  */
573                 if (!(vm_flags & VM_MAYREAD))
574                         return -EFAULT;
575         }
576         /*
577          * gups are always data accesses, not instruction
578          * fetches, so execute=false here
579          */
580         if (!arch_vma_access_permitted(vma, write, false, foreign))
581                 return -EFAULT;
582         return 0;
583 }
584 
585 /**
586  * __get_user_pages() - pin user pages in memory
587  * @tsk:        task_struct of target task
588  * @mm:         mm_struct of target mm
589  * @start:      starting user address
590  * @nr_pages:   number of pages from start to pin
591  * @gup_flags:  flags modifying pin behaviour
592  * @pages:      array that receives pointers to the pages pinned.
593  *              Should be at least nr_pages long. Or NULL, if caller
594  *              only intends to ensure the pages are faulted in.
595  * @vmas:       array of pointers to vmas corresponding to each page.
596  *              Or NULL if the caller does not require them.
597  * @nonblocking: whether waiting for disk IO or mmap_sem contention
598  *
599  * Returns number of pages pinned. This may be fewer than the number
600  * requested. If nr_pages is 0 or negative, returns 0. If no pages
601  * were pinned, returns -errno. Each page returned must be released
602  * with a put_page() call when it is finished with. vmas will only
603  * remain valid while mmap_sem is held.
604  *
605  * Must be called with mmap_sem held.  It may be released.  See below.
606  *
607  * __get_user_pages walks a process's page tables and takes a reference to
608  * each struct page that each user address corresponds to at a given
609  * instant. That is, it takes the page that would be accessed if a user
610  * thread accesses the given user virtual address at that instant.
611  *
612  * This does not guarantee that the page exists in the user mappings when
613  * __get_user_pages returns, and there may even be a completely different
614  * page there in some cases (eg. if mmapped pagecache has been invalidated
615  * and subsequently re faulted). However it does guarantee that the page
616  * won't be freed completely. And mostly callers simply care that the page
617  * contains data that was valid *at some point in time*. Typically, an IO
618  * or similar operation cannot guarantee anything stronger anyway because
619  * locks can't be held over the syscall boundary.
620  *
621  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
622  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
623  * appropriate) must be called after the page is finished with, and
624  * before put_page is called.
625  *
626  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
627  * or mmap_sem contention, and if waiting is needed to pin all pages,
628  * *@nonblocking will be set to 0.  Further, if @gup_flags does not
629  * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
630  * this case.
631  *
632  * A caller using such a combination of @nonblocking and @gup_flags
633  * must therefore hold the mmap_sem for reading only, and recognize
634  * when it's been released.  Otherwise, it must be held for either
635  * reading or writing and will not be released.
636  *
637  * In most cases, get_user_pages or get_user_pages_fast should be used
638  * instead of __get_user_pages. __get_user_pages should be used only if
639  * you need some special @gup_flags.
640  */
641 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
642                 unsigned long start, unsigned long nr_pages,
643                 unsigned int gup_flags, struct page **pages,
644                 struct vm_area_struct **vmas, int *nonblocking)
645 {
646         long i = 0;
647         unsigned int page_mask;
648         struct vm_area_struct *vma = NULL;
649 
650         if (!nr_pages)
651                 return 0;
652 
653         VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
654 
655         /*
656          * If FOLL_FORCE is set then do not force a full fault as the hinting
657          * fault information is unrelated to the reference behaviour of a task
658          * using the address space
659          */
660         if (!(gup_flags & FOLL_FORCE))
661                 gup_flags |= FOLL_NUMA;
662 
663         do {
664                 struct page *page;
665                 unsigned int foll_flags = gup_flags;
666                 unsigned int page_increm;
667 
668                 /* first iteration or cross vma bound */
669                 if (!vma || start >= vma->vm_end) {
670                         vma = find_extend_vma(mm, start);
671                         if (!vma && in_gate_area(mm, start)) {
672                                 int ret;
673                                 ret = get_gate_page(mm, start & PAGE_MASK,
674                                                 gup_flags, &vma,
675                                                 pages ? &pages[i] : NULL);
676                                 if (ret)
677                                         return i ? : ret;
678                                 page_mask = 0;
679                                 goto next_page;
680                         }
681 
682                         if (!vma || check_vma_flags(vma, gup_flags))
683                                 return i ? : -EFAULT;
684                         if (is_vm_hugetlb_page(vma)) {
685                                 i = follow_hugetlb_page(mm, vma, pages, vmas,
686                                                 &start, &nr_pages, i,
687                                                 gup_flags, nonblocking);
688                                 continue;
689                         }
690                 }
691 retry:
692                 /*
693                  * If we have a pending SIGKILL, don't keep faulting pages and
694                  * potentially allocating memory.
695                  */
696                 if (unlikely(fatal_signal_pending(current)))
697                         return i ? i : -ERESTARTSYS;
698                 cond_resched();
699                 page = follow_page_mask(vma, start, foll_flags, &page_mask);
700                 if (!page) {
701                         int ret;
702                         ret = faultin_page(tsk, vma, start, &foll_flags,
703                                         nonblocking);
704                         switch (ret) {
705                         case 0:
706                                 goto retry;
707                         case -EFAULT:
708                         case -ENOMEM:
709                         case -EHWPOISON:
710                                 return i ? i : ret;
711                         case -EBUSY:
712                                 return i;
713                         case -ENOENT:
714                                 goto next_page;
715                         }
716                         BUG();
717                 } else if (PTR_ERR(page) == -EEXIST) {
718                         /*
719                          * Proper page table entry exists, but no corresponding
720                          * struct page.
721                          */
722                         goto next_page;
723                 } else if (IS_ERR(page)) {
724                         return i ? i : PTR_ERR(page);
725                 }
726                 if (pages) {
727                         pages[i] = page;
728                         flush_anon_page(vma, page, start);
729                         flush_dcache_page(page);
730                         page_mask = 0;
731                 }
732 next_page:
733                 if (vmas) {
734                         vmas[i] = vma;
735                         page_mask = 0;
736                 }
737                 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
738                 if (page_increm > nr_pages)
739                         page_increm = nr_pages;
740                 i += page_increm;
741                 start += page_increm * PAGE_SIZE;
742                 nr_pages -= page_increm;
743         } while (nr_pages);
744         return i;
745 }
746 
747 static bool vma_permits_fault(struct vm_area_struct *vma,
748                               unsigned int fault_flags)
749 {
750         bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
751         bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
752         vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
753 
754         if (!(vm_flags & vma->vm_flags))
755                 return false;
756 
757         /*
758          * The architecture might have a hardware protection
759          * mechanism other than read/write that can deny access.
760          *
761          * gup always represents data access, not instruction
762          * fetches, so execute=false here:
763          */
764         if (!arch_vma_access_permitted(vma, write, false, foreign))
765                 return false;
766 
767         return true;
768 }
769 
770 /*
771  * fixup_user_fault() - manually resolve a user page fault
772  * @tsk:        the task_struct to use for page fault accounting, or
773  *              NULL if faults are not to be recorded.
774  * @mm:         mm_struct of target mm
775  * @address:    user address
776  * @fault_flags:flags to pass down to handle_mm_fault()
777  * @unlocked:   did we unlock the mmap_sem while retrying, maybe NULL if caller
778  *              does not allow retry
779  *
780  * This is meant to be called in the specific scenario where for locking reasons
781  * we try to access user memory in atomic context (within a pagefault_disable()
782  * section), this returns -EFAULT, and we want to resolve the user fault before
783  * trying again.
784  *
785  * Typically this is meant to be used by the futex code.
786  *
787  * The main difference with get_user_pages() is that this function will
788  * unconditionally call handle_mm_fault() which will in turn perform all the
789  * necessary SW fixup of the dirty and young bits in the PTE, while
790  * get_user_pages() only guarantees to update these in the struct page.
791  *
792  * This is important for some architectures where those bits also gate the
793  * access permission to the page because they are maintained in software.  On
794  * such architectures, gup() will not be enough to make a subsequent access
795  * succeed.
796  *
797  * This function will not return with an unlocked mmap_sem. So it has not the
798  * same semantics wrt the @mm->mmap_sem as does filemap_fault().
799  */
800 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
801                      unsigned long address, unsigned int fault_flags,
802                      bool *unlocked)
803 {
804         struct vm_area_struct *vma;
805         int ret, major = 0;
806 
807         if (unlocked)
808                 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
809 
810 retry:
811         vma = find_extend_vma(mm, address);
812         if (!vma || address < vma->vm_start)
813                 return -EFAULT;
814 
815         if (!vma_permits_fault(vma, fault_flags))
816                 return -EFAULT;
817 
818         ret = handle_mm_fault(vma, address, fault_flags);
819         major |= ret & VM_FAULT_MAJOR;
820         if (ret & VM_FAULT_ERROR) {
821                 int err = vm_fault_to_errno(ret, 0);
822 
823                 if (err)
824                         return err;
825                 BUG();
826         }
827 
828         if (ret & VM_FAULT_RETRY) {
829                 down_read(&mm->mmap_sem);
830                 if (!(fault_flags & FAULT_FLAG_TRIED)) {
831                         *unlocked = true;
832                         fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
833                         fault_flags |= FAULT_FLAG_TRIED;
834                         goto retry;
835                 }
836         }
837 
838         if (tsk) {
839                 if (major)
840                         tsk->maj_flt++;
841                 else
842                         tsk->min_flt++;
843         }
844         return 0;
845 }
846 EXPORT_SYMBOL_GPL(fixup_user_fault);
847 
848 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
849                                                 struct mm_struct *mm,
850                                                 unsigned long start,
851                                                 unsigned long nr_pages,
852                                                 struct page **pages,
853                                                 struct vm_area_struct **vmas,
854                                                 int *locked,
855                                                 unsigned int flags)
856 {
857         long ret, pages_done;
858         bool lock_dropped;
859 
860         if (locked) {
861                 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
862                 BUG_ON(vmas);
863                 /* check caller initialized locked */
864                 BUG_ON(*locked != 1);
865         }
866 
867         if (pages)
868                 flags |= FOLL_GET;
869 
870         pages_done = 0;
871         lock_dropped = false;
872         for (;;) {
873                 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
874                                        vmas, locked);
875                 if (!locked)
876                         /* VM_FAULT_RETRY couldn't trigger, bypass */
877                         return ret;
878 
879                 /* VM_FAULT_RETRY cannot return errors */
880                 if (!*locked) {
881                         BUG_ON(ret < 0);
882                         BUG_ON(ret >= nr_pages);
883                 }
884 
885                 if (!pages)
886                         /* If it's a prefault don't insist harder */
887                         return ret;
888 
889                 if (ret > 0) {
890                         nr_pages -= ret;
891                         pages_done += ret;
892                         if (!nr_pages)
893                                 break;
894                 }
895                 if (*locked) {
896                         /*
897                          * VM_FAULT_RETRY didn't trigger or it was a
898                          * FOLL_NOWAIT.
899                          */
900                         if (!pages_done)
901                                 pages_done = ret;
902                         break;
903                 }
904                 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
905                 pages += ret;
906                 start += ret << PAGE_SHIFT;
907 
908                 /*
909                  * Repeat on the address that fired VM_FAULT_RETRY
910                  * without FAULT_FLAG_ALLOW_RETRY but with
911                  * FAULT_FLAG_TRIED.
912                  */
913                 *locked = 1;
914                 lock_dropped = true;
915                 down_read(&mm->mmap_sem);
916                 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
917                                        pages, NULL, NULL);
918                 if (ret != 1) {
919                         BUG_ON(ret > 1);
920                         if (!pages_done)
921                                 pages_done = ret;
922                         break;
923                 }
924                 nr_pages--;
925                 pages_done++;
926                 if (!nr_pages)
927                         break;
928                 pages++;
929                 start += PAGE_SIZE;
930         }
931         if (lock_dropped && *locked) {
932                 /*
933                  * We must let the caller know we temporarily dropped the lock
934                  * and so the critical section protected by it was lost.
935                  */
936                 up_read(&mm->mmap_sem);
937                 *locked = 0;
938         }
939         return pages_done;
940 }
941 
942 /*
943  * We can leverage the VM_FAULT_RETRY functionality in the page fault
944  * paths better by using either get_user_pages_locked() or
945  * get_user_pages_unlocked().
946  *
947  * get_user_pages_locked() is suitable to replace the form:
948  *
949  *      down_read(&mm->mmap_sem);
950  *      do_something()
951  *      get_user_pages(tsk, mm, ..., pages, NULL);
952  *      up_read(&mm->mmap_sem);
953  *
954  *  to:
955  *
956  *      int locked = 1;
957  *      down_read(&mm->mmap_sem);
958  *      do_something()
959  *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
960  *      if (locked)
961  *          up_read(&mm->mmap_sem);
962  */
963 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
964                            unsigned int gup_flags, struct page **pages,
965                            int *locked)
966 {
967         return __get_user_pages_locked(current, current->mm, start, nr_pages,
968                                        pages, NULL, locked,
969                                        gup_flags | FOLL_TOUCH);
970 }
971 EXPORT_SYMBOL(get_user_pages_locked);
972 
973 /*
974  * get_user_pages_unlocked() is suitable to replace the form:
975  *
976  *      down_read(&mm->mmap_sem);
977  *      get_user_pages(tsk, mm, ..., pages, NULL);
978  *      up_read(&mm->mmap_sem);
979  *
980  *  with:
981  *
982  *      get_user_pages_unlocked(tsk, mm, ..., pages);
983  *
984  * It is functionally equivalent to get_user_pages_fast so
985  * get_user_pages_fast should be used instead if specific gup_flags
986  * (e.g. FOLL_FORCE) are not required.
987  */
988 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
989                              struct page **pages, unsigned int gup_flags)
990 {
991         struct mm_struct *mm = current->mm;
992         int locked = 1;
993         long ret;
994 
995         down_read(&mm->mmap_sem);
996         ret = __get_user_pages_locked(current, mm, start, nr_pages, pages, NULL,
997                                       &locked, gup_flags | FOLL_TOUCH);
998         if (locked)
999                 up_read(&mm->mmap_sem);
1000         return ret;
1001 }
1002 EXPORT_SYMBOL(get_user_pages_unlocked);
1003 
1004 /*
1005  * get_user_pages_remote() - pin user pages in memory
1006  * @tsk:        the task_struct to use for page fault accounting, or
1007  *              NULL if faults are not to be recorded.
1008  * @mm:         mm_struct of target mm
1009  * @start:      starting user address
1010  * @nr_pages:   number of pages from start to pin
1011  * @gup_flags:  flags modifying lookup behaviour
1012  * @pages:      array that receives pointers to the pages pinned.
1013  *              Should be at least nr_pages long. Or NULL, if caller
1014  *              only intends to ensure the pages are faulted in.
1015  * @vmas:       array of pointers to vmas corresponding to each page.
1016  *              Or NULL if the caller does not require them.
1017  * @locked:     pointer to lock flag indicating whether lock is held and
1018  *              subsequently whether VM_FAULT_RETRY functionality can be
1019  *              utilised. Lock must initially be held.
1020  *
1021  * Returns number of pages pinned. This may be fewer than the number
1022  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1023  * were pinned, returns -errno. Each page returned must be released
1024  * with a put_page() call when it is finished with. vmas will only
1025  * remain valid while mmap_sem is held.
1026  *
1027  * Must be called with mmap_sem held for read or write.
1028  *
1029  * get_user_pages walks a process's page tables and takes a reference to
1030  * each struct page that each user address corresponds to at a given
1031  * instant. That is, it takes the page that would be accessed if a user
1032  * thread accesses the given user virtual address at that instant.
1033  *
1034  * This does not guarantee that the page exists in the user mappings when
1035  * get_user_pages returns, and there may even be a completely different
1036  * page there in some cases (eg. if mmapped pagecache has been invalidated
1037  * and subsequently re faulted). However it does guarantee that the page
1038  * won't be freed completely. And mostly callers simply care that the page
1039  * contains data that was valid *at some point in time*. Typically, an IO
1040  * or similar operation cannot guarantee anything stronger anyway because
1041  * locks can't be held over the syscall boundary.
1042  *
1043  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
1044  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
1045  * be called after the page is finished with, and before put_page is called.
1046  *
1047  * get_user_pages is typically used for fewer-copy IO operations, to get a
1048  * handle on the memory by some means other than accesses via the user virtual
1049  * addresses. The pages may be submitted for DMA to devices or accessed via
1050  * their kernel linear mapping (via the kmap APIs). Care should be taken to
1051  * use the correct cache flushing APIs.
1052  *
1053  * See also get_user_pages_fast, for performance critical applications.
1054  *
1055  * get_user_pages should be phased out in favor of
1056  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
1057  * should use get_user_pages because it cannot pass
1058  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
1059  */
1060 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
1061                 unsigned long start, unsigned long nr_pages,
1062                 unsigned int gup_flags, struct page **pages,
1063                 struct vm_area_struct **vmas, int *locked)
1064 {
1065         return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
1066                                        locked,
1067                                        gup_flags | FOLL_TOUCH | FOLL_REMOTE);
1068 }
1069 EXPORT_SYMBOL(get_user_pages_remote);
1070 
1071 /*
1072  * This is the same as get_user_pages_remote(), just with a
1073  * less-flexible calling convention where we assume that the task
1074  * and mm being operated on are the current task's and don't allow
1075  * passing of a locked parameter.  We also obviously don't pass
1076  * FOLL_REMOTE in here.
1077  */
1078 long get_user_pages(unsigned long start, unsigned long nr_pages,
1079                 unsigned int gup_flags, struct page **pages,
1080                 struct vm_area_struct **vmas)
1081 {
1082         return __get_user_pages_locked(current, current->mm, start, nr_pages,
1083                                        pages, vmas, NULL,
1084                                        gup_flags | FOLL_TOUCH);
1085 }
1086 EXPORT_SYMBOL(get_user_pages);
1087 
1088 #ifdef CONFIG_FS_DAX
1089 /*
1090  * This is the same as get_user_pages() in that it assumes we are
1091  * operating on the current task's mm, but it goes further to validate
1092  * that the vmas associated with the address range are suitable for
1093  * longterm elevated page reference counts. For example, filesystem-dax
1094  * mappings are subject to the lifetime enforced by the filesystem and
1095  * we need guarantees that longterm users like RDMA and V4L2 only
1096  * establish mappings that have a kernel enforced revocation mechanism.
1097  *
1098  * "longterm" == userspace controlled elevated page count lifetime.
1099  * Contrast this to iov_iter_get_pages() usages which are transient.
1100  */
1101 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
1102                 unsigned int gup_flags, struct page **pages,
1103                 struct vm_area_struct **vmas_arg)
1104 {
1105         struct vm_area_struct **vmas = vmas_arg;
1106         struct vm_area_struct *vma_prev = NULL;
1107         long rc, i;
1108 
1109         if (!pages)
1110                 return -EINVAL;
1111 
1112         if (!vmas) {
1113                 vmas = kcalloc(nr_pages, sizeof(struct vm_area_struct *),
1114                                GFP_KERNEL);
1115                 if (!vmas)
1116                         return -ENOMEM;
1117         }
1118 
1119         rc = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
1120 
1121         for (i = 0; i < rc; i++) {
1122                 struct vm_area_struct *vma = vmas[i];
1123 
1124                 if (vma == vma_prev)
1125                         continue;
1126 
1127                 vma_prev = vma;
1128 
1129                 if (vma_is_fsdax(vma))
1130                         break;
1131         }
1132 
1133         /*
1134          * Either get_user_pages() failed, or the vma validation
1135          * succeeded, in either case we don't need to put_page() before
1136          * returning.
1137          */
1138         if (i >= rc)
1139                 goto out;
1140 
1141         for (i = 0; i < rc; i++)
1142                 put_page(pages[i]);
1143         rc = -EOPNOTSUPP;
1144 out:
1145         if (vmas != vmas_arg)
1146                 kfree(vmas);
1147         return rc;
1148 }
1149 EXPORT_SYMBOL(get_user_pages_longterm);
1150 #endif /* CONFIG_FS_DAX */
1151 
1152 /**
1153  * populate_vma_page_range() -  populate a range of pages in the vma.
1154  * @vma:   target vma
1155  * @start: start address
1156  * @end:   end address
1157  * @nonblocking:
1158  *
1159  * This takes care of mlocking the pages too if VM_LOCKED is set.
1160  *
1161  * return 0 on success, negative error code on error.
1162  *
1163  * vma->vm_mm->mmap_sem must be held.
1164  *
1165  * If @nonblocking is NULL, it may be held for read or write and will
1166  * be unperturbed.
1167  *
1168  * If @nonblocking is non-NULL, it must held for read only and may be
1169  * released.  If it's released, *@nonblocking will be set to 0.
1170  */
1171 long populate_vma_page_range(struct vm_area_struct *vma,
1172                 unsigned long start, unsigned long end, int *nonblocking)
1173 {
1174         struct mm_struct *mm = vma->vm_mm;
1175         unsigned long nr_pages = (end - start) / PAGE_SIZE;
1176         int gup_flags;
1177 
1178         VM_BUG_ON(start & ~PAGE_MASK);
1179         VM_BUG_ON(end   & ~PAGE_MASK);
1180         VM_BUG_ON_VMA(start < vma->vm_start, vma);
1181         VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1182         VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1183 
1184         gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1185         if (vma->vm_flags & VM_LOCKONFAULT)
1186                 gup_flags &= ~FOLL_POPULATE;
1187         /*
1188          * We want to touch writable mappings with a write fault in order
1189          * to break COW, except for shared mappings because these don't COW
1190          * and we would not want to dirty them for nothing.
1191          */
1192         if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1193                 gup_flags |= FOLL_WRITE;
1194 
1195         /*
1196          * We want mlock to succeed for regions that have any permissions
1197          * other than PROT_NONE.
1198          */
1199         if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1200                 gup_flags |= FOLL_FORCE;
1201 
1202         /*
1203          * We made sure addr is within a VMA, so the following will
1204          * not result in a stack expansion that recurses back here.
1205          */
1206         return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1207                                 NULL, NULL, nonblocking);
1208 }
1209 
1210 /*
1211  * __mm_populate - populate and/or mlock pages within a range of address space.
1212  *
1213  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1214  * flags. VMAs must be already marked with the desired vm_flags, and
1215  * mmap_sem must not be held.
1216  */
1217 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1218 {
1219         struct mm_struct *mm = current->mm;
1220         unsigned long end, nstart, nend;
1221         struct vm_area_struct *vma = NULL;
1222         int locked = 0;
1223         long ret = 0;
1224 
1225         end = start + len;
1226 
1227         for (nstart = start; nstart < end; nstart = nend) {
1228                 /*
1229                  * We want to fault in pages for [nstart; end) address range.
1230                  * Find first corresponding VMA.
1231                  */
1232                 if (!locked) {
1233                         locked = 1;
1234                         down_read(&mm->mmap_sem);
1235                         vma = find_vma(mm, nstart);
1236                 } else if (nstart >= vma->vm_end)
1237                         vma = vma->vm_next;
1238                 if (!vma || vma->vm_start >= end)
1239                         break;
1240                 /*
1241                  * Set [nstart; nend) to intersection of desired address
1242                  * range with the first VMA. Also, skip undesirable VMA types.
1243                  */
1244                 nend = min(end, vma->vm_end);
1245                 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1246                         continue;
1247                 if (nstart < vma->vm_start)
1248                         nstart = vma->vm_start;
1249                 /*
1250                  * Now fault in a range of pages. populate_vma_page_range()
1251                  * double checks the vma flags, so that it won't mlock pages
1252                  * if the vma was already munlocked.
1253                  */
1254                 ret = populate_vma_page_range(vma, nstart, nend, &locked);
1255                 if (ret < 0) {
1256                         if (ignore_errors) {
1257                                 ret = 0;
1258                                 continue;       /* continue at next VMA */
1259                         }
1260                         break;
1261                 }
1262                 nend = nstart + ret * PAGE_SIZE;
1263                 ret = 0;
1264         }
1265         if (locked)
1266                 up_read(&mm->mmap_sem);
1267         return ret;     /* 0 or negative error code */
1268 }
1269 
1270 /**
1271  * get_dump_page() - pin user page in memory while writing it to core dump
1272  * @addr: user address
1273  *
1274  * Returns struct page pointer of user page pinned for dump,
1275  * to be freed afterwards by put_page().
1276  *
1277  * Returns NULL on any kind of failure - a hole must then be inserted into
1278  * the corefile, to preserve alignment with its headers; and also returns
1279  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1280  * allowing a hole to be left in the corefile to save diskspace.
1281  *
1282  * Called without mmap_sem, but after all other threads have been killed.
1283  */
1284 #ifdef CONFIG_ELF_CORE
1285 struct page *get_dump_page(unsigned long addr)
1286 {
1287         struct vm_area_struct *vma;
1288         struct page *page;
1289 
1290         if (__get_user_pages(current, current->mm, addr, 1,
1291                              FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1292                              NULL) < 1)
1293                 return NULL;
1294         flush_cache_page(vma, addr, page_to_pfn(page));
1295         return page;
1296 }
1297 #endif /* CONFIG_ELF_CORE */
1298 
1299 /*
1300  * Generic Fast GUP
1301  *
1302  * get_user_pages_fast attempts to pin user pages by walking the page
1303  * tables directly and avoids taking locks. Thus the walker needs to be
1304  * protected from page table pages being freed from under it, and should
1305  * block any THP splits.
1306  *
1307  * One way to achieve this is to have the walker disable interrupts, and
1308  * rely on IPIs from the TLB flushing code blocking before the page table
1309  * pages are freed. This is unsuitable for architectures that do not need
1310  * to broadcast an IPI when invalidating TLBs.
1311  *
1312  * Another way to achieve this is to batch up page table containing pages
1313  * belonging to more than one mm_user, then rcu_sched a callback to free those
1314  * pages. Disabling interrupts will allow the fast_gup walker to both block
1315  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1316  * (which is a relatively rare event). The code below adopts this strategy.
1317  *
1318  * Before activating this code, please be aware that the following assumptions
1319  * are currently made:
1320  *
1321  *  *) Either HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table() is used to
1322  *  free pages containing page tables or TLB flushing requires IPI broadcast.
1323  *
1324  *  *) ptes can be read atomically by the architecture.
1325  *
1326  *  *) access_ok is sufficient to validate userspace address ranges.
1327  *
1328  * The last two assumptions can be relaxed by the addition of helper functions.
1329  *
1330  * This code is based heavily on the PowerPC implementation by Nick Piggin.
1331  */
1332 #ifdef CONFIG_HAVE_GENERIC_GUP
1333 
1334 #ifndef gup_get_pte
1335 /*
1336  * We assume that the PTE can be read atomically. If this is not the case for
1337  * your architecture, please provide the helper.
1338  */
1339 static inline pte_t gup_get_pte(pte_t *ptep)
1340 {
1341         return READ_ONCE(*ptep);
1342 }
1343 #endif
1344 
1345 static void undo_dev_pagemap(int *nr, int nr_start, struct page **pages)
1346 {
1347         while ((*nr) - nr_start) {
1348                 struct page *page = pages[--(*nr)];
1349 
1350                 ClearPageReferenced(page);
1351                 put_page(page);
1352         }
1353 }
1354 
1355 #ifdef __HAVE_ARCH_PTE_SPECIAL
1356 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1357                          int write, struct page **pages, int *nr)
1358 {
1359         struct dev_pagemap *pgmap = NULL;
1360         int nr_start = *nr, ret = 0;
1361         pte_t *ptep, *ptem;
1362 
1363         ptem = ptep = pte_offset_map(&pmd, addr);
1364         do {
1365                 pte_t pte = gup_get_pte(ptep);
1366                 struct page *head, *page;
1367 
1368                 /*
1369                  * Similar to the PMD case below, NUMA hinting must take slow
1370                  * path using the pte_protnone check.
1371                  */
1372                 if (pte_protnone(pte))
1373                         goto pte_unmap;
1374 
1375                 if (!pte_access_permitted(pte, write))
1376                         goto pte_unmap;
1377 
1378                 if (pte_devmap(pte)) {
1379                         pgmap = get_dev_pagemap(pte_pfn(pte), pgmap);
1380                         if (unlikely(!pgmap)) {
1381                                 undo_dev_pagemap(nr, nr_start, pages);
1382                                 goto pte_unmap;
1383                         }
1384                 } else if (pte_special(pte))
1385                         goto pte_unmap;
1386 
1387                 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1388                 page = pte_page(pte);
1389                 head = compound_head(page);
1390 
1391                 if (!page_cache_get_speculative(head))
1392                         goto pte_unmap;
1393 
1394                 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1395                         put_page(head);
1396                         goto pte_unmap;
1397                 }
1398 
1399                 VM_BUG_ON_PAGE(compound_head(page) != head, page);
1400 
1401                 SetPageReferenced(page);
1402                 pages[*nr] = page;
1403                 (*nr)++;
1404 
1405         } while (ptep++, addr += PAGE_SIZE, addr != end);
1406 
1407         ret = 1;
1408 
1409 pte_unmap:
1410         if (pgmap)
1411                 put_dev_pagemap(pgmap);
1412         pte_unmap(ptem);
1413         return ret;
1414 }
1415 #else
1416 
1417 /*
1418  * If we can't determine whether or not a pte is special, then fail immediately
1419  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1420  * to be special.
1421  *
1422  * For a futex to be placed on a THP tail page, get_futex_key requires a
1423  * __get_user_pages_fast implementation that can pin pages. Thus it's still
1424  * useful to have gup_huge_pmd even if we can't operate on ptes.
1425  */
1426 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1427                          int write, struct page **pages, int *nr)
1428 {
1429         return 0;
1430 }
1431 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1432 
1433 #if defined(__HAVE_ARCH_PTE_DEVMAP) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1434 static int __gup_device_huge(unsigned long pfn, unsigned long addr,
1435                 unsigned long end, struct page **pages, int *nr)
1436 {
1437         int nr_start = *nr;
1438         struct dev_pagemap *pgmap = NULL;
1439 
1440         do {
1441                 struct page *page = pfn_to_page(pfn);
1442 
1443                 pgmap = get_dev_pagemap(pfn, pgmap);
1444                 if (unlikely(!pgmap)) {
1445                         undo_dev_pagemap(nr, nr_start, pages);
1446                         return 0;
1447                 }
1448                 SetPageReferenced(page);
1449                 pages[*nr] = page;
1450                 get_page(page);
1451                 (*nr)++;
1452                 pfn++;
1453         } while (addr += PAGE_SIZE, addr != end);
1454 
1455         if (pgmap)
1456                 put_dev_pagemap(pgmap);
1457         return 1;
1458 }
1459 
1460 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1461                 unsigned long end, struct page **pages, int *nr)
1462 {
1463         unsigned long fault_pfn;
1464         int nr_start = *nr;
1465 
1466         fault_pfn = pmd_pfn(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1467         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1468                 return 0;
1469 
1470         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1471                 undo_dev_pagemap(nr, nr_start, pages);
1472                 return 0;
1473         }
1474         return 1;
1475 }
1476 
1477 static int __gup_device_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1478                 unsigned long end, struct page **pages, int *nr)
1479 {
1480         unsigned long fault_pfn;
1481         int nr_start = *nr;
1482 
1483         fault_pfn = pud_pfn(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1484         if (!__gup_device_huge(fault_pfn, addr, end, pages, nr))
1485                 return 0;
1486 
1487         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1488                 undo_dev_pagemap(nr, nr_start, pages);
1489                 return 0;
1490         }
1491         return 1;
1492 }
1493 #else
1494 static int __gup_device_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1495                 unsigned long end, struct page **pages, int *nr)
1496 {
1497         BUILD_BUG();
1498         return 0;
1499 }
1500 
1501 static int __gup_device_huge_pud(pud_t pud, pud_t *pudp, unsigned long addr,
1502                 unsigned long end, struct page **pages, int *nr)
1503 {
1504         BUILD_BUG();
1505         return 0;
1506 }
1507 #endif
1508 
1509 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1510                 unsigned long end, int write, struct page **pages, int *nr)
1511 {
1512         struct page *head, *page;
1513         int refs;
1514 
1515         if (!pmd_access_permitted(orig, write))
1516                 return 0;
1517 
1518         if (pmd_devmap(orig))
1519                 return __gup_device_huge_pmd(orig, pmdp, addr, end, pages, nr);
1520 
1521         refs = 0;
1522         page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1523         do {
1524                 pages[*nr] = page;
1525                 (*nr)++;
1526                 page++;
1527                 refs++;
1528         } while (addr += PAGE_SIZE, addr != end);
1529 
1530         head = compound_head(pmd_page(orig));
1531         if (!page_cache_add_speculative(head, refs)) {
1532                 *nr -= refs;
1533                 return 0;
1534         }
1535 
1536         if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1537                 *nr -= refs;
1538                 while (refs--)
1539                         put_page(head);
1540                 return 0;
1541         }
1542 
1543         SetPageReferenced(head);
1544         return 1;
1545 }
1546 
1547 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1548                 unsigned long end, int write, struct page **pages, int *nr)
1549 {
1550         struct page *head, *page;
1551         int refs;
1552 
1553         if (!pud_access_permitted(orig, write))
1554                 return 0;
1555 
1556         if (pud_devmap(orig))
1557                 return __gup_device_huge_pud(orig, pudp, addr, end, pages, nr);
1558 
1559         refs = 0;
1560         page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1561         do {
1562                 pages[*nr] = page;
1563                 (*nr)++;
1564                 page++;
1565                 refs++;
1566         } while (addr += PAGE_SIZE, addr != end);
1567 
1568         head = compound_head(pud_page(orig));
1569         if (!page_cache_add_speculative(head, refs)) {
1570                 *nr -= refs;
1571                 return 0;
1572         }
1573 
1574         if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1575                 *nr -= refs;
1576                 while (refs--)
1577                         put_page(head);
1578                 return 0;
1579         }
1580 
1581         SetPageReferenced(head);
1582         return 1;
1583 }
1584 
1585 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1586                         unsigned long end, int write,
1587                         struct page **pages, int *nr)
1588 {
1589         int refs;
1590         struct page *head, *page;
1591 
1592         if (!pgd_access_permitted(orig, write))
1593                 return 0;
1594 
1595         BUILD_BUG_ON(pgd_devmap(orig));
1596         refs = 0;
1597         page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1598         do {
1599                 pages[*nr] = page;
1600                 (*nr)++;
1601                 page++;
1602                 refs++;
1603         } while (addr += PAGE_SIZE, addr != end);
1604 
1605         head = compound_head(pgd_page(orig));
1606         if (!page_cache_add_speculative(head, refs)) {
1607                 *nr -= refs;
1608                 return 0;
1609         }
1610 
1611         if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1612                 *nr -= refs;
1613                 while (refs--)
1614                         put_page(head);
1615                 return 0;
1616         }
1617 
1618         SetPageReferenced(head);
1619         return 1;
1620 }
1621 
1622 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1623                 int write, struct page **pages, int *nr)
1624 {
1625         unsigned long next;
1626         pmd_t *pmdp;
1627 
1628         pmdp = pmd_offset(&pud, addr);
1629         do {
1630                 pmd_t pmd = READ_ONCE(*pmdp);
1631 
1632                 next = pmd_addr_end(addr, end);
1633                 if (!pmd_present(pmd))
1634                         return 0;
1635 
1636                 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1637                         /*
1638                          * NUMA hinting faults need to be handled in the GUP
1639                          * slowpath for accounting purposes and so that they
1640                          * can be serialised against THP migration.
1641                          */
1642                         if (pmd_protnone(pmd))
1643                                 return 0;
1644 
1645                         if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1646                                 pages, nr))
1647                                 return 0;
1648 
1649                 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1650                         /*
1651                          * architecture have different format for hugetlbfs
1652                          * pmd format and THP pmd format
1653                          */
1654                         if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1655                                          PMD_SHIFT, next, write, pages, nr))
1656                                 return 0;
1657                 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1658                         return 0;
1659         } while (pmdp++, addr = next, addr != end);
1660 
1661         return 1;
1662 }
1663 
1664 static int gup_pud_range(p4d_t p4d, unsigned long addr, unsigned long end,
1665                          int write, struct page **pages, int *nr)
1666 {
1667         unsigned long next;
1668         pud_t *pudp;
1669 
1670         pudp = pud_offset(&p4d, addr);
1671         do {
1672                 pud_t pud = READ_ONCE(*pudp);
1673 
1674                 next = pud_addr_end(addr, end);
1675                 if (pud_none(pud))
1676                         return 0;
1677                 if (unlikely(pud_huge(pud))) {
1678                         if (!gup_huge_pud(pud, pudp, addr, next, write,
1679                                           pages, nr))
1680                                 return 0;
1681                 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1682                         if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1683                                          PUD_SHIFT, next, write, pages, nr))
1684                                 return 0;
1685                 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1686                         return 0;
1687         } while (pudp++, addr = next, addr != end);
1688 
1689         return 1;
1690 }
1691 
1692 static int gup_p4d_range(pgd_t pgd, unsigned long addr, unsigned long end,
1693                          int write, struct page **pages, int *nr)
1694 {
1695         unsigned long next;
1696         p4d_t *p4dp;
1697 
1698         p4dp = p4d_offset(&pgd, addr);
1699         do {
1700                 p4d_t p4d = READ_ONCE(*p4dp);
1701 
1702                 next = p4d_addr_end(addr, end);
1703                 if (p4d_none(p4d))
1704                         return 0;
1705                 BUILD_BUG_ON(p4d_huge(p4d));
1706                 if (unlikely(is_hugepd(__hugepd(p4d_val(p4d))))) {
1707                         if (!gup_huge_pd(__hugepd(p4d_val(p4d)), addr,
1708                                          P4D_SHIFT, next, write, pages, nr))
1709                                 return 0;
1710                 } else if (!gup_pud_range(p4d, addr, next, write, pages, nr))
1711                         return 0;
1712         } while (p4dp++, addr = next, addr != end);
1713 
1714         return 1;
1715 }
1716 
1717 static void gup_pgd_range(unsigned long addr, unsigned long end,
1718                 int write, struct page **pages, int *nr)
1719 {
1720         unsigned long next;
1721         pgd_t *pgdp;
1722 
1723         pgdp = pgd_offset(current->mm, addr);
1724         do {
1725                 pgd_t pgd = READ_ONCE(*pgdp);
1726 
1727                 next = pgd_addr_end(addr, end);
1728                 if (pgd_none(pgd))
1729                         return;
1730                 if (unlikely(pgd_huge(pgd))) {
1731                         if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1732                                           pages, nr))
1733                                 return;
1734                 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1735                         if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1736                                          PGDIR_SHIFT, next, write, pages, nr))
1737                                 return;
1738                 } else if (!gup_p4d_range(pgd, addr, next, write, pages, nr))
1739                         return;
1740         } while (pgdp++, addr = next, addr != end);
1741 }
1742 
1743 #ifndef gup_fast_permitted
1744 /*
1745  * Check if it's allowed to use __get_user_pages_fast() for the range, or
1746  * we need to fall back to the slow version:
1747  */
1748 bool gup_fast_permitted(unsigned long start, int nr_pages, int write)
1749 {
1750         unsigned long len, end;
1751 
1752         len = (unsigned long) nr_pages << PAGE_SHIFT;
1753         end = start + len;
1754         return end >= start;
1755 }
1756 #endif
1757 
1758 /*
1759  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1760  * the regular GUP.
1761  * Note a difference with get_user_pages_fast: this always returns the
1762  * number of pages pinned, 0 if no pages were pinned.
1763  */
1764 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1765                           struct page **pages)
1766 {
1767         unsigned long addr, len, end;
1768         unsigned long flags;
1769         int nr = 0;
1770 
1771         start &= PAGE_MASK;
1772         addr = start;
1773         len = (unsigned long) nr_pages << PAGE_SHIFT;
1774         end = start + len;
1775 
1776         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1777                                         (void __user *)start, len)))
1778                 return 0;
1779 
1780         /*
1781          * Disable interrupts.  We use the nested form as we can already have
1782          * interrupts disabled by get_futex_key.
1783          *
1784          * With interrupts disabled, we block page table pages from being
1785          * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1786          * for more details.
1787          *
1788          * We do not adopt an rcu_read_lock(.) here as we also want to
1789          * block IPIs that come from THPs splitting.
1790          */
1791 
1792         if (gup_fast_permitted(start, nr_pages, write)) {
1793                 local_irq_save(flags);
1794                 gup_pgd_range(addr, end, write, pages, &nr);
1795                 local_irq_restore(flags);
1796         }
1797 
1798         return nr;
1799 }
1800 
1801 /**
1802  * get_user_pages_fast() - pin user pages in memory
1803  * @start:      starting user address
1804  * @nr_pages:   number of pages from start to pin
1805  * @write:      whether pages will be written to
1806  * @pages:      array that receives pointers to the pages pinned.
1807  *              Should be at least nr_pages long.
1808  *
1809  * Attempt to pin user pages in memory without taking mm->mmap_sem.
1810  * If not successful, it will fall back to taking the lock and
1811  * calling get_user_pages().
1812  *
1813  * Returns number of pages pinned. This may be fewer than the number
1814  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1815  * were pinned, returns -errno.
1816  */
1817 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1818                         struct page **pages)
1819 {
1820         unsigned long addr, len, end;
1821         int nr = 0, ret = 0;
1822 
1823         start &= PAGE_MASK;
1824         addr = start;
1825         len = (unsigned long) nr_pages << PAGE_SHIFT;
1826         end = start + len;
1827 
1828         if (nr_pages <= 0)
1829                 return 0;
1830 
1831         if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1832                                         (void __user *)start, len)))
1833                 return -EFAULT;
1834 
1835         if (gup_fast_permitted(start, nr_pages, write)) {
1836                 local_irq_disable();
1837                 gup_pgd_range(addr, end, write, pages, &nr);
1838                 local_irq_enable();
1839                 ret = nr;
1840         }
1841 
1842         if (nr < nr_pages) {
1843                 /* Try to get the remaining pages with get_user_pages */
1844                 start += nr << PAGE_SHIFT;
1845                 pages += nr;
1846 
1847                 ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1848                                 write ? FOLL_WRITE : 0);
1849 
1850                 /* Have to be a bit careful with return values */
1851                 if (nr > 0) {
1852                         if (ret < 0)
1853                                 ret = nr;
1854                         else
1855                                 ret += nr;
1856                 }
1857         }
1858 
1859         return ret;
1860 }
1861 
1862 #endif /* CONFIG_HAVE_GENERIC_GUP */
1863 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp