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/pagemap.h>
8 #include <linux/rmap.h>
9 #include <linux/swap.h>
10 #include <linux/swapops.h>
11
12 #include <linux/sched.h>
13 #include <linux/rwsem.h>
14 #include <linux/hugetlb.h>
15
16 #include <asm/pgtable.h>
17 #include <asm/tlbflush.h>
18
19 #include "internal.h"
20
no_page_table(struct vm_area_struct * vma,unsigned int flags)21 static struct page *no_page_table(struct vm_area_struct *vma,
22 unsigned int flags)
23 {
24 /*
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
follow_pfn_pte(struct vm_area_struct * vma,unsigned long address,pte_t * pte,unsigned int flags)37 static 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
61 /*
62 * FOLL_FORCE or a forced COW break can write even to unwritable pte's,
63 * but only after we've gone through a COW cycle and they are dirty.
64 */
can_follow_write_pte(pte_t pte,unsigned int flags)65 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
66 {
67 return pte_write(pte) || ((flags & FOLL_COW) && pte_dirty(pte));
68 }
69
70 /*
71 * A (separate) COW fault might break the page the other way and
72 * get_user_pages() would return the page from what is now the wrong
73 * VM. So we need to force a COW break at GUP time even for reads.
74 */
should_force_cow_break(struct vm_area_struct * vma,unsigned int flags)75 static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned int flags)
76 {
77 return is_cow_mapping(vma->vm_flags) && (flags & FOLL_GET);
78 }
79
follow_page_pte(struct vm_area_struct * vma,unsigned long address,pmd_t * pmd,unsigned int flags)80 static struct page *follow_page_pte(struct vm_area_struct *vma,
81 unsigned long address, pmd_t *pmd, unsigned int flags)
82 {
83 struct mm_struct *mm = vma->vm_mm;
84 struct page *page;
85 spinlock_t *ptl;
86 pte_t *ptep, pte;
87
88 retry:
89 if (unlikely(pmd_bad(*pmd)))
90 return no_page_table(vma, flags);
91
92 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
93 pte = *ptep;
94 if (!pte_present(pte)) {
95 swp_entry_t entry;
96 /*
97 * KSM's break_ksm() relies upon recognizing a ksm page
98 * even while it is being migrated, so for that case we
99 * need migration_entry_wait().
100 */
101 if (likely(!(flags & FOLL_MIGRATION)))
102 goto no_page;
103 if (pte_none(pte))
104 goto no_page;
105 entry = pte_to_swp_entry(pte);
106 if (!is_migration_entry(entry))
107 goto no_page;
108 pte_unmap_unlock(ptep, ptl);
109 migration_entry_wait(mm, pmd, address);
110 goto retry;
111 }
112 if ((flags & FOLL_NUMA) && pte_protnone(pte))
113 goto no_page;
114 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
115 pte_unmap_unlock(ptep, ptl);
116 return NULL;
117 }
118
119 page = vm_normal_page(vma, address, pte);
120 if (unlikely(!page)) {
121 if (flags & FOLL_DUMP) {
122 /* Avoid special (like zero) pages in core dumps */
123 page = ERR_PTR(-EFAULT);
124 goto out;
125 }
126
127 if (is_zero_pfn(pte_pfn(pte))) {
128 page = pte_page(pte);
129 } else {
130 int ret;
131
132 ret = follow_pfn_pte(vma, address, ptep, flags);
133 page = ERR_PTR(ret);
134 goto out;
135 }
136 }
137
138 if (flags & FOLL_GET) {
139 if (unlikely(!try_get_page_foll(page))) {
140 page = ERR_PTR(-ENOMEM);
141 goto out;
142 }
143 }
144 if (flags & FOLL_TOUCH) {
145 if ((flags & FOLL_WRITE) &&
146 !pte_dirty(pte) && !PageDirty(page))
147 set_page_dirty(page);
148 /*
149 * pte_mkyoung() would be more correct here, but atomic care
150 * is needed to avoid losing the dirty bit: it is easier to use
151 * mark_page_accessed().
152 */
153 mark_page_accessed(page);
154 }
155 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
156 /*
157 * The preliminary mapping check is mainly to avoid the
158 * pointless overhead of lock_page on the ZERO_PAGE
159 * which might bounce very badly if there is contention.
160 *
161 * If the page is already locked, we don't need to
162 * handle it now - vmscan will handle it later if and
163 * when it attempts to reclaim the page.
164 */
165 if (page->mapping && trylock_page(page)) {
166 lru_add_drain(); /* push cached pages to LRU */
167 /*
168 * Because we lock page here, and migration is
169 * blocked by the pte's page reference, and we
170 * know the page is still mapped, we don't even
171 * need to check for file-cache page truncation.
172 */
173 mlock_vma_page(page);
174 unlock_page(page);
175 }
176 }
177 out:
178 pte_unmap_unlock(ptep, ptl);
179 return page;
180 no_page:
181 pte_unmap_unlock(ptep, ptl);
182 if (!pte_none(pte))
183 return NULL;
184 return no_page_table(vma, flags);
185 }
186
187 /**
188 * follow_page_mask - look up a page descriptor from a user-virtual address
189 * @vma: vm_area_struct mapping @address
190 * @address: virtual address to look up
191 * @flags: flags modifying lookup behaviour
192 * @page_mask: on output, *page_mask is set according to the size of the page
193 *
194 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
195 *
196 * Returns the mapped (struct page *), %NULL if no mapping exists, or
197 * an error pointer if there is a mapping to something not represented
198 * by a page descriptor (see also vm_normal_page()).
199 */
follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,unsigned int * page_mask)200 struct page *follow_page_mask(struct vm_area_struct *vma,
201 unsigned long address, unsigned int flags,
202 unsigned int *page_mask)
203 {
204 pgd_t *pgd;
205 pud_t *pud;
206 pmd_t *pmd;
207 spinlock_t *ptl;
208 struct page *page;
209 struct mm_struct *mm = vma->vm_mm;
210
211 *page_mask = 0;
212
213 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
214 if (!IS_ERR(page)) {
215 BUG_ON(flags & FOLL_GET);
216 return page;
217 }
218
219 pgd = pgd_offset(mm, address);
220 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
221 return no_page_table(vma, flags);
222
223 pud = pud_offset(pgd, address);
224 if (pud_none(*pud))
225 return no_page_table(vma, flags);
226 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
227 page = follow_huge_pud(mm, address, pud, flags);
228 if (page)
229 return page;
230 return no_page_table(vma, flags);
231 }
232 if (unlikely(pud_bad(*pud)))
233 return no_page_table(vma, flags);
234
235 pmd = pmd_offset(pud, address);
236 if (pmd_none(*pmd))
237 return no_page_table(vma, flags);
238 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
239 page = follow_huge_pmd(mm, address, pmd, flags);
240 if (page)
241 return page;
242 return no_page_table(vma, flags);
243 }
244 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
245 return no_page_table(vma, flags);
246 if (pmd_trans_huge(*pmd)) {
247 if (flags & FOLL_SPLIT) {
248 split_huge_page_pmd(vma, address, pmd);
249 return follow_page_pte(vma, address, pmd, flags);
250 }
251 ptl = pmd_lock(mm, pmd);
252 if (likely(pmd_trans_huge(*pmd))) {
253 if (unlikely(pmd_trans_splitting(*pmd))) {
254 spin_unlock(ptl);
255 wait_split_huge_page(vma->anon_vma, pmd);
256 } else {
257 page = follow_trans_huge_pmd(vma, address,
258 pmd, flags);
259 spin_unlock(ptl);
260 *page_mask = HPAGE_PMD_NR - 1;
261 return page;
262 }
263 } else
264 spin_unlock(ptl);
265 }
266 return follow_page_pte(vma, address, pmd, flags);
267 }
268
get_gate_page(struct mm_struct * mm,unsigned long address,unsigned int gup_flags,struct vm_area_struct ** vma,struct page ** page)269 static int get_gate_page(struct mm_struct *mm, unsigned long address,
270 unsigned int gup_flags, struct vm_area_struct **vma,
271 struct page **page)
272 {
273 pgd_t *pgd;
274 pud_t *pud;
275 pmd_t *pmd;
276 pte_t *pte;
277 int ret = -EFAULT;
278
279 /* user gate pages are read-only */
280 if (gup_flags & FOLL_WRITE)
281 return -EFAULT;
282 if (address > TASK_SIZE)
283 pgd = pgd_offset_k(address);
284 else
285 pgd = pgd_offset_gate(mm, address);
286 BUG_ON(pgd_none(*pgd));
287 pud = pud_offset(pgd, address);
288 BUG_ON(pud_none(*pud));
289 pmd = pmd_offset(pud, address);
290 if (pmd_none(*pmd))
291 return -EFAULT;
292 VM_BUG_ON(pmd_trans_huge(*pmd));
293 pte = pte_offset_map(pmd, address);
294 if (pte_none(*pte))
295 goto unmap;
296 *vma = get_gate_vma(mm);
297 if (!page)
298 goto out;
299 *page = vm_normal_page(*vma, address, *pte);
300 if (!*page) {
301 if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
302 goto unmap;
303 *page = pte_page(*pte);
304 }
305 if (unlikely(!try_get_page(*page))) {
306 ret = -ENOMEM;
307 goto unmap;
308 }
309 out:
310 ret = 0;
311 unmap:
312 pte_unmap(pte);
313 return ret;
314 }
315
316 /*
317 * mmap_sem must be held on entry. If @nonblocking != NULL and
318 * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
319 * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
320 */
faultin_page(struct task_struct * tsk,struct vm_area_struct * vma,unsigned long address,unsigned int * flags,int * nonblocking)321 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
322 unsigned long address, unsigned int *flags, int *nonblocking)
323 {
324 struct mm_struct *mm = vma->vm_mm;
325 unsigned int fault_flags = 0;
326 int ret;
327
328 /* mlock all present pages, but do not fault in new pages */
329 if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
330 return -ENOENT;
331 if (*flags & FOLL_WRITE)
332 fault_flags |= FAULT_FLAG_WRITE;
333 if (nonblocking)
334 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
335 if (*flags & FOLL_NOWAIT)
336 fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
337 if (*flags & FOLL_TRIED) {
338 VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
339 fault_flags |= FAULT_FLAG_TRIED;
340 }
341
342 ret = handle_mm_fault(mm, vma, address, fault_flags);
343 if (ret & VM_FAULT_ERROR) {
344 if (ret & VM_FAULT_OOM)
345 return -ENOMEM;
346 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
347 return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
348 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
349 return -EFAULT;
350 BUG();
351 }
352
353 if (tsk) {
354 if (ret & VM_FAULT_MAJOR)
355 tsk->maj_flt++;
356 else
357 tsk->min_flt++;
358 }
359
360 if (ret & VM_FAULT_RETRY) {
361 if (nonblocking)
362 *nonblocking = 0;
363 return -EBUSY;
364 }
365
366 /*
367 * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
368 * necessary, even if maybe_mkwrite decided not to set pte_write. We
369 * can thus safely do subsequent page lookups as if they were reads.
370 * But only do so when looping for pte_write is futile: in some cases
371 * userspace may also be wanting to write to the gotten user page,
372 * which a read fault here might prevent (a readonly page might get
373 * reCOWed by userspace write).
374 */
375 if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
376 *flags |= FOLL_COW;
377 return 0;
378 }
379
check_vma_flags(struct vm_area_struct * vma,unsigned long gup_flags)380 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
381 {
382 vm_flags_t vm_flags = vma->vm_flags;
383
384 if (vm_flags & (VM_IO | VM_PFNMAP))
385 return -EFAULT;
386
387 if (gup_flags & FOLL_ANON && !vma_is_anonymous(vma))
388 return -EFAULT;
389
390 if (gup_flags & FOLL_WRITE) {
391 if (!(vm_flags & VM_WRITE)) {
392 if (!(gup_flags & FOLL_FORCE))
393 return -EFAULT;
394 /*
395 * We used to let the write,force case do COW in a
396 * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
397 * set a breakpoint in a read-only mapping of an
398 * executable, without corrupting the file (yet only
399 * when that file had been opened for writing!).
400 * Anon pages in shared mappings are surprising: now
401 * just reject it.
402 */
403 if (!is_cow_mapping(vm_flags)) {
404 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
405 return -EFAULT;
406 }
407 }
408 } else if (!(vm_flags & VM_READ)) {
409 if (!(gup_flags & FOLL_FORCE))
410 return -EFAULT;
411 /*
412 * Is there actually any vma we can reach here which does not
413 * have VM_MAYREAD set?
414 */
415 if (!(vm_flags & VM_MAYREAD))
416 return -EFAULT;
417 }
418 return 0;
419 }
420
421 /**
422 * __get_user_pages() - pin user pages in memory
423 * @tsk: task_struct of target task
424 * @mm: mm_struct of target mm
425 * @start: starting user address
426 * @nr_pages: number of pages from start to pin
427 * @gup_flags: flags modifying pin behaviour
428 * @pages: array that receives pointers to the pages pinned.
429 * Should be at least nr_pages long. Or NULL, if caller
430 * only intends to ensure the pages are faulted in.
431 * @vmas: array of pointers to vmas corresponding to each page.
432 * Or NULL if the caller does not require them.
433 * @nonblocking: whether waiting for disk IO or mmap_sem contention
434 *
435 * Returns number of pages pinned. This may be fewer than the number
436 * requested. If nr_pages is 0 or negative, returns 0. If no pages
437 * were pinned, returns -errno. Each page returned must be released
438 * with a put_page() call when it is finished with. vmas will only
439 * remain valid while mmap_sem is held.
440 *
441 * Must be called with mmap_sem held. It may be released. See below.
442 *
443 * __get_user_pages walks a process's page tables and takes a reference to
444 * each struct page that each user address corresponds to at a given
445 * instant. That is, it takes the page that would be accessed if a user
446 * thread accesses the given user virtual address at that instant.
447 *
448 * This does not guarantee that the page exists in the user mappings when
449 * __get_user_pages returns, and there may even be a completely different
450 * page there in some cases (eg. if mmapped pagecache has been invalidated
451 * and subsequently re faulted). However it does guarantee that the page
452 * won't be freed completely. And mostly callers simply care that the page
453 * contains data that was valid *at some point in time*. Typically, an IO
454 * or similar operation cannot guarantee anything stronger anyway because
455 * locks can't be held over the syscall boundary.
456 *
457 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
458 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
459 * appropriate) must be called after the page is finished with, and
460 * before put_page is called.
461 *
462 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
463 * or mmap_sem contention, and if waiting is needed to pin all pages,
464 * *@nonblocking will be set to 0. Further, if @gup_flags does not
465 * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
466 * this case.
467 *
468 * A caller using such a combination of @nonblocking and @gup_flags
469 * must therefore hold the mmap_sem for reading only, and recognize
470 * when it's been released. Otherwise, it must be held for either
471 * reading or writing and will not be released.
472 *
473 * In most cases, get_user_pages or get_user_pages_fast should be used
474 * instead of __get_user_pages. __get_user_pages should be used only if
475 * you need some special @gup_flags.
476 */
__get_user_pages(struct task_struct * tsk,struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas,int * nonblocking)477 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
478 unsigned long start, unsigned long nr_pages,
479 unsigned int gup_flags, struct page **pages,
480 struct vm_area_struct **vmas, int *nonblocking)
481 {
482 long i = 0;
483 unsigned int page_mask;
484 struct vm_area_struct *vma = NULL;
485
486 if (!nr_pages)
487 return 0;
488
489 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
490
491 /*
492 * If FOLL_FORCE is set then do not force a full fault as the hinting
493 * fault information is unrelated to the reference behaviour of a task
494 * using the address space
495 */
496 if (!(gup_flags & FOLL_FORCE))
497 gup_flags |= FOLL_NUMA;
498
499 do {
500 struct page *page;
501 unsigned int foll_flags = gup_flags;
502 unsigned int page_increm;
503
504 /* first iteration or cross vma bound */
505 if (!vma || start >= vma->vm_end) {
506 vma = find_extend_vma(mm, start);
507 if (!vma && in_gate_area(mm, start)) {
508 int ret;
509 ret = get_gate_page(mm, start & PAGE_MASK,
510 gup_flags, &vma,
511 pages ? &pages[i] : NULL);
512 if (ret)
513 return i ? : ret;
514 page_mask = 0;
515 goto next_page;
516 }
517
518 if (!vma || check_vma_flags(vma, gup_flags))
519 return i ? : -EFAULT;
520 if (is_vm_hugetlb_page(vma)) {
521 if (should_force_cow_break(vma, foll_flags))
522 foll_flags |= FOLL_WRITE;
523 i = follow_hugetlb_page(mm, vma, pages, vmas,
524 &start, &nr_pages, i,
525 foll_flags);
526 continue;
527 }
528 }
529
530 if (should_force_cow_break(vma, foll_flags))
531 foll_flags |= FOLL_WRITE;
532
533 retry:
534 /*
535 * If we have a pending SIGKILL, don't keep faulting pages and
536 * potentially allocating memory.
537 */
538 if (unlikely(fatal_signal_pending(current)))
539 return i ? i : -ERESTARTSYS;
540 cond_resched();
541 page = follow_page_mask(vma, start, foll_flags, &page_mask);
542 if (!page) {
543 int ret;
544 ret = faultin_page(tsk, vma, start, &foll_flags,
545 nonblocking);
546 switch (ret) {
547 case 0:
548 goto retry;
549 case -EFAULT:
550 case -ENOMEM:
551 case -EHWPOISON:
552 return i ? i : ret;
553 case -EBUSY:
554 return i;
555 case -ENOENT:
556 goto next_page;
557 }
558 BUG();
559 } else if (PTR_ERR(page) == -EEXIST) {
560 /*
561 * Proper page table entry exists, but no corresponding
562 * struct page.
563 */
564 goto next_page;
565 } else if (IS_ERR(page)) {
566 return i ? i : PTR_ERR(page);
567 }
568 if (pages) {
569 pages[i] = page;
570 flush_anon_page(vma, page, start);
571 flush_dcache_page(page);
572 page_mask = 0;
573 }
574 next_page:
575 if (vmas) {
576 vmas[i] = vma;
577 page_mask = 0;
578 }
579 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
580 if (page_increm > nr_pages)
581 page_increm = nr_pages;
582 i += page_increm;
583 start += page_increm * PAGE_SIZE;
584 nr_pages -= page_increm;
585 } while (nr_pages);
586 return i;
587 }
588 EXPORT_SYMBOL(__get_user_pages);
589
590 /*
591 * fixup_user_fault() - manually resolve a user page fault
592 * @tsk: the task_struct to use for page fault accounting, or
593 * NULL if faults are not to be recorded.
594 * @mm: mm_struct of target mm
595 * @address: user address
596 * @fault_flags:flags to pass down to handle_mm_fault()
597 *
598 * This is meant to be called in the specific scenario where for locking reasons
599 * we try to access user memory in atomic context (within a pagefault_disable()
600 * section), this returns -EFAULT, and we want to resolve the user fault before
601 * trying again.
602 *
603 * Typically this is meant to be used by the futex code.
604 *
605 * The main difference with get_user_pages() is that this function will
606 * unconditionally call handle_mm_fault() which will in turn perform all the
607 * necessary SW fixup of the dirty and young bits in the PTE, while
608 * handle_mm_fault() only guarantees to update these in the struct page.
609 *
610 * This is important for some architectures where those bits also gate the
611 * access permission to the page because they are maintained in software. On
612 * such architectures, gup() will not be enough to make a subsequent access
613 * succeed.
614 *
615 * This has the same semantics wrt the @mm->mmap_sem as does filemap_fault().
616 */
fixup_user_fault(struct task_struct * tsk,struct mm_struct * mm,unsigned long address,unsigned int fault_flags)617 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
618 unsigned long address, unsigned int fault_flags)
619 {
620 struct vm_area_struct *vma;
621 vm_flags_t vm_flags;
622 int ret;
623
624 vma = find_extend_vma(mm, address);
625 if (!vma || address < vma->vm_start)
626 return -EFAULT;
627
628 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
629 if (!(vm_flags & vma->vm_flags))
630 return -EFAULT;
631
632 ret = handle_mm_fault(mm, vma, address, fault_flags);
633 if (ret & VM_FAULT_ERROR) {
634 if (ret & VM_FAULT_OOM)
635 return -ENOMEM;
636 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
637 return -EHWPOISON;
638 if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
639 return -EFAULT;
640 BUG();
641 }
642 if (tsk) {
643 if (ret & VM_FAULT_MAJOR)
644 tsk->maj_flt++;
645 else
646 tsk->min_flt++;
647 }
648 return 0;
649 }
650
__get_user_pages_locked(struct task_struct * tsk,struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,struct vm_area_struct ** vmas,int * locked,bool notify_drop,unsigned int flags)651 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
652 struct mm_struct *mm,
653 unsigned long start,
654 unsigned long nr_pages,
655 struct page **pages,
656 struct vm_area_struct **vmas,
657 int *locked, bool notify_drop,
658 unsigned int flags)
659 {
660 long ret, pages_done;
661 bool lock_dropped;
662
663 if (locked) {
664 /* if VM_FAULT_RETRY can be returned, vmas become invalid */
665 BUG_ON(vmas);
666 /* check caller initialized locked */
667 BUG_ON(*locked != 1);
668 }
669
670 if (pages)
671 flags |= FOLL_GET;
672
673 pages_done = 0;
674 lock_dropped = false;
675 for (;;) {
676 ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
677 vmas, locked);
678 if (!locked)
679 /* VM_FAULT_RETRY couldn't trigger, bypass */
680 return ret;
681
682 /* VM_FAULT_RETRY cannot return errors */
683 if (!*locked) {
684 BUG_ON(ret < 0);
685 BUG_ON(ret >= nr_pages);
686 }
687
688 if (!pages)
689 /* If it's a prefault don't insist harder */
690 return ret;
691
692 if (ret > 0) {
693 nr_pages -= ret;
694 pages_done += ret;
695 if (!nr_pages)
696 break;
697 }
698 if (*locked) {
699 /* VM_FAULT_RETRY didn't trigger */
700 if (!pages_done)
701 pages_done = ret;
702 break;
703 }
704 /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
705 pages += ret;
706 start += ret << PAGE_SHIFT;
707
708 /*
709 * Repeat on the address that fired VM_FAULT_RETRY
710 * without FAULT_FLAG_ALLOW_RETRY but with
711 * FAULT_FLAG_TRIED.
712 */
713 *locked = 1;
714 lock_dropped = true;
715 down_read(&mm->mmap_sem);
716 ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
717 pages, NULL, NULL);
718 if (ret != 1) {
719 BUG_ON(ret > 1);
720 if (!pages_done)
721 pages_done = ret;
722 break;
723 }
724 nr_pages--;
725 pages_done++;
726 if (!nr_pages)
727 break;
728 pages++;
729 start += PAGE_SIZE;
730 }
731 if (notify_drop && lock_dropped && *locked) {
732 /*
733 * We must let the caller know we temporarily dropped the lock
734 * and so the critical section protected by it was lost.
735 */
736 up_read(&mm->mmap_sem);
737 *locked = 0;
738 }
739 return pages_done;
740 }
741
742 /*
743 * We can leverage the VM_FAULT_RETRY functionality in the page fault
744 * paths better by using either get_user_pages_locked() or
745 * get_user_pages_unlocked().
746 *
747 * get_user_pages_locked() is suitable to replace the form:
748 *
749 * down_read(&mm->mmap_sem);
750 * do_something()
751 * get_user_pages(tsk, mm, ..., pages, NULL);
752 * up_read(&mm->mmap_sem);
753 *
754 * to:
755 *
756 * int locked = 1;
757 * down_read(&mm->mmap_sem);
758 * do_something()
759 * get_user_pages_locked(tsk, mm, ..., pages, &locked);
760 * if (locked)
761 * up_read(&mm->mmap_sem);
762 */
get_user_pages_locked(struct task_struct * tsk,struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,int * locked)763 long get_user_pages_locked(struct task_struct *tsk, struct mm_struct *mm,
764 unsigned long start, unsigned long nr_pages,
765 unsigned int gup_flags, struct page **pages,
766 int *locked)
767 {
768 return __get_user_pages_locked(tsk, mm, start, nr_pages,
769 pages, NULL, locked, true,
770 gup_flags | FOLL_TOUCH);
771 }
772 EXPORT_SYMBOL(get_user_pages_locked);
773
774 /*
775 * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows to
776 * pass additional gup_flags as last parameter (like FOLL_HWPOISON).
777 *
778 * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
779 * caller if required (just like with __get_user_pages). "FOLL_GET",
780 * "FOLL_WRITE" and "FOLL_FORCE" are set implicitly as needed
781 * according to the parameters "pages", "write", "force"
782 * respectively.
783 */
__get_user_pages_unlocked(struct task_struct * tsk,struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)784 __always_inline long __get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
785 unsigned long start, unsigned long nr_pages,
786 struct page **pages, unsigned int gup_flags)
787 {
788 long ret;
789 int locked = 1;
790
791 down_read(&mm->mmap_sem);
792 ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
793 &locked, false, gup_flags);
794 if (locked)
795 up_read(&mm->mmap_sem);
796 return ret;
797 }
798 EXPORT_SYMBOL(__get_user_pages_unlocked);
799
800 /*
801 * get_user_pages_unlocked() is suitable to replace the form:
802 *
803 * down_read(&mm->mmap_sem);
804 * get_user_pages(tsk, mm, ..., pages, NULL);
805 * up_read(&mm->mmap_sem);
806 *
807 * with:
808 *
809 * get_user_pages_unlocked(tsk, mm, ..., pages);
810 *
811 * It is functionally equivalent to get_user_pages_fast so
812 * get_user_pages_fast should be used instead, if the two parameters
813 * "tsk" and "mm" are respectively equal to current and current->mm,
814 * or if "force" shall be set to 1 (get_user_pages_fast misses the
815 * "force" parameter).
816 */
get_user_pages_unlocked(struct task_struct * tsk,struct mm_struct * mm,unsigned long start,unsigned long nr_pages,struct page ** pages,unsigned int gup_flags)817 long get_user_pages_unlocked(struct task_struct *tsk, struct mm_struct *mm,
818 unsigned long start, unsigned long nr_pages,
819 struct page **pages, unsigned int gup_flags)
820 {
821 return __get_user_pages_unlocked(tsk, mm, start, nr_pages,
822 pages, gup_flags | FOLL_TOUCH);
823 }
824 EXPORT_SYMBOL(get_user_pages_unlocked);
825
826 /*
827 * get_user_pages() - pin user pages in memory
828 * @tsk: the task_struct to use for page fault accounting, or
829 * NULL if faults are not to be recorded.
830 * @mm: mm_struct of target mm
831 * @start: starting user address
832 * @nr_pages: number of pages from start to pin
833 * @write: whether pages will be written to by the caller
834 * @force: whether to force access even when user mapping is currently
835 * protected (but never forces write access to shared mapping).
836 * @pages: array that receives pointers to the pages pinned.
837 * Should be at least nr_pages long. Or NULL, if caller
838 * only intends to ensure the pages are faulted in.
839 * @vmas: array of pointers to vmas corresponding to each page.
840 * Or NULL if the caller does not require them.
841 *
842 * Returns number of pages pinned. This may be fewer than the number
843 * requested. If nr_pages is 0 or negative, returns 0. If no pages
844 * were pinned, returns -errno. Each page returned must be released
845 * with a put_page() call when it is finished with. vmas will only
846 * remain valid while mmap_sem is held.
847 *
848 * Must be called with mmap_sem held for read or write.
849 *
850 * get_user_pages walks a process's page tables and takes a reference to
851 * each struct page that each user address corresponds to at a given
852 * instant. That is, it takes the page that would be accessed if a user
853 * thread accesses the given user virtual address at that instant.
854 *
855 * This does not guarantee that the page exists in the user mappings when
856 * get_user_pages returns, and there may even be a completely different
857 * page there in some cases (eg. if mmapped pagecache has been invalidated
858 * and subsequently re faulted). However it does guarantee that the page
859 * won't be freed completely. And mostly callers simply care that the page
860 * contains data that was valid *at some point in time*. Typically, an IO
861 * or similar operation cannot guarantee anything stronger anyway because
862 * locks can't be held over the syscall boundary.
863 *
864 * If write=0, the page must not be written to. If the page is written to,
865 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
866 * after the page is finished with, and before put_page is called.
867 *
868 * get_user_pages is typically used for fewer-copy IO operations, to get a
869 * handle on the memory by some means other than accesses via the user virtual
870 * addresses. The pages may be submitted for DMA to devices or accessed via
871 * their kernel linear mapping (via the kmap APIs). Care should be taken to
872 * use the correct cache flushing APIs.
873 *
874 * See also get_user_pages_fast, for performance critical applications.
875 *
876 * get_user_pages should be phased out in favor of
877 * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
878 * should use get_user_pages because it cannot pass
879 * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
880 */
get_user_pages(struct task_struct * tsk,struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas)881 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
882 unsigned long start, unsigned long nr_pages,
883 unsigned int gup_flags, struct page **pages,
884 struct vm_area_struct **vmas)
885 {
886 return __get_user_pages_locked(tsk, mm, start, nr_pages,
887 pages, vmas, NULL, false,
888 gup_flags | FOLL_TOUCH);
889 }
890 EXPORT_SYMBOL(get_user_pages);
891
892 /**
893 * populate_vma_page_range() - populate a range of pages in the vma.
894 * @vma: target vma
895 * @start: start address
896 * @end: end address
897 * @nonblocking:
898 *
899 * This takes care of mlocking the pages too if VM_LOCKED is set.
900 *
901 * return 0 on success, negative error code on error.
902 *
903 * vma->vm_mm->mmap_sem must be held.
904 *
905 * If @nonblocking is NULL, it may be held for read or write and will
906 * be unperturbed.
907 *
908 * If @nonblocking is non-NULL, it must held for read only and may be
909 * released. If it's released, *@nonblocking will be set to 0.
910 */
populate_vma_page_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,int * nonblocking)911 long populate_vma_page_range(struct vm_area_struct *vma,
912 unsigned long start, unsigned long end, int *nonblocking)
913 {
914 struct mm_struct *mm = vma->vm_mm;
915 unsigned long nr_pages = (end - start) / PAGE_SIZE;
916 int gup_flags;
917
918 VM_BUG_ON(start & ~PAGE_MASK);
919 VM_BUG_ON(end & ~PAGE_MASK);
920 VM_BUG_ON_VMA(start < vma->vm_start, vma);
921 VM_BUG_ON_VMA(end > vma->vm_end, vma);
922 VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
923
924 gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
925 if (vma->vm_flags & VM_LOCKONFAULT)
926 gup_flags &= ~FOLL_POPULATE;
927
928 /*
929 * We want to touch writable mappings with a write fault in order
930 * to break COW, except for shared mappings because these don't COW
931 * and we would not want to dirty them for nothing.
932 */
933 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
934 gup_flags |= FOLL_WRITE;
935
936 /*
937 * We want mlock to succeed for regions that have any permissions
938 * other than PROT_NONE.
939 */
940 if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
941 gup_flags |= FOLL_FORCE;
942
943 /*
944 * We made sure addr is within a VMA, so the following will
945 * not result in a stack expansion that recurses back here.
946 */
947 return __get_user_pages(current, mm, start, nr_pages, gup_flags,
948 NULL, NULL, nonblocking);
949 }
950
951 /*
952 * __mm_populate - populate and/or mlock pages within a range of address space.
953 *
954 * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
955 * flags. VMAs must be already marked with the desired vm_flags, and
956 * mmap_sem must not be held.
957 */
__mm_populate(unsigned long start,unsigned long len,int ignore_errors)958 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
959 {
960 struct mm_struct *mm = current->mm;
961 unsigned long end, nstart, nend;
962 struct vm_area_struct *vma = NULL;
963 int locked = 0;
964 long ret = 0;
965
966 end = start + len;
967
968 for (nstart = start; nstart < end; nstart = nend) {
969 /*
970 * We want to fault in pages for [nstart; end) address range.
971 * Find first corresponding VMA.
972 */
973 if (!locked) {
974 locked = 1;
975 down_read(&mm->mmap_sem);
976 vma = find_vma(mm, nstart);
977 } else if (nstart >= vma->vm_end)
978 vma = vma->vm_next;
979 if (!vma || vma->vm_start >= end)
980 break;
981 /*
982 * Set [nstart; nend) to intersection of desired address
983 * range with the first VMA. Also, skip undesirable VMA types.
984 */
985 nend = min(end, vma->vm_end);
986 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
987 continue;
988 if (nstart < vma->vm_start)
989 nstart = vma->vm_start;
990 /*
991 * Now fault in a range of pages. populate_vma_page_range()
992 * double checks the vma flags, so that it won't mlock pages
993 * if the vma was already munlocked.
994 */
995 ret = populate_vma_page_range(vma, nstart, nend, &locked);
996 if (ret < 0) {
997 if (ignore_errors) {
998 ret = 0;
999 continue; /* continue at next VMA */
1000 }
1001 break;
1002 }
1003 nend = nstart + ret * PAGE_SIZE;
1004 ret = 0;
1005 }
1006 if (locked)
1007 up_read(&mm->mmap_sem);
1008 return ret; /* 0 or negative error code */
1009 }
1010
1011 /**
1012 * get_dump_page() - pin user page in memory while writing it to core dump
1013 * @addr: user address
1014 *
1015 * Returns struct page pointer of user page pinned for dump,
1016 * to be freed afterwards by page_cache_release() or put_page().
1017 *
1018 * Returns NULL on any kind of failure - a hole must then be inserted into
1019 * the corefile, to preserve alignment with its headers; and also returns
1020 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1021 * allowing a hole to be left in the corefile to save diskspace.
1022 *
1023 * Called without mmap_sem, but after all other threads have been killed.
1024 */
1025 #ifdef CONFIG_ELF_CORE
get_dump_page(unsigned long addr)1026 struct page *get_dump_page(unsigned long addr)
1027 {
1028 struct vm_area_struct *vma;
1029 struct page *page;
1030
1031 if (__get_user_pages(current, current->mm, addr, 1,
1032 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1033 NULL) < 1)
1034 return NULL;
1035 flush_cache_page(vma, addr, page_to_pfn(page));
1036 return page;
1037 }
1038 #endif /* CONFIG_ELF_CORE */
1039
1040 /*
1041 * Generic RCU Fast GUP
1042 *
1043 * get_user_pages_fast attempts to pin user pages by walking the page
1044 * tables directly and avoids taking locks. Thus the walker needs to be
1045 * protected from page table pages being freed from under it, and should
1046 * block any THP splits.
1047 *
1048 * One way to achieve this is to have the walker disable interrupts, and
1049 * rely on IPIs from the TLB flushing code blocking before the page table
1050 * pages are freed. This is unsuitable for architectures that do not need
1051 * to broadcast an IPI when invalidating TLBs.
1052 *
1053 * Another way to achieve this is to batch up page table containing pages
1054 * belonging to more than one mm_user, then rcu_sched a callback to free those
1055 * pages. Disabling interrupts will allow the fast_gup walker to both block
1056 * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1057 * (which is a relatively rare event). The code below adopts this strategy.
1058 *
1059 * Before activating this code, please be aware that the following assumptions
1060 * are currently made:
1061 *
1062 * *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1063 * pages containing page tables.
1064 *
1065 * *) THP splits will broadcast an IPI, this can be achieved by overriding
1066 * pmdp_splitting_flush.
1067 *
1068 * *) ptes can be read atomically by the architecture.
1069 *
1070 * *) access_ok is sufficient to validate userspace address ranges.
1071 *
1072 * The last two assumptions can be relaxed by the addition of helper functions.
1073 *
1074 * This code is based heavily on the PowerPC implementation by Nick Piggin.
1075 */
1076 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1077
1078 /*
1079 * Return the compund head page with ref appropriately incremented,
1080 * or NULL if that failed.
1081 */
try_get_compound_head(struct page * page,int refs)1082 static inline struct page *try_get_compound_head(struct page *page, int refs)
1083 {
1084 struct page *head = compound_head(page);
1085 if (WARN_ON_ONCE(atomic_read(&head->_count) < 0))
1086 return NULL;
1087 if (unlikely(!page_cache_add_speculative(head, refs)))
1088 return NULL;
1089 return head;
1090 }
1091
1092 #ifdef __HAVE_ARCH_PTE_SPECIAL
gup_pte_range(pmd_t pmd,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)1093 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1094 int write, struct page **pages, int *nr)
1095 {
1096 pte_t *ptep, *ptem;
1097 int ret = 0;
1098
1099 ptem = ptep = pte_offset_map(&pmd, addr);
1100 do {
1101 /*
1102 * In the line below we are assuming that the pte can be read
1103 * atomically. If this is not the case for your architecture,
1104 * please wrap this in a helper function!
1105 *
1106 * for an example see gup_get_pte in arch/x86/mm/gup.c
1107 */
1108 pte_t pte = READ_ONCE(*ptep);
1109 struct page *page;
1110
1111 /*
1112 * Similar to the PMD case below, NUMA hinting must take slow
1113 * path using the pte_protnone check.
1114 */
1115 if (!pte_present(pte) || pte_special(pte) ||
1116 pte_protnone(pte) || (write && !pte_write(pte)))
1117 goto pte_unmap;
1118
1119 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1120 page = pte_page(pte);
1121
1122 if (WARN_ON_ONCE(page_ref_count(page) < 0))
1123 goto pte_unmap;
1124
1125 if (!page_cache_get_speculative(page))
1126 goto pte_unmap;
1127
1128 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1129 put_page(page);
1130 goto pte_unmap;
1131 }
1132
1133 pages[*nr] = page;
1134 (*nr)++;
1135
1136 } while (ptep++, addr += PAGE_SIZE, addr != end);
1137
1138 ret = 1;
1139
1140 pte_unmap:
1141 pte_unmap(ptem);
1142 return ret;
1143 }
1144 #else
1145
1146 /*
1147 * If we can't determine whether or not a pte is special, then fail immediately
1148 * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1149 * to be special.
1150 *
1151 * For a futex to be placed on a THP tail page, get_futex_key requires a
1152 * __get_user_pages_fast implementation that can pin pages. Thus it's still
1153 * useful to have gup_huge_pmd even if we can't operate on ptes.
1154 */
gup_pte_range(pmd_t pmd,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)1155 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1156 int write, struct page **pages, int *nr)
1157 {
1158 return 0;
1159 }
1160 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1161
gup_huge_pmd(pmd_t orig,pmd_t * pmdp,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)1162 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1163 unsigned long end, int write, struct page **pages, int *nr)
1164 {
1165 struct page *head, *page, *tail;
1166 int refs;
1167
1168 if (write && !pmd_write(orig))
1169 return 0;
1170
1171 refs = 0;
1172 page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1173 tail = page;
1174 do {
1175 pages[*nr] = page;
1176 (*nr)++;
1177 page++;
1178 refs++;
1179 } while (addr += PAGE_SIZE, addr != end);
1180
1181 head = try_get_compound_head(pmd_page(orig), refs);
1182 if (!head) {
1183 *nr -= refs;
1184 return 0;
1185 }
1186
1187 if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1188 *nr -= refs;
1189 while (refs--)
1190 put_page(head);
1191 return 0;
1192 }
1193
1194 /*
1195 * Any tail pages need their mapcount reference taken before we
1196 * return. (This allows the THP code to bump their ref count when
1197 * they are split into base pages).
1198 */
1199 while (refs--) {
1200 if (PageTail(tail))
1201 get_huge_page_tail(tail);
1202 tail++;
1203 }
1204
1205 return 1;
1206 }
1207
gup_huge_pud(pud_t orig,pud_t * pudp,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)1208 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1209 unsigned long end, int write, struct page **pages, int *nr)
1210 {
1211 struct page *head, *page, *tail;
1212 int refs;
1213
1214 if (write && !pud_write(orig))
1215 return 0;
1216
1217 refs = 0;
1218 page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1219 tail = page;
1220 do {
1221 pages[*nr] = page;
1222 (*nr)++;
1223 page++;
1224 refs++;
1225 } while (addr += PAGE_SIZE, addr != end);
1226
1227 head = try_get_compound_head(pud_page(orig), refs);
1228 if (!head) {
1229 *nr -= refs;
1230 return 0;
1231 }
1232
1233 if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1234 *nr -= refs;
1235 while (refs--)
1236 put_page(head);
1237 return 0;
1238 }
1239
1240 while (refs--) {
1241 if (PageTail(tail))
1242 get_huge_page_tail(tail);
1243 tail++;
1244 }
1245
1246 return 1;
1247 }
1248
gup_huge_pgd(pgd_t orig,pgd_t * pgdp,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)1249 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1250 unsigned long end, int write,
1251 struct page **pages, int *nr)
1252 {
1253 int refs;
1254 struct page *head, *page, *tail;
1255
1256 if (write && !pgd_write(orig))
1257 return 0;
1258
1259 refs = 0;
1260 page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1261 tail = page;
1262 do {
1263 pages[*nr] = page;
1264 (*nr)++;
1265 page++;
1266 refs++;
1267 } while (addr += PAGE_SIZE, addr != end);
1268
1269 head = try_get_compound_head(pgd_page(orig), refs);
1270 if (!head) {
1271 *nr -= refs;
1272 return 0;
1273 }
1274
1275 if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1276 *nr -= refs;
1277 while (refs--)
1278 put_page(head);
1279 return 0;
1280 }
1281
1282 while (refs--) {
1283 if (PageTail(tail))
1284 get_huge_page_tail(tail);
1285 tail++;
1286 }
1287
1288 return 1;
1289 }
1290
gup_pmd_range(pud_t pud,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)1291 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1292 int write, struct page **pages, int *nr)
1293 {
1294 unsigned long next;
1295 pmd_t *pmdp;
1296
1297 pmdp = pmd_offset(&pud, addr);
1298 do {
1299 pmd_t pmd = READ_ONCE(*pmdp);
1300
1301 next = pmd_addr_end(addr, end);
1302 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1303 return 0;
1304
1305 if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1306 /*
1307 * NUMA hinting faults need to be handled in the GUP
1308 * slowpath for accounting purposes and so that they
1309 * can be serialised against THP migration.
1310 */
1311 if (pmd_protnone(pmd))
1312 return 0;
1313
1314 if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1315 pages, nr))
1316 return 0;
1317
1318 } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1319 /*
1320 * architecture have different format for hugetlbfs
1321 * pmd format and THP pmd format
1322 */
1323 if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1324 PMD_SHIFT, next, write, pages, nr))
1325 return 0;
1326 } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1327 return 0;
1328 } while (pmdp++, addr = next, addr != end);
1329
1330 return 1;
1331 }
1332
gup_pud_range(pgd_t pgd,unsigned long addr,unsigned long end,int write,struct page ** pages,int * nr)1333 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1334 int write, struct page **pages, int *nr)
1335 {
1336 unsigned long next;
1337 pud_t *pudp;
1338
1339 pudp = pud_offset(&pgd, addr);
1340 do {
1341 pud_t pud = READ_ONCE(*pudp);
1342
1343 next = pud_addr_end(addr, end);
1344 if (pud_none(pud))
1345 return 0;
1346 if (unlikely(pud_huge(pud))) {
1347 if (!gup_huge_pud(pud, pudp, addr, next, write,
1348 pages, nr))
1349 return 0;
1350 } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1351 if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1352 PUD_SHIFT, next, write, pages, nr))
1353 return 0;
1354 } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1355 return 0;
1356 } while (pudp++, addr = next, addr != end);
1357
1358 return 1;
1359 }
1360
1361 /*
1362 * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1363 * the regular GUP. It will only return non-negative values.
1364 *
1365 * Careful, careful! COW breaking can go either way, so a non-write
1366 * access can get ambiguous page results. If you call this function without
1367 * 'write' set, you'd better be sure that you're ok with that ambiguity.
1368 */
__get_user_pages_fast(unsigned long start,int nr_pages,int write,struct page ** pages)1369 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1370 struct page **pages)
1371 {
1372 struct mm_struct *mm = current->mm;
1373 unsigned long addr, len, end;
1374 unsigned long next, flags;
1375 pgd_t *pgdp;
1376 int nr = 0;
1377
1378 start &= PAGE_MASK;
1379 addr = start;
1380 len = (unsigned long) nr_pages << PAGE_SHIFT;
1381 end = start + len;
1382
1383 if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1384 start, len)))
1385 return 0;
1386
1387 /*
1388 * Disable interrupts. We use the nested form as we can already have
1389 * interrupts disabled by get_futex_key.
1390 *
1391 * With interrupts disabled, we block page table pages from being
1392 * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1393 * for more details.
1394 *
1395 * We do not adopt an rcu_read_lock(.) here as we also want to
1396 * block IPIs that come from THPs splitting.
1397 *
1398 * NOTE! We allow read-only gup_fast() here, but you'd better be
1399 * careful about possible COW pages. You'll get _a_ COW page, but
1400 * not necessarily the one you intended to get depending on what
1401 * COW event happens after this. COW may break the page copy in a
1402 * random direction.
1403 */
1404
1405 local_irq_save(flags);
1406 pgdp = pgd_offset(mm, addr);
1407 do {
1408 pgd_t pgd = READ_ONCE(*pgdp);
1409
1410 next = pgd_addr_end(addr, end);
1411 if (pgd_none(pgd))
1412 break;
1413 /*
1414 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
1415 * because get_user_pages() may need to cause an early COW in
1416 * order to avoid confusing the normal COW routines. So only
1417 * targets that are already writable are safe to do by just
1418 * looking at the page tables.
1419 */
1420 if (unlikely(pgd_huge(pgd))) {
1421 if (!gup_huge_pgd(pgd, pgdp, addr, next, 1,
1422 pages, &nr))
1423 break;
1424 } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1425 if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1426 PGDIR_SHIFT, next, 1, pages, &nr))
1427 break;
1428 } else if (!gup_pud_range(pgd, addr, next, 1, pages, &nr))
1429 break;
1430 } while (pgdp++, addr = next, addr != end);
1431 local_irq_restore(flags);
1432
1433 return nr;
1434 }
1435
1436 /**
1437 * get_user_pages_fast() - pin user pages in memory
1438 * @start: starting user address
1439 * @nr_pages: number of pages from start to pin
1440 * @write: whether pages will be written to
1441 * @pages: array that receives pointers to the pages pinned.
1442 * Should be at least nr_pages long.
1443 *
1444 * Attempt to pin user pages in memory without taking mm->mmap_sem.
1445 * If not successful, it will fall back to taking the lock and
1446 * calling get_user_pages().
1447 *
1448 * Returns number of pages pinned. This may be fewer than the number
1449 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1450 * were pinned, returns -errno.
1451 */
get_user_pages_fast(unsigned long start,int nr_pages,int write,struct page ** pages)1452 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1453 struct page **pages)
1454 {
1455 struct mm_struct *mm = current->mm;
1456 int nr, ret;
1457
1458 start &= PAGE_MASK;
1459 nr = __get_user_pages_fast(start, nr_pages, write, pages);
1460 ret = nr;
1461
1462 if (nr < nr_pages) {
1463 /* Try to get the remaining pages with get_user_pages */
1464 start += nr << PAGE_SHIFT;
1465 pages += nr;
1466
1467 ret = get_user_pages_unlocked(current, mm, start,
1468 nr_pages - nr, pages,
1469 write ? FOLL_WRITE : 0);
1470
1471 /* Have to be a bit careful with return values */
1472 if (nr > 0) {
1473 if (ret < 0)
1474 ret = nr;
1475 else
1476 ret += nr;
1477 }
1478 }
1479
1480 return ret;
1481 }
1482
1483 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */
1484