1/*
2 * mm/userfaultfd.c
3 *
4 * Copyright (C) 2015 Red Hat, Inc.
5 *
6 * This work is licensed under the terms of the GNU GPL, version 2. See
7 * the COPYING file in the top-level directory.
8 */
9
10#include <linux/mm.h>
11#include <linux/sched/signal.h>
12#include <linux/pagemap.h>
13#include <linux/rmap.h>
14#include <linux/swap.h>
15#include <linux/swapops.h>
16#include <linux/userfaultfd_k.h>
17#include <linux/mmu_notifier.h>
18#include <linux/hugetlb.h>
19#include <linux/shmem_fs.h>
20#include <asm/tlbflush.h>
21#include "internal.h"
22
23static int mcopy_atomic_pte(struct mm_struct *dst_mm,
24 pmd_t *dst_pmd,
25 struct vm_area_struct *dst_vma,
26 unsigned long dst_addr,
27 unsigned long src_addr,
28 struct page **pagep)
29{
30 struct mem_cgroup *memcg;
31 pte_t _dst_pte, *dst_pte;
32 spinlock_t *ptl;
33 void *page_kaddr;
34 int ret;
35 struct page *page;
36 pgoff_t offset, max_off;
37 struct inode *inode;
38
39 if (!*pagep) {
40 ret = -ENOMEM;
41 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
42 if (!page)
43 goto out;
44
45 page_kaddr = kmap_atomic(page);
46 ret = copy_from_user(page_kaddr,
47 (const void __user *) src_addr,
48 PAGE_SIZE);
49 kunmap_atomic(page_kaddr);
50
51 /* fallback to copy_from_user outside mmap_sem */
52 if (unlikely(ret)) {
53 ret = -ENOENT;
54 *pagep = page;
55 /* don't free the page */
56 goto out;
57 }
58 } else {
59 page = *pagep;
60 *pagep = NULL;
61 }
62
63 /*
64 * The memory barrier inside __SetPageUptodate makes sure that
65 * preceeding stores to the page contents become visible before
66 * the set_pte_at() write.
67 */
68 __SetPageUptodate(page);
69
70 ret = -ENOMEM;
71 if (mem_cgroup_try_charge(page, dst_mm, GFP_KERNEL, &memcg, false))
72 goto out_release;
73
74 _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
75 if (dst_vma->vm_flags & VM_WRITE)
76 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
77
78 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
79 if (dst_vma->vm_file) {
80 /* the shmem MAP_PRIVATE case requires checking the i_size */
81 inode = dst_vma->vm_file->f_inode;
82 offset = linear_page_index(dst_vma, dst_addr);
83 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
84 ret = -EFAULT;
85 if (unlikely(offset >= max_off))
86 goto out_release_uncharge_unlock;
87 }
88 ret = -EEXIST;
89 if (!pte_none(*dst_pte))
90 goto out_release_uncharge_unlock;
91
92 inc_mm_counter(dst_mm, MM_ANONPAGES);
93 page_add_new_anon_rmap(page, dst_vma, dst_addr, false);
94 mem_cgroup_commit_charge(page, memcg, false, false);
95 lru_cache_add_active_or_unevictable(page, dst_vma);
96
97 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
98
99 /* No need to invalidate - it was non-present before */
100 update_mmu_cache(dst_vma, dst_addr, dst_pte);
101
102 pte_unmap_unlock(dst_pte, ptl);
103 ret = 0;
104out:
105 return ret;
106out_release_uncharge_unlock:
107 pte_unmap_unlock(dst_pte, ptl);
108 mem_cgroup_cancel_charge(page, memcg, false);
109out_release:
110 put_page(page);
111 goto out;
112}
113
114static int mfill_zeropage_pte(struct mm_struct *dst_mm,
115 pmd_t *dst_pmd,
116 struct vm_area_struct *dst_vma,
117 unsigned long dst_addr)
118{
119 pte_t _dst_pte, *dst_pte;
120 spinlock_t *ptl;
121 int ret;
122 pgoff_t offset, max_off;
123 struct inode *inode;
124
125 _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
126 dst_vma->vm_page_prot));
127 dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
128 if (dst_vma->vm_file) {
129 /* the shmem MAP_PRIVATE case requires checking the i_size */
130 inode = dst_vma->vm_file->f_inode;
131 offset = linear_page_index(dst_vma, dst_addr);
132 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
133 ret = -EFAULT;
134 if (unlikely(offset >= max_off))
135 goto out_unlock;
136 }
137 ret = -EEXIST;
138 if (!pte_none(*dst_pte))
139 goto out_unlock;
140 set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
141 /* No need to invalidate - it was non-present before */
142 update_mmu_cache(dst_vma, dst_addr, dst_pte);
143 ret = 0;
144out_unlock:
145 pte_unmap_unlock(dst_pte, ptl);
146 return ret;
147}
148
149static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
150{
151 pgd_t *pgd;
152 p4d_t *p4d;
153 pud_t *pud;
154
155 pgd = pgd_offset(mm, address);
156 p4d = p4d_alloc(mm, pgd, address);
157 if (!p4d)
158 return NULL;
159 pud = pud_alloc(mm, p4d, address);
160 if (!pud)
161 return NULL;
162 /*
163 * Note that we didn't run this because the pmd was
164 * missing, the *pmd may be already established and in
165 * turn it may also be a trans_huge_pmd.
166 */
167 return pmd_alloc(mm, pud, address);
168}
169
170#ifdef CONFIG_HUGETLB_PAGE
171/*
172 * __mcopy_atomic processing for HUGETLB vmas. Note that this routine is
173 * called with mmap_sem held, it will release mmap_sem before returning.
174 */
175static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
176 struct vm_area_struct *dst_vma,
177 unsigned long dst_start,
178 unsigned long src_start,
179 unsigned long len,
180 bool zeropage)
181{
182 int vm_alloc_shared = dst_vma->vm_flags & VM_SHARED;
183 int vm_shared = dst_vma->vm_flags & VM_SHARED;
184 ssize_t err;
185 pte_t *dst_pte;
186 unsigned long src_addr, dst_addr;
187 long copied;
188 struct page *page;
189 struct hstate *h;
190 unsigned long vma_hpagesize;
191 pgoff_t idx;
192 u32 hash;
193 struct address_space *mapping;
194
195 /*
196 * There is no default zero huge page for all huge page sizes as
197 * supported by hugetlb. A PMD_SIZE huge pages may exist as used
198 * by THP. Since we can not reliably insert a zero page, this
199 * feature is not supported.
200 */
201 if (zeropage) {
202 up_read(&dst_mm->mmap_sem);
203 return -EINVAL;
204 }
205
206 src_addr = src_start;
207 dst_addr = dst_start;
208 copied = 0;
209 page = NULL;
210 vma_hpagesize = vma_kernel_pagesize(dst_vma);
211
212 /*
213 * Validate alignment based on huge page size
214 */
215 err = -EINVAL;
216 if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
217 goto out_unlock;
218
219retry:
220 /*
221 * On routine entry dst_vma is set. If we had to drop mmap_sem and
222 * retry, dst_vma will be set to NULL and we must lookup again.
223 */
224 if (!dst_vma) {
225 err = -ENOENT;
226 dst_vma = find_vma(dst_mm, dst_start);
227 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
228 goto out_unlock;
229 /*
230 * Check the vma is registered in uffd, this is
231 * required to enforce the VM_MAYWRITE check done at
232 * uffd registration time.
233 */
234 if (!dst_vma->vm_userfaultfd_ctx.ctx)
235 goto out_unlock;
236
237 if (dst_start < dst_vma->vm_start ||
238 dst_start + len > dst_vma->vm_end)
239 goto out_unlock;
240
241 err = -EINVAL;
242 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
243 goto out_unlock;
244
245 vm_shared = dst_vma->vm_flags & VM_SHARED;
246 }
247
248 if (WARN_ON(dst_addr & (vma_hpagesize - 1) ||
249 (len - copied) & (vma_hpagesize - 1)))
250 goto out_unlock;
251
252 /*
253 * If not shared, ensure the dst_vma has a anon_vma.
254 */
255 err = -ENOMEM;
256 if (!vm_shared) {
257 if (unlikely(anon_vma_prepare(dst_vma)))
258 goto out_unlock;
259 }
260
261 h = hstate_vma(dst_vma);
262
263 while (src_addr < src_start + len) {
264 pte_t dst_pteval;
265
266 BUG_ON(dst_addr >= dst_start + len);
267 VM_BUG_ON(dst_addr & ~huge_page_mask(h));
268
269 /*
270 * Serialize via hugetlb_fault_mutex
271 */
272 idx = linear_page_index(dst_vma, dst_addr);
273 mapping = dst_vma->vm_file->f_mapping;
274 hash = hugetlb_fault_mutex_hash(h, dst_mm, dst_vma, mapping,
275 idx, dst_addr);
276 mutex_lock(&hugetlb_fault_mutex_table[hash]);
277
278 err = -ENOMEM;
279 dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
280 if (!dst_pte) {
281 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
282 goto out_unlock;
283 }
284
285 err = -EEXIST;
286 dst_pteval = huge_ptep_get(dst_pte);
287 if (!huge_pte_none(dst_pteval)) {
288 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
289 goto out_unlock;
290 }
291
292 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
293 dst_addr, src_addr, &page);
294
295 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
296 vm_alloc_shared = vm_shared;
297
298 cond_resched();
299
300 if (unlikely(err == -ENOENT)) {
301 up_read(&dst_mm->mmap_sem);
302 BUG_ON(!page);
303
304 err = copy_huge_page_from_user(page,
305 (const void __user *)src_addr,
306 pages_per_huge_page(h), true);
307 if (unlikely(err)) {
308 err = -EFAULT;
309 goto out;
310 }
311 down_read(&dst_mm->mmap_sem);
312
313 dst_vma = NULL;
314 goto retry;
315 } else
316 BUG_ON(page);
317
318 if (!err) {
319 dst_addr += vma_hpagesize;
320 src_addr += vma_hpagesize;
321 copied += vma_hpagesize;
322
323 if (fatal_signal_pending(current))
324 err = -EINTR;
325 }
326 if (err)
327 break;
328 }
329
330out_unlock:
331 up_read(&dst_mm->mmap_sem);
332out:
333 if (page) {
334 /*
335 * We encountered an error and are about to free a newly
336 * allocated huge page.
337 *
338 * Reservation handling is very subtle, and is different for
339 * private and shared mappings. See the routine
340 * restore_reserve_on_error for details. Unfortunately, we
341 * can not call restore_reserve_on_error now as it would
342 * require holding mmap_sem.
343 *
344 * If a reservation for the page existed in the reservation
345 * map of a private mapping, the map was modified to indicate
346 * the reservation was consumed when the page was allocated.
347 * We clear the PagePrivate flag now so that the global
348 * reserve count will not be incremented in free_huge_page.
349 * The reservation map will still indicate the reservation
350 * was consumed and possibly prevent later page allocation.
351 * This is better than leaking a global reservation. If no
352 * reservation existed, it is still safe to clear PagePrivate
353 * as no adjustments to reservation counts were made during
354 * allocation.
355 *
356 * The reservation map for shared mappings indicates which
357 * pages have reservations. When a huge page is allocated
358 * for an address with a reservation, no change is made to
359 * the reserve map. In this case PagePrivate will be set
360 * to indicate that the global reservation count should be
361 * incremented when the page is freed. This is the desired
362 * behavior. However, when a huge page is allocated for an
363 * address without a reservation a reservation entry is added
364 * to the reservation map, and PagePrivate will not be set.
365 * When the page is freed, the global reserve count will NOT
366 * be incremented and it will appear as though we have leaked
367 * reserved page. In this case, set PagePrivate so that the
368 * global reserve count will be incremented to match the
369 * reservation map entry which was created.
370 *
371 * Note that vm_alloc_shared is based on the flags of the vma
372 * for which the page was originally allocated. dst_vma could
373 * be different or NULL on error.
374 */
375 if (vm_alloc_shared)
376 SetPagePrivate(page);
377 else
378 ClearPagePrivate(page);
379 put_page(page);
380 }
381 BUG_ON(copied < 0);
382 BUG_ON(err > 0);
383 BUG_ON(!copied && !err);
384 return copied ? copied : err;
385}
386#else /* !CONFIG_HUGETLB_PAGE */
387/* fail at build time if gcc attempts to use this */
388extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
389 struct vm_area_struct *dst_vma,
390 unsigned long dst_start,
391 unsigned long src_start,
392 unsigned long len,
393 bool zeropage);
394#endif /* CONFIG_HUGETLB_PAGE */
395
396static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
397 pmd_t *dst_pmd,
398 struct vm_area_struct *dst_vma,
399 unsigned long dst_addr,
400 unsigned long src_addr,
401 struct page **page,
402 bool zeropage)
403{
404 ssize_t err;
405
406 /*
407 * The normal page fault path for a shmem will invoke the
408 * fault, fill the hole in the file and COW it right away. The
409 * result generates plain anonymous memory. So when we are
410 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
411 * generate anonymous memory directly without actually filling
412 * the hole. For the MAP_PRIVATE case the robustness check
413 * only happens in the pagetable (to verify it's still none)
414 * and not in the radix tree.
415 */
416 if (!(dst_vma->vm_flags & VM_SHARED)) {
417 if (!zeropage)
418 err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
419 dst_addr, src_addr, page);
420 else
421 err = mfill_zeropage_pte(dst_mm, dst_pmd,
422 dst_vma, dst_addr);
423 } else {
424 if (!zeropage)
425 err = shmem_mcopy_atomic_pte(dst_mm, dst_pmd,
426 dst_vma, dst_addr,
427 src_addr, page);
428 else
429 err = shmem_mfill_zeropage_pte(dst_mm, dst_pmd,
430 dst_vma, dst_addr);
431 }
432
433 return err;
434}
435
436static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
437 unsigned long dst_start,
438 unsigned long src_start,
439 unsigned long len,
440 bool zeropage,
441 bool *mmap_changing)
442{
443 struct vm_area_struct *dst_vma;
444 ssize_t err;
445 pmd_t *dst_pmd;
446 unsigned long src_addr, dst_addr;
447 long copied;
448 struct page *page;
449
450 /*
451 * Sanitize the command parameters:
452 */
453 BUG_ON(dst_start & ~PAGE_MASK);
454 BUG_ON(len & ~PAGE_MASK);
455
456 /* Does the address range wrap, or is the span zero-sized? */
457 BUG_ON(src_start + len <= src_start);
458 BUG_ON(dst_start + len <= dst_start);
459
460 src_addr = src_start;
461 dst_addr = dst_start;
462 copied = 0;
463 page = NULL;
464retry:
465 down_read(&dst_mm->mmap_sem);
466
467 /*
468 * If memory mappings are changing because of non-cooperative
469 * operation (e.g. mremap) running in parallel, bail out and
470 * request the user to retry later
471 */
472 err = -EAGAIN;
473 if (mmap_changing && READ_ONCE(*mmap_changing))
474 goto out_unlock;
475
476 /*
477 * Make sure the vma is not shared, that the dst range is
478 * both valid and fully within a single existing vma.
479 */
480 err = -ENOENT;
481 dst_vma = find_vma(dst_mm, dst_start);
482 if (!dst_vma)
483 goto out_unlock;
484 /*
485 * Check the vma is registered in uffd, this is required to
486 * enforce the VM_MAYWRITE check done at uffd registration
487 * time.
488 */
489 if (!dst_vma->vm_userfaultfd_ctx.ctx)
490 goto out_unlock;
491
492 if (dst_start < dst_vma->vm_start ||
493 dst_start + len > dst_vma->vm_end)
494 goto out_unlock;
495
496 err = -EINVAL;
497 /*
498 * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
499 * it will overwrite vm_ops, so vma_is_anonymous must return false.
500 */
501 if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
502 dst_vma->vm_flags & VM_SHARED))
503 goto out_unlock;
504
505 /*
506 * If this is a HUGETLB vma, pass off to appropriate routine
507 */
508 if (is_vm_hugetlb_page(dst_vma))
509 return __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
510 src_start, len, zeropage);
511
512 if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
513 goto out_unlock;
514
515 /*
516 * Ensure the dst_vma has a anon_vma or this page
517 * would get a NULL anon_vma when moved in the
518 * dst_vma.
519 */
520 err = -ENOMEM;
521 if (!(dst_vma->vm_flags & VM_SHARED) &&
522 unlikely(anon_vma_prepare(dst_vma)))
523 goto out_unlock;
524
525 while (src_addr < src_start + len) {
526 pmd_t dst_pmdval;
527
528 BUG_ON(dst_addr >= dst_start + len);
529
530 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
531 if (unlikely(!dst_pmd)) {
532 err = -ENOMEM;
533 break;
534 }
535
536 dst_pmdval = pmd_read_atomic(dst_pmd);
537 /*
538 * If the dst_pmd is mapped as THP don't
539 * override it and just be strict.
540 */
541 if (unlikely(pmd_trans_huge(dst_pmdval))) {
542 err = -EEXIST;
543 break;
544 }
545 if (unlikely(pmd_none(dst_pmdval)) &&
546 unlikely(__pte_alloc(dst_mm, dst_pmd))) {
547 err = -ENOMEM;
548 break;
549 }
550 /* If an huge pmd materialized from under us fail */
551 if (unlikely(pmd_trans_huge(*dst_pmd))) {
552 err = -EFAULT;
553 break;
554 }
555
556 BUG_ON(pmd_none(*dst_pmd));
557 BUG_ON(pmd_trans_huge(*dst_pmd));
558
559 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
560 src_addr, &page, zeropage);
561 cond_resched();
562
563 if (unlikely(err == -ENOENT)) {
564 void *page_kaddr;
565
566 up_read(&dst_mm->mmap_sem);
567 BUG_ON(!page);
568
569 page_kaddr = kmap(page);
570 err = copy_from_user(page_kaddr,
571 (const void __user *) src_addr,
572 PAGE_SIZE);
573 kunmap(page);
574 if (unlikely(err)) {
575 err = -EFAULT;
576 goto out;
577 }
578 goto retry;
579 } else
580 BUG_ON(page);
581
582 if (!err) {
583 dst_addr += PAGE_SIZE;
584 src_addr += PAGE_SIZE;
585 copied += PAGE_SIZE;
586
587 if (fatal_signal_pending(current))
588 err = -EINTR;
589 }
590 if (err)
591 break;
592 }
593
594out_unlock:
595 up_read(&dst_mm->mmap_sem);
596out:
597 if (page)
598 put_page(page);
599 BUG_ON(copied < 0);
600 BUG_ON(err > 0);
601 BUG_ON(!copied && !err);
602 return copied ? copied : err;
603}
604
605ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
606 unsigned long src_start, unsigned long len,
607 bool *mmap_changing)
608{
609 return __mcopy_atomic(dst_mm, dst_start, src_start, len, false,
610 mmap_changing);
611}
612
613ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
614 unsigned long len, bool *mmap_changing)
615{
616 return __mcopy_atomic(dst_mm, start, 0, len, true, mmap_changing);
617}
618