Merge tag 'linux-kselftest-next-6.6-rc1' of git://git.kernel.org/pub/scm/linux/kernel...
[platform/kernel/linux-rpi.git] / fs / hugetlbfs / inode.c
1 /*
2  * hugetlbpage-backed filesystem.  Based on ramfs.
3  *
4  * Nadia Yvette Chambers, 2002
5  *
6  * Copyright (C) 2002 Linus Torvalds.
7  * License: GPL
8  */
9
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/thread_info.h>
13 #include <asm/current.h>
14 #include <linux/falloc.h>
15 #include <linux/fs.h>
16 #include <linux/mount.h>
17 #include <linux/file.h>
18 #include <linux/kernel.h>
19 #include <linux/writeback.h>
20 #include <linux/pagemap.h>
21 #include <linux/highmem.h>
22 #include <linux/init.h>
23 #include <linux/string.h>
24 #include <linux/capability.h>
25 #include <linux/ctype.h>
26 #include <linux/backing-dev.h>
27 #include <linux/hugetlb.h>
28 #include <linux/pagevec.h>
29 #include <linux/fs_parser.h>
30 #include <linux/mman.h>
31 #include <linux/slab.h>
32 #include <linux/dnotify.h>
33 #include <linux/statfs.h>
34 #include <linux/security.h>
35 #include <linux/magic.h>
36 #include <linux/migrate.h>
37 #include <linux/uio.h>
38
39 #include <linux/uaccess.h>
40 #include <linux/sched/mm.h>
41
42 static const struct address_space_operations hugetlbfs_aops;
43 const struct file_operations hugetlbfs_file_operations;
44 static const struct inode_operations hugetlbfs_dir_inode_operations;
45 static const struct inode_operations hugetlbfs_inode_operations;
46
47 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
48
49 struct hugetlbfs_fs_context {
50         struct hstate           *hstate;
51         unsigned long long      max_size_opt;
52         unsigned long long      min_size_opt;
53         long                    max_hpages;
54         long                    nr_inodes;
55         long                    min_hpages;
56         enum hugetlbfs_size_type max_val_type;
57         enum hugetlbfs_size_type min_val_type;
58         kuid_t                  uid;
59         kgid_t                  gid;
60         umode_t                 mode;
61 };
62
63 int sysctl_hugetlb_shm_group;
64
65 enum hugetlb_param {
66         Opt_gid,
67         Opt_min_size,
68         Opt_mode,
69         Opt_nr_inodes,
70         Opt_pagesize,
71         Opt_size,
72         Opt_uid,
73 };
74
75 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
76         fsparam_u32   ("gid",           Opt_gid),
77         fsparam_string("min_size",      Opt_min_size),
78         fsparam_u32oct("mode",          Opt_mode),
79         fsparam_string("nr_inodes",     Opt_nr_inodes),
80         fsparam_string("pagesize",      Opt_pagesize),
81         fsparam_string("size",          Opt_size),
82         fsparam_u32   ("uid",           Opt_uid),
83         {}
84 };
85
86 #ifdef CONFIG_NUMA
87 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
88                                         struct inode *inode, pgoff_t index)
89 {
90         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
91                                                         index);
92 }
93
94 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
95 {
96         mpol_cond_put(vma->vm_policy);
97 }
98 #else
99 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
100                                         struct inode *inode, pgoff_t index)
101 {
102 }
103
104 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
105 {
106 }
107 #endif
108
109 /*
110  * Mask used when checking the page offset value passed in via system
111  * calls.  This value will be converted to a loff_t which is signed.
112  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
113  * value.  The extra bit (- 1 in the shift value) is to take the sign
114  * bit into account.
115  */
116 #define PGOFF_LOFFT_MAX \
117         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
118
119 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
120 {
121         struct inode *inode = file_inode(file);
122         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
123         loff_t len, vma_len;
124         int ret;
125         struct hstate *h = hstate_file(file);
126
127         /*
128          * vma address alignment (but not the pgoff alignment) has
129          * already been checked by prepare_hugepage_range.  If you add
130          * any error returns here, do so after setting VM_HUGETLB, so
131          * is_vm_hugetlb_page tests below unmap_region go the right
132          * way when do_mmap unwinds (may be important on powerpc
133          * and ia64).
134          */
135         vm_flags_set(vma, VM_HUGETLB | VM_DONTEXPAND);
136         vma->vm_ops = &hugetlb_vm_ops;
137
138         ret = seal_check_future_write(info->seals, vma);
139         if (ret)
140                 return ret;
141
142         /*
143          * page based offset in vm_pgoff could be sufficiently large to
144          * overflow a loff_t when converted to byte offset.  This can
145          * only happen on architectures where sizeof(loff_t) ==
146          * sizeof(unsigned long).  So, only check in those instances.
147          */
148         if (sizeof(unsigned long) == sizeof(loff_t)) {
149                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
150                         return -EINVAL;
151         }
152
153         /* must be huge page aligned */
154         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
155                 return -EINVAL;
156
157         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
158         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
159         /* check for overflow */
160         if (len < vma_len)
161                 return -EINVAL;
162
163         inode_lock(inode);
164         file_accessed(file);
165
166         ret = -ENOMEM;
167         if (!hugetlb_reserve_pages(inode,
168                                 vma->vm_pgoff >> huge_page_order(h),
169                                 len >> huge_page_shift(h), vma,
170                                 vma->vm_flags))
171                 goto out;
172
173         ret = 0;
174         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
175                 i_size_write(inode, len);
176 out:
177         inode_unlock(inode);
178
179         return ret;
180 }
181
182 /*
183  * Called under mmap_write_lock(mm).
184  */
185
186 static unsigned long
187 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
188                 unsigned long len, unsigned long pgoff, unsigned long flags)
189 {
190         struct hstate *h = hstate_file(file);
191         struct vm_unmapped_area_info info;
192
193         info.flags = 0;
194         info.length = len;
195         info.low_limit = current->mm->mmap_base;
196         info.high_limit = arch_get_mmap_end(addr, len, flags);
197         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
198         info.align_offset = 0;
199         return vm_unmapped_area(&info);
200 }
201
202 static unsigned long
203 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
204                 unsigned long len, unsigned long pgoff, unsigned long flags)
205 {
206         struct hstate *h = hstate_file(file);
207         struct vm_unmapped_area_info info;
208
209         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
210         info.length = len;
211         info.low_limit = PAGE_SIZE;
212         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
213         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
214         info.align_offset = 0;
215         addr = vm_unmapped_area(&info);
216
217         /*
218          * A failed mmap() very likely causes application failure,
219          * so fall back to the bottom-up function here. This scenario
220          * can happen with large stack limits and large mmap()
221          * allocations.
222          */
223         if (unlikely(offset_in_page(addr))) {
224                 VM_BUG_ON(addr != -ENOMEM);
225                 info.flags = 0;
226                 info.low_limit = current->mm->mmap_base;
227                 info.high_limit = arch_get_mmap_end(addr, len, flags);
228                 addr = vm_unmapped_area(&info);
229         }
230
231         return addr;
232 }
233
234 unsigned long
235 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
236                                   unsigned long len, unsigned long pgoff,
237                                   unsigned long flags)
238 {
239         struct mm_struct *mm = current->mm;
240         struct vm_area_struct *vma;
241         struct hstate *h = hstate_file(file);
242         const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
243
244         if (len & ~huge_page_mask(h))
245                 return -EINVAL;
246         if (len > TASK_SIZE)
247                 return -ENOMEM;
248
249         if (flags & MAP_FIXED) {
250                 if (prepare_hugepage_range(file, addr, len))
251                         return -EINVAL;
252                 return addr;
253         }
254
255         if (addr) {
256                 addr = ALIGN(addr, huge_page_size(h));
257                 vma = find_vma(mm, addr);
258                 if (mmap_end - len >= addr &&
259                     (!vma || addr + len <= vm_start_gap(vma)))
260                         return addr;
261         }
262
263         /*
264          * Use mm->get_unmapped_area value as a hint to use topdown routine.
265          * If architectures have special needs, they should define their own
266          * version of hugetlb_get_unmapped_area.
267          */
268         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
269                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
270                                 pgoff, flags);
271         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
272                         pgoff, flags);
273 }
274
275 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
276 static unsigned long
277 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
278                           unsigned long len, unsigned long pgoff,
279                           unsigned long flags)
280 {
281         return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
282 }
283 #endif
284
285 /*
286  * Support for read() - Find the page attached to f_mapping and copy out the
287  * data. This provides functionality similar to filemap_read().
288  */
289 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
290 {
291         struct file *file = iocb->ki_filp;
292         struct hstate *h = hstate_file(file);
293         struct address_space *mapping = file->f_mapping;
294         struct inode *inode = mapping->host;
295         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
296         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
297         unsigned long end_index;
298         loff_t isize;
299         ssize_t retval = 0;
300
301         while (iov_iter_count(to)) {
302                 struct page *page;
303                 size_t nr, copied;
304
305                 /* nr is the maximum number of bytes to copy from this page */
306                 nr = huge_page_size(h);
307                 isize = i_size_read(inode);
308                 if (!isize)
309                         break;
310                 end_index = (isize - 1) >> huge_page_shift(h);
311                 if (index > end_index)
312                         break;
313                 if (index == end_index) {
314                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
315                         if (nr <= offset)
316                                 break;
317                 }
318                 nr = nr - offset;
319
320                 /* Find the page */
321                 page = find_lock_page(mapping, index);
322                 if (unlikely(page == NULL)) {
323                         /*
324                          * We have a HOLE, zero out the user-buffer for the
325                          * length of the hole or request.
326                          */
327                         copied = iov_iter_zero(nr, to);
328                 } else {
329                         unlock_page(page);
330
331                         if (PageHWPoison(page)) {
332                                 put_page(page);
333                                 retval = -EIO;
334                                 break;
335                         }
336
337                         /*
338                          * We have the page, copy it to user space buffer.
339                          */
340                         copied = copy_page_to_iter(page, offset, nr, to);
341                         put_page(page);
342                 }
343                 offset += copied;
344                 retval += copied;
345                 if (copied != nr && iov_iter_count(to)) {
346                         if (!retval)
347                                 retval = -EFAULT;
348                         break;
349                 }
350                 index += offset >> huge_page_shift(h);
351                 offset &= ~huge_page_mask(h);
352         }
353         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
354         return retval;
355 }
356
357 static int hugetlbfs_write_begin(struct file *file,
358                         struct address_space *mapping,
359                         loff_t pos, unsigned len,
360                         struct page **pagep, void **fsdata)
361 {
362         return -EINVAL;
363 }
364
365 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
366                         loff_t pos, unsigned len, unsigned copied,
367                         struct page *page, void *fsdata)
368 {
369         BUG();
370         return -EINVAL;
371 }
372
373 static void hugetlb_delete_from_page_cache(struct folio *folio)
374 {
375         folio_clear_dirty(folio);
376         folio_clear_uptodate(folio);
377         filemap_remove_folio(folio);
378 }
379
380 /*
381  * Called with i_mmap_rwsem held for inode based vma maps.  This makes
382  * sure vma (and vm_mm) will not go away.  We also hold the hugetlb fault
383  * mutex for the page in the mapping.  So, we can not race with page being
384  * faulted into the vma.
385  */
386 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
387                                 unsigned long addr, struct page *page)
388 {
389         pte_t *ptep, pte;
390
391         ptep = hugetlb_walk(vma, addr, huge_page_size(hstate_vma(vma)));
392         if (!ptep)
393                 return false;
394
395         pte = huge_ptep_get(ptep);
396         if (huge_pte_none(pte) || !pte_present(pte))
397                 return false;
398
399         if (pte_page(pte) == page)
400                 return true;
401
402         return false;
403 }
404
405 /*
406  * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
407  * No, because the interval tree returns us only those vmas
408  * which overlap the truncated area starting at pgoff,
409  * and no vma on a 32-bit arch can span beyond the 4GB.
410  */
411 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
412 {
413         unsigned long offset = 0;
414
415         if (vma->vm_pgoff < start)
416                 offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
417
418         return vma->vm_start + offset;
419 }
420
421 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
422 {
423         unsigned long t_end;
424
425         if (!end)
426                 return vma->vm_end;
427
428         t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
429         if (t_end > vma->vm_end)
430                 t_end = vma->vm_end;
431         return t_end;
432 }
433
434 /*
435  * Called with hugetlb fault mutex held.  Therefore, no more mappings to
436  * this folio can be created while executing the routine.
437  */
438 static void hugetlb_unmap_file_folio(struct hstate *h,
439                                         struct address_space *mapping,
440                                         struct folio *folio, pgoff_t index)
441 {
442         struct rb_root_cached *root = &mapping->i_mmap;
443         struct hugetlb_vma_lock *vma_lock;
444         struct page *page = &folio->page;
445         struct vm_area_struct *vma;
446         unsigned long v_start;
447         unsigned long v_end;
448         pgoff_t start, end;
449
450         start = index * pages_per_huge_page(h);
451         end = (index + 1) * pages_per_huge_page(h);
452
453         i_mmap_lock_write(mapping);
454 retry:
455         vma_lock = NULL;
456         vma_interval_tree_foreach(vma, root, start, end - 1) {
457                 v_start = vma_offset_start(vma, start);
458                 v_end = vma_offset_end(vma, end);
459
460                 if (!hugetlb_vma_maps_page(vma, v_start, page))
461                         continue;
462
463                 if (!hugetlb_vma_trylock_write(vma)) {
464                         vma_lock = vma->vm_private_data;
465                         /*
466                          * If we can not get vma lock, we need to drop
467                          * immap_sema and take locks in order.  First,
468                          * take a ref on the vma_lock structure so that
469                          * we can be guaranteed it will not go away when
470                          * dropping immap_sema.
471                          */
472                         kref_get(&vma_lock->refs);
473                         break;
474                 }
475
476                 unmap_hugepage_range(vma, v_start, v_end, NULL,
477                                      ZAP_FLAG_DROP_MARKER);
478                 hugetlb_vma_unlock_write(vma);
479         }
480
481         i_mmap_unlock_write(mapping);
482
483         if (vma_lock) {
484                 /*
485                  * Wait on vma_lock.  We know it is still valid as we have
486                  * a reference.  We must 'open code' vma locking as we do
487                  * not know if vma_lock is still attached to vma.
488                  */
489                 down_write(&vma_lock->rw_sema);
490                 i_mmap_lock_write(mapping);
491
492                 vma = vma_lock->vma;
493                 if (!vma) {
494                         /*
495                          * If lock is no longer attached to vma, then just
496                          * unlock, drop our reference and retry looking for
497                          * other vmas.
498                          */
499                         up_write(&vma_lock->rw_sema);
500                         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
501                         goto retry;
502                 }
503
504                 /*
505                  * vma_lock is still attached to vma.  Check to see if vma
506                  * still maps page and if so, unmap.
507                  */
508                 v_start = vma_offset_start(vma, start);
509                 v_end = vma_offset_end(vma, end);
510                 if (hugetlb_vma_maps_page(vma, v_start, page))
511                         unmap_hugepage_range(vma, v_start, v_end, NULL,
512                                              ZAP_FLAG_DROP_MARKER);
513
514                 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
515                 hugetlb_vma_unlock_write(vma);
516
517                 goto retry;
518         }
519 }
520
521 static void
522 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
523                       zap_flags_t zap_flags)
524 {
525         struct vm_area_struct *vma;
526
527         /*
528          * end == 0 indicates that the entire range after start should be
529          * unmapped.  Note, end is exclusive, whereas the interval tree takes
530          * an inclusive "last".
531          */
532         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
533                 unsigned long v_start;
534                 unsigned long v_end;
535
536                 if (!hugetlb_vma_trylock_write(vma))
537                         continue;
538
539                 v_start = vma_offset_start(vma, start);
540                 v_end = vma_offset_end(vma, end);
541
542                 unmap_hugepage_range(vma, v_start, v_end, NULL, zap_flags);
543
544                 /*
545                  * Note that vma lock only exists for shared/non-private
546                  * vmas.  Therefore, lock is not held when calling
547                  * unmap_hugepage_range for private vmas.
548                  */
549                 hugetlb_vma_unlock_write(vma);
550         }
551 }
552
553 /*
554  * Called with hugetlb fault mutex held.
555  * Returns true if page was actually removed, false otherwise.
556  */
557 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
558                                         struct address_space *mapping,
559                                         struct folio *folio, pgoff_t index,
560                                         bool truncate_op)
561 {
562         bool ret = false;
563
564         /*
565          * If folio is mapped, it was faulted in after being
566          * unmapped in caller.  Unmap (again) while holding
567          * the fault mutex.  The mutex will prevent faults
568          * until we finish removing the folio.
569          */
570         if (unlikely(folio_mapped(folio)))
571                 hugetlb_unmap_file_folio(h, mapping, folio, index);
572
573         folio_lock(folio);
574         /*
575          * We must remove the folio from page cache before removing
576          * the region/ reserve map (hugetlb_unreserve_pages).  In
577          * rare out of memory conditions, removal of the region/reserve
578          * map could fail.  Correspondingly, the subpool and global
579          * reserve usage count can need to be adjusted.
580          */
581         VM_BUG_ON_FOLIO(folio_test_hugetlb_restore_reserve(folio), folio);
582         hugetlb_delete_from_page_cache(folio);
583         ret = true;
584         if (!truncate_op) {
585                 if (unlikely(hugetlb_unreserve_pages(inode, index,
586                                                         index + 1, 1)))
587                         hugetlb_fix_reserve_counts(inode);
588         }
589
590         folio_unlock(folio);
591         return ret;
592 }
593
594 /*
595  * remove_inode_hugepages handles two distinct cases: truncation and hole
596  * punch.  There are subtle differences in operation for each case.
597  *
598  * truncation is indicated by end of range being LLONG_MAX
599  *      In this case, we first scan the range and release found pages.
600  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
601  *      maps and global counts.  Page faults can race with truncation.
602  *      During faults, hugetlb_no_page() checks i_size before page allocation,
603  *      and again after obtaining page table lock.  It will 'back out'
604  *      allocations in the truncated range.
605  * hole punch is indicated if end is not LLONG_MAX
606  *      In the hole punch case we scan the range and release found pages.
607  *      Only when releasing a page is the associated region/reserve map
608  *      deleted.  The region/reserve map for ranges without associated
609  *      pages are not modified.  Page faults can race with hole punch.
610  *      This is indicated if we find a mapped page.
611  * Note: If the passed end of range value is beyond the end of file, but
612  * not LLONG_MAX this routine still performs a hole punch operation.
613  */
614 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
615                                    loff_t lend)
616 {
617         struct hstate *h = hstate_inode(inode);
618         struct address_space *mapping = &inode->i_data;
619         const pgoff_t start = lstart >> huge_page_shift(h);
620         const pgoff_t end = lend >> huge_page_shift(h);
621         struct folio_batch fbatch;
622         pgoff_t next, index;
623         int i, freed = 0;
624         bool truncate_op = (lend == LLONG_MAX);
625
626         folio_batch_init(&fbatch);
627         next = start;
628         while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
629                 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
630                         struct folio *folio = fbatch.folios[i];
631                         u32 hash = 0;
632
633                         index = folio->index;
634                         hash = hugetlb_fault_mutex_hash(mapping, index);
635                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
636
637                         /*
638                          * Remove folio that was part of folio_batch.
639                          */
640                         if (remove_inode_single_folio(h, inode, mapping, folio,
641                                                         index, truncate_op))
642                                 freed++;
643
644                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
645                 }
646                 folio_batch_release(&fbatch);
647                 cond_resched();
648         }
649
650         if (truncate_op)
651                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
652 }
653
654 static void hugetlbfs_evict_inode(struct inode *inode)
655 {
656         struct resv_map *resv_map;
657
658         remove_inode_hugepages(inode, 0, LLONG_MAX);
659
660         /*
661          * Get the resv_map from the address space embedded in the inode.
662          * This is the address space which points to any resv_map allocated
663          * at inode creation time.  If this is a device special inode,
664          * i_mapping may not point to the original address space.
665          */
666         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
667         /* Only regular and link inodes have associated reserve maps */
668         if (resv_map)
669                 resv_map_release(&resv_map->refs);
670         clear_inode(inode);
671 }
672
673 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
674 {
675         pgoff_t pgoff;
676         struct address_space *mapping = inode->i_mapping;
677         struct hstate *h = hstate_inode(inode);
678
679         BUG_ON(offset & ~huge_page_mask(h));
680         pgoff = offset >> PAGE_SHIFT;
681
682         i_size_write(inode, offset);
683         i_mmap_lock_write(mapping);
684         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
685                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
686                                       ZAP_FLAG_DROP_MARKER);
687         i_mmap_unlock_write(mapping);
688         remove_inode_hugepages(inode, offset, LLONG_MAX);
689 }
690
691 static void hugetlbfs_zero_partial_page(struct hstate *h,
692                                         struct address_space *mapping,
693                                         loff_t start,
694                                         loff_t end)
695 {
696         pgoff_t idx = start >> huge_page_shift(h);
697         struct folio *folio;
698
699         folio = filemap_lock_folio(mapping, idx);
700         if (IS_ERR(folio))
701                 return;
702
703         start = start & ~huge_page_mask(h);
704         end = end & ~huge_page_mask(h);
705         if (!end)
706                 end = huge_page_size(h);
707
708         folio_zero_segment(folio, (size_t)start, (size_t)end);
709
710         folio_unlock(folio);
711         folio_put(folio);
712 }
713
714 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
715 {
716         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
717         struct address_space *mapping = inode->i_mapping;
718         struct hstate *h = hstate_inode(inode);
719         loff_t hpage_size = huge_page_size(h);
720         loff_t hole_start, hole_end;
721
722         /*
723          * hole_start and hole_end indicate the full pages within the hole.
724          */
725         hole_start = round_up(offset, hpage_size);
726         hole_end = round_down(offset + len, hpage_size);
727
728         inode_lock(inode);
729
730         /* protected by i_rwsem */
731         if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
732                 inode_unlock(inode);
733                 return -EPERM;
734         }
735
736         i_mmap_lock_write(mapping);
737
738         /* If range starts before first full page, zero partial page. */
739         if (offset < hole_start)
740                 hugetlbfs_zero_partial_page(h, mapping,
741                                 offset, min(offset + len, hole_start));
742
743         /* Unmap users of full pages in the hole. */
744         if (hole_end > hole_start) {
745                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
746                         hugetlb_vmdelete_list(&mapping->i_mmap,
747                                               hole_start >> PAGE_SHIFT,
748                                               hole_end >> PAGE_SHIFT, 0);
749         }
750
751         /* If range extends beyond last full page, zero partial page. */
752         if ((offset + len) > hole_end && (offset + len) > hole_start)
753                 hugetlbfs_zero_partial_page(h, mapping,
754                                 hole_end, offset + len);
755
756         i_mmap_unlock_write(mapping);
757
758         /* Remove full pages from the file. */
759         if (hole_end > hole_start)
760                 remove_inode_hugepages(inode, hole_start, hole_end);
761
762         inode_unlock(inode);
763
764         return 0;
765 }
766
767 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
768                                 loff_t len)
769 {
770         struct inode *inode = file_inode(file);
771         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
772         struct address_space *mapping = inode->i_mapping;
773         struct hstate *h = hstate_inode(inode);
774         struct vm_area_struct pseudo_vma;
775         struct mm_struct *mm = current->mm;
776         loff_t hpage_size = huge_page_size(h);
777         unsigned long hpage_shift = huge_page_shift(h);
778         pgoff_t start, index, end;
779         int error;
780         u32 hash;
781
782         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
783                 return -EOPNOTSUPP;
784
785         if (mode & FALLOC_FL_PUNCH_HOLE)
786                 return hugetlbfs_punch_hole(inode, offset, len);
787
788         /*
789          * Default preallocate case.
790          * For this range, start is rounded down and end is rounded up
791          * as well as being converted to page offsets.
792          */
793         start = offset >> hpage_shift;
794         end = (offset + len + hpage_size - 1) >> hpage_shift;
795
796         inode_lock(inode);
797
798         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
799         error = inode_newsize_ok(inode, offset + len);
800         if (error)
801                 goto out;
802
803         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
804                 error = -EPERM;
805                 goto out;
806         }
807
808         /*
809          * Initialize a pseudo vma as this is required by the huge page
810          * allocation routines.  If NUMA is configured, use page index
811          * as input to create an allocation policy.
812          */
813         vma_init(&pseudo_vma, mm);
814         vm_flags_init(&pseudo_vma, VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
815         pseudo_vma.vm_file = file;
816
817         for (index = start; index < end; index++) {
818                 /*
819                  * This is supposed to be the vaddr where the page is being
820                  * faulted in, but we have no vaddr here.
821                  */
822                 struct folio *folio;
823                 unsigned long addr;
824
825                 cond_resched();
826
827                 /*
828                  * fallocate(2) manpage permits EINTR; we may have been
829                  * interrupted because we are using up too much memory.
830                  */
831                 if (signal_pending(current)) {
832                         error = -EINTR;
833                         break;
834                 }
835
836                 /* addr is the offset within the file (zero based) */
837                 addr = index * hpage_size;
838
839                 /* mutex taken here, fault path and hole punch */
840                 hash = hugetlb_fault_mutex_hash(mapping, index);
841                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
842
843                 /* See if already present in mapping to avoid alloc/free */
844                 folio = filemap_get_folio(mapping, index);
845                 if (!IS_ERR(folio)) {
846                         folio_put(folio);
847                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
848                         continue;
849                 }
850
851                 /*
852                  * Allocate folio without setting the avoid_reserve argument.
853                  * There certainly are no reserves associated with the
854                  * pseudo_vma.  However, there could be shared mappings with
855                  * reserves for the file at the inode level.  If we fallocate
856                  * folios in these areas, we need to consume the reserves
857                  * to keep reservation accounting consistent.
858                  */
859                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
860                 folio = alloc_hugetlb_folio(&pseudo_vma, addr, 0);
861                 hugetlb_drop_vma_policy(&pseudo_vma);
862                 if (IS_ERR(folio)) {
863                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
864                         error = PTR_ERR(folio);
865                         goto out;
866                 }
867                 clear_huge_page(&folio->page, addr, pages_per_huge_page(h));
868                 __folio_mark_uptodate(folio);
869                 error = hugetlb_add_to_page_cache(folio, mapping, index);
870                 if (unlikely(error)) {
871                         restore_reserve_on_error(h, &pseudo_vma, addr, folio);
872                         folio_put(folio);
873                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
874                         goto out;
875                 }
876
877                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
878
879                 folio_set_hugetlb_migratable(folio);
880                 /*
881                  * folio_unlock because locked by hugetlb_add_to_page_cache()
882                  * folio_put() due to reference from alloc_hugetlb_folio()
883                  */
884                 folio_unlock(folio);
885                 folio_put(folio);
886         }
887
888         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
889                 i_size_write(inode, offset + len);
890         inode_set_ctime_current(inode);
891 out:
892         inode_unlock(inode);
893         return error;
894 }
895
896 static int hugetlbfs_setattr(struct mnt_idmap *idmap,
897                              struct dentry *dentry, struct iattr *attr)
898 {
899         struct inode *inode = d_inode(dentry);
900         struct hstate *h = hstate_inode(inode);
901         int error;
902         unsigned int ia_valid = attr->ia_valid;
903         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
904
905         error = setattr_prepare(&nop_mnt_idmap, dentry, attr);
906         if (error)
907                 return error;
908
909         if (ia_valid & ATTR_SIZE) {
910                 loff_t oldsize = inode->i_size;
911                 loff_t newsize = attr->ia_size;
912
913                 if (newsize & ~huge_page_mask(h))
914                         return -EINVAL;
915                 /* protected by i_rwsem */
916                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
917                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
918                         return -EPERM;
919                 hugetlb_vmtruncate(inode, newsize);
920         }
921
922         setattr_copy(&nop_mnt_idmap, inode, attr);
923         mark_inode_dirty(inode);
924         return 0;
925 }
926
927 static struct inode *hugetlbfs_get_root(struct super_block *sb,
928                                         struct hugetlbfs_fs_context *ctx)
929 {
930         struct inode *inode;
931
932         inode = new_inode(sb);
933         if (inode) {
934                 inode->i_ino = get_next_ino();
935                 inode->i_mode = S_IFDIR | ctx->mode;
936                 inode->i_uid = ctx->uid;
937                 inode->i_gid = ctx->gid;
938                 inode->i_atime = inode->i_mtime = inode_set_ctime_current(inode);
939                 inode->i_op = &hugetlbfs_dir_inode_operations;
940                 inode->i_fop = &simple_dir_operations;
941                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
942                 inc_nlink(inode);
943                 lockdep_annotate_inode_mutex_key(inode);
944         }
945         return inode;
946 }
947
948 /*
949  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
950  * be taken from reclaim -- unlike regular filesystems. This needs an
951  * annotation because huge_pmd_share() does an allocation under hugetlb's
952  * i_mmap_rwsem.
953  */
954 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
955
956 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
957                                         struct inode *dir,
958                                         umode_t mode, dev_t dev)
959 {
960         struct inode *inode;
961         struct resv_map *resv_map = NULL;
962
963         /*
964          * Reserve maps are only needed for inodes that can have associated
965          * page allocations.
966          */
967         if (S_ISREG(mode) || S_ISLNK(mode)) {
968                 resv_map = resv_map_alloc();
969                 if (!resv_map)
970                         return NULL;
971         }
972
973         inode = new_inode(sb);
974         if (inode) {
975                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
976
977                 inode->i_ino = get_next_ino();
978                 inode_init_owner(&nop_mnt_idmap, inode, dir, mode);
979                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
980                                 &hugetlbfs_i_mmap_rwsem_key);
981                 inode->i_mapping->a_ops = &hugetlbfs_aops;
982                 inode->i_atime = inode->i_mtime = inode_set_ctime_current(inode);
983                 inode->i_mapping->private_data = resv_map;
984                 info->seals = F_SEAL_SEAL;
985                 switch (mode & S_IFMT) {
986                 default:
987                         init_special_inode(inode, mode, dev);
988                         break;
989                 case S_IFREG:
990                         inode->i_op = &hugetlbfs_inode_operations;
991                         inode->i_fop = &hugetlbfs_file_operations;
992                         break;
993                 case S_IFDIR:
994                         inode->i_op = &hugetlbfs_dir_inode_operations;
995                         inode->i_fop = &simple_dir_operations;
996
997                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
998                         inc_nlink(inode);
999                         break;
1000                 case S_IFLNK:
1001                         inode->i_op = &page_symlink_inode_operations;
1002                         inode_nohighmem(inode);
1003                         break;
1004                 }
1005                 lockdep_annotate_inode_mutex_key(inode);
1006         } else {
1007                 if (resv_map)
1008                         kref_put(&resv_map->refs, resv_map_release);
1009         }
1010
1011         return inode;
1012 }
1013
1014 /*
1015  * File creation. Allocate an inode, and we're done..
1016  */
1017 static int hugetlbfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
1018                            struct dentry *dentry, umode_t mode, dev_t dev)
1019 {
1020         struct inode *inode;
1021
1022         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
1023         if (!inode)
1024                 return -ENOSPC;
1025         dir->i_mtime = inode_set_ctime_current(dir);
1026         d_instantiate(dentry, inode);
1027         dget(dentry);/* Extra count - pin the dentry in core */
1028         return 0;
1029 }
1030
1031 static int hugetlbfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1032                            struct dentry *dentry, umode_t mode)
1033 {
1034         int retval = hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry,
1035                                      mode | S_IFDIR, 0);
1036         if (!retval)
1037                 inc_nlink(dir);
1038         return retval;
1039 }
1040
1041 static int hugetlbfs_create(struct mnt_idmap *idmap,
1042                             struct inode *dir, struct dentry *dentry,
1043                             umode_t mode, bool excl)
1044 {
1045         return hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry, mode | S_IFREG, 0);
1046 }
1047
1048 static int hugetlbfs_tmpfile(struct mnt_idmap *idmap,
1049                              struct inode *dir, struct file *file,
1050                              umode_t mode)
1051 {
1052         struct inode *inode;
1053
1054         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode | S_IFREG, 0);
1055         if (!inode)
1056                 return -ENOSPC;
1057         dir->i_mtime = inode_set_ctime_current(dir);
1058         d_tmpfile(file, inode);
1059         return finish_open_simple(file, 0);
1060 }
1061
1062 static int hugetlbfs_symlink(struct mnt_idmap *idmap,
1063                              struct inode *dir, struct dentry *dentry,
1064                              const char *symname)
1065 {
1066         struct inode *inode;
1067         int error = -ENOSPC;
1068
1069         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
1070         if (inode) {
1071                 int l = strlen(symname)+1;
1072                 error = page_symlink(inode, symname, l);
1073                 if (!error) {
1074                         d_instantiate(dentry, inode);
1075                         dget(dentry);
1076                 } else
1077                         iput(inode);
1078         }
1079         dir->i_mtime = inode_set_ctime_current(dir);
1080
1081         return error;
1082 }
1083
1084 #ifdef CONFIG_MIGRATION
1085 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1086                                 struct folio *dst, struct folio *src,
1087                                 enum migrate_mode mode)
1088 {
1089         int rc;
1090
1091         rc = migrate_huge_page_move_mapping(mapping, dst, src);
1092         if (rc != MIGRATEPAGE_SUCCESS)
1093                 return rc;
1094
1095         if (hugetlb_folio_subpool(src)) {
1096                 hugetlb_set_folio_subpool(dst,
1097                                         hugetlb_folio_subpool(src));
1098                 hugetlb_set_folio_subpool(src, NULL);
1099         }
1100
1101         if (mode != MIGRATE_SYNC_NO_COPY)
1102                 folio_migrate_copy(dst, src);
1103         else
1104                 folio_migrate_flags(dst, src);
1105
1106         return MIGRATEPAGE_SUCCESS;
1107 }
1108 #else
1109 #define hugetlbfs_migrate_folio NULL
1110 #endif
1111
1112 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1113                                 struct page *page)
1114 {
1115         return 0;
1116 }
1117
1118 /*
1119  * Display the mount options in /proc/mounts.
1120  */
1121 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1122 {
1123         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1124         struct hugepage_subpool *spool = sbinfo->spool;
1125         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1126         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1127         char mod;
1128
1129         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1130                 seq_printf(m, ",uid=%u",
1131                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1132         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1133                 seq_printf(m, ",gid=%u",
1134                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1135         if (sbinfo->mode != 0755)
1136                 seq_printf(m, ",mode=%o", sbinfo->mode);
1137         if (sbinfo->max_inodes != -1)
1138                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1139
1140         hpage_size /= 1024;
1141         mod = 'K';
1142         if (hpage_size >= 1024) {
1143                 hpage_size /= 1024;
1144                 mod = 'M';
1145         }
1146         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1147         if (spool) {
1148                 if (spool->max_hpages != -1)
1149                         seq_printf(m, ",size=%llu",
1150                                    (unsigned long long)spool->max_hpages << hpage_shift);
1151                 if (spool->min_hpages != -1)
1152                         seq_printf(m, ",min_size=%llu",
1153                                    (unsigned long long)spool->min_hpages << hpage_shift);
1154         }
1155         return 0;
1156 }
1157
1158 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1159 {
1160         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1161         struct hstate *h = hstate_inode(d_inode(dentry));
1162
1163         buf->f_type = HUGETLBFS_MAGIC;
1164         buf->f_bsize = huge_page_size(h);
1165         if (sbinfo) {
1166                 spin_lock(&sbinfo->stat_lock);
1167                 /* If no limits set, just report 0 or -1 for max/free/used
1168                  * blocks, like simple_statfs() */
1169                 if (sbinfo->spool) {
1170                         long free_pages;
1171
1172                         spin_lock_irq(&sbinfo->spool->lock);
1173                         buf->f_blocks = sbinfo->spool->max_hpages;
1174                         free_pages = sbinfo->spool->max_hpages
1175                                 - sbinfo->spool->used_hpages;
1176                         buf->f_bavail = buf->f_bfree = free_pages;
1177                         spin_unlock_irq(&sbinfo->spool->lock);
1178                         buf->f_files = sbinfo->max_inodes;
1179                         buf->f_ffree = sbinfo->free_inodes;
1180                 }
1181                 spin_unlock(&sbinfo->stat_lock);
1182         }
1183         buf->f_namelen = NAME_MAX;
1184         return 0;
1185 }
1186
1187 static void hugetlbfs_put_super(struct super_block *sb)
1188 {
1189         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1190
1191         if (sbi) {
1192                 sb->s_fs_info = NULL;
1193
1194                 if (sbi->spool)
1195                         hugepage_put_subpool(sbi->spool);
1196
1197                 kfree(sbi);
1198         }
1199 }
1200
1201 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1202 {
1203         if (sbinfo->free_inodes >= 0) {
1204                 spin_lock(&sbinfo->stat_lock);
1205                 if (unlikely(!sbinfo->free_inodes)) {
1206                         spin_unlock(&sbinfo->stat_lock);
1207                         return 0;
1208                 }
1209                 sbinfo->free_inodes--;
1210                 spin_unlock(&sbinfo->stat_lock);
1211         }
1212
1213         return 1;
1214 }
1215
1216 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1217 {
1218         if (sbinfo->free_inodes >= 0) {
1219                 spin_lock(&sbinfo->stat_lock);
1220                 sbinfo->free_inodes++;
1221                 spin_unlock(&sbinfo->stat_lock);
1222         }
1223 }
1224
1225
1226 static struct kmem_cache *hugetlbfs_inode_cachep;
1227
1228 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1229 {
1230         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1231         struct hugetlbfs_inode_info *p;
1232
1233         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1234                 return NULL;
1235         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1236         if (unlikely(!p)) {
1237                 hugetlbfs_inc_free_inodes(sbinfo);
1238                 return NULL;
1239         }
1240
1241         /*
1242          * Any time after allocation, hugetlbfs_destroy_inode can be called
1243          * for the inode.  mpol_free_shared_policy is unconditionally called
1244          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1245          * in case of a quick call to destroy.
1246          *
1247          * Note that the policy is initialized even if we are creating a
1248          * private inode.  This simplifies hugetlbfs_destroy_inode.
1249          */
1250         mpol_shared_policy_init(&p->policy, NULL);
1251
1252         return &p->vfs_inode;
1253 }
1254
1255 static void hugetlbfs_free_inode(struct inode *inode)
1256 {
1257         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1258 }
1259
1260 static void hugetlbfs_destroy_inode(struct inode *inode)
1261 {
1262         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1263         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1264 }
1265
1266 static const struct address_space_operations hugetlbfs_aops = {
1267         .write_begin    = hugetlbfs_write_begin,
1268         .write_end      = hugetlbfs_write_end,
1269         .dirty_folio    = noop_dirty_folio,
1270         .migrate_folio  = hugetlbfs_migrate_folio,
1271         .error_remove_page      = hugetlbfs_error_remove_page,
1272 };
1273
1274
1275 static void init_once(void *foo)
1276 {
1277         struct hugetlbfs_inode_info *ei = foo;
1278
1279         inode_init_once(&ei->vfs_inode);
1280 }
1281
1282 const struct file_operations hugetlbfs_file_operations = {
1283         .read_iter              = hugetlbfs_read_iter,
1284         .mmap                   = hugetlbfs_file_mmap,
1285         .fsync                  = noop_fsync,
1286         .get_unmapped_area      = hugetlb_get_unmapped_area,
1287         .llseek                 = default_llseek,
1288         .fallocate              = hugetlbfs_fallocate,
1289 };
1290
1291 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1292         .create         = hugetlbfs_create,
1293         .lookup         = simple_lookup,
1294         .link           = simple_link,
1295         .unlink         = simple_unlink,
1296         .symlink        = hugetlbfs_symlink,
1297         .mkdir          = hugetlbfs_mkdir,
1298         .rmdir          = simple_rmdir,
1299         .mknod          = hugetlbfs_mknod,
1300         .rename         = simple_rename,
1301         .setattr        = hugetlbfs_setattr,
1302         .tmpfile        = hugetlbfs_tmpfile,
1303 };
1304
1305 static const struct inode_operations hugetlbfs_inode_operations = {
1306         .setattr        = hugetlbfs_setattr,
1307 };
1308
1309 static const struct super_operations hugetlbfs_ops = {
1310         .alloc_inode    = hugetlbfs_alloc_inode,
1311         .free_inode     = hugetlbfs_free_inode,
1312         .destroy_inode  = hugetlbfs_destroy_inode,
1313         .evict_inode    = hugetlbfs_evict_inode,
1314         .statfs         = hugetlbfs_statfs,
1315         .put_super      = hugetlbfs_put_super,
1316         .show_options   = hugetlbfs_show_options,
1317 };
1318
1319 /*
1320  * Convert size option passed from command line to number of huge pages
1321  * in the pool specified by hstate.  Size option could be in bytes
1322  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1323  */
1324 static long
1325 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1326                          enum hugetlbfs_size_type val_type)
1327 {
1328         if (val_type == NO_SIZE)
1329                 return -1;
1330
1331         if (val_type == SIZE_PERCENT) {
1332                 size_opt <<= huge_page_shift(h);
1333                 size_opt *= h->max_huge_pages;
1334                 do_div(size_opt, 100);
1335         }
1336
1337         size_opt >>= huge_page_shift(h);
1338         return size_opt;
1339 }
1340
1341 /*
1342  * Parse one mount parameter.
1343  */
1344 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1345 {
1346         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1347         struct fs_parse_result result;
1348         char *rest;
1349         unsigned long ps;
1350         int opt;
1351
1352         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1353         if (opt < 0)
1354                 return opt;
1355
1356         switch (opt) {
1357         case Opt_uid:
1358                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1359                 if (!uid_valid(ctx->uid))
1360                         goto bad_val;
1361                 return 0;
1362
1363         case Opt_gid:
1364                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1365                 if (!gid_valid(ctx->gid))
1366                         goto bad_val;
1367                 return 0;
1368
1369         case Opt_mode:
1370                 ctx->mode = result.uint_32 & 01777U;
1371                 return 0;
1372
1373         case Opt_size:
1374                 /* memparse() will accept a K/M/G without a digit */
1375                 if (!param->string || !isdigit(param->string[0]))
1376                         goto bad_val;
1377                 ctx->max_size_opt = memparse(param->string, &rest);
1378                 ctx->max_val_type = SIZE_STD;
1379                 if (*rest == '%')
1380                         ctx->max_val_type = SIZE_PERCENT;
1381                 return 0;
1382
1383         case Opt_nr_inodes:
1384                 /* memparse() will accept a K/M/G without a digit */
1385                 if (!param->string || !isdigit(param->string[0]))
1386                         goto bad_val;
1387                 ctx->nr_inodes = memparse(param->string, &rest);
1388                 return 0;
1389
1390         case Opt_pagesize:
1391                 ps = memparse(param->string, &rest);
1392                 ctx->hstate = size_to_hstate(ps);
1393                 if (!ctx->hstate) {
1394                         pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1395                         return -EINVAL;
1396                 }
1397                 return 0;
1398
1399         case Opt_min_size:
1400                 /* memparse() will accept a K/M/G without a digit */
1401                 if (!param->string || !isdigit(param->string[0]))
1402                         goto bad_val;
1403                 ctx->min_size_opt = memparse(param->string, &rest);
1404                 ctx->min_val_type = SIZE_STD;
1405                 if (*rest == '%')
1406                         ctx->min_val_type = SIZE_PERCENT;
1407                 return 0;
1408
1409         default:
1410                 return -EINVAL;
1411         }
1412
1413 bad_val:
1414         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1415                       param->string, param->key);
1416 }
1417
1418 /*
1419  * Validate the parsed options.
1420  */
1421 static int hugetlbfs_validate(struct fs_context *fc)
1422 {
1423         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1424
1425         /*
1426          * Use huge page pool size (in hstate) to convert the size
1427          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1428          */
1429         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1430                                                    ctx->max_size_opt,
1431                                                    ctx->max_val_type);
1432         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1433                                                    ctx->min_size_opt,
1434                                                    ctx->min_val_type);
1435
1436         /*
1437          * If max_size was specified, then min_size must be smaller
1438          */
1439         if (ctx->max_val_type > NO_SIZE &&
1440             ctx->min_hpages > ctx->max_hpages) {
1441                 pr_err("Minimum size can not be greater than maximum size\n");
1442                 return -EINVAL;
1443         }
1444
1445         return 0;
1446 }
1447
1448 static int
1449 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1450 {
1451         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1452         struct hugetlbfs_sb_info *sbinfo;
1453
1454         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1455         if (!sbinfo)
1456                 return -ENOMEM;
1457         sb->s_fs_info = sbinfo;
1458         spin_lock_init(&sbinfo->stat_lock);
1459         sbinfo->hstate          = ctx->hstate;
1460         sbinfo->max_inodes      = ctx->nr_inodes;
1461         sbinfo->free_inodes     = ctx->nr_inodes;
1462         sbinfo->spool           = NULL;
1463         sbinfo->uid             = ctx->uid;
1464         sbinfo->gid             = ctx->gid;
1465         sbinfo->mode            = ctx->mode;
1466
1467         /*
1468          * Allocate and initialize subpool if maximum or minimum size is
1469          * specified.  Any needed reservations (for minimum size) are taken
1470          * when the subpool is created.
1471          */
1472         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1473                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1474                                                      ctx->max_hpages,
1475                                                      ctx->min_hpages);
1476                 if (!sbinfo->spool)
1477                         goto out_free;
1478         }
1479         sb->s_maxbytes = MAX_LFS_FILESIZE;
1480         sb->s_blocksize = huge_page_size(ctx->hstate);
1481         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1482         sb->s_magic = HUGETLBFS_MAGIC;
1483         sb->s_op = &hugetlbfs_ops;
1484         sb->s_time_gran = 1;
1485
1486         /*
1487          * Due to the special and limited functionality of hugetlbfs, it does
1488          * not work well as a stacking filesystem.
1489          */
1490         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1491         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1492         if (!sb->s_root)
1493                 goto out_free;
1494         return 0;
1495 out_free:
1496         kfree(sbinfo->spool);
1497         kfree(sbinfo);
1498         return -ENOMEM;
1499 }
1500
1501 static int hugetlbfs_get_tree(struct fs_context *fc)
1502 {
1503         int err = hugetlbfs_validate(fc);
1504         if (err)
1505                 return err;
1506         return get_tree_nodev(fc, hugetlbfs_fill_super);
1507 }
1508
1509 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1510 {
1511         kfree(fc->fs_private);
1512 }
1513
1514 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1515         .free           = hugetlbfs_fs_context_free,
1516         .parse_param    = hugetlbfs_parse_param,
1517         .get_tree       = hugetlbfs_get_tree,
1518 };
1519
1520 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1521 {
1522         struct hugetlbfs_fs_context *ctx;
1523
1524         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1525         if (!ctx)
1526                 return -ENOMEM;
1527
1528         ctx->max_hpages = -1; /* No limit on size by default */
1529         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1530         ctx->uid        = current_fsuid();
1531         ctx->gid        = current_fsgid();
1532         ctx->mode       = 0755;
1533         ctx->hstate     = &default_hstate;
1534         ctx->min_hpages = -1; /* No default minimum size */
1535         ctx->max_val_type = NO_SIZE;
1536         ctx->min_val_type = NO_SIZE;
1537         fc->fs_private = ctx;
1538         fc->ops = &hugetlbfs_fs_context_ops;
1539         return 0;
1540 }
1541
1542 static struct file_system_type hugetlbfs_fs_type = {
1543         .name                   = "hugetlbfs",
1544         .init_fs_context        = hugetlbfs_init_fs_context,
1545         .parameters             = hugetlb_fs_parameters,
1546         .kill_sb                = kill_litter_super,
1547 };
1548
1549 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1550
1551 static int can_do_hugetlb_shm(void)
1552 {
1553         kgid_t shm_group;
1554         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1555         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1556 }
1557
1558 static int get_hstate_idx(int page_size_log)
1559 {
1560         struct hstate *h = hstate_sizelog(page_size_log);
1561
1562         if (!h)
1563                 return -1;
1564         return hstate_index(h);
1565 }
1566
1567 /*
1568  * Note that size should be aligned to proper hugepage size in caller side,
1569  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1570  */
1571 struct file *hugetlb_file_setup(const char *name, size_t size,
1572                                 vm_flags_t acctflag, int creat_flags,
1573                                 int page_size_log)
1574 {
1575         struct inode *inode;
1576         struct vfsmount *mnt;
1577         int hstate_idx;
1578         struct file *file;
1579
1580         hstate_idx = get_hstate_idx(page_size_log);
1581         if (hstate_idx < 0)
1582                 return ERR_PTR(-ENODEV);
1583
1584         mnt = hugetlbfs_vfsmount[hstate_idx];
1585         if (!mnt)
1586                 return ERR_PTR(-ENOENT);
1587
1588         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1589                 struct ucounts *ucounts = current_ucounts();
1590
1591                 if (user_shm_lock(size, ucounts)) {
1592                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1593                                 current->comm, current->pid);
1594                         user_shm_unlock(size, ucounts);
1595                 }
1596                 return ERR_PTR(-EPERM);
1597         }
1598
1599         file = ERR_PTR(-ENOSPC);
1600         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1601         if (!inode)
1602                 goto out;
1603         if (creat_flags == HUGETLB_SHMFS_INODE)
1604                 inode->i_flags |= S_PRIVATE;
1605
1606         inode->i_size = size;
1607         clear_nlink(inode);
1608
1609         if (!hugetlb_reserve_pages(inode, 0,
1610                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1611                         acctflag))
1612                 file = ERR_PTR(-ENOMEM);
1613         else
1614                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1615                                         &hugetlbfs_file_operations);
1616         if (!IS_ERR(file))
1617                 return file;
1618
1619         iput(inode);
1620 out:
1621         return file;
1622 }
1623
1624 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1625 {
1626         struct fs_context *fc;
1627         struct vfsmount *mnt;
1628
1629         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1630         if (IS_ERR(fc)) {
1631                 mnt = ERR_CAST(fc);
1632         } else {
1633                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1634                 ctx->hstate = h;
1635                 mnt = fc_mount(fc);
1636                 put_fs_context(fc);
1637         }
1638         if (IS_ERR(mnt))
1639                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1640                        huge_page_size(h) / SZ_1K);
1641         return mnt;
1642 }
1643
1644 static int __init init_hugetlbfs_fs(void)
1645 {
1646         struct vfsmount *mnt;
1647         struct hstate *h;
1648         int error;
1649         int i;
1650
1651         if (!hugepages_supported()) {
1652                 pr_info("disabling because there are no supported hugepage sizes\n");
1653                 return -ENOTSUPP;
1654         }
1655
1656         error = -ENOMEM;
1657         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1658                                         sizeof(struct hugetlbfs_inode_info),
1659                                         0, SLAB_ACCOUNT, init_once);
1660         if (hugetlbfs_inode_cachep == NULL)
1661                 goto out;
1662
1663         error = register_filesystem(&hugetlbfs_fs_type);
1664         if (error)
1665                 goto out_free;
1666
1667         /* default hstate mount is required */
1668         mnt = mount_one_hugetlbfs(&default_hstate);
1669         if (IS_ERR(mnt)) {
1670                 error = PTR_ERR(mnt);
1671                 goto out_unreg;
1672         }
1673         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1674
1675         /* other hstates are optional */
1676         i = 0;
1677         for_each_hstate(h) {
1678                 if (i == default_hstate_idx) {
1679                         i++;
1680                         continue;
1681                 }
1682
1683                 mnt = mount_one_hugetlbfs(h);
1684                 if (IS_ERR(mnt))
1685                         hugetlbfs_vfsmount[i] = NULL;
1686                 else
1687                         hugetlbfs_vfsmount[i] = mnt;
1688                 i++;
1689         }
1690
1691         return 0;
1692
1693  out_unreg:
1694         (void)unregister_filesystem(&hugetlbfs_fs_type);
1695  out_free:
1696         kmem_cache_destroy(hugetlbfs_inode_cachep);
1697  out:
1698         return error;
1699 }
1700 fs_initcall(init_hugetlbfs_fs)