Merge branch 'next' into for-linus
[linux-2.6-microblaze.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/sched/signal.h>         /* remove ASAP */
15 #include <linux/falloc.h>
16 #include <linux/fs.h>
17 #include <linux/mount.h>
18 #include <linux/file.h>
19 #include <linux/kernel.h>
20 #include <linux/writeback.h>
21 #include <linux/pagemap.h>
22 #include <linux/highmem.h>
23 #include <linux/init.h>
24 #include <linux/string.h>
25 #include <linux/capability.h>
26 #include <linux/ctype.h>
27 #include <linux/backing-dev.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagevec.h>
30 #include <linux/fs_parser.h>
31 #include <linux/mman.h>
32 #include <linux/slab.h>
33 #include <linux/dnotify.h>
34 #include <linux/statfs.h>
35 #include <linux/security.h>
36 #include <linux/magic.h>
37 #include <linux/migrate.h>
38 #include <linux/uio.h>
39
40 #include <linux/uaccess.h>
41 #include <linux/sched/mm.h>
42
43 static const struct super_operations hugetlbfs_ops;
44 static const struct address_space_operations hugetlbfs_aops;
45 const struct file_operations hugetlbfs_file_operations;
46 static const struct inode_operations hugetlbfs_dir_inode_operations;
47 static const struct inode_operations hugetlbfs_inode_operations;
48
49 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
50
51 struct hugetlbfs_fs_context {
52         struct hstate           *hstate;
53         unsigned long long      max_size_opt;
54         unsigned long long      min_size_opt;
55         long                    max_hpages;
56         long                    nr_inodes;
57         long                    min_hpages;
58         enum hugetlbfs_size_type max_val_type;
59         enum hugetlbfs_size_type min_val_type;
60         kuid_t                  uid;
61         kgid_t                  gid;
62         umode_t                 mode;
63 };
64
65 int sysctl_hugetlb_shm_group;
66
67 enum hugetlb_param {
68         Opt_gid,
69         Opt_min_size,
70         Opt_mode,
71         Opt_nr_inodes,
72         Opt_pagesize,
73         Opt_size,
74         Opt_uid,
75 };
76
77 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
78         fsparam_u32   ("gid",           Opt_gid),
79         fsparam_string("min_size",      Opt_min_size),
80         fsparam_u32   ("mode",          Opt_mode),
81         fsparam_string("nr_inodes",     Opt_nr_inodes),
82         fsparam_string("pagesize",      Opt_pagesize),
83         fsparam_string("size",          Opt_size),
84         fsparam_u32   ("uid",           Opt_uid),
85         {}
86 };
87
88 #ifdef CONFIG_NUMA
89 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
90                                         struct inode *inode, pgoff_t index)
91 {
92         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
93                                                         index);
94 }
95
96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
97 {
98         mpol_cond_put(vma->vm_policy);
99 }
100 #else
101 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
102                                         struct inode *inode, pgoff_t index)
103 {
104 }
105
106 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
107 {
108 }
109 #endif
110
111 static void huge_pagevec_release(struct pagevec *pvec)
112 {
113         int i;
114
115         for (i = 0; i < pagevec_count(pvec); ++i)
116                 put_page(pvec->pages[i]);
117
118         pagevec_reinit(pvec);
119 }
120
121 /*
122  * Mask used when checking the page offset value passed in via system
123  * calls.  This value will be converted to a loff_t which is signed.
124  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
125  * value.  The extra bit (- 1 in the shift value) is to take the sign
126  * bit into account.
127  */
128 #define PGOFF_LOFFT_MAX \
129         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
130
131 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
132 {
133         struct inode *inode = file_inode(file);
134         loff_t len, vma_len;
135         int ret;
136         struct hstate *h = hstate_file(file);
137
138         /*
139          * vma address alignment (but not the pgoff alignment) has
140          * already been checked by prepare_hugepage_range.  If you add
141          * any error returns here, do so after setting VM_HUGETLB, so
142          * is_vm_hugetlb_page tests below unmap_region go the right
143          * way when do_mmap unwinds (may be important on powerpc
144          * and ia64).
145          */
146         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
147         vma->vm_ops = &hugetlb_vm_ops;
148
149         /*
150          * page based offset in vm_pgoff could be sufficiently large to
151          * overflow a loff_t when converted to byte offset.  This can
152          * only happen on architectures where sizeof(loff_t) ==
153          * sizeof(unsigned long).  So, only check in those instances.
154          */
155         if (sizeof(unsigned long) == sizeof(loff_t)) {
156                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
157                         return -EINVAL;
158         }
159
160         /* must be huge page aligned */
161         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
162                 return -EINVAL;
163
164         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
165         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
166         /* check for overflow */
167         if (len < vma_len)
168                 return -EINVAL;
169
170         inode_lock(inode);
171         file_accessed(file);
172
173         ret = -ENOMEM;
174         if (hugetlb_reserve_pages(inode,
175                                 vma->vm_pgoff >> huge_page_order(h),
176                                 len >> huge_page_shift(h), vma,
177                                 vma->vm_flags))
178                 goto out;
179
180         ret = 0;
181         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
182                 i_size_write(inode, len);
183 out:
184         inode_unlock(inode);
185
186         return ret;
187 }
188
189 /*
190  * Called under mmap_write_lock(mm).
191  */
192
193 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
194 static unsigned long
195 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
196                 unsigned long len, unsigned long pgoff, unsigned long flags)
197 {
198         struct hstate *h = hstate_file(file);
199         struct vm_unmapped_area_info info;
200
201         info.flags = 0;
202         info.length = len;
203         info.low_limit = current->mm->mmap_base;
204         info.high_limit = TASK_SIZE;
205         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
206         info.align_offset = 0;
207         return vm_unmapped_area(&info);
208 }
209
210 static unsigned long
211 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
212                 unsigned long len, unsigned long pgoff, unsigned long flags)
213 {
214         struct hstate *h = hstate_file(file);
215         struct vm_unmapped_area_info info;
216
217         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
218         info.length = len;
219         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
220         info.high_limit = current->mm->mmap_base;
221         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
222         info.align_offset = 0;
223         addr = vm_unmapped_area(&info);
224
225         /*
226          * A failed mmap() very likely causes application failure,
227          * so fall back to the bottom-up function here. This scenario
228          * can happen with large stack limits and large mmap()
229          * allocations.
230          */
231         if (unlikely(offset_in_page(addr))) {
232                 VM_BUG_ON(addr != -ENOMEM);
233                 info.flags = 0;
234                 info.low_limit = current->mm->mmap_base;
235                 info.high_limit = TASK_SIZE;
236                 addr = vm_unmapped_area(&info);
237         }
238
239         return addr;
240 }
241
242 static unsigned long
243 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
244                 unsigned long len, unsigned long pgoff, unsigned long flags)
245 {
246         struct mm_struct *mm = current->mm;
247         struct vm_area_struct *vma;
248         struct hstate *h = hstate_file(file);
249
250         if (len & ~huge_page_mask(h))
251                 return -EINVAL;
252         if (len > TASK_SIZE)
253                 return -ENOMEM;
254
255         if (flags & MAP_FIXED) {
256                 if (prepare_hugepage_range(file, addr, len))
257                         return -EINVAL;
258                 return addr;
259         }
260
261         if (addr) {
262                 addr = ALIGN(addr, huge_page_size(h));
263                 vma = find_vma(mm, addr);
264                 if (TASK_SIZE - len >= addr &&
265                     (!vma || addr + len <= vm_start_gap(vma)))
266                         return addr;
267         }
268
269         /*
270          * Use mm->get_unmapped_area value as a hint to use topdown routine.
271          * If architectures have special needs, they should define their own
272          * version of hugetlb_get_unmapped_area.
273          */
274         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
275                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
276                                 pgoff, flags);
277         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
278                         pgoff, flags);
279 }
280 #endif
281
282 static size_t
283 hugetlbfs_read_actor(struct page *page, unsigned long offset,
284                         struct iov_iter *to, unsigned long size)
285 {
286         size_t copied = 0;
287         int i, chunksize;
288
289         /* Find which 4k chunk and offset with in that chunk */
290         i = offset >> PAGE_SHIFT;
291         offset = offset & ~PAGE_MASK;
292
293         while (size) {
294                 size_t n;
295                 chunksize = PAGE_SIZE;
296                 if (offset)
297                         chunksize -= offset;
298                 if (chunksize > size)
299                         chunksize = size;
300                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
301                 copied += n;
302                 if (n != chunksize)
303                         return copied;
304                 offset = 0;
305                 size -= chunksize;
306                 i++;
307         }
308         return copied;
309 }
310
311 /*
312  * Support for read() - Find the page attached to f_mapping and copy out the
313  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
314  * since it has PAGE_SIZE assumptions.
315  */
316 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
317 {
318         struct file *file = iocb->ki_filp;
319         struct hstate *h = hstate_file(file);
320         struct address_space *mapping = file->f_mapping;
321         struct inode *inode = mapping->host;
322         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
323         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
324         unsigned long end_index;
325         loff_t isize;
326         ssize_t retval = 0;
327
328         while (iov_iter_count(to)) {
329                 struct page *page;
330                 size_t nr, copied;
331
332                 /* nr is the maximum number of bytes to copy from this page */
333                 nr = huge_page_size(h);
334                 isize = i_size_read(inode);
335                 if (!isize)
336                         break;
337                 end_index = (isize - 1) >> huge_page_shift(h);
338                 if (index > end_index)
339                         break;
340                 if (index == end_index) {
341                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
342                         if (nr <= offset)
343                                 break;
344                 }
345                 nr = nr - offset;
346
347                 /* Find the page */
348                 page = find_lock_page(mapping, index);
349                 if (unlikely(page == NULL)) {
350                         /*
351                          * We have a HOLE, zero out the user-buffer for the
352                          * length of the hole or request.
353                          */
354                         copied = iov_iter_zero(nr, to);
355                 } else {
356                         unlock_page(page);
357
358                         /*
359                          * We have the page, copy it to user space buffer.
360                          */
361                         copied = hugetlbfs_read_actor(page, offset, to, nr);
362                         put_page(page);
363                 }
364                 offset += copied;
365                 retval += copied;
366                 if (copied != nr && iov_iter_count(to)) {
367                         if (!retval)
368                                 retval = -EFAULT;
369                         break;
370                 }
371                 index += offset >> huge_page_shift(h);
372                 offset &= ~huge_page_mask(h);
373         }
374         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
375         return retval;
376 }
377
378 static int hugetlbfs_write_begin(struct file *file,
379                         struct address_space *mapping,
380                         loff_t pos, unsigned len, unsigned flags,
381                         struct page **pagep, void **fsdata)
382 {
383         return -EINVAL;
384 }
385
386 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
387                         loff_t pos, unsigned len, unsigned copied,
388                         struct page *page, void *fsdata)
389 {
390         BUG();
391         return -EINVAL;
392 }
393
394 static void remove_huge_page(struct page *page)
395 {
396         ClearPageDirty(page);
397         ClearPageUptodate(page);
398         delete_from_page_cache(page);
399 }
400
401 static void
402 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
403 {
404         struct vm_area_struct *vma;
405
406         /*
407          * end == 0 indicates that the entire range after
408          * start should be unmapped.
409          */
410         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
411                 unsigned long v_offset;
412                 unsigned long v_end;
413
414                 /*
415                  * Can the expression below overflow on 32-bit arches?
416                  * No, because the interval tree returns us only those vmas
417                  * which overlap the truncated area starting at pgoff,
418                  * and no vma on a 32-bit arch can span beyond the 4GB.
419                  */
420                 if (vma->vm_pgoff < start)
421                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
422                 else
423                         v_offset = 0;
424
425                 if (!end)
426                         v_end = vma->vm_end;
427                 else {
428                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
429                                                         + vma->vm_start;
430                         if (v_end > vma->vm_end)
431                                 v_end = vma->vm_end;
432                 }
433
434                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
435                                                                         NULL);
436         }
437 }
438
439 /*
440  * remove_inode_hugepages handles two distinct cases: truncation and hole
441  * punch.  There are subtle differences in operation for each case.
442  *
443  * truncation is indicated by end of range being LLONG_MAX
444  *      In this case, we first scan the range and release found pages.
445  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
446  *      maps and global counts.  Page faults can not race with truncation
447  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
448  *      page faults in the truncated range by checking i_size.  i_size is
449  *      modified while holding i_mmap_rwsem.
450  * hole punch is indicated if end is not LLONG_MAX
451  *      In the hole punch case we scan the range and release found pages.
452  *      Only when releasing a page is the associated region/reserv map
453  *      deleted.  The region/reserv map for ranges without associated
454  *      pages are not modified.  Page faults can race with hole punch.
455  *      This is indicated if we find a mapped page.
456  * Note: If the passed end of range value is beyond the end of file, but
457  * not LLONG_MAX this routine still performs a hole punch operation.
458  */
459 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
460                                    loff_t lend)
461 {
462         struct hstate *h = hstate_inode(inode);
463         struct address_space *mapping = &inode->i_data;
464         const pgoff_t start = lstart >> huge_page_shift(h);
465         const pgoff_t end = lend >> huge_page_shift(h);
466         struct vm_area_struct pseudo_vma;
467         struct pagevec pvec;
468         pgoff_t next, index;
469         int i, freed = 0;
470         bool truncate_op = (lend == LLONG_MAX);
471
472         vma_init(&pseudo_vma, current->mm);
473         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
474         pagevec_init(&pvec);
475         next = start;
476         while (next < end) {
477                 /*
478                  * When no more pages are found, we are done.
479                  */
480                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
481                         break;
482
483                 for (i = 0; i < pagevec_count(&pvec); ++i) {
484                         struct page *page = pvec.pages[i];
485                         u32 hash;
486
487                         index = page->index;
488                         hash = hugetlb_fault_mutex_hash(mapping, index);
489                         if (!truncate_op) {
490                                 /*
491                                  * Only need to hold the fault mutex in the
492                                  * hole punch case.  This prevents races with
493                                  * page faults.  Races are not possible in the
494                                  * case of truncation.
495                                  */
496                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
497                         }
498
499                         /*
500                          * If page is mapped, it was faulted in after being
501                          * unmapped in caller.  Unmap (again) now after taking
502                          * the fault mutex.  The mutex will prevent faults
503                          * until we finish removing the page.
504                          *
505                          * This race can only happen in the hole punch case.
506                          * Getting here in a truncate operation is a bug.
507                          */
508                         if (unlikely(page_mapped(page))) {
509                                 BUG_ON(truncate_op);
510
511                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
512                                 i_mmap_lock_write(mapping);
513                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
514                                 hugetlb_vmdelete_list(&mapping->i_mmap,
515                                         index * pages_per_huge_page(h),
516                                         (index + 1) * pages_per_huge_page(h));
517                                 i_mmap_unlock_write(mapping);
518                         }
519
520                         lock_page(page);
521                         /*
522                          * We must free the huge page and remove from page
523                          * cache (remove_huge_page) BEFORE removing the
524                          * region/reserve map (hugetlb_unreserve_pages).  In
525                          * rare out of memory conditions, removal of the
526                          * region/reserve map could fail. Correspondingly,
527                          * the subpool and global reserve usage count can need
528                          * to be adjusted.
529                          */
530                         VM_BUG_ON(PagePrivate(page));
531                         remove_huge_page(page);
532                         freed++;
533                         if (!truncate_op) {
534                                 if (unlikely(hugetlb_unreserve_pages(inode,
535                                                         index, index + 1, 1)))
536                                         hugetlb_fix_reserve_counts(inode);
537                         }
538
539                         unlock_page(page);
540                         if (!truncate_op)
541                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
542                 }
543                 huge_pagevec_release(&pvec);
544                 cond_resched();
545         }
546
547         if (truncate_op)
548                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
549 }
550
551 static void hugetlbfs_evict_inode(struct inode *inode)
552 {
553         struct resv_map *resv_map;
554
555         remove_inode_hugepages(inode, 0, LLONG_MAX);
556
557         /*
558          * Get the resv_map from the address space embedded in the inode.
559          * This is the address space which points to any resv_map allocated
560          * at inode creation time.  If this is a device special inode,
561          * i_mapping may not point to the original address space.
562          */
563         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
564         /* Only regular and link inodes have associated reserve maps */
565         if (resv_map)
566                 resv_map_release(&resv_map->refs);
567         clear_inode(inode);
568 }
569
570 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
571 {
572         pgoff_t pgoff;
573         struct address_space *mapping = inode->i_mapping;
574         struct hstate *h = hstate_inode(inode);
575
576         BUG_ON(offset & ~huge_page_mask(h));
577         pgoff = offset >> PAGE_SHIFT;
578
579         i_mmap_lock_write(mapping);
580         i_size_write(inode, offset);
581         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
582                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
583         i_mmap_unlock_write(mapping);
584         remove_inode_hugepages(inode, offset, LLONG_MAX);
585         return 0;
586 }
587
588 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
589 {
590         struct hstate *h = hstate_inode(inode);
591         loff_t hpage_size = huge_page_size(h);
592         loff_t hole_start, hole_end;
593
594         /*
595          * For hole punch round up the beginning offset of the hole and
596          * round down the end.
597          */
598         hole_start = round_up(offset, hpage_size);
599         hole_end = round_down(offset + len, hpage_size);
600
601         if (hole_end > hole_start) {
602                 struct address_space *mapping = inode->i_mapping;
603                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
604
605                 inode_lock(inode);
606
607                 /* protected by i_mutex */
608                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
609                         inode_unlock(inode);
610                         return -EPERM;
611                 }
612
613                 i_mmap_lock_write(mapping);
614                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
615                         hugetlb_vmdelete_list(&mapping->i_mmap,
616                                                 hole_start >> PAGE_SHIFT,
617                                                 hole_end  >> PAGE_SHIFT);
618                 i_mmap_unlock_write(mapping);
619                 remove_inode_hugepages(inode, hole_start, hole_end);
620                 inode_unlock(inode);
621         }
622
623         return 0;
624 }
625
626 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
627                                 loff_t len)
628 {
629         struct inode *inode = file_inode(file);
630         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
631         struct address_space *mapping = inode->i_mapping;
632         struct hstate *h = hstate_inode(inode);
633         struct vm_area_struct pseudo_vma;
634         struct mm_struct *mm = current->mm;
635         loff_t hpage_size = huge_page_size(h);
636         unsigned long hpage_shift = huge_page_shift(h);
637         pgoff_t start, index, end;
638         int error;
639         u32 hash;
640
641         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
642                 return -EOPNOTSUPP;
643
644         if (mode & FALLOC_FL_PUNCH_HOLE)
645                 return hugetlbfs_punch_hole(inode, offset, len);
646
647         /*
648          * Default preallocate case.
649          * For this range, start is rounded down and end is rounded up
650          * as well as being converted to page offsets.
651          */
652         start = offset >> hpage_shift;
653         end = (offset + len + hpage_size - 1) >> hpage_shift;
654
655         inode_lock(inode);
656
657         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
658         error = inode_newsize_ok(inode, offset + len);
659         if (error)
660                 goto out;
661
662         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
663                 error = -EPERM;
664                 goto out;
665         }
666
667         /*
668          * Initialize a pseudo vma as this is required by the huge page
669          * allocation routines.  If NUMA is configured, use page index
670          * as input to create an allocation policy.
671          */
672         vma_init(&pseudo_vma, mm);
673         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
674         pseudo_vma.vm_file = file;
675
676         for (index = start; index < end; index++) {
677                 /*
678                  * This is supposed to be the vaddr where the page is being
679                  * faulted in, but we have no vaddr here.
680                  */
681                 struct page *page;
682                 unsigned long addr;
683                 int avoid_reserve = 0;
684
685                 cond_resched();
686
687                 /*
688                  * fallocate(2) manpage permits EINTR; we may have been
689                  * interrupted because we are using up too much memory.
690                  */
691                 if (signal_pending(current)) {
692                         error = -EINTR;
693                         break;
694                 }
695
696                 /* Set numa allocation policy based on index */
697                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
698
699                 /* addr is the offset within the file (zero based) */
700                 addr = index * hpage_size;
701
702                 /*
703                  * fault mutex taken here, protects against fault path
704                  * and hole punch.  inode_lock previously taken protects
705                  * against truncation.
706                  */
707                 hash = hugetlb_fault_mutex_hash(mapping, index);
708                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
709
710                 /* See if already present in mapping to avoid alloc/free */
711                 page = find_get_page(mapping, index);
712                 if (page) {
713                         put_page(page);
714                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
715                         hugetlb_drop_vma_policy(&pseudo_vma);
716                         continue;
717                 }
718
719                 /* Allocate page and add to page cache */
720                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
721                 hugetlb_drop_vma_policy(&pseudo_vma);
722                 if (IS_ERR(page)) {
723                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
724                         error = PTR_ERR(page);
725                         goto out;
726                 }
727                 clear_huge_page(page, addr, pages_per_huge_page(h));
728                 __SetPageUptodate(page);
729                 error = huge_add_to_page_cache(page, mapping, index);
730                 if (unlikely(error)) {
731                         put_page(page);
732                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
733                         goto out;
734                 }
735
736                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
737
738                 /*
739                  * unlock_page because locked by add_to_page_cache()
740                  * page_put due to reference from alloc_huge_page()
741                  */
742                 unlock_page(page);
743                 put_page(page);
744         }
745
746         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
747                 i_size_write(inode, offset + len);
748         inode->i_ctime = current_time(inode);
749 out:
750         inode_unlock(inode);
751         return error;
752 }
753
754 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
755 {
756         struct inode *inode = d_inode(dentry);
757         struct hstate *h = hstate_inode(inode);
758         int error;
759         unsigned int ia_valid = attr->ia_valid;
760         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
761
762         BUG_ON(!inode);
763
764         error = setattr_prepare(dentry, attr);
765         if (error)
766                 return error;
767
768         if (ia_valid & ATTR_SIZE) {
769                 loff_t oldsize = inode->i_size;
770                 loff_t newsize = attr->ia_size;
771
772                 if (newsize & ~huge_page_mask(h))
773                         return -EINVAL;
774                 /* protected by i_mutex */
775                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
776                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
777                         return -EPERM;
778                 error = hugetlb_vmtruncate(inode, newsize);
779                 if (error)
780                         return error;
781         }
782
783         setattr_copy(inode, attr);
784         mark_inode_dirty(inode);
785         return 0;
786 }
787
788 static struct inode *hugetlbfs_get_root(struct super_block *sb,
789                                         struct hugetlbfs_fs_context *ctx)
790 {
791         struct inode *inode;
792
793         inode = new_inode(sb);
794         if (inode) {
795                 inode->i_ino = get_next_ino();
796                 inode->i_mode = S_IFDIR | ctx->mode;
797                 inode->i_uid = ctx->uid;
798                 inode->i_gid = ctx->gid;
799                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
800                 inode->i_op = &hugetlbfs_dir_inode_operations;
801                 inode->i_fop = &simple_dir_operations;
802                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
803                 inc_nlink(inode);
804                 lockdep_annotate_inode_mutex_key(inode);
805         }
806         return inode;
807 }
808
809 /*
810  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
811  * be taken from reclaim -- unlike regular filesystems. This needs an
812  * annotation because huge_pmd_share() does an allocation under hugetlb's
813  * i_mmap_rwsem.
814  */
815 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
816
817 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
818                                         struct inode *dir,
819                                         umode_t mode, dev_t dev)
820 {
821         struct inode *inode;
822         struct resv_map *resv_map = NULL;
823
824         /*
825          * Reserve maps are only needed for inodes that can have associated
826          * page allocations.
827          */
828         if (S_ISREG(mode) || S_ISLNK(mode)) {
829                 resv_map = resv_map_alloc();
830                 if (!resv_map)
831                         return NULL;
832         }
833
834         inode = new_inode(sb);
835         if (inode) {
836                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
837
838                 inode->i_ino = get_next_ino();
839                 inode_init_owner(inode, dir, mode);
840                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
841                                 &hugetlbfs_i_mmap_rwsem_key);
842                 inode->i_mapping->a_ops = &hugetlbfs_aops;
843                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
844                 inode->i_mapping->private_data = resv_map;
845                 info->seals = F_SEAL_SEAL;
846                 switch (mode & S_IFMT) {
847                 default:
848                         init_special_inode(inode, mode, dev);
849                         break;
850                 case S_IFREG:
851                         inode->i_op = &hugetlbfs_inode_operations;
852                         inode->i_fop = &hugetlbfs_file_operations;
853                         break;
854                 case S_IFDIR:
855                         inode->i_op = &hugetlbfs_dir_inode_operations;
856                         inode->i_fop = &simple_dir_operations;
857
858                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
859                         inc_nlink(inode);
860                         break;
861                 case S_IFLNK:
862                         inode->i_op = &page_symlink_inode_operations;
863                         inode_nohighmem(inode);
864                         break;
865                 }
866                 lockdep_annotate_inode_mutex_key(inode);
867         } else {
868                 if (resv_map)
869                         kref_put(&resv_map->refs, resv_map_release);
870         }
871
872         return inode;
873 }
874
875 /*
876  * File creation. Allocate an inode, and we're done..
877  */
878 static int do_hugetlbfs_mknod(struct inode *dir,
879                         struct dentry *dentry,
880                         umode_t mode,
881                         dev_t dev,
882                         bool tmpfile)
883 {
884         struct inode *inode;
885         int error = -ENOSPC;
886
887         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
888         if (inode) {
889                 dir->i_ctime = dir->i_mtime = current_time(dir);
890                 if (tmpfile) {
891                         d_tmpfile(dentry, inode);
892                 } else {
893                         d_instantiate(dentry, inode);
894                         dget(dentry);/* Extra count - pin the dentry in core */
895                 }
896                 error = 0;
897         }
898         return error;
899 }
900
901 static int hugetlbfs_mknod(struct inode *dir,
902                         struct dentry *dentry, umode_t mode, dev_t dev)
903 {
904         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
905 }
906
907 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
908 {
909         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
910         if (!retval)
911                 inc_nlink(dir);
912         return retval;
913 }
914
915 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
916 {
917         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
918 }
919
920 static int hugetlbfs_tmpfile(struct inode *dir,
921                         struct dentry *dentry, umode_t mode)
922 {
923         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
924 }
925
926 static int hugetlbfs_symlink(struct inode *dir,
927                         struct dentry *dentry, const char *symname)
928 {
929         struct inode *inode;
930         int error = -ENOSPC;
931
932         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
933         if (inode) {
934                 int l = strlen(symname)+1;
935                 error = page_symlink(inode, symname, l);
936                 if (!error) {
937                         d_instantiate(dentry, inode);
938                         dget(dentry);
939                 } else
940                         iput(inode);
941         }
942         dir->i_ctime = dir->i_mtime = current_time(dir);
943
944         return error;
945 }
946
947 /*
948  * mark the head page dirty
949  */
950 static int hugetlbfs_set_page_dirty(struct page *page)
951 {
952         struct page *head = compound_head(page);
953
954         SetPageDirty(head);
955         return 0;
956 }
957
958 static int hugetlbfs_migrate_page(struct address_space *mapping,
959                                 struct page *newpage, struct page *page,
960                                 enum migrate_mode mode)
961 {
962         int rc;
963
964         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
965         if (rc != MIGRATEPAGE_SUCCESS)
966                 return rc;
967
968         /*
969          * page_private is subpool pointer in hugetlb pages.  Transfer to
970          * new page.  PagePrivate is not associated with page_private for
971          * hugetlb pages and can not be set here as only page_huge_active
972          * pages can be migrated.
973          */
974         if (page_private(page)) {
975                 set_page_private(newpage, page_private(page));
976                 set_page_private(page, 0);
977         }
978
979         if (mode != MIGRATE_SYNC_NO_COPY)
980                 migrate_page_copy(newpage, page);
981         else
982                 migrate_page_states(newpage, page);
983
984         return MIGRATEPAGE_SUCCESS;
985 }
986
987 static int hugetlbfs_error_remove_page(struct address_space *mapping,
988                                 struct page *page)
989 {
990         struct inode *inode = mapping->host;
991         pgoff_t index = page->index;
992
993         remove_huge_page(page);
994         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
995                 hugetlb_fix_reserve_counts(inode);
996
997         return 0;
998 }
999
1000 /*
1001  * Display the mount options in /proc/mounts.
1002  */
1003 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1004 {
1005         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1006         struct hugepage_subpool *spool = sbinfo->spool;
1007         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1008         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1009         char mod;
1010
1011         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1012                 seq_printf(m, ",uid=%u",
1013                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1014         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1015                 seq_printf(m, ",gid=%u",
1016                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1017         if (sbinfo->mode != 0755)
1018                 seq_printf(m, ",mode=%o", sbinfo->mode);
1019         if (sbinfo->max_inodes != -1)
1020                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1021
1022         hpage_size /= 1024;
1023         mod = 'K';
1024         if (hpage_size >= 1024) {
1025                 hpage_size /= 1024;
1026                 mod = 'M';
1027         }
1028         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1029         if (spool) {
1030                 if (spool->max_hpages != -1)
1031                         seq_printf(m, ",size=%llu",
1032                                    (unsigned long long)spool->max_hpages << hpage_shift);
1033                 if (spool->min_hpages != -1)
1034                         seq_printf(m, ",min_size=%llu",
1035                                    (unsigned long long)spool->min_hpages << hpage_shift);
1036         }
1037         return 0;
1038 }
1039
1040 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1041 {
1042         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1043         struct hstate *h = hstate_inode(d_inode(dentry));
1044
1045         buf->f_type = HUGETLBFS_MAGIC;
1046         buf->f_bsize = huge_page_size(h);
1047         if (sbinfo) {
1048                 spin_lock(&sbinfo->stat_lock);
1049                 /* If no limits set, just report 0 for max/free/used
1050                  * blocks, like simple_statfs() */
1051                 if (sbinfo->spool) {
1052                         long free_pages;
1053
1054                         spin_lock(&sbinfo->spool->lock);
1055                         buf->f_blocks = sbinfo->spool->max_hpages;
1056                         free_pages = sbinfo->spool->max_hpages
1057                                 - sbinfo->spool->used_hpages;
1058                         buf->f_bavail = buf->f_bfree = free_pages;
1059                         spin_unlock(&sbinfo->spool->lock);
1060                         buf->f_files = sbinfo->max_inodes;
1061                         buf->f_ffree = sbinfo->free_inodes;
1062                 }
1063                 spin_unlock(&sbinfo->stat_lock);
1064         }
1065         buf->f_namelen = NAME_MAX;
1066         return 0;
1067 }
1068
1069 static void hugetlbfs_put_super(struct super_block *sb)
1070 {
1071         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1072
1073         if (sbi) {
1074                 sb->s_fs_info = NULL;
1075
1076                 if (sbi->spool)
1077                         hugepage_put_subpool(sbi->spool);
1078
1079                 kfree(sbi);
1080         }
1081 }
1082
1083 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1084 {
1085         if (sbinfo->free_inodes >= 0) {
1086                 spin_lock(&sbinfo->stat_lock);
1087                 if (unlikely(!sbinfo->free_inodes)) {
1088                         spin_unlock(&sbinfo->stat_lock);
1089                         return 0;
1090                 }
1091                 sbinfo->free_inodes--;
1092                 spin_unlock(&sbinfo->stat_lock);
1093         }
1094
1095         return 1;
1096 }
1097
1098 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1099 {
1100         if (sbinfo->free_inodes >= 0) {
1101                 spin_lock(&sbinfo->stat_lock);
1102                 sbinfo->free_inodes++;
1103                 spin_unlock(&sbinfo->stat_lock);
1104         }
1105 }
1106
1107
1108 static struct kmem_cache *hugetlbfs_inode_cachep;
1109
1110 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1111 {
1112         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1113         struct hugetlbfs_inode_info *p;
1114
1115         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1116                 return NULL;
1117         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1118         if (unlikely(!p)) {
1119                 hugetlbfs_inc_free_inodes(sbinfo);
1120                 return NULL;
1121         }
1122
1123         /*
1124          * Any time after allocation, hugetlbfs_destroy_inode can be called
1125          * for the inode.  mpol_free_shared_policy is unconditionally called
1126          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1127          * in case of a quick call to destroy.
1128          *
1129          * Note that the policy is initialized even if we are creating a
1130          * private inode.  This simplifies hugetlbfs_destroy_inode.
1131          */
1132         mpol_shared_policy_init(&p->policy, NULL);
1133
1134         return &p->vfs_inode;
1135 }
1136
1137 static void hugetlbfs_free_inode(struct inode *inode)
1138 {
1139         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1140 }
1141
1142 static void hugetlbfs_destroy_inode(struct inode *inode)
1143 {
1144         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1145         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1146 }
1147
1148 static const struct address_space_operations hugetlbfs_aops = {
1149         .write_begin    = hugetlbfs_write_begin,
1150         .write_end      = hugetlbfs_write_end,
1151         .set_page_dirty = hugetlbfs_set_page_dirty,
1152         .migratepage    = hugetlbfs_migrate_page,
1153         .error_remove_page      = hugetlbfs_error_remove_page,
1154 };
1155
1156
1157 static void init_once(void *foo)
1158 {
1159         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1160
1161         inode_init_once(&ei->vfs_inode);
1162 }
1163
1164 const struct file_operations hugetlbfs_file_operations = {
1165         .read_iter              = hugetlbfs_read_iter,
1166         .mmap                   = hugetlbfs_file_mmap,
1167         .fsync                  = noop_fsync,
1168         .get_unmapped_area      = hugetlb_get_unmapped_area,
1169         .llseek                 = default_llseek,
1170         .fallocate              = hugetlbfs_fallocate,
1171 };
1172
1173 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1174         .create         = hugetlbfs_create,
1175         .lookup         = simple_lookup,
1176         .link           = simple_link,
1177         .unlink         = simple_unlink,
1178         .symlink        = hugetlbfs_symlink,
1179         .mkdir          = hugetlbfs_mkdir,
1180         .rmdir          = simple_rmdir,
1181         .mknod          = hugetlbfs_mknod,
1182         .rename         = simple_rename,
1183         .setattr        = hugetlbfs_setattr,
1184         .tmpfile        = hugetlbfs_tmpfile,
1185 };
1186
1187 static const struct inode_operations hugetlbfs_inode_operations = {
1188         .setattr        = hugetlbfs_setattr,
1189 };
1190
1191 static const struct super_operations hugetlbfs_ops = {
1192         .alloc_inode    = hugetlbfs_alloc_inode,
1193         .free_inode     = hugetlbfs_free_inode,
1194         .destroy_inode  = hugetlbfs_destroy_inode,
1195         .evict_inode    = hugetlbfs_evict_inode,
1196         .statfs         = hugetlbfs_statfs,
1197         .put_super      = hugetlbfs_put_super,
1198         .show_options   = hugetlbfs_show_options,
1199 };
1200
1201 /*
1202  * Convert size option passed from command line to number of huge pages
1203  * in the pool specified by hstate.  Size option could be in bytes
1204  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1205  */
1206 static long
1207 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1208                          enum hugetlbfs_size_type val_type)
1209 {
1210         if (val_type == NO_SIZE)
1211                 return -1;
1212
1213         if (val_type == SIZE_PERCENT) {
1214                 size_opt <<= huge_page_shift(h);
1215                 size_opt *= h->max_huge_pages;
1216                 do_div(size_opt, 100);
1217         }
1218
1219         size_opt >>= huge_page_shift(h);
1220         return size_opt;
1221 }
1222
1223 /*
1224  * Parse one mount parameter.
1225  */
1226 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1227 {
1228         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1229         struct fs_parse_result result;
1230         char *rest;
1231         unsigned long ps;
1232         int opt;
1233
1234         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1235         if (opt < 0)
1236                 return opt;
1237
1238         switch (opt) {
1239         case Opt_uid:
1240                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1241                 if (!uid_valid(ctx->uid))
1242                         goto bad_val;
1243                 return 0;
1244
1245         case Opt_gid:
1246                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1247                 if (!gid_valid(ctx->gid))
1248                         goto bad_val;
1249                 return 0;
1250
1251         case Opt_mode:
1252                 ctx->mode = result.uint_32 & 01777U;
1253                 return 0;
1254
1255         case Opt_size:
1256                 /* memparse() will accept a K/M/G without a digit */
1257                 if (!isdigit(param->string[0]))
1258                         goto bad_val;
1259                 ctx->max_size_opt = memparse(param->string, &rest);
1260                 ctx->max_val_type = SIZE_STD;
1261                 if (*rest == '%')
1262                         ctx->max_val_type = SIZE_PERCENT;
1263                 return 0;
1264
1265         case Opt_nr_inodes:
1266                 /* memparse() will accept a K/M/G without a digit */
1267                 if (!isdigit(param->string[0]))
1268                         goto bad_val;
1269                 ctx->nr_inodes = memparse(param->string, &rest);
1270                 return 0;
1271
1272         case Opt_pagesize:
1273                 ps = memparse(param->string, &rest);
1274                 ctx->hstate = size_to_hstate(ps);
1275                 if (!ctx->hstate) {
1276                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1277                         return -EINVAL;
1278                 }
1279                 return 0;
1280
1281         case Opt_min_size:
1282                 /* memparse() will accept a K/M/G without a digit */
1283                 if (!isdigit(param->string[0]))
1284                         goto bad_val;
1285                 ctx->min_size_opt = memparse(param->string, &rest);
1286                 ctx->min_val_type = SIZE_STD;
1287                 if (*rest == '%')
1288                         ctx->min_val_type = SIZE_PERCENT;
1289                 return 0;
1290
1291         default:
1292                 return -EINVAL;
1293         }
1294
1295 bad_val:
1296         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1297                       param->string, param->key);
1298 }
1299
1300 /*
1301  * Validate the parsed options.
1302  */
1303 static int hugetlbfs_validate(struct fs_context *fc)
1304 {
1305         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1306
1307         /*
1308          * Use huge page pool size (in hstate) to convert the size
1309          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1310          */
1311         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1312                                                    ctx->max_size_opt,
1313                                                    ctx->max_val_type);
1314         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1315                                                    ctx->min_size_opt,
1316                                                    ctx->min_val_type);
1317
1318         /*
1319          * If max_size was specified, then min_size must be smaller
1320          */
1321         if (ctx->max_val_type > NO_SIZE &&
1322             ctx->min_hpages > ctx->max_hpages) {
1323                 pr_err("Minimum size can not be greater than maximum size\n");
1324                 return -EINVAL;
1325         }
1326
1327         return 0;
1328 }
1329
1330 static int
1331 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1332 {
1333         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1334         struct hugetlbfs_sb_info *sbinfo;
1335
1336         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1337         if (!sbinfo)
1338                 return -ENOMEM;
1339         sb->s_fs_info = sbinfo;
1340         spin_lock_init(&sbinfo->stat_lock);
1341         sbinfo->hstate          = ctx->hstate;
1342         sbinfo->max_inodes      = ctx->nr_inodes;
1343         sbinfo->free_inodes     = ctx->nr_inodes;
1344         sbinfo->spool           = NULL;
1345         sbinfo->uid             = ctx->uid;
1346         sbinfo->gid             = ctx->gid;
1347         sbinfo->mode            = ctx->mode;
1348
1349         /*
1350          * Allocate and initialize subpool if maximum or minimum size is
1351          * specified.  Any needed reservations (for minimim size) are taken
1352          * taken when the subpool is created.
1353          */
1354         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1355                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1356                                                      ctx->max_hpages,
1357                                                      ctx->min_hpages);
1358                 if (!sbinfo->spool)
1359                         goto out_free;
1360         }
1361         sb->s_maxbytes = MAX_LFS_FILESIZE;
1362         sb->s_blocksize = huge_page_size(ctx->hstate);
1363         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1364         sb->s_magic = HUGETLBFS_MAGIC;
1365         sb->s_op = &hugetlbfs_ops;
1366         sb->s_time_gran = 1;
1367
1368         /*
1369          * Due to the special and limited functionality of hugetlbfs, it does
1370          * not work well as a stacking filesystem.
1371          */
1372         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1373         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1374         if (!sb->s_root)
1375                 goto out_free;
1376         return 0;
1377 out_free:
1378         kfree(sbinfo->spool);
1379         kfree(sbinfo);
1380         return -ENOMEM;
1381 }
1382
1383 static int hugetlbfs_get_tree(struct fs_context *fc)
1384 {
1385         int err = hugetlbfs_validate(fc);
1386         if (err)
1387                 return err;
1388         return get_tree_nodev(fc, hugetlbfs_fill_super);
1389 }
1390
1391 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1392 {
1393         kfree(fc->fs_private);
1394 }
1395
1396 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1397         .free           = hugetlbfs_fs_context_free,
1398         .parse_param    = hugetlbfs_parse_param,
1399         .get_tree       = hugetlbfs_get_tree,
1400 };
1401
1402 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1403 {
1404         struct hugetlbfs_fs_context *ctx;
1405
1406         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1407         if (!ctx)
1408                 return -ENOMEM;
1409
1410         ctx->max_hpages = -1; /* No limit on size by default */
1411         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1412         ctx->uid        = current_fsuid();
1413         ctx->gid        = current_fsgid();
1414         ctx->mode       = 0755;
1415         ctx->hstate     = &default_hstate;
1416         ctx->min_hpages = -1; /* No default minimum size */
1417         ctx->max_val_type = NO_SIZE;
1418         ctx->min_val_type = NO_SIZE;
1419         fc->fs_private = ctx;
1420         fc->ops = &hugetlbfs_fs_context_ops;
1421         return 0;
1422 }
1423
1424 static struct file_system_type hugetlbfs_fs_type = {
1425         .name                   = "hugetlbfs",
1426         .init_fs_context        = hugetlbfs_init_fs_context,
1427         .parameters             = hugetlb_fs_parameters,
1428         .kill_sb                = kill_litter_super,
1429 };
1430
1431 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1432
1433 static int can_do_hugetlb_shm(void)
1434 {
1435         kgid_t shm_group;
1436         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1437         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1438 }
1439
1440 static int get_hstate_idx(int page_size_log)
1441 {
1442         struct hstate *h = hstate_sizelog(page_size_log);
1443
1444         if (!h)
1445                 return -1;
1446         return h - hstates;
1447 }
1448
1449 /*
1450  * Note that size should be aligned to proper hugepage size in caller side,
1451  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1452  */
1453 struct file *hugetlb_file_setup(const char *name, size_t size,
1454                                 vm_flags_t acctflag, struct user_struct **user,
1455                                 int creat_flags, int page_size_log)
1456 {
1457         struct inode *inode;
1458         struct vfsmount *mnt;
1459         int hstate_idx;
1460         struct file *file;
1461
1462         hstate_idx = get_hstate_idx(page_size_log);
1463         if (hstate_idx < 0)
1464                 return ERR_PTR(-ENODEV);
1465
1466         *user = NULL;
1467         mnt = hugetlbfs_vfsmount[hstate_idx];
1468         if (!mnt)
1469                 return ERR_PTR(-ENOENT);
1470
1471         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1472                 *user = current_user();
1473                 if (user_shm_lock(size, *user)) {
1474                         task_lock(current);
1475                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1476                                 current->comm, current->pid);
1477                         task_unlock(current);
1478                 } else {
1479                         *user = NULL;
1480                         return ERR_PTR(-EPERM);
1481                 }
1482         }
1483
1484         file = ERR_PTR(-ENOSPC);
1485         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1486         if (!inode)
1487                 goto out;
1488         if (creat_flags == HUGETLB_SHMFS_INODE)
1489                 inode->i_flags |= S_PRIVATE;
1490
1491         inode->i_size = size;
1492         clear_nlink(inode);
1493
1494         if (hugetlb_reserve_pages(inode, 0,
1495                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1496                         acctflag))
1497                 file = ERR_PTR(-ENOMEM);
1498         else
1499                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1500                                         &hugetlbfs_file_operations);
1501         if (!IS_ERR(file))
1502                 return file;
1503
1504         iput(inode);
1505 out:
1506         if (*user) {
1507                 user_shm_unlock(size, *user);
1508                 *user = NULL;
1509         }
1510         return file;
1511 }
1512
1513 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1514 {
1515         struct fs_context *fc;
1516         struct vfsmount *mnt;
1517
1518         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1519         if (IS_ERR(fc)) {
1520                 mnt = ERR_CAST(fc);
1521         } else {
1522                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1523                 ctx->hstate = h;
1524                 mnt = fc_mount(fc);
1525                 put_fs_context(fc);
1526         }
1527         if (IS_ERR(mnt))
1528                 pr_err("Cannot mount internal hugetlbfs for page size %uK",
1529                        1U << (h->order + PAGE_SHIFT - 10));
1530         return mnt;
1531 }
1532
1533 static int __init init_hugetlbfs_fs(void)
1534 {
1535         struct vfsmount *mnt;
1536         struct hstate *h;
1537         int error;
1538         int i;
1539
1540         if (!hugepages_supported()) {
1541                 pr_info("disabling because there are no supported hugepage sizes\n");
1542                 return -ENOTSUPP;
1543         }
1544
1545         error = -ENOMEM;
1546         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1547                                         sizeof(struct hugetlbfs_inode_info),
1548                                         0, SLAB_ACCOUNT, init_once);
1549         if (hugetlbfs_inode_cachep == NULL)
1550                 goto out;
1551
1552         error = register_filesystem(&hugetlbfs_fs_type);
1553         if (error)
1554                 goto out_free;
1555
1556         /* default hstate mount is required */
1557         mnt = mount_one_hugetlbfs(&hstates[default_hstate_idx]);
1558         if (IS_ERR(mnt)) {
1559                 error = PTR_ERR(mnt);
1560                 goto out_unreg;
1561         }
1562         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1563
1564         /* other hstates are optional */
1565         i = 0;
1566         for_each_hstate(h) {
1567                 if (i == default_hstate_idx) {
1568                         i++;
1569                         continue;
1570                 }
1571
1572                 mnt = mount_one_hugetlbfs(h);
1573                 if (IS_ERR(mnt))
1574                         hugetlbfs_vfsmount[i] = NULL;
1575                 else
1576                         hugetlbfs_vfsmount[i] = mnt;
1577                 i++;
1578         }
1579
1580         return 0;
1581
1582  out_unreg:
1583         (void)unregister_filesystem(&hugetlbfs_fs_type);
1584  out_free:
1585         kmem_cache_destroy(hugetlbfs_inode_cachep);
1586  out:
1587         return error;
1588 }
1589 fs_initcall(init_hugetlbfs_fs)