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 static 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_gid ("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_uid ("uid", Opt_uid),
83 {}
84 };
85
86 /*
87 * Mask used when checking the page offset value passed in via system
88 * calls. This value will be converted to a loff_t which is signed.
89 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
90 * value. The extra bit (- 1 in the shift value) is to take the sign
91 * bit into account.
92 */
93 #define PGOFF_LOFFT_MAX \
94 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
95
96 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
97 {
98 struct inode *inode = file_inode(file);
99 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
100 loff_t len, vma_len;
101 int ret;
102 struct hstate *h = hstate_file(file);
103 vm_flags_t vm_flags;
104
105 /*
106 * vma address alignment (but not the pgoff alignment) has
107 * already been checked by prepare_hugepage_range. If you add
108 * any error returns here, do so after setting VM_HUGETLB, so
109 * is_vm_hugetlb_page tests below unmap_region go the right
110 * way when do_mmap unwinds (may be important on powerpc
111 * and ia64).
112 */
113 vm_flags_set(vma, VM_HUGETLB | VM_DONTEXPAND);
114 vma->vm_ops = &hugetlb_vm_ops;
115
116 ret = seal_check_write(info->seals, vma);
117 if (ret)
118 return ret;
119
120 /*
121 * page based offset in vm_pgoff could be sufficiently large to
122 * overflow a loff_t when converted to byte offset. This can
123 * only happen on architectures where sizeof(loff_t) ==
124 * sizeof(unsigned long). So, only check in those instances.
125 */
126 if (sizeof(unsigned long) == sizeof(loff_t)) {
127 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
128 return -EINVAL;
129 }
130
131 /* must be huge page aligned */
132 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
133 return -EINVAL;
134
135 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
136 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
137 /* check for overflow */
138 if (len < vma_len)
139 return -EINVAL;
140
141 inode_lock(inode);
142 file_accessed(file);
143
144 ret = -ENOMEM;
145
146 vm_flags = vma->vm_flags;
147 /*
148 * for SHM_HUGETLB, the pages are reserved in the shmget() call so skip
149 * reserving here. Note: only for SHM hugetlbfs file, the inode
150 * flag S_PRIVATE is set.
151 */
152 if (inode->i_flags & S_PRIVATE)
153 vm_flags |= VM_NORESERVE;
154
155 if (!hugetlb_reserve_pages(inode,
156 vma->vm_pgoff >> huge_page_order(h),
157 len >> huge_page_shift(h), vma,
158 vm_flags))
159 goto out;
160
161 ret = 0;
162 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
163 i_size_write(inode, len);
164 out:
165 inode_unlock(inode);
166
167 return ret;
168 }
169
170 /*
171 * Called under mmap_write_lock(mm).
172 */
173
174 static unsigned long
175 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
176 unsigned long len, unsigned long pgoff, unsigned long flags)
177 {
178 struct hstate *h = hstate_file(file);
179 struct vm_unmapped_area_info info = {};
180
181 info.length = len;
182 info.low_limit = current->mm->mmap_base;
183 info.high_limit = arch_get_mmap_end(addr, len, flags);
184 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
185 return vm_unmapped_area(&info);
186 }
187
188 static unsigned long
189 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
190 unsigned long len, unsigned long pgoff, unsigned long flags)
191 {
192 struct hstate *h = hstate_file(file);
193 struct vm_unmapped_area_info info = {};
194
195 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
196 info.length = len;
197 info.low_limit = PAGE_SIZE;
198 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
199 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
200 addr = vm_unmapped_area(&info);
201
202 /*
203 * A failed mmap() very likely causes application failure,
204 * so fall back to the bottom-up function here. This scenario
205 * can happen with large stack limits and large mmap()
206 * allocations.
207 */
208 if (unlikely(offset_in_page(addr))) {
209 VM_BUG_ON(addr != -ENOMEM);
210 info.flags = 0;
211 info.low_limit = current->mm->mmap_base;
212 info.high_limit = arch_get_mmap_end(addr, len, flags);
213 addr = vm_unmapped_area(&info);
214 }
215
216 return addr;
217 }
218
219 unsigned long
220 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
221 unsigned long len, unsigned long pgoff,
222 unsigned long flags)
223 {
224 struct mm_struct *mm = current->mm;
225 struct vm_area_struct *vma, *prev;
226 struct hstate *h = hstate_file(file);
227 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
228
229 if (len & ~huge_page_mask(h))
230 return -EINVAL;
231 if (len > mmap_end - mmap_min_addr)
232 return -ENOMEM;
233
234 if (flags & MAP_FIXED) {
235 if (prepare_hugepage_range(file, addr, len))
236 return -EINVAL;
237 return addr;
238 }
239
240 if (addr) {
241 addr = ALIGN(addr, huge_page_size(h));
242 vma = find_vma_prev(mm, addr, &prev);
243 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
244 (!vma || addr + len <= vm_start_gap(vma)) &&
245 (!prev || addr >= vm_end_gap(prev)))
246 return addr;
247 }
248
249 /*
250 * Use MMF_TOPDOWN flag as a hint to use topdown routine.
251 * If architectures have special needs, they should define their own
252 * version of hugetlb_get_unmapped_area.
253 */
254 if (test_bit(MMF_TOPDOWN, &mm->flags))
255 return hugetlb_get_unmapped_area_topdown(file, addr, len,
256 pgoff, flags);
257 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
258 pgoff, flags);
259 }
260
261 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
262 static unsigned long
263 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
264 unsigned long len, unsigned long pgoff,
265 unsigned long flags)
266 {
267 return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
268 }
269 #endif
270
271 /*
272 * Someone wants to read @bytes from a HWPOISON hugetlb @page from @offset.
273 * Returns the maximum number of bytes one can read without touching the 1st raw
274 * HWPOISON subpage.
275 *
276 * The implementation borrows the iteration logic from copy_page_to_iter*.
277 */
278 static size_t adjust_range_hwpoison(struct page *page, size_t offset, size_t bytes)
279 {
280 size_t n = 0;
281 size_t res = 0;
282
283 /* First subpage to start the loop. */
284 page = nth_page(page, offset / PAGE_SIZE);
285 offset %= PAGE_SIZE;
286 while (1) {
287 if (is_raw_hwpoison_page_in_hugepage(page))
288 break;
289
290 /* Safe to read n bytes without touching HWPOISON subpage. */
291 n = min(bytes, (size_t)PAGE_SIZE - offset);
292 res += n;
293 bytes -= n;
294 if (!bytes || !n)
295 break;
296 offset += n;
297 if (offset == PAGE_SIZE) {
298 page = nth_page(page, 1);
299 offset = 0;
300 }
301 }
302
303 return res;
304 }
305
306 /*
307 * Support for read() - Find the page attached to f_mapping and copy out the
308 * data. This provides functionality similar to filemap_read().
309 */
310 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
311 {
312 struct file *file = iocb->ki_filp;
313 struct hstate *h = hstate_file(file);
314 struct address_space *mapping = file->f_mapping;
315 struct inode *inode = mapping->host;
316 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
317 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
318 unsigned long end_index;
319 loff_t isize;
320 ssize_t retval = 0;
321
322 while (iov_iter_count(to)) {
323 struct folio *folio;
324 size_t nr, copied, want;
325
326 /* nr is the maximum number of bytes to copy from this page */
327 nr = huge_page_size(h);
328 isize = i_size_read(inode);
329 if (!isize)
330 break;
331 end_index = (isize - 1) >> huge_page_shift(h);
332 if (index > end_index)
333 break;
334 if (index == end_index) {
335 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
336 if (nr <= offset)
337 break;
338 }
339 nr = nr - offset;
340
341 /* Find the folio */
342 folio = filemap_lock_hugetlb_folio(h, mapping, index);
343 if (IS_ERR(folio)) {
344 /*
345 * We have a HOLE, zero out the user-buffer for the
346 * length of the hole or request.
347 */
348 copied = iov_iter_zero(nr, to);
349 } else {
350 folio_unlock(folio);
351
352 if (!folio_test_hwpoison(folio))
353 want = nr;
354 else {
355 /*
356 * Adjust how many bytes safe to read without
357 * touching the 1st raw HWPOISON subpage after
358 * offset.
359 */
360 want = adjust_range_hwpoison(&folio->page, offset, nr);
361 if (want == 0) {
362 folio_put(folio);
363 retval = -EIO;
364 break;
365 }
366 }
367
368 /*
369 * We have the folio, copy it to user space buffer.
370 */
371 copied = copy_folio_to_iter(folio, offset, want, to);
372 folio_put(folio);
373 }
374 offset += copied;
375 retval += copied;
376 if (copied != nr && iov_iter_count(to)) {
377 if (!retval)
378 retval = -EFAULT;
379 break;
380 }
381 index += offset >> huge_page_shift(h);
382 offset &= ~huge_page_mask(h);
383 }
384 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
385 return retval;
386 }
387
388 static int hugetlbfs_write_begin(struct file *file,
389 struct address_space *mapping,
390 loff_t pos, unsigned len,
391 struct page **pagep, void **fsdata)
392 {
393 return -EINVAL;
394 }
395
396 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
397 loff_t pos, unsigned len, unsigned copied,
398 struct page *page, void *fsdata)
399 {
400 BUG();
401 return -EINVAL;
402 }
403
404 static void hugetlb_delete_from_page_cache(struct folio *folio)
405 {
406 folio_clear_dirty(folio);
407 folio_clear_uptodate(folio);
408 filemap_remove_folio(folio);
409 }
410
411 /*
412 * Called with i_mmap_rwsem held for inode based vma maps. This makes
413 * sure vma (and vm_mm) will not go away. We also hold the hugetlb fault
414 * mutex for the page in the mapping. So, we can not race with page being
415 * faulted into the vma.
416 */
417 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
418 unsigned long addr, struct page *page)
419 {
420 pte_t *ptep, pte;
421
422 ptep = hugetlb_walk(vma, addr, huge_page_size(hstate_vma(vma)));
423 if (!ptep)
424 return false;
425
426 pte = huge_ptep_get(vma->vm_mm, addr, ptep);
427 if (huge_pte_none(pte) || !pte_present(pte))
428 return false;
429
430 if (pte_page(pte) == page)
431 return true;
432
433 return false;
434 }
435
436 /*
437 * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
438 * No, because the interval tree returns us only those vmas
439 * which overlap the truncated area starting at pgoff,
440 * and no vma on a 32-bit arch can span beyond the 4GB.
441 */
442 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
443 {
444 unsigned long offset = 0;
445
446 if (vma->vm_pgoff < start)
447 offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
448
449 return vma->vm_start + offset;
450 }
451
452 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
453 {
454 unsigned long t_end;
455
456 if (!end)
457 return vma->vm_end;
458
459 t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
460 if (t_end > vma->vm_end)
461 t_end = vma->vm_end;
462 return t_end;
463 }
464
465 /*
466 * Called with hugetlb fault mutex held. Therefore, no more mappings to
467 * this folio can be created while executing the routine.
468 */
469 static void hugetlb_unmap_file_folio(struct hstate *h,
470 struct address_space *mapping,
471 struct folio *folio, pgoff_t index)
472 {
473 struct rb_root_cached *root = &mapping->i_mmap;
474 struct hugetlb_vma_lock *vma_lock;
475 struct page *page = &folio->page;
476 struct vm_area_struct *vma;
477 unsigned long v_start;
478 unsigned long v_end;
479 pgoff_t start, end;
480
481 start = index * pages_per_huge_page(h);
482 end = (index + 1) * pages_per_huge_page(h);
483
484 i_mmap_lock_write(mapping);
485 retry:
486 vma_lock = NULL;
487 vma_interval_tree_foreach(vma, root, start, end - 1) {
488 v_start = vma_offset_start(vma, start);
489 v_end = vma_offset_end(vma, end);
490
491 if (!hugetlb_vma_maps_page(vma, v_start, page))
492 continue;
493
494 if (!hugetlb_vma_trylock_write(vma)) {
495 vma_lock = vma->vm_private_data;
496 /*
497 * If we can not get vma lock, we need to drop
498 * immap_sema and take locks in order. First,
499 * take a ref on the vma_lock structure so that
500 * we can be guaranteed it will not go away when
501 * dropping immap_sema.
502 */
503 kref_get(&vma_lock->refs);
504 break;
505 }
506
507 unmap_hugepage_range(vma, v_start, v_end, NULL,
508 ZAP_FLAG_DROP_MARKER);
509 hugetlb_vma_unlock_write(vma);
510 }
511
512 i_mmap_unlock_write(mapping);
513
514 if (vma_lock) {
515 /*
516 * Wait on vma_lock. We know it is still valid as we have
517 * a reference. We must 'open code' vma locking as we do
518 * not know if vma_lock is still attached to vma.
519 */
520 down_write(&vma_lock->rw_sema);
521 i_mmap_lock_write(mapping);
522
523 vma = vma_lock->vma;
524 if (!vma) {
525 /*
526 * If lock is no longer attached to vma, then just
527 * unlock, drop our reference and retry looking for
528 * other vmas.
529 */
530 up_write(&vma_lock->rw_sema);
531 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
532 goto retry;
533 }
534
535 /*
536 * vma_lock is still attached to vma. Check to see if vma
537 * still maps page and if so, unmap.
538 */
539 v_start = vma_offset_start(vma, start);
540 v_end = vma_offset_end(vma, end);
541 if (hugetlb_vma_maps_page(vma, v_start, page))
542 unmap_hugepage_range(vma, v_start, v_end, NULL,
543 ZAP_FLAG_DROP_MARKER);
544
545 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
546 hugetlb_vma_unlock_write(vma);
547
548 goto retry;
549 }
550 }
551
552 static void
553 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
554 zap_flags_t zap_flags)
555 {
556 struct vm_area_struct *vma;
557
558 /*
559 * end == 0 indicates that the entire range after start should be
560 * unmapped. Note, end is exclusive, whereas the interval tree takes
561 * an inclusive "last".
562 */
563 vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
564 unsigned long v_start;
565 unsigned long v_end;
566
567 if (!hugetlb_vma_trylock_write(vma))
568 continue;
569
570 v_start = vma_offset_start(vma, start);
571 v_end = vma_offset_end(vma, end);
572
573 unmap_hugepage_range(vma, v_start, v_end, NULL, zap_flags);
574
575 /*
576 * Note that vma lock only exists for shared/non-private
577 * vmas. Therefore, lock is not held when calling
578 * unmap_hugepage_range for private vmas.
579 */
580 hugetlb_vma_unlock_write(vma);
581 }
582 }
583
584 /*
585 * Called with hugetlb fault mutex held.
586 * Returns true if page was actually removed, false otherwise.
587 */
588 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
589 struct address_space *mapping,
590 struct folio *folio, pgoff_t index,
591 bool truncate_op)
592 {
593 bool ret = false;
594
595 /*
596 * If folio is mapped, it was faulted in after being
597 * unmapped in caller. Unmap (again) while holding
598 * the fault mutex. The mutex will prevent faults
599 * until we finish removing the folio.
600 */
601 if (unlikely(folio_mapped(folio)))
602 hugetlb_unmap_file_folio(h, mapping, folio, index);
603
604 folio_lock(folio);
605 /*
606 * We must remove the folio from page cache before removing
607 * the region/ reserve map (hugetlb_unreserve_pages). In
608 * rare out of memory conditions, removal of the region/reserve
609 * map could fail. Correspondingly, the subpool and global
610 * reserve usage count can need to be adjusted.
611 */
612 VM_BUG_ON_FOLIO(folio_test_hugetlb_restore_reserve(folio), folio);
613 hugetlb_delete_from_page_cache(folio);
614 ret = true;
615 if (!truncate_op) {
616 if (unlikely(hugetlb_unreserve_pages(inode, index,
617 index + 1, 1)))
618 hugetlb_fix_reserve_counts(inode);
619 }
620
621 folio_unlock(folio);
622 return ret;
623 }
624
625 /*
626 * remove_inode_hugepages handles two distinct cases: truncation and hole
627 * punch. There are subtle differences in operation for each case.
628 *
629 * truncation is indicated by end of range being LLONG_MAX
630 * In this case, we first scan the range and release found pages.
631 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
632 * maps and global counts. Page faults can race with truncation.
633 * During faults, hugetlb_no_page() checks i_size before page allocation,
634 * and again after obtaining page table lock. It will 'back out'
635 * allocations in the truncated range.
636 * hole punch is indicated if end is not LLONG_MAX
637 * In the hole punch case we scan the range and release found pages.
638 * Only when releasing a page is the associated region/reserve map
639 * deleted. The region/reserve map for ranges without associated
640 * pages are not modified. Page faults can race with hole punch.
641 * This is indicated if we find a mapped page.
642 * Note: If the passed end of range value is beyond the end of file, but
643 * not LLONG_MAX this routine still performs a hole punch operation.
644 */
645 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
646 loff_t lend)
647 {
648 struct hstate *h = hstate_inode(inode);
649 struct address_space *mapping = &inode->i_data;
650 const pgoff_t end = lend >> PAGE_SHIFT;
651 struct folio_batch fbatch;
652 pgoff_t next, index;
653 int i, freed = 0;
654 bool truncate_op = (lend == LLONG_MAX);
655
656 folio_batch_init(&fbatch);
657 next = lstart >> PAGE_SHIFT;
658 while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
659 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
660 struct folio *folio = fbatch.folios[i];
661 u32 hash = 0;
662
663 index = folio->index >> huge_page_order(h);
664 hash = hugetlb_fault_mutex_hash(mapping, index);
665 mutex_lock(&hugetlb_fault_mutex_table[hash]);
666
667 /*
668 * Remove folio that was part of folio_batch.
669 */
670 if (remove_inode_single_folio(h, inode, mapping, folio,
671 index, truncate_op))
672 freed++;
673
674 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
675 }
676 folio_batch_release(&fbatch);
677 cond_resched();
678 }
679
680 if (truncate_op)
681 (void)hugetlb_unreserve_pages(inode,
682 lstart >> huge_page_shift(h),
683 LONG_MAX, freed);
684 }
685
686 static void hugetlbfs_evict_inode(struct inode *inode)
687 {
688 struct resv_map *resv_map;
689
690 remove_inode_hugepages(inode, 0, LLONG_MAX);
691
692 /*
693 * Get the resv_map from the address space embedded in the inode.
694 * This is the address space which points to any resv_map allocated
695 * at inode creation time. If this is a device special inode,
696 * i_mapping may not point to the original address space.
697 */
698 resv_map = (struct resv_map *)(&inode->i_data)->i_private_data;
699 /* Only regular and link inodes have associated reserve maps */
700 if (resv_map)
701 resv_map_release(&resv_map->refs);
702 clear_inode(inode);
703 }
704
705 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
706 {
707 pgoff_t pgoff;
708 struct address_space *mapping = inode->i_mapping;
709 struct hstate *h = hstate_inode(inode);
710
711 BUG_ON(offset & ~huge_page_mask(h));
712 pgoff = offset >> PAGE_SHIFT;
713
714 i_size_write(inode, offset);
715 i_mmap_lock_write(mapping);
716 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
717 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
718 ZAP_FLAG_DROP_MARKER);
719 i_mmap_unlock_write(mapping);
720 remove_inode_hugepages(inode, offset, LLONG_MAX);
721 }
722
723 static void hugetlbfs_zero_partial_page(struct hstate *h,
724 struct address_space *mapping,
725 loff_t start,
726 loff_t end)
727 {
728 pgoff_t idx = start >> huge_page_shift(h);
729 struct folio *folio;
730
731 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
732 if (IS_ERR(folio))
733 return;
734
735 start = start & ~huge_page_mask(h);
736 end = end & ~huge_page_mask(h);
737 if (!end)
738 end = huge_page_size(h);
739
740 folio_zero_segment(folio, (size_t)start, (size_t)end);
741
742 folio_unlock(folio);
743 folio_put(folio);
744 }
745
746 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
747 {
748 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
749 struct address_space *mapping = inode->i_mapping;
750 struct hstate *h = hstate_inode(inode);
751 loff_t hpage_size = huge_page_size(h);
752 loff_t hole_start, hole_end;
753
754 /*
755 * hole_start and hole_end indicate the full pages within the hole.
756 */
757 hole_start = round_up(offset, hpage_size);
758 hole_end = round_down(offset + len, hpage_size);
759
760 inode_lock(inode);
761
762 /* protected by i_rwsem */
763 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
764 inode_unlock(inode);
765 return -EPERM;
766 }
767
768 i_mmap_lock_write(mapping);
769
770 /* If range starts before first full page, zero partial page. */
771 if (offset < hole_start)
772 hugetlbfs_zero_partial_page(h, mapping,
773 offset, min(offset + len, hole_start));
774
775 /* Unmap users of full pages in the hole. */
776 if (hole_end > hole_start) {
777 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
778 hugetlb_vmdelete_list(&mapping->i_mmap,
779 hole_start >> PAGE_SHIFT,
780 hole_end >> PAGE_SHIFT, 0);
781 }
782
783 /* If range extends beyond last full page, zero partial page. */
784 if ((offset + len) > hole_end && (offset + len) > hole_start)
785 hugetlbfs_zero_partial_page(h, mapping,
786 hole_end, offset + len);
787
788 i_mmap_unlock_write(mapping);
789
790 /* Remove full pages from the file. */
791 if (hole_end > hole_start)
792 remove_inode_hugepages(inode, hole_start, hole_end);
793
794 inode_unlock(inode);
795
796 return 0;
797 }
798
799 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
800 loff_t len)
801 {
802 struct inode *inode = file_inode(file);
803 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
804 struct address_space *mapping = inode->i_mapping;
805 struct hstate *h = hstate_inode(inode);
806 struct vm_area_struct pseudo_vma;
807 struct mm_struct *mm = current->mm;
808 loff_t hpage_size = huge_page_size(h);
809 unsigned long hpage_shift = huge_page_shift(h);
810 pgoff_t start, index, end;
811 int error;
812 u32 hash;
813
814 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
815 return -EOPNOTSUPP;
816
817 if (mode & FALLOC_FL_PUNCH_HOLE)
818 return hugetlbfs_punch_hole(inode, offset, len);
819
820 /*
821 * Default preallocate case.
822 * For this range, start is rounded down and end is rounded up
823 * as well as being converted to page offsets.
824 */
825 start = offset >> hpage_shift;
826 end = (offset + len + hpage_size - 1) >> hpage_shift;
827
828 inode_lock(inode);
829
830 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
831 error = inode_newsize_ok(inode, offset + len);
832 if (error)
833 goto out;
834
835 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
836 error = -EPERM;
837 goto out;
838 }
839
840 /*
841 * Initialize a pseudo vma as this is required by the huge page
842 * allocation routines.
843 */
844 vma_init(&pseudo_vma, mm);
845 vm_flags_init(&pseudo_vma, VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
846 pseudo_vma.vm_file = file;
847
848 for (index = start; index < end; index++) {
849 /*
850 * This is supposed to be the vaddr where the page is being
851 * faulted in, but we have no vaddr here.
852 */
853 struct folio *folio;
854 unsigned long addr;
855
856 cond_resched();
857
858 /*
859 * fallocate(2) manpage permits EINTR; we may have been
860 * interrupted because we are using up too much memory.
861 */
862 if (signal_pending(current)) {
863 error = -EINTR;
864 break;
865 }
866
867 /* addr is the offset within the file (zero based) */
868 addr = index * hpage_size;
869
870 /* mutex taken here, fault path and hole punch */
871 hash = hugetlb_fault_mutex_hash(mapping, index);
872 mutex_lock(&hugetlb_fault_mutex_table[hash]);
873
874 /* See if already present in mapping to avoid alloc/free */
875 folio = filemap_get_folio(mapping, index << huge_page_order(h));
876 if (!IS_ERR(folio)) {
877 folio_put(folio);
878 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
879 continue;
880 }
881
882 /*
883 * Allocate folio without setting the avoid_reserve argument.
884 * There certainly are no reserves associated with the
885 * pseudo_vma. However, there could be shared mappings with
886 * reserves for the file at the inode level. If we fallocate
887 * folios in these areas, we need to consume the reserves
888 * to keep reservation accounting consistent.
889 */
890 folio = alloc_hugetlb_folio(&pseudo_vma, addr, 0);
891 if (IS_ERR(folio)) {
892 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
893 error = PTR_ERR(folio);
894 goto out;
895 }
896 folio_zero_user(folio, ALIGN_DOWN(addr, hpage_size));
897 __folio_mark_uptodate(folio);
898 error = hugetlb_add_to_page_cache(folio, mapping, index);
899 if (unlikely(error)) {
900 restore_reserve_on_error(h, &pseudo_vma, addr, folio);
901 folio_put(folio);
902 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
903 goto out;
904 }
905
906 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
907
908 folio_set_hugetlb_migratable(folio);
909 /*
910 * folio_unlock because locked by hugetlb_add_to_page_cache()
911 * folio_put() due to reference from alloc_hugetlb_folio()
912 */
913 folio_unlock(folio);
914 folio_put(folio);
915 }
916
917 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
918 i_size_write(inode, offset + len);
919 inode_set_ctime_current(inode);
920 out:
921 inode_unlock(inode);
922 return error;
923 }
924
925 static int hugetlbfs_setattr(struct mnt_idmap *idmap,
926 struct dentry *dentry, struct iattr *attr)
927 {
928 struct inode *inode = d_inode(dentry);
929 struct hstate *h = hstate_inode(inode);
930 int error;
931 unsigned int ia_valid = attr->ia_valid;
932 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
933
934 error = setattr_prepare(idmap, dentry, attr);
935 if (error)
936 return error;
937
938 if (ia_valid & ATTR_SIZE) {
939 loff_t oldsize = inode->i_size;
940 loff_t newsize = attr->ia_size;
941
942 if (newsize & ~huge_page_mask(h))
943 return -EINVAL;
944 /* protected by i_rwsem */
945 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
946 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
947 return -EPERM;
948 hugetlb_vmtruncate(inode, newsize);
949 }
950
951 setattr_copy(idmap, inode, attr);
952 mark_inode_dirty(inode);
953 return 0;
954 }
955
956 static struct inode *hugetlbfs_get_root(struct super_block *sb,
957 struct hugetlbfs_fs_context *ctx)
958 {
959 struct inode *inode;
960
961 inode = new_inode(sb);
962 if (inode) {
963 inode->i_ino = get_next_ino();
964 inode->i_mode = S_IFDIR | ctx->mode;
965 inode->i_uid = ctx->uid;
966 inode->i_gid = ctx->gid;
967 simple_inode_init_ts(inode);
968 inode->i_op = &hugetlbfs_dir_inode_operations;
969 inode->i_fop = &simple_dir_operations;
970 /* directory inodes start off with i_nlink == 2 (for "." entry) */
971 inc_nlink(inode);
972 lockdep_annotate_inode_mutex_key(inode);
973 }
974 return inode;
975 }
976
977 /*
978 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
979 * be taken from reclaim -- unlike regular filesystems. This needs an
980 * annotation because huge_pmd_share() does an allocation under hugetlb's
981 * i_mmap_rwsem.
982 */
983 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
984
985 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
986 struct mnt_idmap *idmap,
987 struct inode *dir,
988 umode_t mode, dev_t dev)
989 {
990 struct inode *inode;
991 struct resv_map *resv_map = NULL;
992
993 /*
994 * Reserve maps are only needed for inodes that can have associated
995 * page allocations.
996 */
997 if (S_ISREG(mode) || S_ISLNK(mode)) {
998 resv_map = resv_map_alloc();
999 if (!resv_map)
1000 return NULL;
1001 }
1002
1003 inode = new_inode(sb);
1004 if (inode) {
1005 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
1006
1007 inode->i_ino = get_next_ino();
1008 inode_init_owner(idmap, inode, dir, mode);
1009 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
1010 &hugetlbfs_i_mmap_rwsem_key);
1011 inode->i_mapping->a_ops = &hugetlbfs_aops;
1012 simple_inode_init_ts(inode);
1013 inode->i_mapping->i_private_data = resv_map;
1014 info->seals = F_SEAL_SEAL;
1015 switch (mode & S_IFMT) {
1016 default:
1017 init_special_inode(inode, mode, dev);
1018 break;
1019 case S_IFREG:
1020 inode->i_op = &hugetlbfs_inode_operations;
1021 inode->i_fop = &hugetlbfs_file_operations;
1022 break;
1023 case S_IFDIR:
1024 inode->i_op = &hugetlbfs_dir_inode_operations;
1025 inode->i_fop = &simple_dir_operations;
1026
1027 /* directory inodes start off with i_nlink == 2 (for "." entry) */
1028 inc_nlink(inode);
1029 break;
1030 case S_IFLNK:
1031 inode->i_op = &page_symlink_inode_operations;
1032 inode_nohighmem(inode);
1033 break;
1034 }
1035 lockdep_annotate_inode_mutex_key(inode);
1036 } else {
1037 if (resv_map)
1038 kref_put(&resv_map->refs, resv_map_release);
1039 }
1040
1041 return inode;
1042 }
1043
1044 /*
1045 * File creation. Allocate an inode, and we're done..
1046 */
1047 static int hugetlbfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
1048 struct dentry *dentry, umode_t mode, dev_t dev)
1049 {
1050 struct inode *inode;
1051
1052 inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode, dev);
1053 if (!inode)
1054 return -ENOSPC;
1055 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1056 d_instantiate(dentry, inode);
1057 dget(dentry);/* Extra count - pin the dentry in core */
1058 return 0;
1059 }
1060
1061 static int hugetlbfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1062 struct dentry *dentry, umode_t mode)
1063 {
1064 int retval = hugetlbfs_mknod(idmap, dir, dentry,
1065 mode | S_IFDIR, 0);
1066 if (!retval)
1067 inc_nlink(dir);
1068 return retval;
1069 }
1070
1071 static int hugetlbfs_create(struct mnt_idmap *idmap,
1072 struct inode *dir, struct dentry *dentry,
1073 umode_t mode, bool excl)
1074 {
1075 return hugetlbfs_mknod(idmap, dir, dentry, mode | S_IFREG, 0);
1076 }
1077
1078 static int hugetlbfs_tmpfile(struct mnt_idmap *idmap,
1079 struct inode *dir, struct file *file,
1080 umode_t mode)
1081 {
1082 struct inode *inode;
1083
1084 inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode | S_IFREG, 0);
1085 if (!inode)
1086 return -ENOSPC;
1087 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1088 d_tmpfile(file, inode);
1089 return finish_open_simple(file, 0);
1090 }
1091
1092 static int hugetlbfs_symlink(struct mnt_idmap *idmap,
1093 struct inode *dir, struct dentry *dentry,
1094 const char *symname)
1095 {
1096 const umode_t mode = S_IFLNK|S_IRWXUGO;
1097 struct inode *inode;
1098 int error = -ENOSPC;
1099
1100 inode = hugetlbfs_get_inode(dir->i_sb, idmap, dir, mode, 0);
1101 if (inode) {
1102 int l = strlen(symname)+1;
1103 error = page_symlink(inode, symname, l);
1104 if (!error) {
1105 d_instantiate(dentry, inode);
1106 dget(dentry);
1107 } else
1108 iput(inode);
1109 }
1110 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1111
1112 return error;
1113 }
1114
1115 #ifdef CONFIG_MIGRATION
1116 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1117 struct folio *dst, struct folio *src,
1118 enum migrate_mode mode)
1119 {
1120 int rc;
1121
1122 rc = migrate_huge_page_move_mapping(mapping, dst, src);
1123 if (rc != MIGRATEPAGE_SUCCESS)
1124 return rc;
1125
1126 if (hugetlb_folio_subpool(src)) {
1127 hugetlb_set_folio_subpool(dst,
1128 hugetlb_folio_subpool(src));
1129 hugetlb_set_folio_subpool(src, NULL);
1130 }
1131
1132 folio_migrate_flags(dst, src);
1133
1134 return MIGRATEPAGE_SUCCESS;
1135 }
1136 #else
1137 #define hugetlbfs_migrate_folio NULL
1138 #endif
1139
1140 static int hugetlbfs_error_remove_folio(struct address_space *mapping,
1141 struct folio *folio)
1142 {
1143 return 0;
1144 }
1145
1146 /*
1147 * Display the mount options in /proc/mounts.
1148 */
1149 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1150 {
1151 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1152 struct hugepage_subpool *spool = sbinfo->spool;
1153 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1154 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1155 char mod;
1156
1157 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1158 seq_printf(m, ",uid=%u",
1159 from_kuid_munged(&init_user_ns, sbinfo->uid));
1160 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1161 seq_printf(m, ",gid=%u",
1162 from_kgid_munged(&init_user_ns, sbinfo->gid));
1163 if (sbinfo->mode != 0755)
1164 seq_printf(m, ",mode=%o", sbinfo->mode);
1165 if (sbinfo->max_inodes != -1)
1166 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1167
1168 hpage_size /= 1024;
1169 mod = 'K';
1170 if (hpage_size >= 1024) {
1171 hpage_size /= 1024;
1172 mod = 'M';
1173 }
1174 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1175 if (spool) {
1176 if (spool->max_hpages != -1)
1177 seq_printf(m, ",size=%llu",
1178 (unsigned long long)spool->max_hpages << hpage_shift);
1179 if (spool->min_hpages != -1)
1180 seq_printf(m, ",min_size=%llu",
1181 (unsigned long long)spool->min_hpages << hpage_shift);
1182 }
1183 return 0;
1184 }
1185
1186 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1187 {
1188 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1189 struct hstate *h = hstate_inode(d_inode(dentry));
1190 u64 id = huge_encode_dev(dentry->d_sb->s_dev);
1191
1192 buf->f_fsid = u64_to_fsid(id);
1193 buf->f_type = HUGETLBFS_MAGIC;
1194 buf->f_bsize = huge_page_size(h);
1195 if (sbinfo) {
1196 spin_lock(&sbinfo->stat_lock);
1197 /* If no limits set, just report 0 or -1 for max/free/used
1198 * blocks, like simple_statfs() */
1199 if (sbinfo->spool) {
1200 long free_pages;
1201
1202 spin_lock_irq(&sbinfo->spool->lock);
1203 buf->f_blocks = sbinfo->spool->max_hpages;
1204 free_pages = sbinfo->spool->max_hpages
1205 - sbinfo->spool->used_hpages;
1206 buf->f_bavail = buf->f_bfree = free_pages;
1207 spin_unlock_irq(&sbinfo->spool->lock);
1208 buf->f_files = sbinfo->max_inodes;
1209 buf->f_ffree = sbinfo->free_inodes;
1210 }
1211 spin_unlock(&sbinfo->stat_lock);
1212 }
1213 buf->f_namelen = NAME_MAX;
1214 return 0;
1215 }
1216
1217 static void hugetlbfs_put_super(struct super_block *sb)
1218 {
1219 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1220
1221 if (sbi) {
1222 sb->s_fs_info = NULL;
1223
1224 if (sbi->spool)
1225 hugepage_put_subpool(sbi->spool);
1226
1227 kfree(sbi);
1228 }
1229 }
1230
1231 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1232 {
1233 if (sbinfo->free_inodes >= 0) {
1234 spin_lock(&sbinfo->stat_lock);
1235 if (unlikely(!sbinfo->free_inodes)) {
1236 spin_unlock(&sbinfo->stat_lock);
1237 return 0;
1238 }
1239 sbinfo->free_inodes--;
1240 spin_unlock(&sbinfo->stat_lock);
1241 }
1242
1243 return 1;
1244 }
1245
1246 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1247 {
1248 if (sbinfo->free_inodes >= 0) {
1249 spin_lock(&sbinfo->stat_lock);
1250 sbinfo->free_inodes++;
1251 spin_unlock(&sbinfo->stat_lock);
1252 }
1253 }
1254
1255
1256 static struct kmem_cache *hugetlbfs_inode_cachep;
1257
1258 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1259 {
1260 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1261 struct hugetlbfs_inode_info *p;
1262
1263 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1264 return NULL;
1265 p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1266 if (unlikely(!p)) {
1267 hugetlbfs_inc_free_inodes(sbinfo);
1268 return NULL;
1269 }
1270 return &p->vfs_inode;
1271 }
1272
1273 static void hugetlbfs_free_inode(struct inode *inode)
1274 {
1275 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1276 }
1277
1278 static void hugetlbfs_destroy_inode(struct inode *inode)
1279 {
1280 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1281 }
1282
1283 static const struct address_space_operations hugetlbfs_aops = {
1284 .write_begin = hugetlbfs_write_begin,
1285 .write_end = hugetlbfs_write_end,
1286 .dirty_folio = noop_dirty_folio,
1287 .migrate_folio = hugetlbfs_migrate_folio,
1288 .error_remove_folio = hugetlbfs_error_remove_folio,
1289 };
1290
1291
1292 static void init_once(void *foo)
1293 {
1294 struct hugetlbfs_inode_info *ei = foo;
1295
1296 inode_init_once(&ei->vfs_inode);
1297 }
1298
1299 static const struct file_operations hugetlbfs_file_operations = {
1300 .read_iter = hugetlbfs_read_iter,
1301 .mmap = hugetlbfs_file_mmap,
1302 .fsync = noop_fsync,
1303 .get_unmapped_area = hugetlb_get_unmapped_area,
1304 .llseek = default_llseek,
1305 .fallocate = hugetlbfs_fallocate,
1306 .fop_flags = FOP_HUGE_PAGES,
1307 };
1308
1309 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1310 .create = hugetlbfs_create,
1311 .lookup = simple_lookup,
1312 .link = simple_link,
1313 .unlink = simple_unlink,
1314 .symlink = hugetlbfs_symlink,
1315 .mkdir = hugetlbfs_mkdir,
1316 .rmdir = simple_rmdir,
1317 .mknod = hugetlbfs_mknod,
1318 .rename = simple_rename,
1319 .setattr = hugetlbfs_setattr,
1320 .tmpfile = hugetlbfs_tmpfile,
1321 };
1322
1323 static const struct inode_operations hugetlbfs_inode_operations = {
1324 .setattr = hugetlbfs_setattr,
1325 };
1326
1327 static const struct super_operations hugetlbfs_ops = {
1328 .alloc_inode = hugetlbfs_alloc_inode,
1329 .free_inode = hugetlbfs_free_inode,
1330 .destroy_inode = hugetlbfs_destroy_inode,
1331 .evict_inode = hugetlbfs_evict_inode,
1332 .statfs = hugetlbfs_statfs,
1333 .put_super = hugetlbfs_put_super,
1334 .show_options = hugetlbfs_show_options,
1335 };
1336
1337 /*
1338 * Convert size option passed from command line to number of huge pages
1339 * in the pool specified by hstate. Size option could be in bytes
1340 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1341 */
1342 static long
1343 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1344 enum hugetlbfs_size_type val_type)
1345 {
1346 if (val_type == NO_SIZE)
1347 return -1;
1348
1349 if (val_type == SIZE_PERCENT) {
1350 size_opt <<= huge_page_shift(h);
1351 size_opt *= h->max_huge_pages;
1352 do_div(size_opt, 100);
1353 }
1354
1355 size_opt >>= huge_page_shift(h);
1356 return size_opt;
1357 }
1358
1359 /*
1360 * Parse one mount parameter.
1361 */
1362 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1363 {
1364 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1365 struct fs_parse_result result;
1366 struct hstate *h;
1367 char *rest;
1368 unsigned long ps;
1369 int opt;
1370
1371 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1372 if (opt < 0)
1373 return opt;
1374
1375 switch (opt) {
1376 case Opt_uid:
1377 ctx->uid = result.uid;
1378 return 0;
1379
1380 case Opt_gid:
1381 ctx->gid = result.gid;
1382 return 0;
1383
1384 case Opt_mode:
1385 ctx->mode = result.uint_32 & 01777U;
1386 return 0;
1387
1388 case Opt_size:
1389 /* memparse() will accept a K/M/G without a digit */
1390 if (!param->string || !isdigit(param->string[0]))
1391 goto bad_val;
1392 ctx->max_size_opt = memparse(param->string, &rest);
1393 ctx->max_val_type = SIZE_STD;
1394 if (*rest == '%')
1395 ctx->max_val_type = SIZE_PERCENT;
1396 return 0;
1397
1398 case Opt_nr_inodes:
1399 /* memparse() will accept a K/M/G without a digit */
1400 if (!param->string || !isdigit(param->string[0]))
1401 goto bad_val;
1402 ctx->nr_inodes = memparse(param->string, &rest);
1403 return 0;
1404
1405 case Opt_pagesize:
1406 ps = memparse(param->string, &rest);
1407 h = size_to_hstate(ps);
1408 if (!h) {
1409 pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1410 return -EINVAL;
1411 }
1412 ctx->hstate = h;
1413 return 0;
1414
1415 case Opt_min_size:
1416 /* memparse() will accept a K/M/G without a digit */
1417 if (!param->string || !isdigit(param->string[0]))
1418 goto bad_val;
1419 ctx->min_size_opt = memparse(param->string, &rest);
1420 ctx->min_val_type = SIZE_STD;
1421 if (*rest == '%')
1422 ctx->min_val_type = SIZE_PERCENT;
1423 return 0;
1424
1425 default:
1426 return -EINVAL;
1427 }
1428
1429 bad_val:
1430 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1431 param->string, param->key);
1432 }
1433
1434 /*
1435 * Validate the parsed options.
1436 */
1437 static int hugetlbfs_validate(struct fs_context *fc)
1438 {
1439 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1440
1441 /*
1442 * Use huge page pool size (in hstate) to convert the size
1443 * options to number of huge pages. If NO_SIZE, -1 is returned.
1444 */
1445 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1446 ctx->max_size_opt,
1447 ctx->max_val_type);
1448 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1449 ctx->min_size_opt,
1450 ctx->min_val_type);
1451
1452 /*
1453 * If max_size was specified, then min_size must be smaller
1454 */
1455 if (ctx->max_val_type > NO_SIZE &&
1456 ctx->min_hpages > ctx->max_hpages) {
1457 pr_err("Minimum size can not be greater than maximum size\n");
1458 return -EINVAL;
1459 }
1460
1461 return 0;
1462 }
1463
1464 static int
1465 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1466 {
1467 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1468 struct hugetlbfs_sb_info *sbinfo;
1469
1470 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1471 if (!sbinfo)
1472 return -ENOMEM;
1473 sb->s_fs_info = sbinfo;
1474 spin_lock_init(&sbinfo->stat_lock);
1475 sbinfo->hstate = ctx->hstate;
1476 sbinfo->max_inodes = ctx->nr_inodes;
1477 sbinfo->free_inodes = ctx->nr_inodes;
1478 sbinfo->spool = NULL;
1479 sbinfo->uid = ctx->uid;
1480 sbinfo->gid = ctx->gid;
1481 sbinfo->mode = ctx->mode;
1482
1483 /*
1484 * Allocate and initialize subpool if maximum or minimum size is
1485 * specified. Any needed reservations (for minimum size) are taken
1486 * when the subpool is created.
1487 */
1488 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1489 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1490 ctx->max_hpages,
1491 ctx->min_hpages);
1492 if (!sbinfo->spool)
1493 goto out_free;
1494 }
1495 sb->s_maxbytes = MAX_LFS_FILESIZE;
1496 sb->s_blocksize = huge_page_size(ctx->hstate);
1497 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1498 sb->s_magic = HUGETLBFS_MAGIC;
1499 sb->s_op = &hugetlbfs_ops;
1500 sb->s_time_gran = 1;
1501
1502 /*
1503 * Due to the special and limited functionality of hugetlbfs, it does
1504 * not work well as a stacking filesystem.
1505 */
1506 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1507 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1508 if (!sb->s_root)
1509 goto out_free;
1510 return 0;
1511 out_free:
1512 kfree(sbinfo->spool);
1513 kfree(sbinfo);
1514 return -ENOMEM;
1515 }
1516
1517 static int hugetlbfs_get_tree(struct fs_context *fc)
1518 {
1519 int err = hugetlbfs_validate(fc);
1520 if (err)
1521 return err;
1522 return get_tree_nodev(fc, hugetlbfs_fill_super);
1523 }
1524
1525 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1526 {
1527 kfree(fc->fs_private);
1528 }
1529
1530 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1531 .free = hugetlbfs_fs_context_free,
1532 .parse_param = hugetlbfs_parse_param,
1533 .get_tree = hugetlbfs_get_tree,
1534 };
1535
1536 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1537 {
1538 struct hugetlbfs_fs_context *ctx;
1539
1540 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1541 if (!ctx)
1542 return -ENOMEM;
1543
1544 ctx->max_hpages = -1; /* No limit on size by default */
1545 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1546 ctx->uid = current_fsuid();
1547 ctx->gid = current_fsgid();
1548 ctx->mode = 0755;
1549 ctx->hstate = &default_hstate;
1550 ctx->min_hpages = -1; /* No default minimum size */
1551 ctx->max_val_type = NO_SIZE;
1552 ctx->min_val_type = NO_SIZE;
1553 fc->fs_private = ctx;
1554 fc->ops = &hugetlbfs_fs_context_ops;
1555 return 0;
1556 }
1557
1558 static struct file_system_type hugetlbfs_fs_type = {
1559 .name = "hugetlbfs",
1560 .init_fs_context = hugetlbfs_init_fs_context,
1561 .parameters = hugetlb_fs_parameters,
1562 .kill_sb = kill_litter_super,
1563 .fs_flags = FS_ALLOW_IDMAP,
1564 };
1565
1566 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1567
1568 static int can_do_hugetlb_shm(void)
1569 {
1570 kgid_t shm_group;
1571 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1572 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1573 }
1574
1575 static int get_hstate_idx(int page_size_log)
1576 {
1577 struct hstate *h = hstate_sizelog(page_size_log);
1578
1579 if (!h)
1580 return -1;
1581 return hstate_index(h);
1582 }
1583
1584 /*
1585 * Note that size should be aligned to proper hugepage size in caller side,
1586 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1587 */
1588 struct file *hugetlb_file_setup(const char *name, size_t size,
1589 vm_flags_t acctflag, int creat_flags,
1590 int page_size_log)
1591 {
1592 struct inode *inode;
1593 struct vfsmount *mnt;
1594 int hstate_idx;
1595 struct file *file;
1596
1597 hstate_idx = get_hstate_idx(page_size_log);
1598 if (hstate_idx < 0)
1599 return ERR_PTR(-ENODEV);
1600
1601 mnt = hugetlbfs_vfsmount[hstate_idx];
1602 if (!mnt)
1603 return ERR_PTR(-ENOENT);
1604
1605 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1606 struct ucounts *ucounts = current_ucounts();
1607
1608 if (user_shm_lock(size, ucounts)) {
1609 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1610 current->comm, current->pid);
1611 user_shm_unlock(size, ucounts);
1612 }
1613 return ERR_PTR(-EPERM);
1614 }
1615
1616 file = ERR_PTR(-ENOSPC);
1617 /* hugetlbfs_vfsmount[] mounts do not use idmapped mounts. */
1618 inode = hugetlbfs_get_inode(mnt->mnt_sb, &nop_mnt_idmap, NULL,
1619 S_IFREG | S_IRWXUGO, 0);
1620 if (!inode)
1621 goto out;
1622 if (creat_flags == HUGETLB_SHMFS_INODE)
1623 inode->i_flags |= S_PRIVATE;
1624
1625 inode->i_size = size;
1626 clear_nlink(inode);
1627
1628 if (!hugetlb_reserve_pages(inode, 0,
1629 size >> huge_page_shift(hstate_inode(inode)), NULL,
1630 acctflag))
1631 file = ERR_PTR(-ENOMEM);
1632 else
1633 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1634 &hugetlbfs_file_operations);
1635 if (!IS_ERR(file))
1636 return file;
1637
1638 iput(inode);
1639 out:
1640 return file;
1641 }
1642
1643 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1644 {
1645 struct fs_context *fc;
1646 struct vfsmount *mnt;
1647
1648 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1649 if (IS_ERR(fc)) {
1650 mnt = ERR_CAST(fc);
1651 } else {
1652 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1653 ctx->hstate = h;
1654 mnt = fc_mount(fc);
1655 put_fs_context(fc);
1656 }
1657 if (IS_ERR(mnt))
1658 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1659 huge_page_size(h) / SZ_1K);
1660 return mnt;
1661 }
1662
1663 static int __init init_hugetlbfs_fs(void)
1664 {
1665 struct vfsmount *mnt;
1666 struct hstate *h;
1667 int error;
1668 int i;
1669
1670 if (!hugepages_supported()) {
1671 pr_info("disabling because there are no supported hugepage sizes\n");
1672 return -ENOTSUPP;
1673 }
1674
1675 error = -ENOMEM;
1676 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1677 sizeof(struct hugetlbfs_inode_info),
1678 0, SLAB_ACCOUNT, init_once);
1679 if (hugetlbfs_inode_cachep == NULL)
1680 goto out;
1681
1682 error = register_filesystem(&hugetlbfs_fs_type);
1683 if (error)
1684 goto out_free;
1685
1686 /* default hstate mount is required */
1687 mnt = mount_one_hugetlbfs(&default_hstate);
1688 if (IS_ERR(mnt)) {
1689 error = PTR_ERR(mnt);
1690 goto out_unreg;
1691 }
1692 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1693
1694 /* other hstates are optional */
1695 i = 0;
1696 for_each_hstate(h) {
1697 if (i == default_hstate_idx) {
1698 i++;
1699 continue;
1700 }
1701
1702 mnt = mount_one_hugetlbfs(h);
1703 if (IS_ERR(mnt))
1704 hugetlbfs_vfsmount[i] = NULL;
1705 else
1706 hugetlbfs_vfsmount[i] = mnt;
1707 i++;
1708 }
1709
1710 return 0;
1711
1712 out_unreg:
1713 (void)unregister_filesystem(&hugetlbfs_fs_type);
1714 out_free:
1715 kmem_cache_destroy(hugetlbfs_inode_cachep);
1716 out:
1717 return error;
1718 }
1719 fs_initcall(init_hugetlbfs_fs)
1720
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.