xref: /linux/fs/hugetlbfs/inode.c (revision 93d90ad708b8da6efc0e487b66111aa9db7f70c7)
1 /*
2  * hugetlbpage-backed filesystem.  Based on ramfs.
3  *
4  * Nadia Yvette Chambers, 2002
5  *
6  * Copyright (C) 2002 Linus Torvalds.
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/module.h>
12 #include <linux/thread_info.h>
13 #include <asm/current.h>
14 #include <linux/sched.h>		/* remove ASAP */
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/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 
38 #include <asm/uaccess.h>
39 
40 static const struct super_operations hugetlbfs_ops;
41 static const struct address_space_operations hugetlbfs_aops;
42 const struct file_operations hugetlbfs_file_operations;
43 static const struct inode_operations hugetlbfs_dir_inode_operations;
44 static const struct inode_operations hugetlbfs_inode_operations;
45 
46 struct hugetlbfs_config {
47 	kuid_t   uid;
48 	kgid_t   gid;
49 	umode_t mode;
50 	long	nr_blocks;
51 	long	nr_inodes;
52 	struct hstate *hstate;
53 };
54 
55 struct hugetlbfs_inode_info {
56 	struct shared_policy policy;
57 	struct inode vfs_inode;
58 };
59 
60 static inline struct hugetlbfs_inode_info *HUGETLBFS_I(struct inode *inode)
61 {
62 	return container_of(inode, struct hugetlbfs_inode_info, vfs_inode);
63 }
64 
65 static struct backing_dev_info hugetlbfs_backing_dev_info = {
66 	.name		= "hugetlbfs",
67 	.ra_pages	= 0,	/* No readahead */
68 	.capabilities	= BDI_CAP_NO_ACCT_AND_WRITEBACK,
69 };
70 
71 int sysctl_hugetlb_shm_group;
72 
73 enum {
74 	Opt_size, Opt_nr_inodes,
75 	Opt_mode, Opt_uid, Opt_gid,
76 	Opt_pagesize,
77 	Opt_err,
78 };
79 
80 static const match_table_t tokens = {
81 	{Opt_size,	"size=%s"},
82 	{Opt_nr_inodes,	"nr_inodes=%s"},
83 	{Opt_mode,	"mode=%o"},
84 	{Opt_uid,	"uid=%u"},
85 	{Opt_gid,	"gid=%u"},
86 	{Opt_pagesize,	"pagesize=%s"},
87 	{Opt_err,	NULL},
88 };
89 
90 static void huge_pagevec_release(struct pagevec *pvec)
91 {
92 	int i;
93 
94 	for (i = 0; i < pagevec_count(pvec); ++i)
95 		put_page(pvec->pages[i]);
96 
97 	pagevec_reinit(pvec);
98 }
99 
100 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
101 {
102 	struct inode *inode = file_inode(file);
103 	loff_t len, vma_len;
104 	int ret;
105 	struct hstate *h = hstate_file(file);
106 
107 	/*
108 	 * vma address alignment (but not the pgoff alignment) has
109 	 * already been checked by prepare_hugepage_range.  If you add
110 	 * any error returns here, do so after setting VM_HUGETLB, so
111 	 * is_vm_hugetlb_page tests below unmap_region go the right
112 	 * way when do_mmap_pgoff unwinds (may be important on powerpc
113 	 * and ia64).
114 	 */
115 	vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
116 	vma->vm_ops = &hugetlb_vm_ops;
117 
118 	if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
119 		return -EINVAL;
120 
121 	vma_len = (loff_t)(vma->vm_end - vma->vm_start);
122 
123 	mutex_lock(&inode->i_mutex);
124 	file_accessed(file);
125 
126 	ret = -ENOMEM;
127 	len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
128 
129 	if (hugetlb_reserve_pages(inode,
130 				vma->vm_pgoff >> huge_page_order(h),
131 				len >> huge_page_shift(h), vma,
132 				vma->vm_flags))
133 		goto out;
134 
135 	ret = 0;
136 	hugetlb_prefault_arch_hook(vma->vm_mm);
137 	if (vma->vm_flags & VM_WRITE && inode->i_size < len)
138 		inode->i_size = len;
139 out:
140 	mutex_unlock(&inode->i_mutex);
141 
142 	return ret;
143 }
144 
145 /*
146  * Called under down_write(mmap_sem).
147  */
148 
149 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
150 static unsigned long
151 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
152 		unsigned long len, unsigned long pgoff, unsigned long flags)
153 {
154 	struct mm_struct *mm = current->mm;
155 	struct vm_area_struct *vma;
156 	struct hstate *h = hstate_file(file);
157 	struct vm_unmapped_area_info info;
158 
159 	if (len & ~huge_page_mask(h))
160 		return -EINVAL;
161 	if (len > TASK_SIZE)
162 		return -ENOMEM;
163 
164 	if (flags & MAP_FIXED) {
165 		if (prepare_hugepage_range(file, addr, len))
166 			return -EINVAL;
167 		return addr;
168 	}
169 
170 	if (addr) {
171 		addr = ALIGN(addr, huge_page_size(h));
172 		vma = find_vma(mm, addr);
173 		if (TASK_SIZE - len >= addr &&
174 		    (!vma || addr + len <= vma->vm_start))
175 			return addr;
176 	}
177 
178 	info.flags = 0;
179 	info.length = len;
180 	info.low_limit = TASK_UNMAPPED_BASE;
181 	info.high_limit = TASK_SIZE;
182 	info.align_mask = PAGE_MASK & ~huge_page_mask(h);
183 	info.align_offset = 0;
184 	return vm_unmapped_area(&info);
185 }
186 #endif
187 
188 static int
189 hugetlbfs_read_actor(struct page *page, unsigned long offset,
190 			char __user *buf, unsigned long count,
191 			unsigned long size)
192 {
193 	char *kaddr;
194 	unsigned long left, copied = 0;
195 	int i, chunksize;
196 
197 	if (size > count)
198 		size = count;
199 
200 	/* Find which 4k chunk and offset with in that chunk */
201 	i = offset >> PAGE_CACHE_SHIFT;
202 	offset = offset & ~PAGE_CACHE_MASK;
203 
204 	while (size) {
205 		chunksize = PAGE_CACHE_SIZE;
206 		if (offset)
207 			chunksize -= offset;
208 		if (chunksize > size)
209 			chunksize = size;
210 		kaddr = kmap(&page[i]);
211 		left = __copy_to_user(buf, kaddr + offset, chunksize);
212 		kunmap(&page[i]);
213 		if (left) {
214 			copied += (chunksize - left);
215 			break;
216 		}
217 		offset = 0;
218 		size -= chunksize;
219 		buf += chunksize;
220 		copied += chunksize;
221 		i++;
222 	}
223 	return copied ? copied : -EFAULT;
224 }
225 
226 /*
227  * Support for read() - Find the page attached to f_mapping and copy out the
228  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
229  * since it has PAGE_CACHE_SIZE assumptions.
230  */
231 static ssize_t hugetlbfs_read(struct file *filp, char __user *buf,
232 			      size_t len, loff_t *ppos)
233 {
234 	struct hstate *h = hstate_file(filp);
235 	struct address_space *mapping = filp->f_mapping;
236 	struct inode *inode = mapping->host;
237 	unsigned long index = *ppos >> huge_page_shift(h);
238 	unsigned long offset = *ppos & ~huge_page_mask(h);
239 	unsigned long end_index;
240 	loff_t isize;
241 	ssize_t retval = 0;
242 
243 	/* validate length */
244 	if (len == 0)
245 		goto out;
246 
247 	for (;;) {
248 		struct page *page;
249 		unsigned long nr, ret;
250 		int ra;
251 
252 		/* nr is the maximum number of bytes to copy from this page */
253 		nr = huge_page_size(h);
254 		isize = i_size_read(inode);
255 		if (!isize)
256 			goto out;
257 		end_index = (isize - 1) >> huge_page_shift(h);
258 		if (index >= end_index) {
259 			if (index > end_index)
260 				goto out;
261 			nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
262 			if (nr <= offset)
263 				goto out;
264 		}
265 		nr = nr - offset;
266 
267 		/* Find the page */
268 		page = find_lock_page(mapping, index);
269 		if (unlikely(page == NULL)) {
270 			/*
271 			 * We have a HOLE, zero out the user-buffer for the
272 			 * length of the hole or request.
273 			 */
274 			ret = len < nr ? len : nr;
275 			if (clear_user(buf, ret))
276 				ra = -EFAULT;
277 			else
278 				ra = 0;
279 		} else {
280 			unlock_page(page);
281 
282 			/*
283 			 * We have the page, copy it to user space buffer.
284 			 */
285 			ra = hugetlbfs_read_actor(page, offset, buf, len, nr);
286 			ret = ra;
287 			page_cache_release(page);
288 		}
289 		if (ra < 0) {
290 			if (retval == 0)
291 				retval = ra;
292 			goto out;
293 		}
294 
295 		offset += ret;
296 		retval += ret;
297 		len -= ret;
298 		index += offset >> huge_page_shift(h);
299 		offset &= ~huge_page_mask(h);
300 
301 		/* short read or no more work */
302 		if ((ret != nr) || (len == 0))
303 			break;
304 	}
305 out:
306 	*ppos = ((loff_t)index << huge_page_shift(h)) + offset;
307 	return retval;
308 }
309 
310 static int hugetlbfs_write_begin(struct file *file,
311 			struct address_space *mapping,
312 			loff_t pos, unsigned len, unsigned flags,
313 			struct page **pagep, void **fsdata)
314 {
315 	return -EINVAL;
316 }
317 
318 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
319 			loff_t pos, unsigned len, unsigned copied,
320 			struct page *page, void *fsdata)
321 {
322 	BUG();
323 	return -EINVAL;
324 }
325 
326 static void truncate_huge_page(struct page *page)
327 {
328 	cancel_dirty_page(page, /* No IO accounting for huge pages? */0);
329 	ClearPageUptodate(page);
330 	delete_from_page_cache(page);
331 }
332 
333 static void truncate_hugepages(struct inode *inode, loff_t lstart)
334 {
335 	struct hstate *h = hstate_inode(inode);
336 	struct address_space *mapping = &inode->i_data;
337 	const pgoff_t start = lstart >> huge_page_shift(h);
338 	struct pagevec pvec;
339 	pgoff_t next;
340 	int i, freed = 0;
341 
342 	pagevec_init(&pvec, 0);
343 	next = start;
344 	while (1) {
345 		if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
346 			if (next == start)
347 				break;
348 			next = start;
349 			continue;
350 		}
351 
352 		for (i = 0; i < pagevec_count(&pvec); ++i) {
353 			struct page *page = pvec.pages[i];
354 
355 			lock_page(page);
356 			if (page->index > next)
357 				next = page->index;
358 			++next;
359 			truncate_huge_page(page);
360 			unlock_page(page);
361 			freed++;
362 		}
363 		huge_pagevec_release(&pvec);
364 	}
365 	BUG_ON(!lstart && mapping->nrpages);
366 	hugetlb_unreserve_pages(inode, start, freed);
367 }
368 
369 static void hugetlbfs_evict_inode(struct inode *inode)
370 {
371 	struct resv_map *resv_map;
372 
373 	truncate_hugepages(inode, 0);
374 	resv_map = (struct resv_map *)inode->i_mapping->private_data;
375 	/* root inode doesn't have the resv_map, so we should check it */
376 	if (resv_map)
377 		resv_map_release(&resv_map->refs);
378 	clear_inode(inode);
379 }
380 
381 static inline void
382 hugetlb_vmtruncate_list(struct rb_root *root, pgoff_t pgoff)
383 {
384 	struct vm_area_struct *vma;
385 
386 	vma_interval_tree_foreach(vma, root, pgoff, ULONG_MAX) {
387 		unsigned long v_offset;
388 
389 		/*
390 		 * Can the expression below overflow on 32-bit arches?
391 		 * No, because the interval tree returns us only those vmas
392 		 * which overlap the truncated area starting at pgoff,
393 		 * and no vma on a 32-bit arch can span beyond the 4GB.
394 		 */
395 		if (vma->vm_pgoff < pgoff)
396 			v_offset = (pgoff - vma->vm_pgoff) << PAGE_SHIFT;
397 		else
398 			v_offset = 0;
399 
400 		unmap_hugepage_range(vma, vma->vm_start + v_offset,
401 				     vma->vm_end, NULL);
402 	}
403 }
404 
405 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
406 {
407 	pgoff_t pgoff;
408 	struct address_space *mapping = inode->i_mapping;
409 	struct hstate *h = hstate_inode(inode);
410 
411 	BUG_ON(offset & ~huge_page_mask(h));
412 	pgoff = offset >> PAGE_SHIFT;
413 
414 	i_size_write(inode, offset);
415 	i_mmap_lock_write(mapping);
416 	if (!RB_EMPTY_ROOT(&mapping->i_mmap))
417 		hugetlb_vmtruncate_list(&mapping->i_mmap, pgoff);
418 	i_mmap_unlock_write(mapping);
419 	truncate_hugepages(inode, offset);
420 	return 0;
421 }
422 
423 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
424 {
425 	struct inode *inode = dentry->d_inode;
426 	struct hstate *h = hstate_inode(inode);
427 	int error;
428 	unsigned int ia_valid = attr->ia_valid;
429 
430 	BUG_ON(!inode);
431 
432 	error = inode_change_ok(inode, attr);
433 	if (error)
434 		return error;
435 
436 	if (ia_valid & ATTR_SIZE) {
437 		error = -EINVAL;
438 		if (attr->ia_size & ~huge_page_mask(h))
439 			return -EINVAL;
440 		error = hugetlb_vmtruncate(inode, attr->ia_size);
441 		if (error)
442 			return error;
443 	}
444 
445 	setattr_copy(inode, attr);
446 	mark_inode_dirty(inode);
447 	return 0;
448 }
449 
450 static struct inode *hugetlbfs_get_root(struct super_block *sb,
451 					struct hugetlbfs_config *config)
452 {
453 	struct inode *inode;
454 
455 	inode = new_inode(sb);
456 	if (inode) {
457 		struct hugetlbfs_inode_info *info;
458 		inode->i_ino = get_next_ino();
459 		inode->i_mode = S_IFDIR | config->mode;
460 		inode->i_uid = config->uid;
461 		inode->i_gid = config->gid;
462 		inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
463 		info = HUGETLBFS_I(inode);
464 		mpol_shared_policy_init(&info->policy, NULL);
465 		inode->i_op = &hugetlbfs_dir_inode_operations;
466 		inode->i_fop = &simple_dir_operations;
467 		/* directory inodes start off with i_nlink == 2 (for "." entry) */
468 		inc_nlink(inode);
469 		lockdep_annotate_inode_mutex_key(inode);
470 	}
471 	return inode;
472 }
473 
474 /*
475  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
476  * be taken from reclaim -- unlike regular filesystems. This needs an
477  * annotation because huge_pmd_share() does an allocation under
478  * i_mmap_rwsem.
479  */
480 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
481 
482 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
483 					struct inode *dir,
484 					umode_t mode, dev_t dev)
485 {
486 	struct inode *inode;
487 	struct resv_map *resv_map;
488 
489 	resv_map = resv_map_alloc();
490 	if (!resv_map)
491 		return NULL;
492 
493 	inode = new_inode(sb);
494 	if (inode) {
495 		struct hugetlbfs_inode_info *info;
496 		inode->i_ino = get_next_ino();
497 		inode_init_owner(inode, dir, mode);
498 		lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
499 				&hugetlbfs_i_mmap_rwsem_key);
500 		inode->i_mapping->a_ops = &hugetlbfs_aops;
501 		inode->i_mapping->backing_dev_info =&hugetlbfs_backing_dev_info;
502 		inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
503 		inode->i_mapping->private_data = resv_map;
504 		info = HUGETLBFS_I(inode);
505 		/*
506 		 * The policy is initialized here even if we are creating a
507 		 * private inode because initialization simply creates an
508 		 * an empty rb tree and calls spin_lock_init(), later when we
509 		 * call mpol_free_shared_policy() it will just return because
510 		 * the rb tree will still be empty.
511 		 */
512 		mpol_shared_policy_init(&info->policy, NULL);
513 		switch (mode & S_IFMT) {
514 		default:
515 			init_special_inode(inode, mode, dev);
516 			break;
517 		case S_IFREG:
518 			inode->i_op = &hugetlbfs_inode_operations;
519 			inode->i_fop = &hugetlbfs_file_operations;
520 			break;
521 		case S_IFDIR:
522 			inode->i_op = &hugetlbfs_dir_inode_operations;
523 			inode->i_fop = &simple_dir_operations;
524 
525 			/* directory inodes start off with i_nlink == 2 (for "." entry) */
526 			inc_nlink(inode);
527 			break;
528 		case S_IFLNK:
529 			inode->i_op = &page_symlink_inode_operations;
530 			break;
531 		}
532 		lockdep_annotate_inode_mutex_key(inode);
533 	} else
534 		kref_put(&resv_map->refs, resv_map_release);
535 
536 	return inode;
537 }
538 
539 /*
540  * File creation. Allocate an inode, and we're done..
541  */
542 static int hugetlbfs_mknod(struct inode *dir,
543 			struct dentry *dentry, umode_t mode, dev_t dev)
544 {
545 	struct inode *inode;
546 	int error = -ENOSPC;
547 
548 	inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
549 	if (inode) {
550 		dir->i_ctime = dir->i_mtime = CURRENT_TIME;
551 		d_instantiate(dentry, inode);
552 		dget(dentry);	/* Extra count - pin the dentry in core */
553 		error = 0;
554 	}
555 	return error;
556 }
557 
558 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
559 {
560 	int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
561 	if (!retval)
562 		inc_nlink(dir);
563 	return retval;
564 }
565 
566 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
567 {
568 	return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
569 }
570 
571 static int hugetlbfs_symlink(struct inode *dir,
572 			struct dentry *dentry, const char *symname)
573 {
574 	struct inode *inode;
575 	int error = -ENOSPC;
576 
577 	inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
578 	if (inode) {
579 		int l = strlen(symname)+1;
580 		error = page_symlink(inode, symname, l);
581 		if (!error) {
582 			d_instantiate(dentry, inode);
583 			dget(dentry);
584 		} else
585 			iput(inode);
586 	}
587 	dir->i_ctime = dir->i_mtime = CURRENT_TIME;
588 
589 	return error;
590 }
591 
592 /*
593  * mark the head page dirty
594  */
595 static int hugetlbfs_set_page_dirty(struct page *page)
596 {
597 	struct page *head = compound_head(page);
598 
599 	SetPageDirty(head);
600 	return 0;
601 }
602 
603 static int hugetlbfs_migrate_page(struct address_space *mapping,
604 				struct page *newpage, struct page *page,
605 				enum migrate_mode mode)
606 {
607 	int rc;
608 
609 	rc = migrate_huge_page_move_mapping(mapping, newpage, page);
610 	if (rc != MIGRATEPAGE_SUCCESS)
611 		return rc;
612 	migrate_page_copy(newpage, page);
613 
614 	return MIGRATEPAGE_SUCCESS;
615 }
616 
617 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
618 {
619 	struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
620 	struct hstate *h = hstate_inode(dentry->d_inode);
621 
622 	buf->f_type = HUGETLBFS_MAGIC;
623 	buf->f_bsize = huge_page_size(h);
624 	if (sbinfo) {
625 		spin_lock(&sbinfo->stat_lock);
626 		/* If no limits set, just report 0 for max/free/used
627 		 * blocks, like simple_statfs() */
628 		if (sbinfo->spool) {
629 			long free_pages;
630 
631 			spin_lock(&sbinfo->spool->lock);
632 			buf->f_blocks = sbinfo->spool->max_hpages;
633 			free_pages = sbinfo->spool->max_hpages
634 				- sbinfo->spool->used_hpages;
635 			buf->f_bavail = buf->f_bfree = free_pages;
636 			spin_unlock(&sbinfo->spool->lock);
637 			buf->f_files = sbinfo->max_inodes;
638 			buf->f_ffree = sbinfo->free_inodes;
639 		}
640 		spin_unlock(&sbinfo->stat_lock);
641 	}
642 	buf->f_namelen = NAME_MAX;
643 	return 0;
644 }
645 
646 static void hugetlbfs_put_super(struct super_block *sb)
647 {
648 	struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
649 
650 	if (sbi) {
651 		sb->s_fs_info = NULL;
652 
653 		if (sbi->spool)
654 			hugepage_put_subpool(sbi->spool);
655 
656 		kfree(sbi);
657 	}
658 }
659 
660 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
661 {
662 	if (sbinfo->free_inodes >= 0) {
663 		spin_lock(&sbinfo->stat_lock);
664 		if (unlikely(!sbinfo->free_inodes)) {
665 			spin_unlock(&sbinfo->stat_lock);
666 			return 0;
667 		}
668 		sbinfo->free_inodes--;
669 		spin_unlock(&sbinfo->stat_lock);
670 	}
671 
672 	return 1;
673 }
674 
675 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
676 {
677 	if (sbinfo->free_inodes >= 0) {
678 		spin_lock(&sbinfo->stat_lock);
679 		sbinfo->free_inodes++;
680 		spin_unlock(&sbinfo->stat_lock);
681 	}
682 }
683 
684 
685 static struct kmem_cache *hugetlbfs_inode_cachep;
686 
687 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
688 {
689 	struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
690 	struct hugetlbfs_inode_info *p;
691 
692 	if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
693 		return NULL;
694 	p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
695 	if (unlikely(!p)) {
696 		hugetlbfs_inc_free_inodes(sbinfo);
697 		return NULL;
698 	}
699 	return &p->vfs_inode;
700 }
701 
702 static void hugetlbfs_i_callback(struct rcu_head *head)
703 {
704 	struct inode *inode = container_of(head, struct inode, i_rcu);
705 	kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
706 }
707 
708 static void hugetlbfs_destroy_inode(struct inode *inode)
709 {
710 	hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
711 	mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
712 	call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
713 }
714 
715 static const struct address_space_operations hugetlbfs_aops = {
716 	.write_begin	= hugetlbfs_write_begin,
717 	.write_end	= hugetlbfs_write_end,
718 	.set_page_dirty	= hugetlbfs_set_page_dirty,
719 	.migratepage    = hugetlbfs_migrate_page,
720 };
721 
722 
723 static void init_once(void *foo)
724 {
725 	struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
726 
727 	inode_init_once(&ei->vfs_inode);
728 }
729 
730 const struct file_operations hugetlbfs_file_operations = {
731 	.read			= hugetlbfs_read,
732 	.mmap			= hugetlbfs_file_mmap,
733 	.fsync			= noop_fsync,
734 	.get_unmapped_area	= hugetlb_get_unmapped_area,
735 	.llseek		= default_llseek,
736 };
737 
738 static const struct inode_operations hugetlbfs_dir_inode_operations = {
739 	.create		= hugetlbfs_create,
740 	.lookup		= simple_lookup,
741 	.link		= simple_link,
742 	.unlink		= simple_unlink,
743 	.symlink	= hugetlbfs_symlink,
744 	.mkdir		= hugetlbfs_mkdir,
745 	.rmdir		= simple_rmdir,
746 	.mknod		= hugetlbfs_mknod,
747 	.rename		= simple_rename,
748 	.setattr	= hugetlbfs_setattr,
749 };
750 
751 static const struct inode_operations hugetlbfs_inode_operations = {
752 	.setattr	= hugetlbfs_setattr,
753 };
754 
755 static const struct super_operations hugetlbfs_ops = {
756 	.alloc_inode    = hugetlbfs_alloc_inode,
757 	.destroy_inode  = hugetlbfs_destroy_inode,
758 	.evict_inode	= hugetlbfs_evict_inode,
759 	.statfs		= hugetlbfs_statfs,
760 	.put_super	= hugetlbfs_put_super,
761 	.show_options	= generic_show_options,
762 };
763 
764 static int
765 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
766 {
767 	char *p, *rest;
768 	substring_t args[MAX_OPT_ARGS];
769 	int option;
770 	unsigned long long size = 0;
771 	enum { NO_SIZE, SIZE_STD, SIZE_PERCENT } setsize = NO_SIZE;
772 
773 	if (!options)
774 		return 0;
775 
776 	while ((p = strsep(&options, ",")) != NULL) {
777 		int token;
778 		if (!*p)
779 			continue;
780 
781 		token = match_token(p, tokens, args);
782 		switch (token) {
783 		case Opt_uid:
784 			if (match_int(&args[0], &option))
785  				goto bad_val;
786 			pconfig->uid = make_kuid(current_user_ns(), option);
787 			if (!uid_valid(pconfig->uid))
788 				goto bad_val;
789 			break;
790 
791 		case Opt_gid:
792 			if (match_int(&args[0], &option))
793  				goto bad_val;
794 			pconfig->gid = make_kgid(current_user_ns(), option);
795 			if (!gid_valid(pconfig->gid))
796 				goto bad_val;
797 			break;
798 
799 		case Opt_mode:
800 			if (match_octal(&args[0], &option))
801  				goto bad_val;
802 			pconfig->mode = option & 01777U;
803 			break;
804 
805 		case Opt_size: {
806 			/* memparse() will accept a K/M/G without a digit */
807 			if (!isdigit(*args[0].from))
808 				goto bad_val;
809 			size = memparse(args[0].from, &rest);
810 			setsize = SIZE_STD;
811 			if (*rest == '%')
812 				setsize = SIZE_PERCENT;
813 			break;
814 		}
815 
816 		case Opt_nr_inodes:
817 			/* memparse() will accept a K/M/G without a digit */
818 			if (!isdigit(*args[0].from))
819 				goto bad_val;
820 			pconfig->nr_inodes = memparse(args[0].from, &rest);
821 			break;
822 
823 		case Opt_pagesize: {
824 			unsigned long ps;
825 			ps = memparse(args[0].from, &rest);
826 			pconfig->hstate = size_to_hstate(ps);
827 			if (!pconfig->hstate) {
828 				pr_err("Unsupported page size %lu MB\n",
829 					ps >> 20);
830 				return -EINVAL;
831 			}
832 			break;
833 		}
834 
835 		default:
836 			pr_err("Bad mount option: \"%s\"\n", p);
837 			return -EINVAL;
838 			break;
839 		}
840 	}
841 
842 	/* Do size after hstate is set up */
843 	if (setsize > NO_SIZE) {
844 		struct hstate *h = pconfig->hstate;
845 		if (setsize == SIZE_PERCENT) {
846 			size <<= huge_page_shift(h);
847 			size *= h->max_huge_pages;
848 			do_div(size, 100);
849 		}
850 		pconfig->nr_blocks = (size >> huge_page_shift(h));
851 	}
852 
853 	return 0;
854 
855 bad_val:
856 	pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
857  	return -EINVAL;
858 }
859 
860 static int
861 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
862 {
863 	int ret;
864 	struct hugetlbfs_config config;
865 	struct hugetlbfs_sb_info *sbinfo;
866 
867 	save_mount_options(sb, data);
868 
869 	config.nr_blocks = -1; /* No limit on size by default */
870 	config.nr_inodes = -1; /* No limit on number of inodes by default */
871 	config.uid = current_fsuid();
872 	config.gid = current_fsgid();
873 	config.mode = 0755;
874 	config.hstate = &default_hstate;
875 	ret = hugetlbfs_parse_options(data, &config);
876 	if (ret)
877 		return ret;
878 
879 	sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
880 	if (!sbinfo)
881 		return -ENOMEM;
882 	sb->s_fs_info = sbinfo;
883 	sbinfo->hstate = config.hstate;
884 	spin_lock_init(&sbinfo->stat_lock);
885 	sbinfo->max_inodes = config.nr_inodes;
886 	sbinfo->free_inodes = config.nr_inodes;
887 	sbinfo->spool = NULL;
888 	if (config.nr_blocks != -1) {
889 		sbinfo->spool = hugepage_new_subpool(config.nr_blocks);
890 		if (!sbinfo->spool)
891 			goto out_free;
892 	}
893 	sb->s_maxbytes = MAX_LFS_FILESIZE;
894 	sb->s_blocksize = huge_page_size(config.hstate);
895 	sb->s_blocksize_bits = huge_page_shift(config.hstate);
896 	sb->s_magic = HUGETLBFS_MAGIC;
897 	sb->s_op = &hugetlbfs_ops;
898 	sb->s_time_gran = 1;
899 	sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
900 	if (!sb->s_root)
901 		goto out_free;
902 	return 0;
903 out_free:
904 	kfree(sbinfo->spool);
905 	kfree(sbinfo);
906 	return -ENOMEM;
907 }
908 
909 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
910 	int flags, const char *dev_name, void *data)
911 {
912 	return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
913 }
914 
915 static struct file_system_type hugetlbfs_fs_type = {
916 	.name		= "hugetlbfs",
917 	.mount		= hugetlbfs_mount,
918 	.kill_sb	= kill_litter_super,
919 };
920 MODULE_ALIAS_FS("hugetlbfs");
921 
922 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
923 
924 static int can_do_hugetlb_shm(void)
925 {
926 	kgid_t shm_group;
927 	shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
928 	return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
929 }
930 
931 static int get_hstate_idx(int page_size_log)
932 {
933 	struct hstate *h = hstate_sizelog(page_size_log);
934 
935 	if (!h)
936 		return -1;
937 	return h - hstates;
938 }
939 
940 static const struct dentry_operations anon_ops = {
941 	.d_dname = simple_dname
942 };
943 
944 /*
945  * Note that size should be aligned to proper hugepage size in caller side,
946  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
947  */
948 struct file *hugetlb_file_setup(const char *name, size_t size,
949 				vm_flags_t acctflag, struct user_struct **user,
950 				int creat_flags, int page_size_log)
951 {
952 	struct file *file = ERR_PTR(-ENOMEM);
953 	struct inode *inode;
954 	struct path path;
955 	struct super_block *sb;
956 	struct qstr quick_string;
957 	int hstate_idx;
958 
959 	hstate_idx = get_hstate_idx(page_size_log);
960 	if (hstate_idx < 0)
961 		return ERR_PTR(-ENODEV);
962 
963 	*user = NULL;
964 	if (!hugetlbfs_vfsmount[hstate_idx])
965 		return ERR_PTR(-ENOENT);
966 
967 	if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
968 		*user = current_user();
969 		if (user_shm_lock(size, *user)) {
970 			task_lock(current);
971 			pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
972 				current->comm, current->pid);
973 			task_unlock(current);
974 		} else {
975 			*user = NULL;
976 			return ERR_PTR(-EPERM);
977 		}
978 	}
979 
980 	sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
981 	quick_string.name = name;
982 	quick_string.len = strlen(quick_string.name);
983 	quick_string.hash = 0;
984 	path.dentry = d_alloc_pseudo(sb, &quick_string);
985 	if (!path.dentry)
986 		goto out_shm_unlock;
987 
988 	d_set_d_op(path.dentry, &anon_ops);
989 	path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
990 	file = ERR_PTR(-ENOSPC);
991 	inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
992 	if (!inode)
993 		goto out_dentry;
994 
995 	file = ERR_PTR(-ENOMEM);
996 	if (hugetlb_reserve_pages(inode, 0,
997 			size >> huge_page_shift(hstate_inode(inode)), NULL,
998 			acctflag))
999 		goto out_inode;
1000 
1001 	d_instantiate(path.dentry, inode);
1002 	inode->i_size = size;
1003 	clear_nlink(inode);
1004 
1005 	file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1006 			&hugetlbfs_file_operations);
1007 	if (IS_ERR(file))
1008 		goto out_dentry; /* inode is already attached */
1009 
1010 	return file;
1011 
1012 out_inode:
1013 	iput(inode);
1014 out_dentry:
1015 	path_put(&path);
1016 out_shm_unlock:
1017 	if (*user) {
1018 		user_shm_unlock(size, *user);
1019 		*user = NULL;
1020 	}
1021 	return file;
1022 }
1023 
1024 static int __init init_hugetlbfs_fs(void)
1025 {
1026 	struct hstate *h;
1027 	int error;
1028 	int i;
1029 
1030 	if (!hugepages_supported()) {
1031 		pr_info("disabling because there are no supported hugepage sizes\n");
1032 		return -ENOTSUPP;
1033 	}
1034 
1035 	error = bdi_init(&hugetlbfs_backing_dev_info);
1036 	if (error)
1037 		return error;
1038 
1039 	error = -ENOMEM;
1040 	hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1041 					sizeof(struct hugetlbfs_inode_info),
1042 					0, 0, init_once);
1043 	if (hugetlbfs_inode_cachep == NULL)
1044 		goto out2;
1045 
1046 	error = register_filesystem(&hugetlbfs_fs_type);
1047 	if (error)
1048 		goto out;
1049 
1050 	i = 0;
1051 	for_each_hstate(h) {
1052 		char buf[50];
1053 		unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1054 
1055 		snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1056 		hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1057 							buf);
1058 
1059 		if (IS_ERR(hugetlbfs_vfsmount[i])) {
1060 			pr_err("Cannot mount internal hugetlbfs for "
1061 				"page size %uK", ps_kb);
1062 			error = PTR_ERR(hugetlbfs_vfsmount[i]);
1063 			hugetlbfs_vfsmount[i] = NULL;
1064 		}
1065 		i++;
1066 	}
1067 	/* Non default hstates are optional */
1068 	if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1069 		return 0;
1070 
1071  out:
1072 	kmem_cache_destroy(hugetlbfs_inode_cachep);
1073  out2:
1074 	bdi_destroy(&hugetlbfs_backing_dev_info);
1075 	return error;
1076 }
1077 
1078 static void __exit exit_hugetlbfs_fs(void)
1079 {
1080 	struct hstate *h;
1081 	int i;
1082 
1083 
1084 	/*
1085 	 * Make sure all delayed rcu free inodes are flushed before we
1086 	 * destroy cache.
1087 	 */
1088 	rcu_barrier();
1089 	kmem_cache_destroy(hugetlbfs_inode_cachep);
1090 	i = 0;
1091 	for_each_hstate(h)
1092 		kern_unmount(hugetlbfs_vfsmount[i++]);
1093 	unregister_filesystem(&hugetlbfs_fs_type);
1094 	bdi_destroy(&hugetlbfs_backing_dev_info);
1095 }
1096 
1097 module_init(init_hugetlbfs_fs)
1098 module_exit(exit_hugetlbfs_fs)
1099 
1100 MODULE_LICENSE("GPL");
1101