3 * Anonymous Shared Memory Subsystem, ashmem
5 * Copyright (C) 2008 Google, Inc.
9 * This software is licensed under the terms of the GNU General Public
10 * License version 2, as published by the Free Software Foundation, and
11 * may be copied, distributed, and modified under those terms.
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
19 #define pr_fmt(fmt) "ashmem: " fmt
21 #include <linux/init.h>
22 #include <linux/export.h>
23 #include <linux/file.h>
25 #include <linux/falloc.h>
26 #include <linux/miscdevice.h>
27 #include <linux/security.h>
29 #include <linux/mman.h>
30 #include <linux/uaccess.h>
31 #include <linux/personality.h>
32 #include <linux/bitops.h>
33 #include <linux/mutex.h>
34 #include <linux/shmem_fs.h>
37 #define ASHMEM_NAME_PREFIX "dev/ashmem/"
38 #define ASHMEM_NAME_PREFIX_LEN (sizeof(ASHMEM_NAME_PREFIX) - 1)
39 #define ASHMEM_FULL_NAME_LEN (ASHMEM_NAME_LEN + ASHMEM_NAME_PREFIX_LEN)
42 * struct ashmem_area - The anonymous shared memory area
43 * @name: The optional name in /proc/pid/maps
44 * @unpinned_list: The list of all ashmem areas
45 * @file: The shmem-based backing file
46 * @size: The size of the mapping, in bytes
47 * @prot_mask: The allowed protection bits, as vm_flags
49 * The lifecycle of this structure is from our parent file's open() until
50 * its release(). It is also protected by 'ashmem_mutex'
52 * Warning: Mappings do NOT pin this structure; It dies on close()
55 char name[ASHMEM_FULL_NAME_LEN];
56 struct list_head unpinned_list;
59 unsigned long prot_mask;
63 * struct ashmem_range - A range of unpinned/evictable pages
64 * @lru: The entry in the LRU list
65 * @unpinned: The entry in its area's unpinned list
66 * @asma: The associated anonymous shared memory area.
67 * @pgstart: The starting page (inclusive)
68 * @pgend: The ending page (inclusive)
69 * @purged: The purge status (ASHMEM_NOT or ASHMEM_WAS_PURGED)
71 * The lifecycle of this structure is from unpin to pin.
72 * It is protected by 'ashmem_mutex'
76 struct list_head unpinned;
77 struct ashmem_area *asma;
83 /* LRU list of unpinned pages, protected by ashmem_mutex */
84 static LIST_HEAD(ashmem_lru_list);
87 * long lru_count - The count of pages on our LRU list.
89 * This is protected by ashmem_mutex.
91 static unsigned long lru_count;
94 * ashmem_mutex - protects the list of and each individual ashmem_area
96 * Lock Ordering: ashmex_mutex -> i_mutex -> i_alloc_sem
98 static DEFINE_MUTEX(ashmem_mutex);
100 static struct kmem_cache *ashmem_area_cachep __read_mostly;
101 static struct kmem_cache *ashmem_range_cachep __read_mostly;
103 static inline unsigned long range_size(struct ashmem_range *range)
105 return range->pgend - range->pgstart + 1;
108 static inline bool range_on_lru(struct ashmem_range *range)
110 return range->purged == ASHMEM_NOT_PURGED;
113 static inline bool page_range_subsumes_range(struct ashmem_range *range,
114 size_t start, size_t end)
116 return (range->pgstart >= start) && (range->pgend <= end);
119 static inline bool page_range_subsumed_by_range(struct ashmem_range *range,
120 size_t start, size_t end)
122 return (range->pgstart <= start) && (range->pgend >= end);
125 static inline bool page_in_range(struct ashmem_range *range, size_t page)
127 return (range->pgstart <= page) && (range->pgend >= page);
130 static inline bool page_range_in_range(struct ashmem_range *range,
131 size_t start, size_t end)
133 return page_in_range(range, start) || page_in_range(range, end) ||
134 page_range_subsumes_range(range, start, end);
137 static inline bool range_before_page(struct ashmem_range *range, size_t page)
139 return range->pgend < page;
142 #define PROT_MASK (PROT_EXEC | PROT_READ | PROT_WRITE)
145 * lru_add() - Adds a range of memory to the LRU list
146 * @range: The memory range being added.
148 * The range is first added to the end (tail) of the LRU list.
149 * After this, the size of the range is added to @lru_count
151 static inline void lru_add(struct ashmem_range *range)
153 list_add_tail(&range->lru, &ashmem_lru_list);
154 lru_count += range_size(range);
158 * lru_del() - Removes a range of memory from the LRU list
159 * @range: The memory range being removed
161 * The range is first deleted from the LRU list.
162 * After this, the size of the range is removed from @lru_count
164 static inline void lru_del(struct ashmem_range *range)
166 list_del(&range->lru);
167 lru_count -= range_size(range);
171 * range_alloc() - Allocates and initializes a new ashmem_range structure
172 * @asma: The associated ashmem_area
173 * @prev_range: The previous ashmem_range in the sorted asma->unpinned list
174 * @purged: Initial purge status (ASMEM_NOT_PURGED or ASHMEM_WAS_PURGED)
175 * @start: The starting page (inclusive)
176 * @end: The ending page (inclusive)
178 * This function is protected by ashmem_mutex.
180 * Return: 0 if successful, or -ENOMEM if there is an error
182 static int range_alloc(struct ashmem_area *asma,
183 struct ashmem_range *prev_range, unsigned int purged,
184 size_t start, size_t end)
186 struct ashmem_range *range;
188 range = kmem_cache_zalloc(ashmem_range_cachep, GFP_KERNEL);
189 if (unlikely(!range))
193 range->pgstart = start;
195 range->purged = purged;
197 list_add_tail(&range->unpinned, &prev_range->unpinned);
199 if (range_on_lru(range))
206 * range_del() - Deletes and dealloctes an ashmem_range structure
207 * @range: The associated ashmem_range that has previously been allocated
209 static void range_del(struct ashmem_range *range)
211 list_del(&range->unpinned);
212 if (range_on_lru(range))
214 kmem_cache_free(ashmem_range_cachep, range);
218 * range_shrink() - Shrinks an ashmem_range
219 * @range: The associated ashmem_range being shrunk
220 * @start: The starting byte of the new range
221 * @end: The ending byte of the new range
223 * This does not modify the data inside the existing range in any way - It
224 * simply shrinks the boundaries of the range.
226 * Theoretically, with a little tweaking, this could eventually be changed
227 * to range_resize, and expand the lru_count if the new range is larger.
229 static inline void range_shrink(struct ashmem_range *range,
230 size_t start, size_t end)
232 size_t pre = range_size(range);
234 range->pgstart = start;
237 if (range_on_lru(range))
238 lru_count -= pre - range_size(range);
242 * ashmem_open() - Opens an Anonymous Shared Memory structure
243 * @inode: The backing file's index node(?)
244 * @file: The backing file
246 * Please note that the ashmem_area is not returned by this function - It is
247 * instead written to "file->private_data".
249 * Return: 0 if successful, or another code if unsuccessful.
251 static int ashmem_open(struct inode *inode, struct file *file)
253 struct ashmem_area *asma;
256 ret = generic_file_open(inode, file);
260 asma = kmem_cache_zalloc(ashmem_area_cachep, GFP_KERNEL);
264 INIT_LIST_HEAD(&asma->unpinned_list);
265 memcpy(asma->name, ASHMEM_NAME_PREFIX, ASHMEM_NAME_PREFIX_LEN);
266 asma->prot_mask = PROT_MASK;
267 file->private_data = asma;
273 * ashmem_release() - Releases an Anonymous Shared Memory structure
274 * @ignored: The backing file's Index Node(?) - It is ignored here.
275 * @file: The backing file
277 * Return: 0 if successful. If it is anything else, go have a coffee and
280 static int ashmem_release(struct inode *ignored, struct file *file)
282 struct ashmem_area *asma = file->private_data;
283 struct ashmem_range *range, *next;
285 mutex_lock(&ashmem_mutex);
286 list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned)
288 mutex_unlock(&ashmem_mutex);
292 kmem_cache_free(ashmem_area_cachep, asma);
298 * ashmem_read() - Reads a set of bytes from an Ashmem-enabled file
299 * @file: The associated backing file.
300 * @buf: The buffer of data being written to
301 * @len: The number of bytes being read
302 * @pos: The position of the first byte to read.
304 * Return: 0 if successful, or another return code if not.
306 static ssize_t ashmem_read(struct file *file, char __user *buf,
307 size_t len, loff_t *pos)
309 struct ashmem_area *asma = file->private_data;
312 mutex_lock(&ashmem_mutex);
314 /* If size is not set, or set to 0, always return EOF. */
323 mutex_unlock(&ashmem_mutex);
326 * asma and asma->file are used outside the lock here. We assume
327 * once asma->file is set it will never be changed, and will not
328 * be destroyed until all references to the file are dropped and
329 * ashmem_release is called.
331 ret = __vfs_read(asma->file, buf, len, pos);
333 /** Update backing file pos, since f_ops->read() doesn't */
334 asma->file->f_pos = *pos;
338 mutex_unlock(&ashmem_mutex);
342 static loff_t ashmem_llseek(struct file *file, loff_t offset, int origin)
344 struct ashmem_area *asma = file->private_data;
347 mutex_lock(&ashmem_mutex);
349 if (asma->size == 0) {
359 ret = vfs_llseek(asma->file, offset, origin);
363 /** Copy f_pos from backing file, since f_ops->llseek() sets it */
364 file->f_pos = asma->file->f_pos;
367 mutex_unlock(&ashmem_mutex);
371 static inline vm_flags_t calc_vm_may_flags(unsigned long prot)
373 return _calc_vm_trans(prot, PROT_READ, VM_MAYREAD) |
374 _calc_vm_trans(prot, PROT_WRITE, VM_MAYWRITE) |
375 _calc_vm_trans(prot, PROT_EXEC, VM_MAYEXEC);
378 static int ashmem_mmap(struct file *file, struct vm_area_struct *vma)
380 struct ashmem_area *asma = file->private_data;
383 mutex_lock(&ashmem_mutex);
385 /* user needs to SET_SIZE before mapping */
386 if (unlikely(!asma->size)) {
391 /* requested protection bits must match our allowed protection mask */
392 if (unlikely((vma->vm_flags & ~calc_vm_prot_bits(asma->prot_mask, 0)) &
393 calc_vm_prot_bits(PROT_MASK, 0))) {
397 vma->vm_flags &= ~calc_vm_may_flags(~asma->prot_mask);
400 char *name = ASHMEM_NAME_DEF;
403 if (asma->name[ASHMEM_NAME_PREFIX_LEN] != '\0')
406 /* ... and allocate the backing shmem file */
407 vmfile = shmem_file_setup(name, asma->size, vma->vm_flags);
408 if (IS_ERR(vmfile)) {
409 ret = PTR_ERR(vmfile);
412 vmfile->f_mode |= FMODE_LSEEK;
415 get_file(asma->file);
418 * XXX - Reworked to use shmem_zero_setup() instead of
419 * shmem_set_file while we're in staging. -jstultz
421 if (vma->vm_flags & VM_SHARED) {
422 ret = shmem_zero_setup(vma);
431 vma->vm_file = asma->file;
434 mutex_unlock(&ashmem_mutex);
439 * ashmem_shrink - our cache shrinker, called from mm/vmscan.c
441 * 'nr_to_scan' is the number of objects to scan for freeing.
443 * 'gfp_mask' is the mask of the allocation that got us into this mess.
445 * Return value is the number of objects freed or -1 if we cannot
446 * proceed without risk of deadlock (due to gfp_mask).
448 * We approximate LRU via least-recently-unpinned, jettisoning unpinned partial
449 * chunks of ashmem regions LRU-wise one-at-a-time until we hit 'nr_to_scan'
453 ashmem_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
455 struct ashmem_range *range, *next;
456 unsigned long freed = 0;
458 /* We might recurse into filesystem code, so bail out if necessary */
459 if (!(sc->gfp_mask & __GFP_FS))
462 if (!mutex_trylock(&ashmem_mutex))
465 list_for_each_entry_safe(range, next, &ashmem_lru_list, lru) {
466 loff_t start = range->pgstart * PAGE_SIZE;
467 loff_t end = (range->pgend + 1) * PAGE_SIZE;
469 vfs_fallocate(range->asma->file,
470 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
472 range->purged = ASHMEM_WAS_PURGED;
475 freed += range_size(range);
476 if (--sc->nr_to_scan <= 0)
479 mutex_unlock(&ashmem_mutex);
484 ashmem_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
487 * note that lru_count is count of pages on the lru, not a count of
488 * objects on the list. This means the scan function needs to return the
489 * number of pages freed, not the number of objects scanned.
494 static struct shrinker ashmem_shrinker = {
495 .count_objects = ashmem_shrink_count,
496 .scan_objects = ashmem_shrink_scan,
498 * XXX (dchinner): I wish people would comment on why they need on
499 * significant changes to the default value here
501 .seeks = DEFAULT_SEEKS * 4,
504 static int set_prot_mask(struct ashmem_area *asma, unsigned long prot)
508 mutex_lock(&ashmem_mutex);
510 /* the user can only remove, not add, protection bits */
511 if (unlikely((asma->prot_mask & prot) != prot)) {
516 /* does the application expect PROT_READ to imply PROT_EXEC? */
517 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
520 asma->prot_mask = prot;
523 mutex_unlock(&ashmem_mutex);
527 static int set_name(struct ashmem_area *asma, void __user *name)
531 char local_name[ASHMEM_NAME_LEN];
534 * Holding the ashmem_mutex while doing a copy_from_user might cause
535 * an data abort which would try to access mmap_sem. If another
536 * thread has invoked ashmem_mmap then it will be holding the
537 * semaphore and will be waiting for ashmem_mutex, there by leading to
538 * deadlock. We'll release the mutex and take the name to a local
539 * variable that does not need protection and later copy the local
540 * variable to the structure member with lock held.
542 len = strncpy_from_user(local_name, name, ASHMEM_NAME_LEN);
545 if (len == ASHMEM_NAME_LEN)
546 local_name[ASHMEM_NAME_LEN - 1] = '\0';
547 mutex_lock(&ashmem_mutex);
548 /* cannot change an existing mapping's name */
549 if (unlikely(asma->file))
552 strcpy(asma->name + ASHMEM_NAME_PREFIX_LEN, local_name);
554 mutex_unlock(&ashmem_mutex);
558 static int get_name(struct ashmem_area *asma, void __user *name)
563 * Have a local variable to which we'll copy the content
564 * from asma with the lock held. Later we can copy this to the user
565 * space safely without holding any locks. So even if we proceed to
566 * wait for mmap_sem, it won't lead to deadlock.
568 char local_name[ASHMEM_NAME_LEN];
570 mutex_lock(&ashmem_mutex);
571 if (asma->name[ASHMEM_NAME_PREFIX_LEN] != '\0') {
573 * Copying only `len', instead of ASHMEM_NAME_LEN, bytes
574 * prevents us from revealing one user's stack to another.
576 len = strlen(asma->name + ASHMEM_NAME_PREFIX_LEN) + 1;
577 memcpy(local_name, asma->name + ASHMEM_NAME_PREFIX_LEN, len);
579 len = sizeof(ASHMEM_NAME_DEF);
580 memcpy(local_name, ASHMEM_NAME_DEF, len);
582 mutex_unlock(&ashmem_mutex);
585 * Now we are just copying from the stack variable to userland
588 if (unlikely(copy_to_user(name, local_name, len)))
594 * ashmem_pin - pin the given ashmem region, returning whether it was
595 * previously purged (ASHMEM_WAS_PURGED) or not (ASHMEM_NOT_PURGED).
597 * Caller must hold ashmem_mutex.
599 static int ashmem_pin(struct ashmem_area *asma, size_t pgstart, size_t pgend)
601 struct ashmem_range *range, *next;
602 int ret = ASHMEM_NOT_PURGED;
604 list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned) {
605 /* moved past last applicable page; we can short circuit */
606 if (range_before_page(range, pgstart))
610 * The user can ask us to pin pages that span multiple ranges,
611 * or to pin pages that aren't even unpinned, so this is messy.
614 * 1. The requested range subsumes an existing range, so we
615 * just remove the entire matching range.
616 * 2. The requested range overlaps the start of an existing
617 * range, so we just update that range.
618 * 3. The requested range overlaps the end of an existing
619 * range, so we just update that range.
620 * 4. The requested range punches a hole in an existing range,
621 * so we have to update one side of the range and then
622 * create a new range for the other side.
624 if (page_range_in_range(range, pgstart, pgend)) {
625 ret |= range->purged;
627 /* Case #1: Easy. Just nuke the whole thing. */
628 if (page_range_subsumes_range(range, pgstart, pgend)) {
633 /* Case #2: We overlap from the start, so adjust it */
634 if (range->pgstart >= pgstart) {
635 range_shrink(range, pgend + 1, range->pgend);
639 /* Case #3: We overlap from the rear, so adjust it */
640 if (range->pgend <= pgend) {
641 range_shrink(range, range->pgstart,
647 * Case #4: We eat a chunk out of the middle. A bit
648 * more complicated, we allocate a new range for the
649 * second half and adjust the first chunk's endpoint.
651 range_alloc(asma, range, range->purged,
652 pgend + 1, range->pgend);
653 range_shrink(range, range->pgstart, pgstart - 1);
662 * ashmem_unpin - unpin the given range of pages. Returns zero on success.
664 * Caller must hold ashmem_mutex.
666 static int ashmem_unpin(struct ashmem_area *asma, size_t pgstart, size_t pgend)
668 struct ashmem_range *range, *next;
669 unsigned int purged = ASHMEM_NOT_PURGED;
672 list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned) {
673 /* short circuit: this is our insertion point */
674 if (range_before_page(range, pgstart))
678 * The user can ask us to unpin pages that are already entirely
679 * or partially pinned. We handle those two cases here.
681 if (page_range_subsumed_by_range(range, pgstart, pgend))
683 if (page_range_in_range(range, pgstart, pgend)) {
684 pgstart = min(range->pgstart, pgstart);
685 pgend = max(range->pgend, pgend);
686 purged |= range->purged;
692 return range_alloc(asma, range, purged, pgstart, pgend);
696 * ashmem_get_pin_status - Returns ASHMEM_IS_UNPINNED if _any_ pages in the
697 * given interval are unpinned and ASHMEM_IS_PINNED otherwise.
699 * Caller must hold ashmem_mutex.
701 static int ashmem_get_pin_status(struct ashmem_area *asma, size_t pgstart,
704 struct ashmem_range *range;
705 int ret = ASHMEM_IS_PINNED;
707 list_for_each_entry(range, &asma->unpinned_list, unpinned) {
708 if (range_before_page(range, pgstart))
710 if (page_range_in_range(range, pgstart, pgend)) {
711 ret = ASHMEM_IS_UNPINNED;
719 static int ashmem_pin_unpin(struct ashmem_area *asma, unsigned long cmd,
722 struct ashmem_pin pin;
723 size_t pgstart, pgend;
726 if (unlikely(!asma->file))
729 if (unlikely(copy_from_user(&pin, p, sizeof(pin))))
732 /* per custom, you can pass zero for len to mean "everything onward" */
734 pin.len = PAGE_ALIGN(asma->size) - pin.offset;
736 if (unlikely((pin.offset | pin.len) & ~PAGE_MASK))
739 if (unlikely(((__u32)-1) - pin.offset < pin.len))
742 if (unlikely(PAGE_ALIGN(asma->size) < pin.offset + pin.len))
745 pgstart = pin.offset / PAGE_SIZE;
746 pgend = pgstart + (pin.len / PAGE_SIZE) - 1;
748 mutex_lock(&ashmem_mutex);
752 ret = ashmem_pin(asma, pgstart, pgend);
755 ret = ashmem_unpin(asma, pgstart, pgend);
757 case ASHMEM_GET_PIN_STATUS:
758 ret = ashmem_get_pin_status(asma, pgstart, pgend);
762 mutex_unlock(&ashmem_mutex);
767 static long ashmem_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
769 struct ashmem_area *asma = file->private_data;
773 case ASHMEM_SET_NAME:
774 ret = set_name(asma, (void __user *)arg);
776 case ASHMEM_GET_NAME:
777 ret = get_name(asma, (void __user *)arg);
779 case ASHMEM_SET_SIZE:
783 asma->size = (size_t)arg;
786 case ASHMEM_GET_SIZE:
789 case ASHMEM_SET_PROT_MASK:
790 ret = set_prot_mask(asma, arg);
792 case ASHMEM_GET_PROT_MASK:
793 ret = asma->prot_mask;
797 case ASHMEM_GET_PIN_STATUS:
798 ret = ashmem_pin_unpin(asma, cmd, (void __user *)arg);
800 case ASHMEM_PURGE_ALL_CACHES:
802 if (capable(CAP_SYS_ADMIN)) {
803 struct shrink_control sc = {
804 .gfp_mask = GFP_KERNEL,
805 .nr_to_scan = LONG_MAX,
807 ret = ashmem_shrink_count(&ashmem_shrinker, &sc);
808 ashmem_shrink_scan(&ashmem_shrinker, &sc);
816 /* support of 32bit userspace on 64bit platforms */
818 static long compat_ashmem_ioctl(struct file *file, unsigned int cmd,
822 case COMPAT_ASHMEM_SET_SIZE:
823 cmd = ASHMEM_SET_SIZE;
825 case COMPAT_ASHMEM_SET_PROT_MASK:
826 cmd = ASHMEM_SET_PROT_MASK;
829 return ashmem_ioctl(file, cmd, arg);
833 static const struct file_operations ashmem_fops = {
834 .owner = THIS_MODULE,
836 .release = ashmem_release,
838 .llseek = ashmem_llseek,
840 .unlocked_ioctl = ashmem_ioctl,
842 .compat_ioctl = compat_ashmem_ioctl,
846 static struct miscdevice ashmem_misc = {
847 .minor = MISC_DYNAMIC_MINOR,
849 .fops = &ashmem_fops,
852 static int __init ashmem_init(void)
856 ashmem_area_cachep = kmem_cache_create("ashmem_area_cache",
857 sizeof(struct ashmem_area),
859 if (unlikely(!ashmem_area_cachep)) {
860 pr_err("failed to create slab cache\n");
864 ashmem_range_cachep = kmem_cache_create("ashmem_range_cache",
865 sizeof(struct ashmem_range),
867 if (unlikely(!ashmem_range_cachep)) {
868 pr_err("failed to create slab cache\n");
872 ret = misc_register(&ashmem_misc);
874 pr_err("failed to register misc device!\n");
878 register_shrinker(&ashmem_shrinker);
880 pr_info("initialized\n");
885 kmem_cache_destroy(ashmem_range_cachep);
887 kmem_cache_destroy(ashmem_area_cachep);
891 device_initcall(ashmem_init);