ashmem.c revision 1d3f8f2da1c28709a3c494f3872b89c871906b2d 
1/* mm/ashmem.c
2**
3** Anonymous Shared Memory Subsystem, ashmem
4**
5** Copyright (C) 2008 Google, Inc.
6**
7** Robert Love <rlove@google.com>
8**
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.
12**
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.
17*/
18
19#include <linux/module.h>
20#include <linux/file.h>
21#include <linux/fs.h>
22#include <linux/miscdevice.h>
23#include <linux/security.h>
24#include <linux/mm.h>
25#include <linux/mman.h>
26#include <linux/uaccess.h>
27#include <linux/personality.h>
28#include <linux/bitops.h>
29#include <linux/mutex.h>
30#include <linux/shmem_fs.h>
31#include "ashmem.h"
32
33#define ASHMEM_NAME_PREFIX "dev/ashmem/"
34#define ASHMEM_NAME_PREFIX_LEN (sizeof(ASHMEM_NAME_PREFIX) - 1)
35#define ASHMEM_FULL_NAME_LEN (ASHMEM_NAME_LEN + ASHMEM_NAME_PREFIX_LEN)
36
37/*
38 * ashmem_area - anonymous shared memory area
39 * Lifecycle: From our parent file's open() until its release()
40 * Locking: Protected by `ashmem_mutex'
41 * Big Note: Mappings do NOT pin this structure; it dies on close()
42 */
43struct ashmem_area {
44	char name[ASHMEM_FULL_NAME_LEN];/* optional name for /proc/pid/maps */
45	struct list_head unpinned_list;	/* list of all ashmem areas */
46	struct file *file;		/* the shmem-based backing file */
47	size_t size;			/* size of the mapping, in bytes */
48	unsigned long prot_mask;	/* allowed prot bits, as vm_flags */
49};
50
51/*
52 * ashmem_range - represents an interval of unpinned (evictable) pages
53 * Lifecycle: From unpin to pin
54 * Locking: Protected by `ashmem_mutex'
55 */
56struct ashmem_range {
57	struct list_head lru;		/* entry in LRU list */
58	struct list_head unpinned;	/* entry in its area's unpinned list */
59	struct ashmem_area *asma;	/* associated area */
60	size_t pgstart;			/* starting page, inclusive */
61	size_t pgend;			/* ending page, inclusive */
62	unsigned int purged;		/* ASHMEM_NOT or ASHMEM_WAS_PURGED */
63};
64
65/* LRU list of unpinned pages, protected by ashmem_mutex */
66static LIST_HEAD(ashmem_lru_list);
67
68/* Count of pages on our LRU list, protected by ashmem_mutex */
69static unsigned long lru_count;
70
71/*
72 * ashmem_mutex - protects the list of and each individual ashmem_area
73 *
74 * Lock Ordering: ashmex_mutex -> i_mutex -> i_alloc_sem
75 */
76static DEFINE_MUTEX(ashmem_mutex);
77
78static struct kmem_cache *ashmem_area_cachep __read_mostly;
79static struct kmem_cache *ashmem_range_cachep __read_mostly;
80
81#define range_size(range) \
82  ((range)->pgend - (range)->pgstart + 1)
83
84#define range_on_lru(range) \
85  ((range)->purged == ASHMEM_NOT_PURGED)
86
87#define page_range_subsumes_range(range, start, end) \
88  (((range)->pgstart >= (start)) && ((range)->pgend <= (end)))
89
90#define page_range_subsumed_by_range(range, start, end) \
91  (((range)->pgstart <= (start)) && ((range)->pgend >= (end)))
92
93#define page_in_range(range, page) \
94 (((range)->pgstart <= (page)) && ((range)->pgend >= (page)))
95
96#define page_range_in_range(range, start, end) \
97  (page_in_range(range, start) || page_in_range(range, end) || \
98   page_range_subsumes_range(range, start, end))
99
100#define range_before_page(range, page) \
101  ((range)->pgend < (page))
102
103#define PROT_MASK		(PROT_EXEC | PROT_READ | PROT_WRITE)
104
105static inline void lru_add(struct ashmem_range *range)
106{
107	list_add_tail(&range->lru, &ashmem_lru_list);
108	lru_count += range_size(range);
109}
110
111static inline void lru_del(struct ashmem_range *range)
112{
113	list_del(&range->lru);
114	lru_count -= range_size(range);
115}
116
117/*
118 * range_alloc - allocate and initialize a new ashmem_range structure
119 *
120 * 'asma' - associated ashmem_area
121 * 'prev_range' - the previous ashmem_range in the sorted asma->unpinned list
122 * 'purged' - initial purge value (ASMEM_NOT_PURGED or ASHMEM_WAS_PURGED)
123 * 'start' - starting page, inclusive
124 * 'end' - ending page, inclusive
125 *
126 * Caller must hold ashmem_mutex.
127 */
128static int range_alloc(struct ashmem_area *asma,
129		       struct ashmem_range *prev_range, unsigned int purged,
130		       size_t start, size_t end)
131{
132	struct ashmem_range *range;
133
134	range = kmem_cache_zalloc(ashmem_range_cachep, GFP_KERNEL);
135	if (unlikely(!range))
136		return -ENOMEM;
137
138	range->asma = asma;
139	range->pgstart = start;
140	range->pgend = end;
141	range->purged = purged;
142
143	list_add_tail(&range->unpinned, &prev_range->unpinned);
144
145	if (range_on_lru(range))
146		lru_add(range);
147
148	return 0;
149}
150
151static void range_del(struct ashmem_range *range)
152{
153	list_del(&range->unpinned);
154	if (range_on_lru(range))
155		lru_del(range);
156	kmem_cache_free(ashmem_range_cachep, range);
157}
158
159/*
160 * range_shrink - shrinks a range
161 *
162 * Caller must hold ashmem_mutex.
163 */
164static inline void range_shrink(struct ashmem_range *range,
165				size_t start, size_t end)
166{
167	size_t pre = range_size(range);
168
169	range->pgstart = start;
170	range->pgend = end;
171
172	if (range_on_lru(range))
173		lru_count -= pre - range_size(range);
174}
175
176static int ashmem_open(struct inode *inode, struct file *file)
177{
178	struct ashmem_area *asma;
179	int ret;
180
181	ret = nonseekable_open(inode, file);
182	if (unlikely(ret))
183		return ret;
184
185	asma = kmem_cache_zalloc(ashmem_area_cachep, GFP_KERNEL);
186	if (unlikely(!asma))
187		return -ENOMEM;
188
189	INIT_LIST_HEAD(&asma->unpinned_list);
190	memcpy(asma->name, ASHMEM_NAME_PREFIX, ASHMEM_NAME_PREFIX_LEN);
191	asma->prot_mask = PROT_MASK;
192	file->private_data = asma;
193
194	return 0;
195}
196
197static int ashmem_release(struct inode *ignored, struct file *file)
198{
199	struct ashmem_area *asma = file->private_data;
200	struct ashmem_range *range, *next;
201
202	mutex_lock(&ashmem_mutex);
203	list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned)
204		range_del(range);
205	mutex_unlock(&ashmem_mutex);
206
207	if (asma->file)
208		fput(asma->file);
209	kmem_cache_free(ashmem_area_cachep, asma);
210
211	return 0;
212}
213
214static ssize_t ashmem_read(struct file *file, char __user *buf,
215			   size_t len, loff_t *pos)
216{
217	struct ashmem_area *asma = file->private_data;
218	int ret = 0;
219
220	mutex_lock(&ashmem_mutex);
221
222	/* If size is not set, or set to 0, always return EOF. */
223	if (asma->size == 0) {
224		goto out;
225        }
226
227	if (!asma->file) {
228		ret = -EBADF;
229		goto out;
230	}
231
232	ret = asma->file->f_op->read(asma->file, buf, len, pos);
233
234out:
235	mutex_unlock(&ashmem_mutex);
236	return ret;
237}
238
239static inline unsigned long
240calc_vm_may_flags(unsigned long prot)
241{
242	return _calc_vm_trans(prot, PROT_READ,  VM_MAYREAD ) |
243	       _calc_vm_trans(prot, PROT_WRITE, VM_MAYWRITE) |
244	       _calc_vm_trans(prot, PROT_EXEC,  VM_MAYEXEC);
245}
246
247static int ashmem_mmap(struct file *file, struct vm_area_struct *vma)
248{
249	struct ashmem_area *asma = file->private_data;
250	int ret = 0;
251
252	mutex_lock(&ashmem_mutex);
253
254	/* user needs to SET_SIZE before mapping */
255	if (unlikely(!asma->size)) {
256		ret = -EINVAL;
257		goto out;
258	}
259
260	/* requested protection bits must match our allowed protection mask */
261	if (unlikely((vma->vm_flags & ~calc_vm_prot_bits(asma->prot_mask)) &
262						calc_vm_prot_bits(PROT_MASK))) {
263		ret = -EPERM;
264		goto out;
265	}
266	vma->vm_flags &= ~calc_vm_may_flags(~asma->prot_mask);
267
268	if (!asma->file) {
269		char *name = ASHMEM_NAME_DEF;
270		struct file *vmfile;
271
272		if (asma->name[ASHMEM_NAME_PREFIX_LEN] != '\0')
273			name = asma->name;
274
275		/* ... and allocate the backing shmem file */
276		vmfile = shmem_file_setup(name, asma->size, vma->vm_flags);
277		if (unlikely(IS_ERR(vmfile))) {
278			ret = PTR_ERR(vmfile);
279			goto out;
280		}
281		asma->file = vmfile;
282	}
283	get_file(asma->file);
284
285	/*
286	 * XXX - Reworked to use shmem_zero_setup() instead of
287	 * shmem_set_file while we're in staging. -jstultz
288	 */
289	if (vma->vm_flags & VM_SHARED) {
290		ret = shmem_zero_setup(vma);
291		if (ret) {
292			fput(asma->file);
293			goto out;
294		}
295	}
296
297	if (vma->vm_file)
298		fput(vma->vm_file);
299	vma->vm_file = asma->file;
300	vma->vm_flags |= VM_CAN_NONLINEAR;
301
302out:
303	mutex_unlock(&ashmem_mutex);
304	return ret;
305}
306
307/*
308 * ashmem_shrink - our cache shrinker, called from mm/vmscan.c :: shrink_slab
309 *
310 * 'nr_to_scan' is the number of objects (pages) to prune, or 0 to query how
311 * many objects (pages) we have in total.
312 *
313 * 'gfp_mask' is the mask of the allocation that got us into this mess.
314 *
315 * Return value is the number of objects (pages) remaining, or -1 if we cannot
316 * proceed without risk of deadlock (due to gfp_mask).
317 *
318 * We approximate LRU via least-recently-unpinned, jettisoning unpinned partial
319 * chunks of ashmem regions LRU-wise one-at-a-time until we hit 'nr_to_scan'
320 * pages freed.
321 */
322static int ashmem_shrink(struct shrinker *s, int nr_to_scan, gfp_t gfp_mask)
323{
324	struct ashmem_range *range, *next;
325
326	/* We might recurse into filesystem code, so bail out if necessary */
327	if (nr_to_scan && !(gfp_mask & __GFP_FS))
328		return -1;
329	if (!nr_to_scan)
330		return lru_count;
331
332	mutex_lock(&ashmem_mutex);
333	list_for_each_entry_safe(range, next, &ashmem_lru_list, lru) {
334		struct inode *inode = range->asma->file->f_dentry->d_inode;
335		loff_t start = range->pgstart * PAGE_SIZE;
336		loff_t end = (range->pgend + 1) * PAGE_SIZE - 1;
337
338		vmtruncate_range(inode, start, end);
339		range->purged = ASHMEM_WAS_PURGED;
340		lru_del(range);
341
342		nr_to_scan -= range_size(range);
343		if (nr_to_scan <= 0)
344			break;
345	}
346	mutex_unlock(&ashmem_mutex);
347
348	return lru_count;
349}
350
351static struct shrinker ashmem_shrinker = {
352	.shrink = ashmem_shrink,
353	.seeks = DEFAULT_SEEKS * 4,
354};
355
356static int set_prot_mask(struct ashmem_area *asma, unsigned long prot)
357{
358	int ret = 0;
359
360	mutex_lock(&ashmem_mutex);
361
362	/* the user can only remove, not add, protection bits */
363	if (unlikely((asma->prot_mask & prot) != prot)) {
364		ret = -EINVAL;
365		goto out;
366	}
367
368	/* does the application expect PROT_READ to imply PROT_EXEC? */
369	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
370		prot |= PROT_EXEC;
371
372	asma->prot_mask = prot;
373
374out:
375	mutex_unlock(&ashmem_mutex);
376	return ret;
377}
378
379static int set_name(struct ashmem_area *asma, void __user *name)
380{
381	int ret = 0;
382
383	mutex_lock(&ashmem_mutex);
384
385	/* cannot change an existing mapping's name */
386	if (unlikely(asma->file)) {
387		ret = -EINVAL;
388		goto out;
389	}
390
391	if (unlikely(copy_from_user(asma->name + ASHMEM_NAME_PREFIX_LEN,
392				    name, ASHMEM_NAME_LEN)))
393		ret = -EFAULT;
394	asma->name[ASHMEM_FULL_NAME_LEN-1] = '\0';
395
396out:
397	mutex_unlock(&ashmem_mutex);
398
399	return ret;
400}
401
402static int get_name(struct ashmem_area *asma, void __user *name)
403{
404	int ret = 0;
405
406	mutex_lock(&ashmem_mutex);
407	if (asma->name[ASHMEM_NAME_PREFIX_LEN] != '\0') {
408		size_t len;
409
410		/*
411		 * Copying only `len', instead of ASHMEM_NAME_LEN, bytes
412		 * prevents us from revealing one user's stack to another.
413		 */
414		len = strlen(asma->name + ASHMEM_NAME_PREFIX_LEN) + 1;
415		if (unlikely(copy_to_user(name,
416				asma->name + ASHMEM_NAME_PREFIX_LEN, len)))
417			ret = -EFAULT;
418	} else {
419		if (unlikely(copy_to_user(name, ASHMEM_NAME_DEF,
420					  sizeof(ASHMEM_NAME_DEF))))
421			ret = -EFAULT;
422	}
423	mutex_unlock(&ashmem_mutex);
424
425	return ret;
426}
427
428/*
429 * ashmem_pin - pin the given ashmem region, returning whether it was
430 * previously purged (ASHMEM_WAS_PURGED) or not (ASHMEM_NOT_PURGED).
431 *
432 * Caller must hold ashmem_mutex.
433 */
434static int ashmem_pin(struct ashmem_area *asma, size_t pgstart, size_t pgend)
435{
436	struct ashmem_range *range, *next;
437	int ret = ASHMEM_NOT_PURGED;
438
439	list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned) {
440		/* moved past last applicable page; we can short circuit */
441		if (range_before_page(range, pgstart))
442			break;
443
444		/*
445		 * The user can ask us to pin pages that span multiple ranges,
446		 * or to pin pages that aren't even unpinned, so this is messy.
447		 *
448		 * Four cases:
449		 * 1. The requested range subsumes an existing range, so we
450		 *    just remove the entire matching range.
451		 * 2. The requested range overlaps the start of an existing
452		 *    range, so we just update that range.
453		 * 3. The requested range overlaps the end of an existing
454		 *    range, so we just update that range.
455		 * 4. The requested range punches a hole in an existing range,
456		 *    so we have to update one side of the range and then
457		 *    create a new range for the other side.
458		 */
459		if (page_range_in_range(range, pgstart, pgend)) {
460			ret |= range->purged;
461
462			/* Case #1: Easy. Just nuke the whole thing. */
463			if (page_range_subsumes_range(range, pgstart, pgend)) {
464				range_del(range);
465				continue;
466			}
467
468			/* Case #2: We overlap from the start, so adjust it */
469			if (range->pgstart >= pgstart) {
470				range_shrink(range, pgend + 1, range->pgend);
471				continue;
472			}
473
474			/* Case #3: We overlap from the rear, so adjust it */
475			if (range->pgend <= pgend) {
476				range_shrink(range, range->pgstart, pgstart-1);
477				continue;
478			}
479
480			/*
481			 * Case #4: We eat a chunk out of the middle. A bit
482			 * more complicated, we allocate a new range for the
483			 * second half and adjust the first chunk's endpoint.
484			 */
485			range_alloc(asma, range, range->purged,
486				    pgend + 1, range->pgend);
487			range_shrink(range, range->pgstart, pgstart - 1);
488			break;
489		}
490	}
491
492	return ret;
493}
494
495/*
496 * ashmem_unpin - unpin the given range of pages. Returns zero on success.
497 *
498 * Caller must hold ashmem_mutex.
499 */
500static int ashmem_unpin(struct ashmem_area *asma, size_t pgstart, size_t pgend)
501{
502	struct ashmem_range *range, *next;
503	unsigned int purged = ASHMEM_NOT_PURGED;
504
505restart:
506	list_for_each_entry_safe(range, next, &asma->unpinned_list, unpinned) {
507		/* short circuit: this is our insertion point */
508		if (range_before_page(range, pgstart))
509			break;
510
511		/*
512		 * The user can ask us to unpin pages that are already entirely
513		 * or partially pinned. We handle those two cases here.
514		 */
515		if (page_range_subsumed_by_range(range, pgstart, pgend))
516			return 0;
517		if (page_range_in_range(range, pgstart, pgend)) {
518			pgstart = min_t(size_t, range->pgstart, pgstart),
519			pgend = max_t(size_t, range->pgend, pgend);
520			purged |= range->purged;
521			range_del(range);
522			goto restart;
523		}
524	}
525
526	return range_alloc(asma, range, purged, pgstart, pgend);
527}
528
529/*
530 * ashmem_get_pin_status - Returns ASHMEM_IS_UNPINNED if _any_ pages in the
531 * given interval are unpinned and ASHMEM_IS_PINNED otherwise.
532 *
533 * Caller must hold ashmem_mutex.
534 */
535static int ashmem_get_pin_status(struct ashmem_area *asma, size_t pgstart,
536				 size_t pgend)
537{
538	struct ashmem_range *range;
539	int ret = ASHMEM_IS_PINNED;
540
541	list_for_each_entry(range, &asma->unpinned_list, unpinned) {
542		if (range_before_page(range, pgstart))
543			break;
544		if (page_range_in_range(range, pgstart, pgend)) {
545			ret = ASHMEM_IS_UNPINNED;
546			break;
547		}
548	}
549
550	return ret;
551}
552
553static int ashmem_pin_unpin(struct ashmem_area *asma, unsigned long cmd,
554			    void __user *p)
555{
556	struct ashmem_pin pin;
557	size_t pgstart, pgend;
558	int ret = -EINVAL;
559
560	if (unlikely(!asma->file))
561		return -EINVAL;
562
563	if (unlikely(copy_from_user(&pin, p, sizeof(pin))))
564		return -EFAULT;
565
566	/* per custom, you can pass zero for len to mean "everything onward" */
567	if (!pin.len)
568		pin.len = PAGE_ALIGN(asma->size) - pin.offset;
569
570	if (unlikely((pin.offset | pin.len) & ~PAGE_MASK))
571		return -EINVAL;
572
573	if (unlikely(((__u32) -1) - pin.offset < pin.len))
574		return -EINVAL;
575
576	if (unlikely(PAGE_ALIGN(asma->size) < pin.offset + pin.len))
577		return -EINVAL;
578
579	pgstart = pin.offset / PAGE_SIZE;
580	pgend = pgstart + (pin.len / PAGE_SIZE) - 1;
581
582	mutex_lock(&ashmem_mutex);
583
584	switch (cmd) {
585	case ASHMEM_PIN:
586		ret = ashmem_pin(asma, pgstart, pgend);
587		break;
588	case ASHMEM_UNPIN:
589		ret = ashmem_unpin(asma, pgstart, pgend);
590		break;
591	case ASHMEM_GET_PIN_STATUS:
592		ret = ashmem_get_pin_status(asma, pgstart, pgend);
593		break;
594	}
595
596	mutex_unlock(&ashmem_mutex);
597
598	return ret;
599}
600
601static long ashmem_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
602{
603	struct ashmem_area *asma = file->private_data;
604	long ret = -ENOTTY;
605
606	switch (cmd) {
607	case ASHMEM_SET_NAME:
608		ret = set_name(asma, (void __user *) arg);
609		break;
610	case ASHMEM_GET_NAME:
611		ret = get_name(asma, (void __user *) arg);
612		break;
613	case ASHMEM_SET_SIZE:
614		ret = -EINVAL;
615		if (!asma->file) {
616			ret = 0;
617			asma->size = (size_t) arg;
618		}
619		break;
620	case ASHMEM_GET_SIZE:
621		ret = asma->size;
622		break;
623	case ASHMEM_SET_PROT_MASK:
624		ret = set_prot_mask(asma, arg);
625		break;
626	case ASHMEM_GET_PROT_MASK:
627		ret = asma->prot_mask;
628		break;
629	case ASHMEM_PIN:
630	case ASHMEM_UNPIN:
631	case ASHMEM_GET_PIN_STATUS:
632		ret = ashmem_pin_unpin(asma, cmd, (void __user *) arg);
633		break;
634	case ASHMEM_PURGE_ALL_CACHES:
635		ret = -EPERM;
636		if (capable(CAP_SYS_ADMIN)) {
637			ret = ashmem_shrink(&ashmem_shrinker, 0, GFP_KERNEL);
638			ashmem_shrink(&ashmem_shrinker, ret, GFP_KERNEL);
639		}
640		break;
641	}
642
643	return ret;
644}
645
646static struct file_operations ashmem_fops = {
647	.owner = THIS_MODULE,
648	.open = ashmem_open,
649	.release = ashmem_release,
650        .read = ashmem_read,
651	.mmap = ashmem_mmap,
652	.unlocked_ioctl = ashmem_ioctl,
653	.compat_ioctl = ashmem_ioctl,
654};
655
656static struct miscdevice ashmem_misc = {
657	.minor = MISC_DYNAMIC_MINOR,
658	.name = "ashmem",
659	.fops = &ashmem_fops,
660};
661
662static int __init ashmem_init(void)
663{
664	int ret;
665
666	ashmem_area_cachep = kmem_cache_create("ashmem_area_cache",
667					  sizeof(struct ashmem_area),
668					  0, 0, NULL);
669	if (unlikely(!ashmem_area_cachep)) {
670		printk(KERN_ERR "ashmem: failed to create slab cache\n");
671		return -ENOMEM;
672	}
673
674	ashmem_range_cachep = kmem_cache_create("ashmem_range_cache",
675					  sizeof(struct ashmem_range),
676					  0, 0, NULL);
677	if (unlikely(!ashmem_range_cachep)) {
678		printk(KERN_ERR "ashmem: failed to create slab cache\n");
679		return -ENOMEM;
680	}
681
682	ret = misc_register(&ashmem_misc);
683	if (unlikely(ret)) {
684		printk(KERN_ERR "ashmem: failed to register misc device!\n");
685		return ret;
686	}
687
688	register_shrinker(&ashmem_shrinker);
689
690	printk(KERN_INFO "ashmem: initialized\n");
691
692	return 0;
693}
694
695static void __exit ashmem_exit(void)
696{
697	int ret;
698
699	unregister_shrinker(&ashmem_shrinker);
700
701	ret = misc_deregister(&ashmem_misc);
702	if (unlikely(ret))
703		printk(KERN_ERR "ashmem: failed to unregister misc device!\n");
704
705	kmem_cache_destroy(ashmem_range_cachep);
706	kmem_cache_destroy(ashmem_area_cachep);
707
708	printk(KERN_INFO "ashmem: unloaded\n");
709}
710
711module_init(ashmem_init);
712module_exit(ashmem_exit);
713
714MODULE_LICENSE("GPL");
715