Loading...
1/*
2 * memfd_create system call and file sealing support
3 *
4 * Code was originally included in shmem.c, and broken out to facilitate
5 * use by hugetlbfs as well as tmpfs.
6 *
7 * This file is released under the GPL.
8 */
9
10#include <linux/fs.h>
11#include <linux/vfs.h>
12#include <linux/pagemap.h>
13#include <linux/file.h>
14#include <linux/mm.h>
15#include <linux/sched/signal.h>
16#include <linux/khugepaged.h>
17#include <linux/syscalls.h>
18#include <linux/hugetlb.h>
19#include <linux/shmem_fs.h>
20#include <linux/memfd.h>
21#include <uapi/linux/memfd.h>
22
23/*
24 * We need a tag: a new tag would expand every xa_node by 8 bytes,
25 * so reuse a tag which we firmly believe is never set or cleared on tmpfs
26 * or hugetlbfs because they are memory only filesystems.
27 */
28#define MEMFD_TAG_PINNED PAGECACHE_TAG_TOWRITE
29#define LAST_SCAN 4 /* about 150ms max */
30
31static void memfd_tag_pins(struct xa_state *xas)
32{
33 struct page *page;
34 int latency = 0;
35 int cache_count;
36
37 lru_add_drain();
38
39 xas_lock_irq(xas);
40 xas_for_each(xas, page, ULONG_MAX) {
41 cache_count = 1;
42 if (!xa_is_value(page) &&
43 PageTransHuge(page) && !PageHuge(page))
44 cache_count = HPAGE_PMD_NR;
45
46 if (!xa_is_value(page) &&
47 page_count(page) - total_mapcount(page) != cache_count)
48 xas_set_mark(xas, MEMFD_TAG_PINNED);
49 if (cache_count != 1)
50 xas_set(xas, page->index + cache_count);
51
52 latency += cache_count;
53 if (latency < XA_CHECK_SCHED)
54 continue;
55 latency = 0;
56
57 xas_pause(xas);
58 xas_unlock_irq(xas);
59 cond_resched();
60 xas_lock_irq(xas);
61 }
62 xas_unlock_irq(xas);
63}
64
65/*
66 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
67 * via get_user_pages(), drivers might have some pending I/O without any active
68 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
69 * and see whether it has an elevated ref-count. If so, we tag them and wait for
70 * them to be dropped.
71 * The caller must guarantee that no new user will acquire writable references
72 * to those pages to avoid races.
73 */
74static int memfd_wait_for_pins(struct address_space *mapping)
75{
76 XA_STATE(xas, &mapping->i_pages, 0);
77 struct page *page;
78 int error, scan;
79
80 memfd_tag_pins(&xas);
81
82 error = 0;
83 for (scan = 0; scan <= LAST_SCAN; scan++) {
84 int latency = 0;
85 int cache_count;
86
87 if (!xas_marked(&xas, MEMFD_TAG_PINNED))
88 break;
89
90 if (!scan)
91 lru_add_drain_all();
92 else if (schedule_timeout_killable((HZ << scan) / 200))
93 scan = LAST_SCAN;
94
95 xas_set(&xas, 0);
96 xas_lock_irq(&xas);
97 xas_for_each_marked(&xas, page, ULONG_MAX, MEMFD_TAG_PINNED) {
98 bool clear = true;
99
100 cache_count = 1;
101 if (!xa_is_value(page) &&
102 PageTransHuge(page) && !PageHuge(page))
103 cache_count = HPAGE_PMD_NR;
104
105 if (!xa_is_value(page) && cache_count !=
106 page_count(page) - total_mapcount(page)) {
107 /*
108 * On the last scan, we clean up all those tags
109 * we inserted; but make a note that we still
110 * found pages pinned.
111 */
112 if (scan == LAST_SCAN)
113 error = -EBUSY;
114 else
115 clear = false;
116 }
117 if (clear)
118 xas_clear_mark(&xas, MEMFD_TAG_PINNED);
119
120 latency += cache_count;
121 if (latency < XA_CHECK_SCHED)
122 continue;
123 latency = 0;
124
125 xas_pause(&xas);
126 xas_unlock_irq(&xas);
127 cond_resched();
128 xas_lock_irq(&xas);
129 }
130 xas_unlock_irq(&xas);
131 }
132
133 return error;
134}
135
136static unsigned int *memfd_file_seals_ptr(struct file *file)
137{
138 if (shmem_file(file))
139 return &SHMEM_I(file_inode(file))->seals;
140
141#ifdef CONFIG_HUGETLBFS
142 if (is_file_hugepages(file))
143 return &HUGETLBFS_I(file_inode(file))->seals;
144#endif
145
146 return NULL;
147}
148
149#define F_ALL_SEALS (F_SEAL_SEAL | \
150 F_SEAL_SHRINK | \
151 F_SEAL_GROW | \
152 F_SEAL_WRITE | \
153 F_SEAL_FUTURE_WRITE)
154
155static int memfd_add_seals(struct file *file, unsigned int seals)
156{
157 struct inode *inode = file_inode(file);
158 unsigned int *file_seals;
159 int error;
160
161 /*
162 * SEALING
163 * Sealing allows multiple parties to share a tmpfs or hugetlbfs file
164 * but restrict access to a specific subset of file operations. Seals
165 * can only be added, but never removed. This way, mutually untrusted
166 * parties can share common memory regions with a well-defined policy.
167 * A malicious peer can thus never perform unwanted operations on a
168 * shared object.
169 *
170 * Seals are only supported on special tmpfs or hugetlbfs files and
171 * always affect the whole underlying inode. Once a seal is set, it
172 * may prevent some kinds of access to the file. Currently, the
173 * following seals are defined:
174 * SEAL_SEAL: Prevent further seals from being set on this file
175 * SEAL_SHRINK: Prevent the file from shrinking
176 * SEAL_GROW: Prevent the file from growing
177 * SEAL_WRITE: Prevent write access to the file
178 *
179 * As we don't require any trust relationship between two parties, we
180 * must prevent seals from being removed. Therefore, sealing a file
181 * only adds a given set of seals to the file, it never touches
182 * existing seals. Furthermore, the "setting seals"-operation can be
183 * sealed itself, which basically prevents any further seal from being
184 * added.
185 *
186 * Semantics of sealing are only defined on volatile files. Only
187 * anonymous tmpfs and hugetlbfs files support sealing. More
188 * importantly, seals are never written to disk. Therefore, there's
189 * no plan to support it on other file types.
190 */
191
192 if (!(file->f_mode & FMODE_WRITE))
193 return -EPERM;
194 if (seals & ~(unsigned int)F_ALL_SEALS)
195 return -EINVAL;
196
197 inode_lock(inode);
198
199 file_seals = memfd_file_seals_ptr(file);
200 if (!file_seals) {
201 error = -EINVAL;
202 goto unlock;
203 }
204
205 if (*file_seals & F_SEAL_SEAL) {
206 error = -EPERM;
207 goto unlock;
208 }
209
210 if ((seals & F_SEAL_WRITE) && !(*file_seals & F_SEAL_WRITE)) {
211 error = mapping_deny_writable(file->f_mapping);
212 if (error)
213 goto unlock;
214
215 error = memfd_wait_for_pins(file->f_mapping);
216 if (error) {
217 mapping_allow_writable(file->f_mapping);
218 goto unlock;
219 }
220 }
221
222 *file_seals |= seals;
223 error = 0;
224
225unlock:
226 inode_unlock(inode);
227 return error;
228}
229
230static int memfd_get_seals(struct file *file)
231{
232 unsigned int *seals = memfd_file_seals_ptr(file);
233
234 return seals ? *seals : -EINVAL;
235}
236
237long memfd_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
238{
239 long error;
240
241 switch (cmd) {
242 case F_ADD_SEALS:
243 /* disallow upper 32bit */
244 if (arg > UINT_MAX)
245 return -EINVAL;
246
247 error = memfd_add_seals(file, arg);
248 break;
249 case F_GET_SEALS:
250 error = memfd_get_seals(file);
251 break;
252 default:
253 error = -EINVAL;
254 break;
255 }
256
257 return error;
258}
259
260#define MFD_NAME_PREFIX "memfd:"
261#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
262#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
263
264#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
265
266SYSCALL_DEFINE2(memfd_create,
267 const char __user *, uname,
268 unsigned int, flags)
269{
270 unsigned int *file_seals;
271 struct file *file;
272 int fd, error;
273 char *name;
274 long len;
275
276 if (!(flags & MFD_HUGETLB)) {
277 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
278 return -EINVAL;
279 } else {
280 /* Allow huge page size encoding in flags. */
281 if (flags & ~(unsigned int)(MFD_ALL_FLAGS |
282 (MFD_HUGE_MASK << MFD_HUGE_SHIFT)))
283 return -EINVAL;
284 }
285
286 /* length includes terminating zero */
287 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
288 if (len <= 0)
289 return -EFAULT;
290 if (len > MFD_NAME_MAX_LEN + 1)
291 return -EINVAL;
292
293 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_KERNEL);
294 if (!name)
295 return -ENOMEM;
296
297 strcpy(name, MFD_NAME_PREFIX);
298 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
299 error = -EFAULT;
300 goto err_name;
301 }
302
303 /* terminating-zero may have changed after strnlen_user() returned */
304 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
305 error = -EFAULT;
306 goto err_name;
307 }
308
309 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
310 if (fd < 0) {
311 error = fd;
312 goto err_name;
313 }
314
315 if (flags & MFD_HUGETLB) {
316 file = hugetlb_file_setup(name, 0, VM_NORESERVE,
317 HUGETLB_ANONHUGE_INODE,
318 (flags >> MFD_HUGE_SHIFT) &
319 MFD_HUGE_MASK);
320 } else
321 file = shmem_file_setup(name, 0, VM_NORESERVE);
322 if (IS_ERR(file)) {
323 error = PTR_ERR(file);
324 goto err_fd;
325 }
326 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
327 file->f_flags |= O_LARGEFILE;
328
329 if (flags & MFD_ALLOW_SEALING) {
330 file_seals = memfd_file_seals_ptr(file);
331 *file_seals &= ~F_SEAL_SEAL;
332 }
333
334 fd_install(fd, file);
335 kfree(name);
336 return fd;
337
338err_fd:
339 put_unused_fd(fd);
340err_name:
341 kfree(name);
342 return error;
343}
1/*
2 * memfd_create system call and file sealing support
3 *
4 * Code was originally included in shmem.c, and broken out to facilitate
5 * use by hugetlbfs as well as tmpfs.
6 *
7 * This file is released under the GPL.
8 */
9
10#include <linux/fs.h>
11#include <linux/vfs.h>
12#include <linux/pagemap.h>
13#include <linux/file.h>
14#include <linux/mm.h>
15#include <linux/sched/signal.h>
16#include <linux/khugepaged.h>
17#include <linux/syscalls.h>
18#include <linux/hugetlb.h>
19#include <linux/shmem_fs.h>
20#include <linux/memfd.h>
21#include <uapi/linux/memfd.h>
22
23/*
24 * We need a tag: a new tag would expand every xa_node by 8 bytes,
25 * so reuse a tag which we firmly believe is never set or cleared on tmpfs
26 * or hugetlbfs because they are memory only filesystems.
27 */
28#define MEMFD_TAG_PINNED PAGECACHE_TAG_TOWRITE
29#define LAST_SCAN 4 /* about 150ms max */
30
31static void memfd_tag_pins(struct xa_state *xas)
32{
33 struct page *page;
34 unsigned int tagged = 0;
35
36 lru_add_drain();
37
38 xas_lock_irq(xas);
39 xas_for_each(xas, page, ULONG_MAX) {
40 if (xa_is_value(page))
41 continue;
42 page = find_subpage(page, xas->xa_index);
43 if (page_count(page) - page_mapcount(page) > 1)
44 xas_set_mark(xas, MEMFD_TAG_PINNED);
45
46 if (++tagged % XA_CHECK_SCHED)
47 continue;
48
49 xas_pause(xas);
50 xas_unlock_irq(xas);
51 cond_resched();
52 xas_lock_irq(xas);
53 }
54 xas_unlock_irq(xas);
55}
56
57/*
58 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
59 * via get_user_pages(), drivers might have some pending I/O without any active
60 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
61 * and see whether it has an elevated ref-count. If so, we tag them and wait for
62 * them to be dropped.
63 * The caller must guarantee that no new user will acquire writable references
64 * to those pages to avoid races.
65 */
66static int memfd_wait_for_pins(struct address_space *mapping)
67{
68 XA_STATE(xas, &mapping->i_pages, 0);
69 struct page *page;
70 int error, scan;
71
72 memfd_tag_pins(&xas);
73
74 error = 0;
75 for (scan = 0; scan <= LAST_SCAN; scan++) {
76 unsigned int tagged = 0;
77
78 if (!xas_marked(&xas, MEMFD_TAG_PINNED))
79 break;
80
81 if (!scan)
82 lru_add_drain_all();
83 else if (schedule_timeout_killable((HZ << scan) / 200))
84 scan = LAST_SCAN;
85
86 xas_set(&xas, 0);
87 xas_lock_irq(&xas);
88 xas_for_each_marked(&xas, page, ULONG_MAX, MEMFD_TAG_PINNED) {
89 bool clear = true;
90 if (xa_is_value(page))
91 continue;
92 page = find_subpage(page, xas.xa_index);
93 if (page_count(page) - page_mapcount(page) != 1) {
94 /*
95 * On the last scan, we clean up all those tags
96 * we inserted; but make a note that we still
97 * found pages pinned.
98 */
99 if (scan == LAST_SCAN)
100 error = -EBUSY;
101 else
102 clear = false;
103 }
104 if (clear)
105 xas_clear_mark(&xas, MEMFD_TAG_PINNED);
106 if (++tagged % XA_CHECK_SCHED)
107 continue;
108
109 xas_pause(&xas);
110 xas_unlock_irq(&xas);
111 cond_resched();
112 xas_lock_irq(&xas);
113 }
114 xas_unlock_irq(&xas);
115 }
116
117 return error;
118}
119
120static unsigned int *memfd_file_seals_ptr(struct file *file)
121{
122 if (shmem_file(file))
123 return &SHMEM_I(file_inode(file))->seals;
124
125#ifdef CONFIG_HUGETLBFS
126 if (is_file_hugepages(file))
127 return &HUGETLBFS_I(file_inode(file))->seals;
128#endif
129
130 return NULL;
131}
132
133#define F_ALL_SEALS (F_SEAL_SEAL | \
134 F_SEAL_SHRINK | \
135 F_SEAL_GROW | \
136 F_SEAL_WRITE | \
137 F_SEAL_FUTURE_WRITE)
138
139static int memfd_add_seals(struct file *file, unsigned int seals)
140{
141 struct inode *inode = file_inode(file);
142 unsigned int *file_seals;
143 int error;
144
145 /*
146 * SEALING
147 * Sealing allows multiple parties to share a tmpfs or hugetlbfs file
148 * but restrict access to a specific subset of file operations. Seals
149 * can only be added, but never removed. This way, mutually untrusted
150 * parties can share common memory regions with a well-defined policy.
151 * A malicious peer can thus never perform unwanted operations on a
152 * shared object.
153 *
154 * Seals are only supported on special tmpfs or hugetlbfs files and
155 * always affect the whole underlying inode. Once a seal is set, it
156 * may prevent some kinds of access to the file. Currently, the
157 * following seals are defined:
158 * SEAL_SEAL: Prevent further seals from being set on this file
159 * SEAL_SHRINK: Prevent the file from shrinking
160 * SEAL_GROW: Prevent the file from growing
161 * SEAL_WRITE: Prevent write access to the file
162 *
163 * As we don't require any trust relationship between two parties, we
164 * must prevent seals from being removed. Therefore, sealing a file
165 * only adds a given set of seals to the file, it never touches
166 * existing seals. Furthermore, the "setting seals"-operation can be
167 * sealed itself, which basically prevents any further seal from being
168 * added.
169 *
170 * Semantics of sealing are only defined on volatile files. Only
171 * anonymous tmpfs and hugetlbfs files support sealing. More
172 * importantly, seals are never written to disk. Therefore, there's
173 * no plan to support it on other file types.
174 */
175
176 if (!(file->f_mode & FMODE_WRITE))
177 return -EPERM;
178 if (seals & ~(unsigned int)F_ALL_SEALS)
179 return -EINVAL;
180
181 inode_lock(inode);
182
183 file_seals = memfd_file_seals_ptr(file);
184 if (!file_seals) {
185 error = -EINVAL;
186 goto unlock;
187 }
188
189 if (*file_seals & F_SEAL_SEAL) {
190 error = -EPERM;
191 goto unlock;
192 }
193
194 if ((seals & F_SEAL_WRITE) && !(*file_seals & F_SEAL_WRITE)) {
195 error = mapping_deny_writable(file->f_mapping);
196 if (error)
197 goto unlock;
198
199 error = memfd_wait_for_pins(file->f_mapping);
200 if (error) {
201 mapping_allow_writable(file->f_mapping);
202 goto unlock;
203 }
204 }
205
206 *file_seals |= seals;
207 error = 0;
208
209unlock:
210 inode_unlock(inode);
211 return error;
212}
213
214static int memfd_get_seals(struct file *file)
215{
216 unsigned int *seals = memfd_file_seals_ptr(file);
217
218 return seals ? *seals : -EINVAL;
219}
220
221long memfd_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
222{
223 long error;
224
225 switch (cmd) {
226 case F_ADD_SEALS:
227 /* disallow upper 32bit */
228 if (arg > UINT_MAX)
229 return -EINVAL;
230
231 error = memfd_add_seals(file, arg);
232 break;
233 case F_GET_SEALS:
234 error = memfd_get_seals(file);
235 break;
236 default:
237 error = -EINVAL;
238 break;
239 }
240
241 return error;
242}
243
244#define MFD_NAME_PREFIX "memfd:"
245#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
246#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
247
248#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
249
250SYSCALL_DEFINE2(memfd_create,
251 const char __user *, uname,
252 unsigned int, flags)
253{
254 unsigned int *file_seals;
255 struct file *file;
256 int fd, error;
257 char *name;
258 long len;
259
260 if (!(flags & MFD_HUGETLB)) {
261 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
262 return -EINVAL;
263 } else {
264 /* Allow huge page size encoding in flags. */
265 if (flags & ~(unsigned int)(MFD_ALL_FLAGS |
266 (MFD_HUGE_MASK << MFD_HUGE_SHIFT)))
267 return -EINVAL;
268 }
269
270 /* length includes terminating zero */
271 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
272 if (len <= 0)
273 return -EFAULT;
274 if (len > MFD_NAME_MAX_LEN + 1)
275 return -EINVAL;
276
277 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_KERNEL);
278 if (!name)
279 return -ENOMEM;
280
281 strcpy(name, MFD_NAME_PREFIX);
282 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
283 error = -EFAULT;
284 goto err_name;
285 }
286
287 /* terminating-zero may have changed after strnlen_user() returned */
288 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
289 error = -EFAULT;
290 goto err_name;
291 }
292
293 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
294 if (fd < 0) {
295 error = fd;
296 goto err_name;
297 }
298
299 if (flags & MFD_HUGETLB) {
300 struct user_struct *user = NULL;
301
302 file = hugetlb_file_setup(name, 0, VM_NORESERVE, &user,
303 HUGETLB_ANONHUGE_INODE,
304 (flags >> MFD_HUGE_SHIFT) &
305 MFD_HUGE_MASK);
306 } else
307 file = shmem_file_setup(name, 0, VM_NORESERVE);
308 if (IS_ERR(file)) {
309 error = PTR_ERR(file);
310 goto err_fd;
311 }
312 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
313 file->f_flags |= O_LARGEFILE;
314
315 if (flags & MFD_ALLOW_SEALING) {
316 file_seals = memfd_file_seals_ptr(file);
317 *file_seals &= ~F_SEAL_SEAL;
318 }
319
320 fd_install(fd, file);
321 kfree(name);
322 return fd;
323
324err_fd:
325 put_unused_fd(fd);
326err_name:
327 kfree(name);
328 return error;
329}