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 <linux/pid_namespace.h>
22#include <uapi/linux/memfd.h>
23
24/*
25 * We need a tag: a new tag would expand every xa_node by 8 bytes,
26 * so reuse a tag which we firmly believe is never set or cleared on tmpfs
27 * or hugetlbfs because they are memory only filesystems.
28 */
29#define MEMFD_TAG_PINNED PAGECACHE_TAG_TOWRITE
30#define LAST_SCAN 4 /* about 150ms max */
31
32static bool memfd_folio_has_extra_refs(struct folio *folio)
33{
34 return folio_ref_count(folio) - folio_mapcount(folio) !=
35 folio_nr_pages(folio);
36}
37
38static void memfd_tag_pins(struct xa_state *xas)
39{
40 struct folio *folio;
41 int latency = 0;
42
43 lru_add_drain();
44
45 xas_lock_irq(xas);
46 xas_for_each(xas, folio, ULONG_MAX) {
47 if (!xa_is_value(folio) && memfd_folio_has_extra_refs(folio))
48 xas_set_mark(xas, MEMFD_TAG_PINNED);
49
50 if (++latency < XA_CHECK_SCHED)
51 continue;
52 latency = 0;
53
54 xas_pause(xas);
55 xas_unlock_irq(xas);
56 cond_resched();
57 xas_lock_irq(xas);
58 }
59 xas_unlock_irq(xas);
60}
61
62/*
63 * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
64 * via get_user_pages(), drivers might have some pending I/O without any active
65 * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all folios
66 * and see whether it has an elevated ref-count. If so, we tag them and wait for
67 * them to be dropped.
68 * The caller must guarantee that no new user will acquire writable references
69 * to those folios to avoid races.
70 */
71static int memfd_wait_for_pins(struct address_space *mapping)
72{
73 XA_STATE(xas, &mapping->i_pages, 0);
74 struct folio *folio;
75 int error, scan;
76
77 memfd_tag_pins(&xas);
78
79 error = 0;
80 for (scan = 0; scan <= LAST_SCAN; scan++) {
81 int latency = 0;
82
83 if (!xas_marked(&xas, MEMFD_TAG_PINNED))
84 break;
85
86 if (!scan)
87 lru_add_drain_all();
88 else if (schedule_timeout_killable((HZ << scan) / 200))
89 scan = LAST_SCAN;
90
91 xas_set(&xas, 0);
92 xas_lock_irq(&xas);
93 xas_for_each_marked(&xas, folio, ULONG_MAX, MEMFD_TAG_PINNED) {
94 bool clear = true;
95
96 if (!xa_is_value(folio) &&
97 memfd_folio_has_extra_refs(folio)) {
98 /*
99 * On the last scan, we clean up all those tags
100 * we inserted; but make a note that we still
101 * found folios pinned.
102 */
103 if (scan == LAST_SCAN)
104 error = -EBUSY;
105 else
106 clear = false;
107 }
108 if (clear)
109 xas_clear_mark(&xas, MEMFD_TAG_PINNED);
110
111 if (++latency < XA_CHECK_SCHED)
112 continue;
113 latency = 0;
114
115 xas_pause(&xas);
116 xas_unlock_irq(&xas);
117 cond_resched();
118 xas_lock_irq(&xas);
119 }
120 xas_unlock_irq(&xas);
121 }
122
123 return error;
124}
125
126static unsigned int *memfd_file_seals_ptr(struct file *file)
127{
128 if (shmem_file(file))
129 return &SHMEM_I(file_inode(file))->seals;
130
131#ifdef CONFIG_HUGETLBFS
132 if (is_file_hugepages(file))
133 return &HUGETLBFS_I(file_inode(file))->seals;
134#endif
135
136 return NULL;
137}
138
139#define F_ALL_SEALS (F_SEAL_SEAL | \
140 F_SEAL_EXEC | \
141 F_SEAL_SHRINK | \
142 F_SEAL_GROW | \
143 F_SEAL_WRITE | \
144 F_SEAL_FUTURE_WRITE)
145
146static int memfd_add_seals(struct file *file, unsigned int seals)
147{
148 struct inode *inode = file_inode(file);
149 unsigned int *file_seals;
150 int error;
151
152 /*
153 * SEALING
154 * Sealing allows multiple parties to share a tmpfs or hugetlbfs file
155 * but restrict access to a specific subset of file operations. Seals
156 * can only be added, but never removed. This way, mutually untrusted
157 * parties can share common memory regions with a well-defined policy.
158 * A malicious peer can thus never perform unwanted operations on a
159 * shared object.
160 *
161 * Seals are only supported on special tmpfs or hugetlbfs files and
162 * always affect the whole underlying inode. Once a seal is set, it
163 * may prevent some kinds of access to the file. Currently, the
164 * following seals are defined:
165 * SEAL_SEAL: Prevent further seals from being set on this file
166 * SEAL_SHRINK: Prevent the file from shrinking
167 * SEAL_GROW: Prevent the file from growing
168 * SEAL_WRITE: Prevent write access to the file
169 * SEAL_EXEC: Prevent modification of the exec bits in the file mode
170 *
171 * As we don't require any trust relationship between two parties, we
172 * must prevent seals from being removed. Therefore, sealing a file
173 * only adds a given set of seals to the file, it never touches
174 * existing seals. Furthermore, the "setting seals"-operation can be
175 * sealed itself, which basically prevents any further seal from being
176 * added.
177 *
178 * Semantics of sealing are only defined on volatile files. Only
179 * anonymous tmpfs and hugetlbfs files support sealing. More
180 * importantly, seals are never written to disk. Therefore, there's
181 * no plan to support it on other file types.
182 */
183
184 if (!(file->f_mode & FMODE_WRITE))
185 return -EPERM;
186 if (seals & ~(unsigned int)F_ALL_SEALS)
187 return -EINVAL;
188
189 inode_lock(inode);
190
191 file_seals = memfd_file_seals_ptr(file);
192 if (!file_seals) {
193 error = -EINVAL;
194 goto unlock;
195 }
196
197 if (*file_seals & F_SEAL_SEAL) {
198 error = -EPERM;
199 goto unlock;
200 }
201
202 if ((seals & F_SEAL_WRITE) && !(*file_seals & F_SEAL_WRITE)) {
203 error = mapping_deny_writable(file->f_mapping);
204 if (error)
205 goto unlock;
206
207 error = memfd_wait_for_pins(file->f_mapping);
208 if (error) {
209 mapping_allow_writable(file->f_mapping);
210 goto unlock;
211 }
212 }
213
214 /*
215 * SEAL_EXEC implys SEAL_WRITE, making W^X from the start.
216 */
217 if (seals & F_SEAL_EXEC && inode->i_mode & 0111)
218 seals |= F_SEAL_SHRINK|F_SEAL_GROW|F_SEAL_WRITE|F_SEAL_FUTURE_WRITE;
219
220 *file_seals |= seals;
221 error = 0;
222
223unlock:
224 inode_unlock(inode);
225 return error;
226}
227
228static int memfd_get_seals(struct file *file)
229{
230 unsigned int *seals = memfd_file_seals_ptr(file);
231
232 return seals ? *seals : -EINVAL;
233}
234
235long memfd_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
236{
237 long error;
238
239 switch (cmd) {
240 case F_ADD_SEALS:
241 error = memfd_add_seals(file, arg);
242 break;
243 case F_GET_SEALS:
244 error = memfd_get_seals(file);
245 break;
246 default:
247 error = -EINVAL;
248 break;
249 }
250
251 return error;
252}
253
254#define MFD_NAME_PREFIX "memfd:"
255#define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
256#define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
257
258#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB | MFD_NOEXEC_SEAL | MFD_EXEC)
259
260static int check_sysctl_memfd_noexec(unsigned int *flags)
261{
262#ifdef CONFIG_SYSCTL
263 struct pid_namespace *ns = task_active_pid_ns(current);
264 int sysctl = pidns_memfd_noexec_scope(ns);
265
266 if (!(*flags & (MFD_EXEC | MFD_NOEXEC_SEAL))) {
267 if (sysctl >= MEMFD_NOEXEC_SCOPE_NOEXEC_SEAL)
268 *flags |= MFD_NOEXEC_SEAL;
269 else
270 *flags |= MFD_EXEC;
271 }
272
273 if (!(*flags & MFD_NOEXEC_SEAL) && sysctl >= MEMFD_NOEXEC_SCOPE_NOEXEC_ENFORCED) {
274 pr_err_ratelimited(
275 "%s[%d]: memfd_create() requires MFD_NOEXEC_SEAL with vm.memfd_noexec=%d\n",
276 current->comm, task_pid_nr(current), sysctl);
277 return -EACCES;
278 }
279#endif
280 return 0;
281}
282
283SYSCALL_DEFINE2(memfd_create,
284 const char __user *, uname,
285 unsigned int, flags)
286{
287 unsigned int *file_seals;
288 struct file *file;
289 int fd, error;
290 char *name;
291 long len;
292
293 if (!(flags & MFD_HUGETLB)) {
294 if (flags & ~(unsigned int)MFD_ALL_FLAGS)
295 return -EINVAL;
296 } else {
297 /* Allow huge page size encoding in flags. */
298 if (flags & ~(unsigned int)(MFD_ALL_FLAGS |
299 (MFD_HUGE_MASK << MFD_HUGE_SHIFT)))
300 return -EINVAL;
301 }
302
303 /* Invalid if both EXEC and NOEXEC_SEAL are set.*/
304 if ((flags & MFD_EXEC) && (flags & MFD_NOEXEC_SEAL))
305 return -EINVAL;
306
307 error = check_sysctl_memfd_noexec(&flags);
308 if (error < 0)
309 return error;
310
311 /* length includes terminating zero */
312 len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
313 if (len <= 0)
314 return -EFAULT;
315 if (len > MFD_NAME_MAX_LEN + 1)
316 return -EINVAL;
317
318 name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_KERNEL);
319 if (!name)
320 return -ENOMEM;
321
322 strcpy(name, MFD_NAME_PREFIX);
323 if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
324 error = -EFAULT;
325 goto err_name;
326 }
327
328 /* terminating-zero may have changed after strnlen_user() returned */
329 if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
330 error = -EFAULT;
331 goto err_name;
332 }
333
334 fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
335 if (fd < 0) {
336 error = fd;
337 goto err_name;
338 }
339
340 if (flags & MFD_HUGETLB) {
341 file = hugetlb_file_setup(name, 0, VM_NORESERVE,
342 HUGETLB_ANONHUGE_INODE,
343 (flags >> MFD_HUGE_SHIFT) &
344 MFD_HUGE_MASK);
345 } else
346 file = shmem_file_setup(name, 0, VM_NORESERVE);
347 if (IS_ERR(file)) {
348 error = PTR_ERR(file);
349 goto err_fd;
350 }
351 file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
352 file->f_flags |= O_LARGEFILE;
353
354 if (flags & MFD_NOEXEC_SEAL) {
355 struct inode *inode = file_inode(file);
356
357 inode->i_mode &= ~0111;
358 file_seals = memfd_file_seals_ptr(file);
359 if (file_seals) {
360 *file_seals &= ~F_SEAL_SEAL;
361 *file_seals |= F_SEAL_EXEC;
362 }
363 } else if (flags & MFD_ALLOW_SEALING) {
364 /* MFD_EXEC and MFD_ALLOW_SEALING are set */
365 file_seals = memfd_file_seals_ptr(file);
366 if (file_seals)
367 *file_seals &= ~F_SEAL_SEAL;
368 }
369
370 fd_install(fd, file);
371 kfree(name);
372 return fd;
373
374err_fd:
375 put_unused_fd(fd);
376err_name:
377 kfree(name);
378 return error;
379}