Commit | Line | Data |
---|---|---|
86039bd3 AA |
1 | /* |
2 | * fs/userfaultfd.c | |
3 | * | |
4 | * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org> | |
5 | * Copyright (C) 2008-2009 Red Hat, Inc. | |
6 | * Copyright (C) 2015 Red Hat, Inc. | |
7 | * | |
8 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
9 | * the COPYING file in the top-level directory. | |
10 | * | |
11 | * Some part derived from fs/eventfd.c (anon inode setup) and | |
12 | * mm/ksm.c (mm hashing). | |
13 | */ | |
14 | ||
15 | #include <linux/hashtable.h> | |
16 | #include <linux/sched.h> | |
17 | #include <linux/mm.h> | |
18 | #include <linux/poll.h> | |
19 | #include <linux/slab.h> | |
20 | #include <linux/seq_file.h> | |
21 | #include <linux/file.h> | |
22 | #include <linux/bug.h> | |
23 | #include <linux/anon_inodes.h> | |
24 | #include <linux/syscalls.h> | |
25 | #include <linux/userfaultfd_k.h> | |
26 | #include <linux/mempolicy.h> | |
27 | #include <linux/ioctl.h> | |
28 | #include <linux/security.h> | |
29 | ||
30 | enum userfaultfd_state { | |
31 | UFFD_STATE_WAIT_API, | |
32 | UFFD_STATE_RUNNING, | |
33 | }; | |
34 | ||
35 | struct userfaultfd_ctx { | |
36 | /* pseudo fd refcounting */ | |
37 | atomic_t refcount; | |
15b726ef AA |
38 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
39 | wait_queue_head_t fault_pending_wqh; | |
40 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
41 | wait_queue_head_t fault_wqh; |
42 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
43 | wait_queue_head_t fd_wqh; | |
44 | /* userfaultfd syscall flags */ | |
45 | unsigned int flags; | |
46 | /* state machine */ | |
47 | enum userfaultfd_state state; | |
48 | /* released */ | |
49 | bool released; | |
50 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ | |
51 | struct mm_struct *mm; | |
52 | }; | |
53 | ||
54 | struct userfaultfd_wait_queue { | |
a9b85f94 | 55 | struct uffd_msg msg; |
86039bd3 | 56 | wait_queue_t wq; |
86039bd3 AA |
57 | struct userfaultfd_ctx *ctx; |
58 | }; | |
59 | ||
60 | struct userfaultfd_wake_range { | |
61 | unsigned long start; | |
62 | unsigned long len; | |
63 | }; | |
64 | ||
65 | static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode, | |
66 | int wake_flags, void *key) | |
67 | { | |
68 | struct userfaultfd_wake_range *range = key; | |
69 | int ret; | |
70 | struct userfaultfd_wait_queue *uwq; | |
71 | unsigned long start, len; | |
72 | ||
73 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
74 | ret = 0; | |
86039bd3 AA |
75 | /* len == 0 means wake all */ |
76 | start = range->start; | |
77 | len = range->len; | |
a9b85f94 AA |
78 | if (len && (start > uwq->msg.arg.pagefault.address || |
79 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 AA |
80 | goto out; |
81 | ret = wake_up_state(wq->private, mode); | |
82 | if (ret) | |
83 | /* | |
84 | * Wake only once, autoremove behavior. | |
85 | * | |
86 | * After the effect of list_del_init is visible to the | |
87 | * other CPUs, the waitqueue may disappear from under | |
88 | * us, see the !list_empty_careful() in | |
89 | * handle_userfault(). try_to_wake_up() has an | |
90 | * implicit smp_mb__before_spinlock, and the | |
91 | * wq->private is read before calling the extern | |
92 | * function "wake_up_state" (which in turns calls | |
93 | * try_to_wake_up). While the spin_lock;spin_unlock; | |
94 | * wouldn't be enough, the smp_mb__before_spinlock is | |
95 | * enough to avoid an explicit smp_mb() here. | |
96 | */ | |
97 | list_del_init(&wq->task_list); | |
98 | out: | |
99 | return ret; | |
100 | } | |
101 | ||
102 | /** | |
103 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
104 | * context. | |
105 | * @ctx: [in] Pointer to the userfaultfd context. | |
106 | * | |
107 | * Returns: In case of success, returns not zero. | |
108 | */ | |
109 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
110 | { | |
111 | if (!atomic_inc_not_zero(&ctx->refcount)) | |
112 | BUG(); | |
113 | } | |
114 | ||
115 | /** | |
116 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
117 | * context. | |
118 | * @ctx: [in] Pointer to userfaultfd context. | |
119 | * | |
120 | * The userfaultfd context reference must have been previously acquired either | |
121 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
122 | */ | |
123 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
124 | { | |
125 | if (atomic_dec_and_test(&ctx->refcount)) { | |
126 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); | |
127 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
128 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
129 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
130 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); | |
131 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
132 | mmput(ctx->mm); | |
133 | kfree(ctx); | |
134 | } | |
135 | } | |
136 | ||
a9b85f94 | 137 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 138 | { |
a9b85f94 AA |
139 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
140 | /* | |
141 | * Must use memset to zero out the paddings or kernel data is | |
142 | * leaked to userland. | |
143 | */ | |
144 | memset(msg, 0, sizeof(struct uffd_msg)); | |
145 | } | |
146 | ||
147 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
148 | unsigned int flags, | |
149 | unsigned long reason) | |
150 | { | |
151 | struct uffd_msg msg; | |
152 | msg_init(&msg); | |
153 | msg.event = UFFD_EVENT_PAGEFAULT; | |
154 | msg.arg.pagefault.address = address; | |
86039bd3 AA |
155 | if (flags & FAULT_FLAG_WRITE) |
156 | /* | |
a9b85f94 AA |
157 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WRITE was set in the |
158 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE | |
159 | * was not set in a UFFD_EVENT_PAGEFAULT, it means it | |
160 | * was a read fault, otherwise if set it means it's | |
161 | * a write fault. | |
86039bd3 | 162 | */ |
a9b85f94 | 163 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 AA |
164 | if (reason & VM_UFFD_WP) |
165 | /* | |
a9b85f94 AA |
166 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
167 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was | |
168 | * not set in a UFFD_EVENT_PAGEFAULT, it means it was | |
169 | * a missing fault, otherwise if set it means it's a | |
170 | * write protect fault. | |
86039bd3 | 171 | */ |
a9b85f94 AA |
172 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
173 | return msg; | |
86039bd3 AA |
174 | } |
175 | ||
176 | /* | |
177 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
178 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
179 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
180 | * recommendation in __lock_page_or_retry is not an understatement. | |
181 | * | |
182 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released | |
183 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is | |
184 | * not set. | |
185 | * | |
186 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
187 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
188 | * fatal_signal_pending()s, and the mmap_sem must be released before | |
189 | * returning it. | |
190 | */ | |
191 | int handle_userfault(struct vm_area_struct *vma, unsigned long address, | |
192 | unsigned int flags, unsigned long reason) | |
193 | { | |
194 | struct mm_struct *mm = vma->vm_mm; | |
195 | struct userfaultfd_ctx *ctx; | |
196 | struct userfaultfd_wait_queue uwq; | |
ba85c702 | 197 | int ret; |
86039bd3 AA |
198 | |
199 | BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
200 | ||
ba85c702 | 201 | ret = VM_FAULT_SIGBUS; |
86039bd3 AA |
202 | ctx = vma->vm_userfaultfd_ctx.ctx; |
203 | if (!ctx) | |
ba85c702 | 204 | goto out; |
86039bd3 AA |
205 | |
206 | BUG_ON(ctx->mm != mm); | |
207 | ||
208 | VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP)); | |
209 | VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP)); | |
210 | ||
211 | /* | |
212 | * If it's already released don't get it. This avoids to loop | |
213 | * in __get_user_pages if userfaultfd_release waits on the | |
214 | * caller of handle_userfault to release the mmap_sem. | |
215 | */ | |
216 | if (unlikely(ACCESS_ONCE(ctx->released))) | |
ba85c702 | 217 | goto out; |
86039bd3 AA |
218 | |
219 | /* | |
220 | * Check that we can return VM_FAULT_RETRY. | |
221 | * | |
222 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
223 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
224 | * -EBUSY failures, if the userfaultfd is to be extended for | |
225 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
226 | * without first stopping userland access to the memory. For | |
227 | * VM_UFFD_MISSING userfaults this is enough for now. | |
228 | */ | |
229 | if (unlikely(!(flags & FAULT_FLAG_ALLOW_RETRY))) { | |
230 | /* | |
231 | * Validate the invariant that nowait must allow retry | |
232 | * to be sure not to return SIGBUS erroneously on | |
233 | * nowait invocations. | |
234 | */ | |
235 | BUG_ON(flags & FAULT_FLAG_RETRY_NOWAIT); | |
236 | #ifdef CONFIG_DEBUG_VM | |
237 | if (printk_ratelimit()) { | |
238 | printk(KERN_WARNING | |
239 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", flags); | |
240 | dump_stack(); | |
241 | } | |
242 | #endif | |
ba85c702 | 243 | goto out; |
86039bd3 AA |
244 | } |
245 | ||
246 | /* | |
247 | * Handle nowait, not much to do other than tell it to retry | |
248 | * and wait. | |
249 | */ | |
ba85c702 | 250 | ret = VM_FAULT_RETRY; |
86039bd3 | 251 | if (flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 252 | goto out; |
86039bd3 AA |
253 | |
254 | /* take the reference before dropping the mmap_sem */ | |
255 | userfaultfd_ctx_get(ctx); | |
256 | ||
257 | /* be gentle and immediately relinquish the mmap_sem */ | |
258 | up_read(&mm->mmap_sem); | |
259 | ||
260 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); | |
261 | uwq.wq.private = current; | |
a9b85f94 | 262 | uwq.msg = userfault_msg(address, flags, reason); |
86039bd3 AA |
263 | uwq.ctx = ctx; |
264 | ||
15b726ef | 265 | spin_lock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
266 | /* |
267 | * After the __add_wait_queue the uwq is visible to userland | |
268 | * through poll/read(). | |
269 | */ | |
15b726ef AA |
270 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
271 | /* | |
272 | * The smp_mb() after __set_current_state prevents the reads | |
273 | * following the spin_unlock to happen before the list_add in | |
274 | * __add_wait_queue. | |
275 | */ | |
ba85c702 | 276 | set_current_state(TASK_KILLABLE); |
15b726ef | 277 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 278 | |
ba85c702 AA |
279 | if (likely(!ACCESS_ONCE(ctx->released) && |
280 | !fatal_signal_pending(current))) { | |
86039bd3 AA |
281 | wake_up_poll(&ctx->fd_wqh, POLLIN); |
282 | schedule(); | |
ba85c702 AA |
283 | ret |= VM_FAULT_MAJOR; |
284 | } | |
86039bd3 | 285 | |
ba85c702 | 286 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
287 | |
288 | /* | |
289 | * Here we race with the list_del; list_add in | |
290 | * userfaultfd_ctx_read(), however because we don't ever run | |
291 | * list_del_init() to refile across the two lists, the prev | |
292 | * and next pointers will never point to self. list_add also | |
293 | * would never let any of the two pointers to point to | |
294 | * self. So list_empty_careful won't risk to see both pointers | |
295 | * pointing to self at any time during the list refile. The | |
296 | * only case where list_del_init() is called is the full | |
297 | * removal in the wake function and there we don't re-list_add | |
298 | * and it's fine not to block on the spinlock. The uwq on this | |
299 | * kernel stack can be released after the list_del_init. | |
300 | */ | |
ba85c702 | 301 | if (!list_empty_careful(&uwq.wq.task_list)) { |
15b726ef AA |
302 | spin_lock(&ctx->fault_pending_wqh.lock); |
303 | /* | |
304 | * No need of list_del_init(), the uwq on the stack | |
305 | * will be freed shortly anyway. | |
306 | */ | |
307 | list_del(&uwq.wq.task_list); | |
308 | spin_unlock(&ctx->fault_pending_wqh.lock); | |
86039bd3 | 309 | } |
86039bd3 AA |
310 | |
311 | /* | |
312 | * ctx may go away after this if the userfault pseudo fd is | |
313 | * already released. | |
314 | */ | |
315 | userfaultfd_ctx_put(ctx); | |
316 | ||
ba85c702 AA |
317 | out: |
318 | return ret; | |
86039bd3 AA |
319 | } |
320 | ||
321 | static int userfaultfd_release(struct inode *inode, struct file *file) | |
322 | { | |
323 | struct userfaultfd_ctx *ctx = file->private_data; | |
324 | struct mm_struct *mm = ctx->mm; | |
325 | struct vm_area_struct *vma, *prev; | |
326 | /* len == 0 means wake all */ | |
327 | struct userfaultfd_wake_range range = { .len = 0, }; | |
328 | unsigned long new_flags; | |
329 | ||
330 | ACCESS_ONCE(ctx->released) = true; | |
331 | ||
332 | /* | |
333 | * Flush page faults out of all CPUs. NOTE: all page faults | |
334 | * must be retried without returning VM_FAULT_SIGBUS if | |
335 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
336 | * changes while handle_userfault released the mmap_sem. So | |
337 | * it's critical that released is set to true (above), before | |
338 | * taking the mmap_sem for writing. | |
339 | */ | |
340 | down_write(&mm->mmap_sem); | |
341 | prev = NULL; | |
342 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
343 | cond_resched(); | |
344 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
345 | !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
346 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { | |
347 | prev = vma; | |
348 | continue; | |
349 | } | |
350 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
351 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, | |
352 | new_flags, vma->anon_vma, | |
353 | vma->vm_file, vma->vm_pgoff, | |
354 | vma_policy(vma), | |
355 | NULL_VM_UFFD_CTX); | |
356 | if (prev) | |
357 | vma = prev; | |
358 | else | |
359 | prev = vma; | |
360 | vma->vm_flags = new_flags; | |
361 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
362 | } | |
363 | up_write(&mm->mmap_sem); | |
364 | ||
365 | /* | |
15b726ef | 366 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 367 | * the last page faults that may have been already waiting on |
15b726ef | 368 | * the fault_*wqh. |
86039bd3 | 369 | */ |
15b726ef AA |
370 | spin_lock(&ctx->fault_pending_wqh.lock); |
371 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, 0, &range); | |
86039bd3 | 372 | __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, 0, &range); |
15b726ef | 373 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
374 | |
375 | wake_up_poll(&ctx->fd_wqh, POLLHUP); | |
376 | userfaultfd_ctx_put(ctx); | |
377 | return 0; | |
378 | } | |
379 | ||
15b726ef AA |
380 | /* fault_pending_wqh.lock must be hold by the caller */ |
381 | static inline struct userfaultfd_wait_queue *find_userfault( | |
382 | struct userfaultfd_ctx *ctx) | |
86039bd3 AA |
383 | { |
384 | wait_queue_t *wq; | |
15b726ef | 385 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 386 | |
15b726ef | 387 | VM_BUG_ON(!spin_is_locked(&ctx->fault_pending_wqh.lock)); |
86039bd3 | 388 | |
15b726ef AA |
389 | uwq = NULL; |
390 | if (!waitqueue_active(&ctx->fault_pending_wqh)) | |
391 | goto out; | |
392 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
393 | wq = list_last_entry(&ctx->fault_pending_wqh.task_list, | |
394 | typeof(*wq), task_list); | |
395 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
396 | out: | |
397 | return uwq; | |
86039bd3 AA |
398 | } |
399 | ||
400 | static unsigned int userfaultfd_poll(struct file *file, poll_table *wait) | |
401 | { | |
402 | struct userfaultfd_ctx *ctx = file->private_data; | |
403 | unsigned int ret; | |
404 | ||
405 | poll_wait(file, &ctx->fd_wqh, wait); | |
406 | ||
407 | switch (ctx->state) { | |
408 | case UFFD_STATE_WAIT_API: | |
409 | return POLLERR; | |
410 | case UFFD_STATE_RUNNING: | |
ba85c702 AA |
411 | /* |
412 | * poll() never guarantees that read won't block. | |
413 | * userfaults can be waken before they're read(). | |
414 | */ | |
415 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
416 | return POLLERR; | |
15b726ef AA |
417 | /* |
418 | * lockless access to see if there are pending faults | |
419 | * __pollwait last action is the add_wait_queue but | |
420 | * the spin_unlock would allow the waitqueue_active to | |
421 | * pass above the actual list_add inside | |
422 | * add_wait_queue critical section. So use a full | |
423 | * memory barrier to serialize the list_add write of | |
424 | * add_wait_queue() with the waitqueue_active read | |
425 | * below. | |
426 | */ | |
427 | ret = 0; | |
428 | smp_mb(); | |
429 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
430 | ret = POLLIN; | |
86039bd3 AA |
431 | return ret; |
432 | default: | |
433 | BUG(); | |
434 | } | |
435 | } | |
436 | ||
437 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, | |
a9b85f94 | 438 | struct uffd_msg *msg) |
86039bd3 AA |
439 | { |
440 | ssize_t ret; | |
441 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 442 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 443 | |
15b726ef | 444 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
86039bd3 AA |
445 | spin_lock(&ctx->fd_wqh.lock); |
446 | __add_wait_queue(&ctx->fd_wqh, &wait); | |
447 | for (;;) { | |
448 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
449 | spin_lock(&ctx->fault_pending_wqh.lock); |
450 | uwq = find_userfault(ctx); | |
451 | if (uwq) { | |
86039bd3 | 452 | /* |
15b726ef AA |
453 | * The fault_pending_wqh.lock prevents the uwq |
454 | * to disappear from under us. | |
455 | * | |
456 | * Refile this userfault from | |
457 | * fault_pending_wqh to fault_wqh, it's not | |
458 | * pending anymore after we read it. | |
459 | * | |
460 | * Use list_del() by hand (as | |
461 | * userfaultfd_wake_function also uses | |
462 | * list_del_init() by hand) to be sure nobody | |
463 | * changes __remove_wait_queue() to use | |
464 | * list_del_init() in turn breaking the | |
465 | * !list_empty_careful() check in | |
466 | * handle_userfault(). The uwq->wq.task_list | |
467 | * must never be empty at any time during the | |
468 | * refile, or the waitqueue could disappear | |
469 | * from under us. The "wait_queue_head_t" | |
470 | * parameter of __remove_wait_queue() is unused | |
471 | * anyway. | |
86039bd3 | 472 | */ |
15b726ef AA |
473 | list_del(&uwq->wq.task_list); |
474 | __add_wait_queue(&ctx->fault_wqh, &uwq->wq); | |
475 | ||
a9b85f94 AA |
476 | /* careful to always initialize msg if ret == 0 */ |
477 | *msg = uwq->msg; | |
15b726ef | 478 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
479 | ret = 0; |
480 | break; | |
481 | } | |
15b726ef | 482 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
483 | if (signal_pending(current)) { |
484 | ret = -ERESTARTSYS; | |
485 | break; | |
486 | } | |
487 | if (no_wait) { | |
488 | ret = -EAGAIN; | |
489 | break; | |
490 | } | |
491 | spin_unlock(&ctx->fd_wqh.lock); | |
492 | schedule(); | |
493 | spin_lock(&ctx->fd_wqh.lock); | |
494 | } | |
495 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
496 | __set_current_state(TASK_RUNNING); | |
497 | spin_unlock(&ctx->fd_wqh.lock); | |
498 | ||
499 | return ret; | |
500 | } | |
501 | ||
502 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
503 | size_t count, loff_t *ppos) | |
504 | { | |
505 | struct userfaultfd_ctx *ctx = file->private_data; | |
506 | ssize_t _ret, ret = 0; | |
a9b85f94 | 507 | struct uffd_msg msg; |
86039bd3 AA |
508 | int no_wait = file->f_flags & O_NONBLOCK; |
509 | ||
510 | if (ctx->state == UFFD_STATE_WAIT_API) | |
511 | return -EINVAL; | |
512 | BUG_ON(ctx->state != UFFD_STATE_RUNNING); | |
513 | ||
514 | for (;;) { | |
a9b85f94 | 515 | if (count < sizeof(msg)) |
86039bd3 | 516 | return ret ? ret : -EINVAL; |
a9b85f94 | 517 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg); |
86039bd3 AA |
518 | if (_ret < 0) |
519 | return ret ? ret : _ret; | |
a9b85f94 | 520 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 521 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
522 | ret += sizeof(msg); |
523 | buf += sizeof(msg); | |
524 | count -= sizeof(msg); | |
86039bd3 AA |
525 | /* |
526 | * Allow to read more than one fault at time but only | |
527 | * block if waiting for the very first one. | |
528 | */ | |
529 | no_wait = O_NONBLOCK; | |
530 | } | |
531 | } | |
532 | ||
533 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
534 | struct userfaultfd_wake_range *range) | |
535 | { | |
536 | unsigned long start, end; | |
537 | ||
538 | start = range->start; | |
539 | end = range->start + range->len; | |
540 | ||
15b726ef | 541 | spin_lock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 542 | /* wake all in the range and autoremove */ |
15b726ef AA |
543 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
544 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, 0, | |
545 | range); | |
546 | if (waitqueue_active(&ctx->fault_wqh)) | |
547 | __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, 0, range); | |
548 | spin_unlock(&ctx->fault_pending_wqh.lock); | |
86039bd3 AA |
549 | } |
550 | ||
551 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
552 | struct userfaultfd_wake_range *range) | |
553 | { | |
554 | /* | |
555 | * To be sure waitqueue_active() is not reordered by the CPU | |
556 | * before the pagetable update, use an explicit SMP memory | |
557 | * barrier here. PT lock release or up_read(mmap_sem) still | |
558 | * have release semantics that can allow the | |
559 | * waitqueue_active() to be reordered before the pte update. | |
560 | */ | |
561 | smp_mb(); | |
562 | ||
563 | /* | |
564 | * Use waitqueue_active because it's very frequent to | |
565 | * change the address space atomically even if there are no | |
566 | * userfaults yet. So we take the spinlock only when we're | |
567 | * sure we've userfaults to wake. | |
568 | */ | |
15b726ef AA |
569 | if (waitqueue_active(&ctx->fault_pending_wqh) || |
570 | waitqueue_active(&ctx->fault_wqh)) | |
86039bd3 AA |
571 | __wake_userfault(ctx, range); |
572 | } | |
573 | ||
574 | static __always_inline int validate_range(struct mm_struct *mm, | |
575 | __u64 start, __u64 len) | |
576 | { | |
577 | __u64 task_size = mm->task_size; | |
578 | ||
579 | if (start & ~PAGE_MASK) | |
580 | return -EINVAL; | |
581 | if (len & ~PAGE_MASK) | |
582 | return -EINVAL; | |
583 | if (!len) | |
584 | return -EINVAL; | |
585 | if (start < mmap_min_addr) | |
586 | return -EINVAL; | |
587 | if (start >= task_size) | |
588 | return -EINVAL; | |
589 | if (len > task_size - start) | |
590 | return -EINVAL; | |
591 | return 0; | |
592 | } | |
593 | ||
594 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, | |
595 | unsigned long arg) | |
596 | { | |
597 | struct mm_struct *mm = ctx->mm; | |
598 | struct vm_area_struct *vma, *prev, *cur; | |
599 | int ret; | |
600 | struct uffdio_register uffdio_register; | |
601 | struct uffdio_register __user *user_uffdio_register; | |
602 | unsigned long vm_flags, new_flags; | |
603 | bool found; | |
604 | unsigned long start, end, vma_end; | |
605 | ||
606 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
607 | ||
608 | ret = -EFAULT; | |
609 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
610 | sizeof(uffdio_register)-sizeof(__u64))) | |
611 | goto out; | |
612 | ||
613 | ret = -EINVAL; | |
614 | if (!uffdio_register.mode) | |
615 | goto out; | |
616 | if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING| | |
617 | UFFDIO_REGISTER_MODE_WP)) | |
618 | goto out; | |
619 | vm_flags = 0; | |
620 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
621 | vm_flags |= VM_UFFD_MISSING; | |
622 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { | |
623 | vm_flags |= VM_UFFD_WP; | |
624 | /* | |
625 | * FIXME: remove the below error constraint by | |
626 | * implementing the wprotect tracking mode. | |
627 | */ | |
628 | ret = -EINVAL; | |
629 | goto out; | |
630 | } | |
631 | ||
632 | ret = validate_range(mm, uffdio_register.range.start, | |
633 | uffdio_register.range.len); | |
634 | if (ret) | |
635 | goto out; | |
636 | ||
637 | start = uffdio_register.range.start; | |
638 | end = start + uffdio_register.range.len; | |
639 | ||
640 | down_write(&mm->mmap_sem); | |
641 | vma = find_vma_prev(mm, start, &prev); | |
642 | ||
643 | ret = -ENOMEM; | |
644 | if (!vma) | |
645 | goto out_unlock; | |
646 | ||
647 | /* check that there's at least one vma in the range */ | |
648 | ret = -EINVAL; | |
649 | if (vma->vm_start >= end) | |
650 | goto out_unlock; | |
651 | ||
652 | /* | |
653 | * Search for not compatible vmas. | |
654 | * | |
655 | * FIXME: this shall be relaxed later so that it doesn't fail | |
656 | * on tmpfs backed vmas (in addition to the current allowance | |
657 | * on anonymous vmas). | |
658 | */ | |
659 | found = false; | |
660 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
661 | cond_resched(); | |
662 | ||
663 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
664 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
665 | ||
666 | /* check not compatible vmas */ | |
667 | ret = -EINVAL; | |
668 | if (cur->vm_ops) | |
669 | goto out_unlock; | |
670 | ||
671 | /* | |
672 | * Check that this vma isn't already owned by a | |
673 | * different userfaultfd. We can't allow more than one | |
674 | * userfaultfd to own a single vma simultaneously or we | |
675 | * wouldn't know which one to deliver the userfaults to. | |
676 | */ | |
677 | ret = -EBUSY; | |
678 | if (cur->vm_userfaultfd_ctx.ctx && | |
679 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
680 | goto out_unlock; | |
681 | ||
682 | found = true; | |
683 | } | |
684 | BUG_ON(!found); | |
685 | ||
686 | if (vma->vm_start < start) | |
687 | prev = vma; | |
688 | ||
689 | ret = 0; | |
690 | do { | |
691 | cond_resched(); | |
692 | ||
693 | BUG_ON(vma->vm_ops); | |
694 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && | |
695 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
696 | ||
697 | /* | |
698 | * Nothing to do: this vma is already registered into this | |
699 | * userfaultfd and with the right tracking mode too. | |
700 | */ | |
701 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
702 | (vma->vm_flags & vm_flags) == vm_flags) | |
703 | goto skip; | |
704 | ||
705 | if (vma->vm_start > start) | |
706 | start = vma->vm_start; | |
707 | vma_end = min(end, vma->vm_end); | |
708 | ||
709 | new_flags = (vma->vm_flags & ~vm_flags) | vm_flags; | |
710 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
711 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
712 | vma_policy(vma), | |
713 | ((struct vm_userfaultfd_ctx){ ctx })); | |
714 | if (prev) { | |
715 | vma = prev; | |
716 | goto next; | |
717 | } | |
718 | if (vma->vm_start < start) { | |
719 | ret = split_vma(mm, vma, start, 1); | |
720 | if (ret) | |
721 | break; | |
722 | } | |
723 | if (vma->vm_end > end) { | |
724 | ret = split_vma(mm, vma, end, 0); | |
725 | if (ret) | |
726 | break; | |
727 | } | |
728 | next: | |
729 | /* | |
730 | * In the vma_merge() successful mprotect-like case 8: | |
731 | * the next vma was merged into the current one and | |
732 | * the current one has not been updated yet. | |
733 | */ | |
734 | vma->vm_flags = new_flags; | |
735 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
736 | ||
737 | skip: | |
738 | prev = vma; | |
739 | start = vma->vm_end; | |
740 | vma = vma->vm_next; | |
741 | } while (vma && vma->vm_start < end); | |
742 | out_unlock: | |
743 | up_write(&mm->mmap_sem); | |
744 | if (!ret) { | |
745 | /* | |
746 | * Now that we scanned all vmas we can already tell | |
747 | * userland which ioctls methods are guaranteed to | |
748 | * succeed on this range. | |
749 | */ | |
750 | if (put_user(UFFD_API_RANGE_IOCTLS, | |
751 | &user_uffdio_register->ioctls)) | |
752 | ret = -EFAULT; | |
753 | } | |
754 | out: | |
755 | return ret; | |
756 | } | |
757 | ||
758 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
759 | unsigned long arg) | |
760 | { | |
761 | struct mm_struct *mm = ctx->mm; | |
762 | struct vm_area_struct *vma, *prev, *cur; | |
763 | int ret; | |
764 | struct uffdio_range uffdio_unregister; | |
765 | unsigned long new_flags; | |
766 | bool found; | |
767 | unsigned long start, end, vma_end; | |
768 | const void __user *buf = (void __user *)arg; | |
769 | ||
770 | ret = -EFAULT; | |
771 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
772 | goto out; | |
773 | ||
774 | ret = validate_range(mm, uffdio_unregister.start, | |
775 | uffdio_unregister.len); | |
776 | if (ret) | |
777 | goto out; | |
778 | ||
779 | start = uffdio_unregister.start; | |
780 | end = start + uffdio_unregister.len; | |
781 | ||
782 | down_write(&mm->mmap_sem); | |
783 | vma = find_vma_prev(mm, start, &prev); | |
784 | ||
785 | ret = -ENOMEM; | |
786 | if (!vma) | |
787 | goto out_unlock; | |
788 | ||
789 | /* check that there's at least one vma in the range */ | |
790 | ret = -EINVAL; | |
791 | if (vma->vm_start >= end) | |
792 | goto out_unlock; | |
793 | ||
794 | /* | |
795 | * Search for not compatible vmas. | |
796 | * | |
797 | * FIXME: this shall be relaxed later so that it doesn't fail | |
798 | * on tmpfs backed vmas (in addition to the current allowance | |
799 | * on anonymous vmas). | |
800 | */ | |
801 | found = false; | |
802 | ret = -EINVAL; | |
803 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
804 | cond_resched(); | |
805 | ||
806 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
807 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
808 | ||
809 | /* | |
810 | * Check not compatible vmas, not strictly required | |
811 | * here as not compatible vmas cannot have an | |
812 | * userfaultfd_ctx registered on them, but this | |
813 | * provides for more strict behavior to notice | |
814 | * unregistration errors. | |
815 | */ | |
816 | if (cur->vm_ops) | |
817 | goto out_unlock; | |
818 | ||
819 | found = true; | |
820 | } | |
821 | BUG_ON(!found); | |
822 | ||
823 | if (vma->vm_start < start) | |
824 | prev = vma; | |
825 | ||
826 | ret = 0; | |
827 | do { | |
828 | cond_resched(); | |
829 | ||
830 | BUG_ON(vma->vm_ops); | |
831 | ||
832 | /* | |
833 | * Nothing to do: this vma is already registered into this | |
834 | * userfaultfd and with the right tracking mode too. | |
835 | */ | |
836 | if (!vma->vm_userfaultfd_ctx.ctx) | |
837 | goto skip; | |
838 | ||
839 | if (vma->vm_start > start) | |
840 | start = vma->vm_start; | |
841 | vma_end = min(end, vma->vm_end); | |
842 | ||
843 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
844 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
845 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
846 | vma_policy(vma), | |
847 | NULL_VM_UFFD_CTX); | |
848 | if (prev) { | |
849 | vma = prev; | |
850 | goto next; | |
851 | } | |
852 | if (vma->vm_start < start) { | |
853 | ret = split_vma(mm, vma, start, 1); | |
854 | if (ret) | |
855 | break; | |
856 | } | |
857 | if (vma->vm_end > end) { | |
858 | ret = split_vma(mm, vma, end, 0); | |
859 | if (ret) | |
860 | break; | |
861 | } | |
862 | next: | |
863 | /* | |
864 | * In the vma_merge() successful mprotect-like case 8: | |
865 | * the next vma was merged into the current one and | |
866 | * the current one has not been updated yet. | |
867 | */ | |
868 | vma->vm_flags = new_flags; | |
869 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
870 | ||
871 | skip: | |
872 | prev = vma; | |
873 | start = vma->vm_end; | |
874 | vma = vma->vm_next; | |
875 | } while (vma && vma->vm_start < end); | |
876 | out_unlock: | |
877 | up_write(&mm->mmap_sem); | |
878 | out: | |
879 | return ret; | |
880 | } | |
881 | ||
882 | /* | |
ba85c702 AA |
883 | * userfaultfd_wake is needed in case an userfault is in flight by the |
884 | * time a UFFDIO_COPY (or other ioctl variants) completes. The page | |
885 | * may be well get mapped and the page fault if repeated wouldn't lead | |
886 | * to a userfault anymore, but before scheduling in TASK_KILLABLE mode | |
887 | * handle_userfault() doesn't recheck the pagetables and it doesn't | |
888 | * serialize against UFFDO_COPY (or other ioctl variants). Ultimately | |
889 | * the knowledge of which pages are mapped is left to userland who is | |
890 | * responsible for handling the race between read() userfaults and | |
891 | * background UFFDIO_COPY (or other ioctl variants), if done by | |
892 | * separate concurrent threads. | |
893 | * | |
894 | * userfaultfd_wake may be used in combination with the | |
895 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
896 | */ |
897 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
898 | unsigned long arg) | |
899 | { | |
900 | int ret; | |
901 | struct uffdio_range uffdio_wake; | |
902 | struct userfaultfd_wake_range range; | |
903 | const void __user *buf = (void __user *)arg; | |
904 | ||
905 | ret = -EFAULT; | |
906 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
907 | goto out; | |
908 | ||
909 | ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len); | |
910 | if (ret) | |
911 | goto out; | |
912 | ||
913 | range.start = uffdio_wake.start; | |
914 | range.len = uffdio_wake.len; | |
915 | ||
916 | /* | |
917 | * len == 0 means wake all and we don't want to wake all here, | |
918 | * so check it again to be sure. | |
919 | */ | |
920 | VM_BUG_ON(!range.len); | |
921 | ||
922 | wake_userfault(ctx, &range); | |
923 | ret = 0; | |
924 | ||
925 | out: | |
926 | return ret; | |
927 | } | |
928 | ||
929 | /* | |
930 | * userland asks for a certain API version and we return which bits | |
931 | * and ioctl commands are implemented in this kernel for such API | |
932 | * version or -EINVAL if unknown. | |
933 | */ | |
934 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
935 | unsigned long arg) | |
936 | { | |
937 | struct uffdio_api uffdio_api; | |
938 | void __user *buf = (void __user *)arg; | |
939 | int ret; | |
940 | ||
941 | ret = -EINVAL; | |
942 | if (ctx->state != UFFD_STATE_WAIT_API) | |
943 | goto out; | |
944 | ret = -EFAULT; | |
a9b85f94 | 945 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 946 | goto out; |
a9b85f94 | 947 | if (uffdio_api.api != UFFD_API || uffdio_api.features) { |
86039bd3 AA |
948 | memset(&uffdio_api, 0, sizeof(uffdio_api)); |
949 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
950 | goto out; | |
951 | ret = -EINVAL; | |
952 | goto out; | |
953 | } | |
3f602d27 | 954 | uffdio_api.features = UFFD_API_FEATURES; |
86039bd3 AA |
955 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
956 | ret = -EFAULT; | |
957 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
958 | goto out; | |
959 | ctx->state = UFFD_STATE_RUNNING; | |
960 | ret = 0; | |
961 | out: | |
962 | return ret; | |
963 | } | |
964 | ||
965 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
966 | unsigned long arg) | |
967 | { | |
968 | int ret = -EINVAL; | |
969 | struct userfaultfd_ctx *ctx = file->private_data; | |
970 | ||
971 | switch(cmd) { | |
972 | case UFFDIO_API: | |
973 | ret = userfaultfd_api(ctx, arg); | |
974 | break; | |
975 | case UFFDIO_REGISTER: | |
976 | ret = userfaultfd_register(ctx, arg); | |
977 | break; | |
978 | case UFFDIO_UNREGISTER: | |
979 | ret = userfaultfd_unregister(ctx, arg); | |
980 | break; | |
981 | case UFFDIO_WAKE: | |
982 | ret = userfaultfd_wake(ctx, arg); | |
983 | break; | |
984 | } | |
985 | return ret; | |
986 | } | |
987 | ||
988 | #ifdef CONFIG_PROC_FS | |
989 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
990 | { | |
991 | struct userfaultfd_ctx *ctx = f->private_data; | |
992 | wait_queue_t *wq; | |
993 | struct userfaultfd_wait_queue *uwq; | |
994 | unsigned long pending = 0, total = 0; | |
995 | ||
15b726ef AA |
996 | spin_lock(&ctx->fault_pending_wqh.lock); |
997 | list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) { | |
998 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
999 | pending++; | |
1000 | total++; | |
1001 | } | |
86039bd3 AA |
1002 | list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) { |
1003 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
86039bd3 AA |
1004 | total++; |
1005 | } | |
15b726ef | 1006 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1007 | |
1008 | /* | |
1009 | * If more protocols will be added, there will be all shown | |
1010 | * separated by a space. Like this: | |
1011 | * protocols: aa:... bb:... | |
1012 | */ | |
1013 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
3f602d27 | 1014 | pending, total, UFFD_API, UFFD_API_FEATURES, |
86039bd3 AA |
1015 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
1016 | } | |
1017 | #endif | |
1018 | ||
1019 | static const struct file_operations userfaultfd_fops = { | |
1020 | #ifdef CONFIG_PROC_FS | |
1021 | .show_fdinfo = userfaultfd_show_fdinfo, | |
1022 | #endif | |
1023 | .release = userfaultfd_release, | |
1024 | .poll = userfaultfd_poll, | |
1025 | .read = userfaultfd_read, | |
1026 | .unlocked_ioctl = userfaultfd_ioctl, | |
1027 | .compat_ioctl = userfaultfd_ioctl, | |
1028 | .llseek = noop_llseek, | |
1029 | }; | |
1030 | ||
1031 | /** | |
1032 | * userfaultfd_file_create - Creates an userfaultfd file pointer. | |
1033 | * @flags: Flags for the userfaultfd file. | |
1034 | * | |
1035 | * This function creates an userfaultfd file pointer, w/out installing | |
1036 | * it into the fd table. This is useful when the userfaultfd file is | |
1037 | * used during the initialization of data structures that require | |
1038 | * extra setup after the userfaultfd creation. So the userfaultfd | |
1039 | * creation is split into the file pointer creation phase, and the | |
1040 | * file descriptor installation phase. In this way races with | |
1041 | * userspace closing the newly installed file descriptor can be | |
1042 | * avoided. Returns an userfaultfd file pointer, or a proper error | |
1043 | * pointer. | |
1044 | */ | |
1045 | static struct file *userfaultfd_file_create(int flags) | |
1046 | { | |
1047 | struct file *file; | |
1048 | struct userfaultfd_ctx *ctx; | |
1049 | ||
1050 | BUG_ON(!current->mm); | |
1051 | ||
1052 | /* Check the UFFD_* constants for consistency. */ | |
1053 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); | |
1054 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
1055 | ||
1056 | file = ERR_PTR(-EINVAL); | |
1057 | if (flags & ~UFFD_SHARED_FCNTL_FLAGS) | |
1058 | goto out; | |
1059 | ||
1060 | file = ERR_PTR(-ENOMEM); | |
1061 | ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); | |
1062 | if (!ctx) | |
1063 | goto out; | |
1064 | ||
1065 | atomic_set(&ctx->refcount, 1); | |
15b726ef | 1066 | init_waitqueue_head(&ctx->fault_pending_wqh); |
86039bd3 AA |
1067 | init_waitqueue_head(&ctx->fault_wqh); |
1068 | init_waitqueue_head(&ctx->fd_wqh); | |
1069 | ctx->flags = flags; | |
1070 | ctx->state = UFFD_STATE_WAIT_API; | |
1071 | ctx->released = false; | |
1072 | ctx->mm = current->mm; | |
1073 | /* prevent the mm struct to be freed */ | |
1074 | atomic_inc(&ctx->mm->mm_users); | |
1075 | ||
1076 | file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx, | |
1077 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS)); | |
1078 | if (IS_ERR(file)) | |
1079 | kfree(ctx); | |
1080 | out: | |
1081 | return file; | |
1082 | } | |
1083 | ||
1084 | SYSCALL_DEFINE1(userfaultfd, int, flags) | |
1085 | { | |
1086 | int fd, error; | |
1087 | struct file *file; | |
1088 | ||
1089 | error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS); | |
1090 | if (error < 0) | |
1091 | return error; | |
1092 | fd = error; | |
1093 | ||
1094 | file = userfaultfd_file_create(flags); | |
1095 | if (IS_ERR(file)) { | |
1096 | error = PTR_ERR(file); | |
1097 | goto err_put_unused_fd; | |
1098 | } | |
1099 | fd_install(fd, file); | |
1100 | ||
1101 | return fd; | |
1102 | ||
1103 | err_put_unused_fd: | |
1104 | put_unused_fd(fd); | |
1105 | ||
1106 | return error; | |
1107 | } |