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