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