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 | ||
9cd75c3c | 15 | #include <linux/list.h> |
86039bd3 AA |
16 | #include <linux/hashtable.h> |
17 | #include <linux/sched.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/poll.h> | |
20 | #include <linux/slab.h> | |
21 | #include <linux/seq_file.h> | |
22 | #include <linux/file.h> | |
23 | #include <linux/bug.h> | |
24 | #include <linux/anon_inodes.h> | |
25 | #include <linux/syscalls.h> | |
26 | #include <linux/userfaultfd_k.h> | |
27 | #include <linux/mempolicy.h> | |
28 | #include <linux/ioctl.h> | |
29 | #include <linux/security.h> | |
30 | ||
3004ec9c AA |
31 | static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly; |
32 | ||
86039bd3 AA |
33 | enum userfaultfd_state { |
34 | UFFD_STATE_WAIT_API, | |
35 | UFFD_STATE_RUNNING, | |
36 | }; | |
37 | ||
3004ec9c AA |
38 | /* |
39 | * Start with fault_pending_wqh and fault_wqh so they're more likely | |
40 | * to be in the same cacheline. | |
41 | */ | |
86039bd3 | 42 | struct userfaultfd_ctx { |
15b726ef AA |
43 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
44 | wait_queue_head_t fault_pending_wqh; | |
45 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
46 | wait_queue_head_t fault_wqh; |
47 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
48 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
49 | /* waitqueue head for events */ |
50 | wait_queue_head_t event_wqh; | |
2c5b7e1b AA |
51 | /* a refile sequence protected by fault_pending_wqh lock */ |
52 | struct seqcount refile_seq; | |
3004ec9c AA |
53 | /* pseudo fd refcounting */ |
54 | atomic_t refcount; | |
86039bd3 AA |
55 | /* userfaultfd syscall flags */ |
56 | unsigned int flags; | |
9cd75c3c PE |
57 | /* features requested from the userspace */ |
58 | unsigned int features; | |
86039bd3 AA |
59 | /* state machine */ |
60 | enum userfaultfd_state state; | |
61 | /* released */ | |
62 | bool released; | |
63 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ | |
64 | struct mm_struct *mm; | |
65 | }; | |
66 | ||
67 | struct userfaultfd_wait_queue { | |
a9b85f94 | 68 | struct uffd_msg msg; |
86039bd3 | 69 | wait_queue_t wq; |
86039bd3 | 70 | struct userfaultfd_ctx *ctx; |
15a77c6f | 71 | bool waken; |
86039bd3 AA |
72 | }; |
73 | ||
74 | struct userfaultfd_wake_range { | |
75 | unsigned long start; | |
76 | unsigned long len; | |
77 | }; | |
78 | ||
79 | static int userfaultfd_wake_function(wait_queue_t *wq, unsigned mode, | |
80 | int wake_flags, void *key) | |
81 | { | |
82 | struct userfaultfd_wake_range *range = key; | |
83 | int ret; | |
84 | struct userfaultfd_wait_queue *uwq; | |
85 | unsigned long start, len; | |
86 | ||
87 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
88 | ret = 0; | |
86039bd3 AA |
89 | /* len == 0 means wake all */ |
90 | start = range->start; | |
91 | len = range->len; | |
a9b85f94 AA |
92 | if (len && (start > uwq->msg.arg.pagefault.address || |
93 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 94 | goto out; |
15a77c6f AA |
95 | WRITE_ONCE(uwq->waken, true); |
96 | /* | |
97 | * The implicit smp_mb__before_spinlock in try_to_wake_up() | |
98 | * renders uwq->waken visible to other CPUs before the task is | |
99 | * waken. | |
100 | */ | |
86039bd3 AA |
101 | ret = wake_up_state(wq->private, mode); |
102 | if (ret) | |
103 | /* | |
104 | * Wake only once, autoremove behavior. | |
105 | * | |
106 | * After the effect of list_del_init is visible to the | |
107 | * other CPUs, the waitqueue may disappear from under | |
108 | * us, see the !list_empty_careful() in | |
109 | * handle_userfault(). try_to_wake_up() has an | |
110 | * implicit smp_mb__before_spinlock, and the | |
111 | * wq->private is read before calling the extern | |
112 | * function "wake_up_state" (which in turns calls | |
113 | * try_to_wake_up). While the spin_lock;spin_unlock; | |
114 | * wouldn't be enough, the smp_mb__before_spinlock is | |
115 | * enough to avoid an explicit smp_mb() here. | |
116 | */ | |
117 | list_del_init(&wq->task_list); | |
118 | out: | |
119 | return ret; | |
120 | } | |
121 | ||
122 | /** | |
123 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
124 | * context. | |
125 | * @ctx: [in] Pointer to the userfaultfd context. | |
126 | * | |
127 | * Returns: In case of success, returns not zero. | |
128 | */ | |
129 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
130 | { | |
131 | if (!atomic_inc_not_zero(&ctx->refcount)) | |
132 | BUG(); | |
133 | } | |
134 | ||
135 | /** | |
136 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
137 | * context. | |
138 | * @ctx: [in] Pointer to userfaultfd context. | |
139 | * | |
140 | * The userfaultfd context reference must have been previously acquired either | |
141 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
142 | */ | |
143 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
144 | { | |
145 | if (atomic_dec_and_test(&ctx->refcount)) { | |
146 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); | |
147 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
148 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
149 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
150 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
151 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
152 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
153 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 154 | mmdrop(ctx->mm); |
3004ec9c | 155 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
156 | } |
157 | } | |
158 | ||
a9b85f94 | 159 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 160 | { |
a9b85f94 AA |
161 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
162 | /* | |
163 | * Must use memset to zero out the paddings or kernel data is | |
164 | * leaked to userland. | |
165 | */ | |
166 | memset(msg, 0, sizeof(struct uffd_msg)); | |
167 | } | |
168 | ||
169 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
170 | unsigned int flags, | |
171 | unsigned long reason) | |
172 | { | |
173 | struct uffd_msg msg; | |
174 | msg_init(&msg); | |
175 | msg.event = UFFD_EVENT_PAGEFAULT; | |
176 | msg.arg.pagefault.address = address; | |
86039bd3 AA |
177 | if (flags & FAULT_FLAG_WRITE) |
178 | /* | |
a4605a61 | 179 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
a9b85f94 AA |
180 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE |
181 | * was not set in a UFFD_EVENT_PAGEFAULT, it means it | |
182 | * was a read fault, otherwise if set it means it's | |
183 | * a write fault. | |
86039bd3 | 184 | */ |
a9b85f94 | 185 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 AA |
186 | if (reason & VM_UFFD_WP) |
187 | /* | |
a9b85f94 AA |
188 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
189 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was | |
190 | * not set in a UFFD_EVENT_PAGEFAULT, it means it was | |
191 | * a missing fault, otherwise if set it means it's a | |
192 | * write protect fault. | |
86039bd3 | 193 | */ |
a9b85f94 AA |
194 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
195 | return msg; | |
86039bd3 AA |
196 | } |
197 | ||
8d2afd96 AA |
198 | /* |
199 | * Verify the pagetables are still not ok after having reigstered into | |
200 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
201 | * userfault that has already been resolved, if userfaultfd_read and | |
202 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
203 | * threads. | |
204 | */ | |
205 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
206 | unsigned long address, | |
207 | unsigned long flags, | |
208 | unsigned long reason) | |
209 | { | |
210 | struct mm_struct *mm = ctx->mm; | |
211 | pgd_t *pgd; | |
212 | pud_t *pud; | |
213 | pmd_t *pmd, _pmd; | |
214 | pte_t *pte; | |
215 | bool ret = true; | |
216 | ||
217 | VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
218 | ||
219 | pgd = pgd_offset(mm, address); | |
220 | if (!pgd_present(*pgd)) | |
221 | goto out; | |
222 | pud = pud_offset(pgd, address); | |
223 | if (!pud_present(*pud)) | |
224 | goto out; | |
225 | pmd = pmd_offset(pud, address); | |
226 | /* | |
227 | * READ_ONCE must function as a barrier with narrower scope | |
228 | * and it must be equivalent to: | |
229 | * _pmd = *pmd; barrier(); | |
230 | * | |
231 | * This is to deal with the instability (as in | |
232 | * pmd_trans_unstable) of the pmd. | |
233 | */ | |
234 | _pmd = READ_ONCE(*pmd); | |
235 | if (!pmd_present(_pmd)) | |
236 | goto out; | |
237 | ||
238 | ret = false; | |
239 | if (pmd_trans_huge(_pmd)) | |
240 | goto out; | |
241 | ||
242 | /* | |
243 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
244 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
245 | */ | |
246 | pte = pte_offset_map(pmd, address); | |
247 | /* | |
248 | * Lockless access: we're in a wait_event so it's ok if it | |
249 | * changes under us. | |
250 | */ | |
251 | if (pte_none(*pte)) | |
252 | ret = true; | |
253 | pte_unmap(pte); | |
254 | ||
255 | out: | |
256 | return ret; | |
257 | } | |
258 | ||
86039bd3 AA |
259 | /* |
260 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
261 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
262 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
263 | * recommendation in __lock_page_or_retry is not an understatement. | |
264 | * | |
265 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released | |
266 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is | |
267 | * not set. | |
268 | * | |
269 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
270 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
271 | * fatal_signal_pending()s, and the mmap_sem must be released before | |
272 | * returning it. | |
273 | */ | |
82b0f8c3 | 274 | int handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 275 | { |
82b0f8c3 | 276 | struct mm_struct *mm = vmf->vma->vm_mm; |
86039bd3 AA |
277 | struct userfaultfd_ctx *ctx; |
278 | struct userfaultfd_wait_queue uwq; | |
ba85c702 | 279 | int ret; |
dfa37dc3 | 280 | bool must_wait, return_to_userland; |
15a77c6f | 281 | long blocking_state; |
86039bd3 AA |
282 | |
283 | BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
284 | ||
ba85c702 | 285 | ret = VM_FAULT_SIGBUS; |
82b0f8c3 | 286 | ctx = vmf->vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 287 | if (!ctx) |
ba85c702 | 288 | goto out; |
86039bd3 AA |
289 | |
290 | BUG_ON(ctx->mm != mm); | |
291 | ||
292 | VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP)); | |
293 | VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP)); | |
294 | ||
295 | /* | |
296 | * If it's already released don't get it. This avoids to loop | |
297 | * in __get_user_pages if userfaultfd_release waits on the | |
298 | * caller of handle_userfault to release the mmap_sem. | |
299 | */ | |
300 | if (unlikely(ACCESS_ONCE(ctx->released))) | |
ba85c702 | 301 | goto out; |
86039bd3 | 302 | |
39680f50 LT |
303 | /* |
304 | * We don't do userfault handling for the final child pid update. | |
305 | */ | |
306 | if (current->flags & PF_EXITING) | |
307 | goto out; | |
308 | ||
86039bd3 AA |
309 | /* |
310 | * Check that we can return VM_FAULT_RETRY. | |
311 | * | |
312 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
313 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
314 | * -EBUSY failures, if the userfaultfd is to be extended for | |
315 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
316 | * without first stopping userland access to the memory. For | |
317 | * VM_UFFD_MISSING userfaults this is enough for now. | |
318 | */ | |
82b0f8c3 | 319 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
320 | /* |
321 | * Validate the invariant that nowait must allow retry | |
322 | * to be sure not to return SIGBUS erroneously on | |
323 | * nowait invocations. | |
324 | */ | |
82b0f8c3 | 325 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
326 | #ifdef CONFIG_DEBUG_VM |
327 | if (printk_ratelimit()) { | |
328 | printk(KERN_WARNING | |
82b0f8c3 JK |
329 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
330 | vmf->flags); | |
86039bd3 AA |
331 | dump_stack(); |
332 | } | |
333 | #endif | |
ba85c702 | 334 | goto out; |
86039bd3 AA |
335 | } |
336 | ||
337 | /* | |
338 | * Handle nowait, not much to do other than tell it to retry | |
339 | * and wait. | |
340 | */ | |
ba85c702 | 341 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 342 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 343 | goto out; |
86039bd3 AA |
344 | |
345 | /* take the reference before dropping the mmap_sem */ | |
346 | userfaultfd_ctx_get(ctx); | |
347 | ||
86039bd3 AA |
348 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
349 | uwq.wq.private = current; | |
82b0f8c3 | 350 | uwq.msg = userfault_msg(vmf->address, vmf->flags, reason); |
86039bd3 | 351 | uwq.ctx = ctx; |
15a77c6f | 352 | uwq.waken = false; |
86039bd3 | 353 | |
bae473a4 | 354 | return_to_userland = |
82b0f8c3 | 355 | (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) == |
dfa37dc3 | 356 | (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE); |
15a77c6f AA |
357 | blocking_state = return_to_userland ? TASK_INTERRUPTIBLE : |
358 | TASK_KILLABLE; | |
dfa37dc3 | 359 | |
15b726ef | 360 | spin_lock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
361 | /* |
362 | * After the __add_wait_queue the uwq is visible to userland | |
363 | * through poll/read(). | |
364 | */ | |
15b726ef AA |
365 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
366 | /* | |
367 | * The smp_mb() after __set_current_state prevents the reads | |
368 | * following the spin_unlock to happen before the list_add in | |
369 | * __add_wait_queue. | |
370 | */ | |
15a77c6f | 371 | set_current_state(blocking_state); |
15b726ef | 372 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 373 | |
82b0f8c3 JK |
374 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, |
375 | reason); | |
8d2afd96 AA |
376 | up_read(&mm->mmap_sem); |
377 | ||
378 | if (likely(must_wait && !ACCESS_ONCE(ctx->released) && | |
dfa37dc3 AA |
379 | (return_to_userland ? !signal_pending(current) : |
380 | !fatal_signal_pending(current)))) { | |
86039bd3 AA |
381 | wake_up_poll(&ctx->fd_wqh, POLLIN); |
382 | schedule(); | |
ba85c702 | 383 | ret |= VM_FAULT_MAJOR; |
15a77c6f AA |
384 | |
385 | /* | |
386 | * False wakeups can orginate even from rwsem before | |
387 | * up_read() however userfaults will wait either for a | |
388 | * targeted wakeup on the specific uwq waitqueue from | |
389 | * wake_userfault() or for signals or for uffd | |
390 | * release. | |
391 | */ | |
392 | while (!READ_ONCE(uwq.waken)) { | |
393 | /* | |
394 | * This needs the full smp_store_mb() | |
395 | * guarantee as the state write must be | |
396 | * visible to other CPUs before reading | |
397 | * uwq.waken from other CPUs. | |
398 | */ | |
399 | set_current_state(blocking_state); | |
400 | if (READ_ONCE(uwq.waken) || | |
401 | READ_ONCE(ctx->released) || | |
402 | (return_to_userland ? signal_pending(current) : | |
403 | fatal_signal_pending(current))) | |
404 | break; | |
405 | schedule(); | |
406 | } | |
ba85c702 | 407 | } |
86039bd3 | 408 | |
ba85c702 | 409 | __set_current_state(TASK_RUNNING); |
15b726ef | 410 | |
dfa37dc3 AA |
411 | if (return_to_userland) { |
412 | if (signal_pending(current) && | |
413 | !fatal_signal_pending(current)) { | |
414 | /* | |
415 | * If we got a SIGSTOP or SIGCONT and this is | |
416 | * a normal userland page fault, just let | |
417 | * userland return so the signal will be | |
418 | * handled and gdb debugging works. The page | |
419 | * fault code immediately after we return from | |
420 | * this function is going to release the | |
421 | * mmap_sem and it's not depending on it | |
422 | * (unlike gup would if we were not to return | |
423 | * VM_FAULT_RETRY). | |
424 | * | |
425 | * If a fatal signal is pending we still take | |
426 | * the streamlined VM_FAULT_RETRY failure path | |
427 | * and there's no need to retake the mmap_sem | |
428 | * in such case. | |
429 | */ | |
430 | down_read(&mm->mmap_sem); | |
431 | ret = 0; | |
432 | } | |
433 | } | |
434 | ||
15b726ef AA |
435 | /* |
436 | * Here we race with the list_del; list_add in | |
437 | * userfaultfd_ctx_read(), however because we don't ever run | |
438 | * list_del_init() to refile across the two lists, the prev | |
439 | * and next pointers will never point to self. list_add also | |
440 | * would never let any of the two pointers to point to | |
441 | * self. So list_empty_careful won't risk to see both pointers | |
442 | * pointing to self at any time during the list refile. The | |
443 | * only case where list_del_init() is called is the full | |
444 | * removal in the wake function and there we don't re-list_add | |
445 | * and it's fine not to block on the spinlock. The uwq on this | |
446 | * kernel stack can be released after the list_del_init. | |
447 | */ | |
ba85c702 | 448 | if (!list_empty_careful(&uwq.wq.task_list)) { |
15b726ef AA |
449 | spin_lock(&ctx->fault_pending_wqh.lock); |
450 | /* | |
451 | * No need of list_del_init(), the uwq on the stack | |
452 | * will be freed shortly anyway. | |
453 | */ | |
454 | list_del(&uwq.wq.task_list); | |
455 | spin_unlock(&ctx->fault_pending_wqh.lock); | |
86039bd3 | 456 | } |
86039bd3 AA |
457 | |
458 | /* | |
459 | * ctx may go away after this if the userfault pseudo fd is | |
460 | * already released. | |
461 | */ | |
462 | userfaultfd_ctx_put(ctx); | |
463 | ||
ba85c702 AA |
464 | out: |
465 | return ret; | |
86039bd3 AA |
466 | } |
467 | ||
9cd75c3c PE |
468 | static int __maybe_unused userfaultfd_event_wait_completion( |
469 | struct userfaultfd_ctx *ctx, | |
470 | struct userfaultfd_wait_queue *ewq) | |
471 | { | |
472 | int ret = 0; | |
473 | ||
474 | ewq->ctx = ctx; | |
475 | init_waitqueue_entry(&ewq->wq, current); | |
476 | ||
477 | spin_lock(&ctx->event_wqh.lock); | |
478 | /* | |
479 | * After the __add_wait_queue the uwq is visible to userland | |
480 | * through poll/read(). | |
481 | */ | |
482 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
483 | for (;;) { | |
484 | set_current_state(TASK_KILLABLE); | |
485 | if (ewq->msg.event == 0) | |
486 | break; | |
487 | if (ACCESS_ONCE(ctx->released) || | |
488 | fatal_signal_pending(current)) { | |
489 | ret = -1; | |
490 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
491 | break; | |
492 | } | |
493 | ||
494 | spin_unlock(&ctx->event_wqh.lock); | |
495 | ||
496 | wake_up_poll(&ctx->fd_wqh, POLLIN); | |
497 | schedule(); | |
498 | ||
499 | spin_lock(&ctx->event_wqh.lock); | |
500 | } | |
501 | __set_current_state(TASK_RUNNING); | |
502 | spin_unlock(&ctx->event_wqh.lock); | |
503 | ||
504 | /* | |
505 | * ctx may go away after this if the userfault pseudo fd is | |
506 | * already released. | |
507 | */ | |
508 | ||
509 | userfaultfd_ctx_put(ctx); | |
510 | return ret; | |
511 | } | |
512 | ||
513 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
514 | struct userfaultfd_wait_queue *ewq) | |
515 | { | |
516 | ewq->msg.event = 0; | |
517 | wake_up_locked(&ctx->event_wqh); | |
518 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
519 | } | |
520 | ||
86039bd3 AA |
521 | static int userfaultfd_release(struct inode *inode, struct file *file) |
522 | { | |
523 | struct userfaultfd_ctx *ctx = file->private_data; | |
524 | struct mm_struct *mm = ctx->mm; | |
525 | struct vm_area_struct *vma, *prev; | |
526 | /* len == 0 means wake all */ | |
527 | struct userfaultfd_wake_range range = { .len = 0, }; | |
528 | unsigned long new_flags; | |
529 | ||
530 | ACCESS_ONCE(ctx->released) = true; | |
531 | ||
d2005e3f ON |
532 | if (!mmget_not_zero(mm)) |
533 | goto wakeup; | |
534 | ||
86039bd3 AA |
535 | /* |
536 | * Flush page faults out of all CPUs. NOTE: all page faults | |
537 | * must be retried without returning VM_FAULT_SIGBUS if | |
538 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
539 | * changes while handle_userfault released the mmap_sem. So | |
540 | * it's critical that released is set to true (above), before | |
541 | * taking the mmap_sem for writing. | |
542 | */ | |
543 | down_write(&mm->mmap_sem); | |
544 | prev = NULL; | |
545 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
546 | cond_resched(); | |
547 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
548 | !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
549 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { | |
550 | prev = vma; | |
551 | continue; | |
552 | } | |
553 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
554 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, | |
555 | new_flags, vma->anon_vma, | |
556 | vma->vm_file, vma->vm_pgoff, | |
557 | vma_policy(vma), | |
558 | NULL_VM_UFFD_CTX); | |
559 | if (prev) | |
560 | vma = prev; | |
561 | else | |
562 | prev = vma; | |
563 | vma->vm_flags = new_flags; | |
564 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
565 | } | |
566 | up_write(&mm->mmap_sem); | |
d2005e3f ON |
567 | mmput(mm); |
568 | wakeup: | |
86039bd3 | 569 | /* |
15b726ef | 570 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 571 | * the last page faults that may have been already waiting on |
15b726ef | 572 | * the fault_*wqh. |
86039bd3 | 573 | */ |
15b726ef | 574 | spin_lock(&ctx->fault_pending_wqh.lock); |
ac5be6b4 AA |
575 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
576 | __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range); | |
15b726ef | 577 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
578 | |
579 | wake_up_poll(&ctx->fd_wqh, POLLHUP); | |
580 | userfaultfd_ctx_put(ctx); | |
581 | return 0; | |
582 | } | |
583 | ||
15b726ef | 584 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
585 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
586 | wait_queue_head_t *wqh) | |
86039bd3 AA |
587 | { |
588 | wait_queue_t *wq; | |
15b726ef | 589 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 590 | |
6dcc27fd | 591 | VM_BUG_ON(!spin_is_locked(&wqh->lock)); |
86039bd3 | 592 | |
15b726ef | 593 | uwq = NULL; |
6dcc27fd | 594 | if (!waitqueue_active(wqh)) |
15b726ef AA |
595 | goto out; |
596 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
6dcc27fd | 597 | wq = list_last_entry(&wqh->task_list, typeof(*wq), task_list); |
15b726ef AA |
598 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
599 | out: | |
600 | return uwq; | |
86039bd3 | 601 | } |
6dcc27fd PE |
602 | |
603 | static inline struct userfaultfd_wait_queue *find_userfault( | |
604 | struct userfaultfd_ctx *ctx) | |
605 | { | |
606 | return find_userfault_in(&ctx->fault_pending_wqh); | |
607 | } | |
86039bd3 | 608 | |
9cd75c3c PE |
609 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
610 | struct userfaultfd_ctx *ctx) | |
611 | { | |
612 | return find_userfault_in(&ctx->event_wqh); | |
613 | } | |
614 | ||
86039bd3 AA |
615 | static unsigned int userfaultfd_poll(struct file *file, poll_table *wait) |
616 | { | |
617 | struct userfaultfd_ctx *ctx = file->private_data; | |
618 | unsigned int ret; | |
619 | ||
620 | poll_wait(file, &ctx->fd_wqh, wait); | |
621 | ||
622 | switch (ctx->state) { | |
623 | case UFFD_STATE_WAIT_API: | |
624 | return POLLERR; | |
625 | case UFFD_STATE_RUNNING: | |
ba85c702 AA |
626 | /* |
627 | * poll() never guarantees that read won't block. | |
628 | * userfaults can be waken before they're read(). | |
629 | */ | |
630 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
631 | return POLLERR; | |
15b726ef AA |
632 | /* |
633 | * lockless access to see if there are pending faults | |
634 | * __pollwait last action is the add_wait_queue but | |
635 | * the spin_unlock would allow the waitqueue_active to | |
636 | * pass above the actual list_add inside | |
637 | * add_wait_queue critical section. So use a full | |
638 | * memory barrier to serialize the list_add write of | |
639 | * add_wait_queue() with the waitqueue_active read | |
640 | * below. | |
641 | */ | |
642 | ret = 0; | |
643 | smp_mb(); | |
644 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
645 | ret = POLLIN; | |
9cd75c3c PE |
646 | else if (waitqueue_active(&ctx->event_wqh)) |
647 | ret = POLLIN; | |
648 | ||
86039bd3 AA |
649 | return ret; |
650 | default: | |
8474901a AA |
651 | WARN_ON_ONCE(1); |
652 | return POLLERR; | |
86039bd3 AA |
653 | } |
654 | } | |
655 | ||
656 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, | |
a9b85f94 | 657 | struct uffd_msg *msg) |
86039bd3 AA |
658 | { |
659 | ssize_t ret; | |
660 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 661 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 662 | |
15b726ef | 663 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
86039bd3 AA |
664 | spin_lock(&ctx->fd_wqh.lock); |
665 | __add_wait_queue(&ctx->fd_wqh, &wait); | |
666 | for (;;) { | |
667 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
668 | spin_lock(&ctx->fault_pending_wqh.lock); |
669 | uwq = find_userfault(ctx); | |
670 | if (uwq) { | |
2c5b7e1b AA |
671 | /* |
672 | * Use a seqcount to repeat the lockless check | |
673 | * in wake_userfault() to avoid missing | |
674 | * wakeups because during the refile both | |
675 | * waitqueue could become empty if this is the | |
676 | * only userfault. | |
677 | */ | |
678 | write_seqcount_begin(&ctx->refile_seq); | |
679 | ||
86039bd3 | 680 | /* |
15b726ef AA |
681 | * The fault_pending_wqh.lock prevents the uwq |
682 | * to disappear from under us. | |
683 | * | |
684 | * Refile this userfault from | |
685 | * fault_pending_wqh to fault_wqh, it's not | |
686 | * pending anymore after we read it. | |
687 | * | |
688 | * Use list_del() by hand (as | |
689 | * userfaultfd_wake_function also uses | |
690 | * list_del_init() by hand) to be sure nobody | |
691 | * changes __remove_wait_queue() to use | |
692 | * list_del_init() in turn breaking the | |
693 | * !list_empty_careful() check in | |
694 | * handle_userfault(). The uwq->wq.task_list | |
695 | * must never be empty at any time during the | |
696 | * refile, or the waitqueue could disappear | |
697 | * from under us. The "wait_queue_head_t" | |
698 | * parameter of __remove_wait_queue() is unused | |
699 | * anyway. | |
86039bd3 | 700 | */ |
15b726ef AA |
701 | list_del(&uwq->wq.task_list); |
702 | __add_wait_queue(&ctx->fault_wqh, &uwq->wq); | |
703 | ||
2c5b7e1b AA |
704 | write_seqcount_end(&ctx->refile_seq); |
705 | ||
a9b85f94 AA |
706 | /* careful to always initialize msg if ret == 0 */ |
707 | *msg = uwq->msg; | |
15b726ef | 708 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
709 | ret = 0; |
710 | break; | |
711 | } | |
15b726ef | 712 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
713 | |
714 | spin_lock(&ctx->event_wqh.lock); | |
715 | uwq = find_userfault_evt(ctx); | |
716 | if (uwq) { | |
717 | *msg = uwq->msg; | |
718 | ||
719 | userfaultfd_event_complete(ctx, uwq); | |
720 | spin_unlock(&ctx->event_wqh.lock); | |
721 | ret = 0; | |
722 | break; | |
723 | } | |
724 | spin_unlock(&ctx->event_wqh.lock); | |
725 | ||
86039bd3 AA |
726 | if (signal_pending(current)) { |
727 | ret = -ERESTARTSYS; | |
728 | break; | |
729 | } | |
730 | if (no_wait) { | |
731 | ret = -EAGAIN; | |
732 | break; | |
733 | } | |
734 | spin_unlock(&ctx->fd_wqh.lock); | |
735 | schedule(); | |
736 | spin_lock(&ctx->fd_wqh.lock); | |
737 | } | |
738 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
739 | __set_current_state(TASK_RUNNING); | |
740 | spin_unlock(&ctx->fd_wqh.lock); | |
741 | ||
742 | return ret; | |
743 | } | |
744 | ||
745 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
746 | size_t count, loff_t *ppos) | |
747 | { | |
748 | struct userfaultfd_ctx *ctx = file->private_data; | |
749 | ssize_t _ret, ret = 0; | |
a9b85f94 | 750 | struct uffd_msg msg; |
86039bd3 AA |
751 | int no_wait = file->f_flags & O_NONBLOCK; |
752 | ||
753 | if (ctx->state == UFFD_STATE_WAIT_API) | |
754 | return -EINVAL; | |
86039bd3 AA |
755 | |
756 | for (;;) { | |
a9b85f94 | 757 | if (count < sizeof(msg)) |
86039bd3 | 758 | return ret ? ret : -EINVAL; |
a9b85f94 | 759 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg); |
86039bd3 AA |
760 | if (_ret < 0) |
761 | return ret ? ret : _ret; | |
a9b85f94 | 762 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 763 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
764 | ret += sizeof(msg); |
765 | buf += sizeof(msg); | |
766 | count -= sizeof(msg); | |
86039bd3 AA |
767 | /* |
768 | * Allow to read more than one fault at time but only | |
769 | * block if waiting for the very first one. | |
770 | */ | |
771 | no_wait = O_NONBLOCK; | |
772 | } | |
773 | } | |
774 | ||
775 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
776 | struct userfaultfd_wake_range *range) | |
777 | { | |
778 | unsigned long start, end; | |
779 | ||
780 | start = range->start; | |
781 | end = range->start + range->len; | |
782 | ||
15b726ef | 783 | spin_lock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 784 | /* wake all in the range and autoremove */ |
15b726ef | 785 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 786 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
787 | range); |
788 | if (waitqueue_active(&ctx->fault_wqh)) | |
ac5be6b4 | 789 | __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range); |
15b726ef | 790 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
791 | } |
792 | ||
793 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
794 | struct userfaultfd_wake_range *range) | |
795 | { | |
2c5b7e1b AA |
796 | unsigned seq; |
797 | bool need_wakeup; | |
798 | ||
86039bd3 AA |
799 | /* |
800 | * To be sure waitqueue_active() is not reordered by the CPU | |
801 | * before the pagetable update, use an explicit SMP memory | |
802 | * barrier here. PT lock release or up_read(mmap_sem) still | |
803 | * have release semantics that can allow the | |
804 | * waitqueue_active() to be reordered before the pte update. | |
805 | */ | |
806 | smp_mb(); | |
807 | ||
808 | /* | |
809 | * Use waitqueue_active because it's very frequent to | |
810 | * change the address space atomically even if there are no | |
811 | * userfaults yet. So we take the spinlock only when we're | |
812 | * sure we've userfaults to wake. | |
813 | */ | |
2c5b7e1b AA |
814 | do { |
815 | seq = read_seqcount_begin(&ctx->refile_seq); | |
816 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
817 | waitqueue_active(&ctx->fault_wqh); | |
818 | cond_resched(); | |
819 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
820 | if (need_wakeup) | |
86039bd3 AA |
821 | __wake_userfault(ctx, range); |
822 | } | |
823 | ||
824 | static __always_inline int validate_range(struct mm_struct *mm, | |
825 | __u64 start, __u64 len) | |
826 | { | |
827 | __u64 task_size = mm->task_size; | |
828 | ||
829 | if (start & ~PAGE_MASK) | |
830 | return -EINVAL; | |
831 | if (len & ~PAGE_MASK) | |
832 | return -EINVAL; | |
833 | if (!len) | |
834 | return -EINVAL; | |
835 | if (start < mmap_min_addr) | |
836 | return -EINVAL; | |
837 | if (start >= task_size) | |
838 | return -EINVAL; | |
839 | if (len > task_size - start) | |
840 | return -EINVAL; | |
841 | return 0; | |
842 | } | |
843 | ||
844 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, | |
845 | unsigned long arg) | |
846 | { | |
847 | struct mm_struct *mm = ctx->mm; | |
848 | struct vm_area_struct *vma, *prev, *cur; | |
849 | int ret; | |
850 | struct uffdio_register uffdio_register; | |
851 | struct uffdio_register __user *user_uffdio_register; | |
852 | unsigned long vm_flags, new_flags; | |
853 | bool found; | |
854 | unsigned long start, end, vma_end; | |
855 | ||
856 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
857 | ||
858 | ret = -EFAULT; | |
859 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
860 | sizeof(uffdio_register)-sizeof(__u64))) | |
861 | goto out; | |
862 | ||
863 | ret = -EINVAL; | |
864 | if (!uffdio_register.mode) | |
865 | goto out; | |
866 | if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING| | |
867 | UFFDIO_REGISTER_MODE_WP)) | |
868 | goto out; | |
869 | vm_flags = 0; | |
870 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
871 | vm_flags |= VM_UFFD_MISSING; | |
872 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { | |
873 | vm_flags |= VM_UFFD_WP; | |
874 | /* | |
875 | * FIXME: remove the below error constraint by | |
876 | * implementing the wprotect tracking mode. | |
877 | */ | |
878 | ret = -EINVAL; | |
879 | goto out; | |
880 | } | |
881 | ||
882 | ret = validate_range(mm, uffdio_register.range.start, | |
883 | uffdio_register.range.len); | |
884 | if (ret) | |
885 | goto out; | |
886 | ||
887 | start = uffdio_register.range.start; | |
888 | end = start + uffdio_register.range.len; | |
889 | ||
d2005e3f ON |
890 | ret = -ENOMEM; |
891 | if (!mmget_not_zero(mm)) | |
892 | goto out; | |
893 | ||
86039bd3 AA |
894 | down_write(&mm->mmap_sem); |
895 | vma = find_vma_prev(mm, start, &prev); | |
86039bd3 AA |
896 | if (!vma) |
897 | goto out_unlock; | |
898 | ||
899 | /* check that there's at least one vma in the range */ | |
900 | ret = -EINVAL; | |
901 | if (vma->vm_start >= end) | |
902 | goto out_unlock; | |
903 | ||
904 | /* | |
905 | * Search for not compatible vmas. | |
906 | * | |
907 | * FIXME: this shall be relaxed later so that it doesn't fail | |
908 | * on tmpfs backed vmas (in addition to the current allowance | |
909 | * on anonymous vmas). | |
910 | */ | |
911 | found = false; | |
912 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
913 | cond_resched(); | |
914 | ||
915 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
916 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
917 | ||
918 | /* check not compatible vmas */ | |
919 | ret = -EINVAL; | |
a94720bf | 920 | if (!vma_is_anonymous(cur)) |
86039bd3 AA |
921 | goto out_unlock; |
922 | ||
923 | /* | |
924 | * Check that this vma isn't already owned by a | |
925 | * different userfaultfd. We can't allow more than one | |
926 | * userfaultfd to own a single vma simultaneously or we | |
927 | * wouldn't know which one to deliver the userfaults to. | |
928 | */ | |
929 | ret = -EBUSY; | |
930 | if (cur->vm_userfaultfd_ctx.ctx && | |
931 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
932 | goto out_unlock; | |
933 | ||
934 | found = true; | |
935 | } | |
936 | BUG_ON(!found); | |
937 | ||
938 | if (vma->vm_start < start) | |
939 | prev = vma; | |
940 | ||
941 | ret = 0; | |
942 | do { | |
943 | cond_resched(); | |
944 | ||
a94720bf | 945 | BUG_ON(!vma_is_anonymous(vma)); |
86039bd3 AA |
946 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
947 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
948 | ||
949 | /* | |
950 | * Nothing to do: this vma is already registered into this | |
951 | * userfaultfd and with the right tracking mode too. | |
952 | */ | |
953 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
954 | (vma->vm_flags & vm_flags) == vm_flags) | |
955 | goto skip; | |
956 | ||
957 | if (vma->vm_start > start) | |
958 | start = vma->vm_start; | |
959 | vma_end = min(end, vma->vm_end); | |
960 | ||
961 | new_flags = (vma->vm_flags & ~vm_flags) | vm_flags; | |
962 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
963 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
964 | vma_policy(vma), | |
965 | ((struct vm_userfaultfd_ctx){ ctx })); | |
966 | if (prev) { | |
967 | vma = prev; | |
968 | goto next; | |
969 | } | |
970 | if (vma->vm_start < start) { | |
971 | ret = split_vma(mm, vma, start, 1); | |
972 | if (ret) | |
973 | break; | |
974 | } | |
975 | if (vma->vm_end > end) { | |
976 | ret = split_vma(mm, vma, end, 0); | |
977 | if (ret) | |
978 | break; | |
979 | } | |
980 | next: | |
981 | /* | |
982 | * In the vma_merge() successful mprotect-like case 8: | |
983 | * the next vma was merged into the current one and | |
984 | * the current one has not been updated yet. | |
985 | */ | |
986 | vma->vm_flags = new_flags; | |
987 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
988 | ||
989 | skip: | |
990 | prev = vma; | |
991 | start = vma->vm_end; | |
992 | vma = vma->vm_next; | |
993 | } while (vma && vma->vm_start < end); | |
994 | out_unlock: | |
995 | up_write(&mm->mmap_sem); | |
d2005e3f | 996 | mmput(mm); |
86039bd3 AA |
997 | if (!ret) { |
998 | /* | |
999 | * Now that we scanned all vmas we can already tell | |
1000 | * userland which ioctls methods are guaranteed to | |
1001 | * succeed on this range. | |
1002 | */ | |
1003 | if (put_user(UFFD_API_RANGE_IOCTLS, | |
1004 | &user_uffdio_register->ioctls)) | |
1005 | ret = -EFAULT; | |
1006 | } | |
1007 | out: | |
1008 | return ret; | |
1009 | } | |
1010 | ||
1011 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1012 | unsigned long arg) | |
1013 | { | |
1014 | struct mm_struct *mm = ctx->mm; | |
1015 | struct vm_area_struct *vma, *prev, *cur; | |
1016 | int ret; | |
1017 | struct uffdio_range uffdio_unregister; | |
1018 | unsigned long new_flags; | |
1019 | bool found; | |
1020 | unsigned long start, end, vma_end; | |
1021 | const void __user *buf = (void __user *)arg; | |
1022 | ||
1023 | ret = -EFAULT; | |
1024 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1025 | goto out; | |
1026 | ||
1027 | ret = validate_range(mm, uffdio_unregister.start, | |
1028 | uffdio_unregister.len); | |
1029 | if (ret) | |
1030 | goto out; | |
1031 | ||
1032 | start = uffdio_unregister.start; | |
1033 | end = start + uffdio_unregister.len; | |
1034 | ||
d2005e3f ON |
1035 | ret = -ENOMEM; |
1036 | if (!mmget_not_zero(mm)) | |
1037 | goto out; | |
1038 | ||
86039bd3 AA |
1039 | down_write(&mm->mmap_sem); |
1040 | vma = find_vma_prev(mm, start, &prev); | |
86039bd3 AA |
1041 | if (!vma) |
1042 | goto out_unlock; | |
1043 | ||
1044 | /* check that there's at least one vma in the range */ | |
1045 | ret = -EINVAL; | |
1046 | if (vma->vm_start >= end) | |
1047 | goto out_unlock; | |
1048 | ||
1049 | /* | |
1050 | * Search for not compatible vmas. | |
1051 | * | |
1052 | * FIXME: this shall be relaxed later so that it doesn't fail | |
1053 | * on tmpfs backed vmas (in addition to the current allowance | |
1054 | * on anonymous vmas). | |
1055 | */ | |
1056 | found = false; | |
1057 | ret = -EINVAL; | |
1058 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
1059 | cond_resched(); | |
1060 | ||
1061 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1062 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1063 | ||
1064 | /* | |
1065 | * Check not compatible vmas, not strictly required | |
1066 | * here as not compatible vmas cannot have an | |
1067 | * userfaultfd_ctx registered on them, but this | |
1068 | * provides for more strict behavior to notice | |
1069 | * unregistration errors. | |
1070 | */ | |
a94720bf | 1071 | if (!vma_is_anonymous(cur)) |
86039bd3 AA |
1072 | goto out_unlock; |
1073 | ||
1074 | found = true; | |
1075 | } | |
1076 | BUG_ON(!found); | |
1077 | ||
1078 | if (vma->vm_start < start) | |
1079 | prev = vma; | |
1080 | ||
1081 | ret = 0; | |
1082 | do { | |
1083 | cond_resched(); | |
1084 | ||
a94720bf | 1085 | BUG_ON(!vma_is_anonymous(vma)); |
86039bd3 AA |
1086 | |
1087 | /* | |
1088 | * Nothing to do: this vma is already registered into this | |
1089 | * userfaultfd and with the right tracking mode too. | |
1090 | */ | |
1091 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1092 | goto skip; | |
1093 | ||
1094 | if (vma->vm_start > start) | |
1095 | start = vma->vm_start; | |
1096 | vma_end = min(end, vma->vm_end); | |
1097 | ||
1098 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
1099 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
1100 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1101 | vma_policy(vma), | |
1102 | NULL_VM_UFFD_CTX); | |
1103 | if (prev) { | |
1104 | vma = prev; | |
1105 | goto next; | |
1106 | } | |
1107 | if (vma->vm_start < start) { | |
1108 | ret = split_vma(mm, vma, start, 1); | |
1109 | if (ret) | |
1110 | break; | |
1111 | } | |
1112 | if (vma->vm_end > end) { | |
1113 | ret = split_vma(mm, vma, end, 0); | |
1114 | if (ret) | |
1115 | break; | |
1116 | } | |
1117 | next: | |
1118 | /* | |
1119 | * In the vma_merge() successful mprotect-like case 8: | |
1120 | * the next vma was merged into the current one and | |
1121 | * the current one has not been updated yet. | |
1122 | */ | |
1123 | vma->vm_flags = new_flags; | |
1124 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
1125 | ||
1126 | skip: | |
1127 | prev = vma; | |
1128 | start = vma->vm_end; | |
1129 | vma = vma->vm_next; | |
1130 | } while (vma && vma->vm_start < end); | |
1131 | out_unlock: | |
1132 | up_write(&mm->mmap_sem); | |
d2005e3f | 1133 | mmput(mm); |
86039bd3 AA |
1134 | out: |
1135 | return ret; | |
1136 | } | |
1137 | ||
1138 | /* | |
ba85c702 AA |
1139 | * userfaultfd_wake may be used in combination with the |
1140 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1141 | */ |
1142 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1143 | unsigned long arg) | |
1144 | { | |
1145 | int ret; | |
1146 | struct uffdio_range uffdio_wake; | |
1147 | struct userfaultfd_wake_range range; | |
1148 | const void __user *buf = (void __user *)arg; | |
1149 | ||
1150 | ret = -EFAULT; | |
1151 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1152 | goto out; | |
1153 | ||
1154 | ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len); | |
1155 | if (ret) | |
1156 | goto out; | |
1157 | ||
1158 | range.start = uffdio_wake.start; | |
1159 | range.len = uffdio_wake.len; | |
1160 | ||
1161 | /* | |
1162 | * len == 0 means wake all and we don't want to wake all here, | |
1163 | * so check it again to be sure. | |
1164 | */ | |
1165 | VM_BUG_ON(!range.len); | |
1166 | ||
1167 | wake_userfault(ctx, &range); | |
1168 | ret = 0; | |
1169 | ||
1170 | out: | |
1171 | return ret; | |
1172 | } | |
1173 | ||
ad465cae AA |
1174 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1175 | unsigned long arg) | |
1176 | { | |
1177 | __s64 ret; | |
1178 | struct uffdio_copy uffdio_copy; | |
1179 | struct uffdio_copy __user *user_uffdio_copy; | |
1180 | struct userfaultfd_wake_range range; | |
1181 | ||
1182 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1183 | ||
1184 | ret = -EFAULT; | |
1185 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1186 | /* don't copy "copy" last field */ | |
1187 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1188 | goto out; | |
1189 | ||
1190 | ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len); | |
1191 | if (ret) | |
1192 | goto out; | |
1193 | /* | |
1194 | * double check for wraparound just in case. copy_from_user() | |
1195 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1196 | * in the userland range. | |
1197 | */ | |
1198 | ret = -EINVAL; | |
1199 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1200 | goto out; | |
1201 | if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE) | |
1202 | goto out; | |
d2005e3f ON |
1203 | if (mmget_not_zero(ctx->mm)) { |
1204 | ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src, | |
1205 | uffdio_copy.len); | |
1206 | mmput(ctx->mm); | |
1207 | } | |
ad465cae AA |
1208 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1209 | return -EFAULT; | |
1210 | if (ret < 0) | |
1211 | goto out; | |
1212 | BUG_ON(!ret); | |
1213 | /* len == 0 would wake all */ | |
1214 | range.len = ret; | |
1215 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1216 | range.start = uffdio_copy.dst; | |
1217 | wake_userfault(ctx, &range); | |
1218 | } | |
1219 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1220 | out: | |
1221 | return ret; | |
1222 | } | |
1223 | ||
1224 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1225 | unsigned long arg) | |
1226 | { | |
1227 | __s64 ret; | |
1228 | struct uffdio_zeropage uffdio_zeropage; | |
1229 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1230 | struct userfaultfd_wake_range range; | |
1231 | ||
1232 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1233 | ||
1234 | ret = -EFAULT; | |
1235 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1236 | /* don't copy "zeropage" last field */ | |
1237 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1238 | goto out; | |
1239 | ||
1240 | ret = validate_range(ctx->mm, uffdio_zeropage.range.start, | |
1241 | uffdio_zeropage.range.len); | |
1242 | if (ret) | |
1243 | goto out; | |
1244 | ret = -EINVAL; | |
1245 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1246 | goto out; | |
1247 | ||
d2005e3f ON |
1248 | if (mmget_not_zero(ctx->mm)) { |
1249 | ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start, | |
1250 | uffdio_zeropage.range.len); | |
1251 | mmput(ctx->mm); | |
1252 | } | |
ad465cae AA |
1253 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1254 | return -EFAULT; | |
1255 | if (ret < 0) | |
1256 | goto out; | |
1257 | /* len == 0 would wake all */ | |
1258 | BUG_ON(!ret); | |
1259 | range.len = ret; | |
1260 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1261 | range.start = uffdio_zeropage.range.start; | |
1262 | wake_userfault(ctx, &range); | |
1263 | } | |
1264 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1265 | out: | |
1266 | return ret; | |
1267 | } | |
1268 | ||
9cd75c3c PE |
1269 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1270 | { | |
1271 | /* | |
1272 | * For the current set of features the bits just coincide | |
1273 | */ | |
1274 | return (unsigned int)user_features; | |
1275 | } | |
1276 | ||
86039bd3 AA |
1277 | /* |
1278 | * userland asks for a certain API version and we return which bits | |
1279 | * and ioctl commands are implemented in this kernel for such API | |
1280 | * version or -EINVAL if unknown. | |
1281 | */ | |
1282 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1283 | unsigned long arg) | |
1284 | { | |
1285 | struct uffdio_api uffdio_api; | |
1286 | void __user *buf = (void __user *)arg; | |
1287 | int ret; | |
65603144 | 1288 | __u64 features; |
86039bd3 AA |
1289 | |
1290 | ret = -EINVAL; | |
1291 | if (ctx->state != UFFD_STATE_WAIT_API) | |
1292 | goto out; | |
1293 | ret = -EFAULT; | |
a9b85f94 | 1294 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1295 | goto out; |
65603144 AA |
1296 | features = uffdio_api.features; |
1297 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) { | |
86039bd3 AA |
1298 | memset(&uffdio_api, 0, sizeof(uffdio_api)); |
1299 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1300 | goto out; | |
1301 | ret = -EINVAL; | |
1302 | goto out; | |
1303 | } | |
65603144 AA |
1304 | /* report all available features and ioctls to userland */ |
1305 | uffdio_api.features = UFFD_API_FEATURES; | |
86039bd3 AA |
1306 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1307 | ret = -EFAULT; | |
1308 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1309 | goto out; | |
1310 | ctx->state = UFFD_STATE_RUNNING; | |
65603144 AA |
1311 | /* only enable the requested features for this uffd context */ |
1312 | ctx->features = uffd_ctx_features(features); | |
86039bd3 AA |
1313 | ret = 0; |
1314 | out: | |
1315 | return ret; | |
1316 | } | |
1317 | ||
1318 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
1319 | unsigned long arg) | |
1320 | { | |
1321 | int ret = -EINVAL; | |
1322 | struct userfaultfd_ctx *ctx = file->private_data; | |
1323 | ||
e6485a47 AA |
1324 | if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API) |
1325 | return -EINVAL; | |
1326 | ||
86039bd3 AA |
1327 | switch(cmd) { |
1328 | case UFFDIO_API: | |
1329 | ret = userfaultfd_api(ctx, arg); | |
1330 | break; | |
1331 | case UFFDIO_REGISTER: | |
1332 | ret = userfaultfd_register(ctx, arg); | |
1333 | break; | |
1334 | case UFFDIO_UNREGISTER: | |
1335 | ret = userfaultfd_unregister(ctx, arg); | |
1336 | break; | |
1337 | case UFFDIO_WAKE: | |
1338 | ret = userfaultfd_wake(ctx, arg); | |
1339 | break; | |
ad465cae AA |
1340 | case UFFDIO_COPY: |
1341 | ret = userfaultfd_copy(ctx, arg); | |
1342 | break; | |
1343 | case UFFDIO_ZEROPAGE: | |
1344 | ret = userfaultfd_zeropage(ctx, arg); | |
1345 | break; | |
86039bd3 AA |
1346 | } |
1347 | return ret; | |
1348 | } | |
1349 | ||
1350 | #ifdef CONFIG_PROC_FS | |
1351 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
1352 | { | |
1353 | struct userfaultfd_ctx *ctx = f->private_data; | |
1354 | wait_queue_t *wq; | |
1355 | struct userfaultfd_wait_queue *uwq; | |
1356 | unsigned long pending = 0, total = 0; | |
1357 | ||
15b726ef AA |
1358 | spin_lock(&ctx->fault_pending_wqh.lock); |
1359 | list_for_each_entry(wq, &ctx->fault_pending_wqh.task_list, task_list) { | |
1360 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
1361 | pending++; | |
1362 | total++; | |
1363 | } | |
86039bd3 AA |
1364 | list_for_each_entry(wq, &ctx->fault_wqh.task_list, task_list) { |
1365 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
86039bd3 AA |
1366 | total++; |
1367 | } | |
15b726ef | 1368 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1369 | |
1370 | /* | |
1371 | * If more protocols will be added, there will be all shown | |
1372 | * separated by a space. Like this: | |
1373 | * protocols: aa:... bb:... | |
1374 | */ | |
1375 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
3f602d27 | 1376 | pending, total, UFFD_API, UFFD_API_FEATURES, |
86039bd3 AA |
1377 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
1378 | } | |
1379 | #endif | |
1380 | ||
1381 | static const struct file_operations userfaultfd_fops = { | |
1382 | #ifdef CONFIG_PROC_FS | |
1383 | .show_fdinfo = userfaultfd_show_fdinfo, | |
1384 | #endif | |
1385 | .release = userfaultfd_release, | |
1386 | .poll = userfaultfd_poll, | |
1387 | .read = userfaultfd_read, | |
1388 | .unlocked_ioctl = userfaultfd_ioctl, | |
1389 | .compat_ioctl = userfaultfd_ioctl, | |
1390 | .llseek = noop_llseek, | |
1391 | }; | |
1392 | ||
3004ec9c AA |
1393 | static void init_once_userfaultfd_ctx(void *mem) |
1394 | { | |
1395 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
1396 | ||
1397 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
1398 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 1399 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 1400 | init_waitqueue_head(&ctx->fd_wqh); |
2c5b7e1b | 1401 | seqcount_init(&ctx->refile_seq); |
3004ec9c AA |
1402 | } |
1403 | ||
86039bd3 AA |
1404 | /** |
1405 | * userfaultfd_file_create - Creates an userfaultfd file pointer. | |
1406 | * @flags: Flags for the userfaultfd file. | |
1407 | * | |
1408 | * This function creates an userfaultfd file pointer, w/out installing | |
1409 | * it into the fd table. This is useful when the userfaultfd file is | |
1410 | * used during the initialization of data structures that require | |
1411 | * extra setup after the userfaultfd creation. So the userfaultfd | |
1412 | * creation is split into the file pointer creation phase, and the | |
1413 | * file descriptor installation phase. In this way races with | |
1414 | * userspace closing the newly installed file descriptor can be | |
1415 | * avoided. Returns an userfaultfd file pointer, or a proper error | |
1416 | * pointer. | |
1417 | */ | |
1418 | static struct file *userfaultfd_file_create(int flags) | |
1419 | { | |
1420 | struct file *file; | |
1421 | struct userfaultfd_ctx *ctx; | |
1422 | ||
1423 | BUG_ON(!current->mm); | |
1424 | ||
1425 | /* Check the UFFD_* constants for consistency. */ | |
1426 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); | |
1427 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
1428 | ||
1429 | file = ERR_PTR(-EINVAL); | |
1430 | if (flags & ~UFFD_SHARED_FCNTL_FLAGS) | |
1431 | goto out; | |
1432 | ||
1433 | file = ERR_PTR(-ENOMEM); | |
3004ec9c | 1434 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 AA |
1435 | if (!ctx) |
1436 | goto out; | |
1437 | ||
1438 | atomic_set(&ctx->refcount, 1); | |
86039bd3 | 1439 | ctx->flags = flags; |
9cd75c3c | 1440 | ctx->features = 0; |
86039bd3 AA |
1441 | ctx->state = UFFD_STATE_WAIT_API; |
1442 | ctx->released = false; | |
1443 | ctx->mm = current->mm; | |
1444 | /* prevent the mm struct to be freed */ | |
d2005e3f | 1445 | atomic_inc(&ctx->mm->mm_count); |
86039bd3 AA |
1446 | |
1447 | file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx, | |
1448 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS)); | |
c03e946f | 1449 | if (IS_ERR(file)) { |
d2005e3f | 1450 | mmdrop(ctx->mm); |
3004ec9c | 1451 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 1452 | } |
86039bd3 AA |
1453 | out: |
1454 | return file; | |
1455 | } | |
1456 | ||
1457 | SYSCALL_DEFINE1(userfaultfd, int, flags) | |
1458 | { | |
1459 | int fd, error; | |
1460 | struct file *file; | |
1461 | ||
1462 | error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS); | |
1463 | if (error < 0) | |
1464 | return error; | |
1465 | fd = error; | |
1466 | ||
1467 | file = userfaultfd_file_create(flags); | |
1468 | if (IS_ERR(file)) { | |
1469 | error = PTR_ERR(file); | |
1470 | goto err_put_unused_fd; | |
1471 | } | |
1472 | fd_install(fd, file); | |
1473 | ||
1474 | return fd; | |
1475 | ||
1476 | err_put_unused_fd: | |
1477 | put_unused_fd(fd); | |
1478 | ||
1479 | return error; | |
1480 | } | |
3004ec9c AA |
1481 | |
1482 | static int __init userfaultfd_init(void) | |
1483 | { | |
1484 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", | |
1485 | sizeof(struct userfaultfd_ctx), | |
1486 | 0, | |
1487 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
1488 | init_once_userfaultfd_ctx); | |
1489 | return 0; | |
1490 | } | |
1491 | __initcall(userfaultfd_init); |