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