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