]>
Commit | Line | Data |
---|---|---|
86039bd3 AA |
1 | /* |
2 | * fs/userfaultfd.c | |
3 | * | |
4 | * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org> | |
5 | * Copyright (C) 2008-2009 Red Hat, Inc. | |
6 | * Copyright (C) 2015 Red Hat, Inc. | |
7 | * | |
8 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
9 | * the COPYING file in the top-level directory. | |
10 | * | |
11 | * Some part derived from fs/eventfd.c (anon inode setup) and | |
12 | * mm/ksm.c (mm hashing). | |
13 | */ | |
14 | ||
9cd75c3c | 15 | #include <linux/list.h> |
86039bd3 | 16 | #include <linux/hashtable.h> |
174cd4b1 | 17 | #include <linux/sched/signal.h> |
6e84f315 | 18 | #include <linux/sched/mm.h> |
86039bd3 AA |
19 | #include <linux/mm.h> |
20 | #include <linux/poll.h> | |
21 | #include <linux/slab.h> | |
22 | #include <linux/seq_file.h> | |
23 | #include <linux/file.h> | |
24 | #include <linux/bug.h> | |
25 | #include <linux/anon_inodes.h> | |
26 | #include <linux/syscalls.h> | |
27 | #include <linux/userfaultfd_k.h> | |
28 | #include <linux/mempolicy.h> | |
29 | #include <linux/ioctl.h> | |
30 | #include <linux/security.h> | |
cab350af | 31 | #include <linux/hugetlb.h> |
86039bd3 | 32 | |
3004ec9c AA |
33 | static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly; |
34 | ||
86039bd3 AA |
35 | enum userfaultfd_state { |
36 | UFFD_STATE_WAIT_API, | |
37 | UFFD_STATE_RUNNING, | |
38 | }; | |
39 | ||
3004ec9c AA |
40 | /* |
41 | * Start with fault_pending_wqh and fault_wqh so they're more likely | |
42 | * to be in the same cacheline. | |
43 | */ | |
86039bd3 | 44 | struct userfaultfd_ctx { |
15b726ef AA |
45 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
46 | wait_queue_head_t fault_pending_wqh; | |
47 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
48 | wait_queue_head_t fault_wqh; |
49 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
50 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
51 | /* waitqueue head for events */ |
52 | wait_queue_head_t event_wqh; | |
2c5b7e1b AA |
53 | /* a refile sequence protected by fault_pending_wqh lock */ |
54 | struct seqcount refile_seq; | |
3004ec9c AA |
55 | /* pseudo fd refcounting */ |
56 | atomic_t refcount; | |
86039bd3 AA |
57 | /* userfaultfd syscall flags */ |
58 | unsigned int flags; | |
9cd75c3c PE |
59 | /* features requested from the userspace */ |
60 | unsigned int features; | |
86039bd3 AA |
61 | /* state machine */ |
62 | enum userfaultfd_state state; | |
63 | /* released */ | |
64 | bool released; | |
65 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ | |
66 | struct mm_struct *mm; | |
67 | }; | |
68 | ||
893e26e6 PE |
69 | struct userfaultfd_fork_ctx { |
70 | struct userfaultfd_ctx *orig; | |
71 | struct userfaultfd_ctx *new; | |
72 | struct list_head list; | |
73 | }; | |
74 | ||
897ab3e0 MR |
75 | struct userfaultfd_unmap_ctx { |
76 | struct userfaultfd_ctx *ctx; | |
77 | unsigned long start; | |
78 | unsigned long end; | |
79 | struct list_head list; | |
80 | }; | |
81 | ||
86039bd3 | 82 | struct userfaultfd_wait_queue { |
a9b85f94 | 83 | struct uffd_msg msg; |
ac6424b9 | 84 | wait_queue_entry_t wq; |
86039bd3 | 85 | struct userfaultfd_ctx *ctx; |
15a77c6f | 86 | bool waken; |
86039bd3 AA |
87 | }; |
88 | ||
89 | struct userfaultfd_wake_range { | |
90 | unsigned long start; | |
91 | unsigned long len; | |
92 | }; | |
93 | ||
ac6424b9 | 94 | static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode, |
86039bd3 AA |
95 | int wake_flags, void *key) |
96 | { | |
97 | struct userfaultfd_wake_range *range = key; | |
98 | int ret; | |
99 | struct userfaultfd_wait_queue *uwq; | |
100 | unsigned long start, len; | |
101 | ||
102 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
103 | ret = 0; | |
86039bd3 AA |
104 | /* len == 0 means wake all */ |
105 | start = range->start; | |
106 | len = range->len; | |
a9b85f94 AA |
107 | if (len && (start > uwq->msg.arg.pagefault.address || |
108 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 109 | goto out; |
15a77c6f AA |
110 | WRITE_ONCE(uwq->waken, true); |
111 | /* | |
a9668cd6 PZ |
112 | * The Program-Order guarantees provided by the scheduler |
113 | * ensure uwq->waken is visible before the task is woken. | |
15a77c6f | 114 | */ |
86039bd3 | 115 | ret = wake_up_state(wq->private, mode); |
a9668cd6 | 116 | if (ret) { |
86039bd3 AA |
117 | /* |
118 | * Wake only once, autoremove behavior. | |
119 | * | |
a9668cd6 PZ |
120 | * After the effect of list_del_init is visible to the other |
121 | * CPUs, the waitqueue may disappear from under us, see the | |
122 | * !list_empty_careful() in handle_userfault(). | |
123 | * | |
124 | * try_to_wake_up() has an implicit smp_mb(), and the | |
125 | * wq->private is read before calling the extern function | |
126 | * "wake_up_state" (which in turns calls try_to_wake_up). | |
86039bd3 | 127 | */ |
2055da97 | 128 | list_del_init(&wq->entry); |
a9668cd6 | 129 | } |
86039bd3 AA |
130 | out: |
131 | return ret; | |
132 | } | |
133 | ||
134 | /** | |
135 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
136 | * context. | |
137 | * @ctx: [in] Pointer to the userfaultfd context. | |
86039bd3 AA |
138 | */ |
139 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
140 | { | |
141 | if (!atomic_inc_not_zero(&ctx->refcount)) | |
142 | BUG(); | |
143 | } | |
144 | ||
145 | /** | |
146 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
147 | * context. | |
148 | * @ctx: [in] Pointer to userfaultfd context. | |
149 | * | |
150 | * The userfaultfd context reference must have been previously acquired either | |
151 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
152 | */ | |
153 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
154 | { | |
155 | if (atomic_dec_and_test(&ctx->refcount)) { | |
156 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); | |
157 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
158 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
159 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
160 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
161 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
162 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
163 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 164 | mmdrop(ctx->mm); |
3004ec9c | 165 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
166 | } |
167 | } | |
168 | ||
a9b85f94 | 169 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 170 | { |
a9b85f94 AA |
171 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
172 | /* | |
173 | * Must use memset to zero out the paddings or kernel data is | |
174 | * leaked to userland. | |
175 | */ | |
176 | memset(msg, 0, sizeof(struct uffd_msg)); | |
177 | } | |
178 | ||
179 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
180 | unsigned int flags, | |
181 | unsigned long reason) | |
182 | { | |
183 | struct uffd_msg msg; | |
184 | msg_init(&msg); | |
185 | msg.event = UFFD_EVENT_PAGEFAULT; | |
186 | msg.arg.pagefault.address = address; | |
86039bd3 AA |
187 | if (flags & FAULT_FLAG_WRITE) |
188 | /* | |
a4605a61 | 189 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
a9b85f94 AA |
190 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE |
191 | * was not set in a UFFD_EVENT_PAGEFAULT, it means it | |
192 | * was a read fault, otherwise if set it means it's | |
193 | * a write fault. | |
86039bd3 | 194 | */ |
a9b85f94 | 195 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 AA |
196 | if (reason & VM_UFFD_WP) |
197 | /* | |
a9b85f94 AA |
198 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
199 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was | |
200 | * not set in a UFFD_EVENT_PAGEFAULT, it means it was | |
201 | * a missing fault, otherwise if set it means it's a | |
202 | * write protect fault. | |
86039bd3 | 203 | */ |
a9b85f94 AA |
204 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
205 | return msg; | |
86039bd3 AA |
206 | } |
207 | ||
369cd212 MK |
208 | #ifdef CONFIG_HUGETLB_PAGE |
209 | /* | |
210 | * Same functionality as userfaultfd_must_wait below with modifications for | |
211 | * hugepmd ranges. | |
212 | */ | |
213 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 214 | struct vm_area_struct *vma, |
369cd212 MK |
215 | unsigned long address, |
216 | unsigned long flags, | |
217 | unsigned long reason) | |
218 | { | |
219 | struct mm_struct *mm = ctx->mm; | |
220 | pte_t *pte; | |
221 | bool ret = true; | |
222 | ||
223 | VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
224 | ||
7868a208 | 225 | pte = huge_pte_offset(mm, address, vma_mmu_pagesize(vma)); |
369cd212 MK |
226 | if (!pte) |
227 | goto out; | |
228 | ||
229 | ret = false; | |
230 | ||
231 | /* | |
232 | * Lockless access: we're in a wait_event so it's ok if it | |
233 | * changes under us. | |
234 | */ | |
235 | if (huge_pte_none(*pte)) | |
236 | ret = true; | |
237 | if (!huge_pte_write(*pte) && (reason & VM_UFFD_WP)) | |
238 | ret = true; | |
239 | out: | |
240 | return ret; | |
241 | } | |
242 | #else | |
243 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 244 | struct vm_area_struct *vma, |
369cd212 MK |
245 | unsigned long address, |
246 | unsigned long flags, | |
247 | unsigned long reason) | |
248 | { | |
249 | return false; /* should never get here */ | |
250 | } | |
251 | #endif /* CONFIG_HUGETLB_PAGE */ | |
252 | ||
8d2afd96 AA |
253 | /* |
254 | * Verify the pagetables are still not ok after having reigstered into | |
255 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
256 | * userfault that has already been resolved, if userfaultfd_read and | |
257 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
258 | * threads. | |
259 | */ | |
260 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
261 | unsigned long address, | |
262 | unsigned long flags, | |
263 | unsigned long reason) | |
264 | { | |
265 | struct mm_struct *mm = ctx->mm; | |
266 | pgd_t *pgd; | |
c2febafc | 267 | p4d_t *p4d; |
8d2afd96 AA |
268 | pud_t *pud; |
269 | pmd_t *pmd, _pmd; | |
270 | pte_t *pte; | |
271 | bool ret = true; | |
272 | ||
273 | VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
274 | ||
275 | pgd = pgd_offset(mm, address); | |
276 | if (!pgd_present(*pgd)) | |
277 | goto out; | |
c2febafc KS |
278 | p4d = p4d_offset(pgd, address); |
279 | if (!p4d_present(*p4d)) | |
280 | goto out; | |
281 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
282 | if (!pud_present(*pud)) |
283 | goto out; | |
284 | pmd = pmd_offset(pud, address); | |
285 | /* | |
286 | * READ_ONCE must function as a barrier with narrower scope | |
287 | * and it must be equivalent to: | |
288 | * _pmd = *pmd; barrier(); | |
289 | * | |
290 | * This is to deal with the instability (as in | |
291 | * pmd_trans_unstable) of the pmd. | |
292 | */ | |
293 | _pmd = READ_ONCE(*pmd); | |
294 | if (!pmd_present(_pmd)) | |
295 | goto out; | |
296 | ||
297 | ret = false; | |
298 | if (pmd_trans_huge(_pmd)) | |
299 | goto out; | |
300 | ||
301 | /* | |
302 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
303 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
304 | */ | |
305 | pte = pte_offset_map(pmd, address); | |
306 | /* | |
307 | * Lockless access: we're in a wait_event so it's ok if it | |
308 | * changes under us. | |
309 | */ | |
310 | if (pte_none(*pte)) | |
311 | ret = true; | |
312 | pte_unmap(pte); | |
313 | ||
314 | out: | |
315 | return ret; | |
316 | } | |
317 | ||
86039bd3 AA |
318 | /* |
319 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
320 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
321 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
322 | * recommendation in __lock_page_or_retry is not an understatement. | |
323 | * | |
324 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released | |
325 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is | |
326 | * not set. | |
327 | * | |
328 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
329 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
330 | * fatal_signal_pending()s, and the mmap_sem must be released before | |
331 | * returning it. | |
332 | */ | |
82b0f8c3 | 333 | int handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 334 | { |
82b0f8c3 | 335 | struct mm_struct *mm = vmf->vma->vm_mm; |
86039bd3 AA |
336 | struct userfaultfd_ctx *ctx; |
337 | struct userfaultfd_wait_queue uwq; | |
ba85c702 | 338 | int ret; |
dfa37dc3 | 339 | bool must_wait, return_to_userland; |
15a77c6f | 340 | long blocking_state; |
86039bd3 | 341 | |
ba85c702 | 342 | ret = VM_FAULT_SIGBUS; |
64c2b203 AA |
343 | |
344 | /* | |
345 | * We don't do userfault handling for the final child pid update. | |
346 | * | |
347 | * We also don't do userfault handling during | |
348 | * coredumping. hugetlbfs has the special | |
349 | * follow_hugetlb_page() to skip missing pages in the | |
350 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
351 | * the no_page_table() helper in follow_page_mask(), but the | |
352 | * shmem_vm_ops->fault method is invoked even during | |
353 | * coredumping without mmap_sem and it ends up here. | |
354 | */ | |
355 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
356 | goto out; | |
357 | ||
358 | /* | |
359 | * Coredumping runs without mmap_sem so we can only check that | |
360 | * the mmap_sem is held, if PF_DUMPCORE was not set. | |
361 | */ | |
362 | WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem)); | |
363 | ||
82b0f8c3 | 364 | ctx = vmf->vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 365 | if (!ctx) |
ba85c702 | 366 | goto out; |
86039bd3 AA |
367 | |
368 | BUG_ON(ctx->mm != mm); | |
369 | ||
370 | VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP)); | |
371 | VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP)); | |
372 | ||
2d6d6f5a PS |
373 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
374 | goto out; | |
375 | ||
86039bd3 AA |
376 | /* |
377 | * If it's already released don't get it. This avoids to loop | |
378 | * in __get_user_pages if userfaultfd_release waits on the | |
379 | * caller of handle_userfault to release the mmap_sem. | |
380 | */ | |
381 | if (unlikely(ACCESS_ONCE(ctx->released))) | |
ba85c702 | 382 | goto out; |
86039bd3 AA |
383 | |
384 | /* | |
385 | * Check that we can return VM_FAULT_RETRY. | |
386 | * | |
387 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
388 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
389 | * -EBUSY failures, if the userfaultfd is to be extended for | |
390 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
391 | * without first stopping userland access to the memory. For | |
392 | * VM_UFFD_MISSING userfaults this is enough for now. | |
393 | */ | |
82b0f8c3 | 394 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
395 | /* |
396 | * Validate the invariant that nowait must allow retry | |
397 | * to be sure not to return SIGBUS erroneously on | |
398 | * nowait invocations. | |
399 | */ | |
82b0f8c3 | 400 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
401 | #ifdef CONFIG_DEBUG_VM |
402 | if (printk_ratelimit()) { | |
403 | printk(KERN_WARNING | |
82b0f8c3 JK |
404 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
405 | vmf->flags); | |
86039bd3 AA |
406 | dump_stack(); |
407 | } | |
408 | #endif | |
ba85c702 | 409 | goto out; |
86039bd3 AA |
410 | } |
411 | ||
412 | /* | |
413 | * Handle nowait, not much to do other than tell it to retry | |
414 | * and wait. | |
415 | */ | |
ba85c702 | 416 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 417 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 418 | goto out; |
86039bd3 AA |
419 | |
420 | /* take the reference before dropping the mmap_sem */ | |
421 | userfaultfd_ctx_get(ctx); | |
422 | ||
86039bd3 AA |
423 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
424 | uwq.wq.private = current; | |
82b0f8c3 | 425 | uwq.msg = userfault_msg(vmf->address, vmf->flags, reason); |
86039bd3 | 426 | uwq.ctx = ctx; |
15a77c6f | 427 | uwq.waken = false; |
86039bd3 | 428 | |
bae473a4 | 429 | return_to_userland = |
82b0f8c3 | 430 | (vmf->flags & (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE)) == |
dfa37dc3 | 431 | (FAULT_FLAG_USER|FAULT_FLAG_KILLABLE); |
15a77c6f AA |
432 | blocking_state = return_to_userland ? TASK_INTERRUPTIBLE : |
433 | TASK_KILLABLE; | |
dfa37dc3 | 434 | |
15b726ef | 435 | spin_lock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
436 | /* |
437 | * After the __add_wait_queue the uwq is visible to userland | |
438 | * through poll/read(). | |
439 | */ | |
15b726ef AA |
440 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
441 | /* | |
442 | * The smp_mb() after __set_current_state prevents the reads | |
443 | * following the spin_unlock to happen before the list_add in | |
444 | * __add_wait_queue. | |
445 | */ | |
15a77c6f | 446 | set_current_state(blocking_state); |
15b726ef | 447 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 448 | |
369cd212 MK |
449 | if (!is_vm_hugetlb_page(vmf->vma)) |
450 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, | |
451 | reason); | |
452 | else | |
7868a208 PA |
453 | must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma, |
454 | vmf->address, | |
369cd212 | 455 | vmf->flags, reason); |
8d2afd96 AA |
456 | up_read(&mm->mmap_sem); |
457 | ||
458 | if (likely(must_wait && !ACCESS_ONCE(ctx->released) && | |
dfa37dc3 AA |
459 | (return_to_userland ? !signal_pending(current) : |
460 | !fatal_signal_pending(current)))) { | |
86039bd3 AA |
461 | wake_up_poll(&ctx->fd_wqh, POLLIN); |
462 | schedule(); | |
ba85c702 | 463 | ret |= VM_FAULT_MAJOR; |
15a77c6f AA |
464 | |
465 | /* | |
466 | * False wakeups can orginate even from rwsem before | |
467 | * up_read() however userfaults will wait either for a | |
468 | * targeted wakeup on the specific uwq waitqueue from | |
469 | * wake_userfault() or for signals or for uffd | |
470 | * release. | |
471 | */ | |
472 | while (!READ_ONCE(uwq.waken)) { | |
473 | /* | |
474 | * This needs the full smp_store_mb() | |
475 | * guarantee as the state write must be | |
476 | * visible to other CPUs before reading | |
477 | * uwq.waken from other CPUs. | |
478 | */ | |
479 | set_current_state(blocking_state); | |
480 | if (READ_ONCE(uwq.waken) || | |
481 | READ_ONCE(ctx->released) || | |
482 | (return_to_userland ? signal_pending(current) : | |
483 | fatal_signal_pending(current))) | |
484 | break; | |
485 | schedule(); | |
486 | } | |
ba85c702 | 487 | } |
86039bd3 | 488 | |
ba85c702 | 489 | __set_current_state(TASK_RUNNING); |
15b726ef | 490 | |
dfa37dc3 AA |
491 | if (return_to_userland) { |
492 | if (signal_pending(current) && | |
493 | !fatal_signal_pending(current)) { | |
494 | /* | |
495 | * If we got a SIGSTOP or SIGCONT and this is | |
496 | * a normal userland page fault, just let | |
497 | * userland return so the signal will be | |
498 | * handled and gdb debugging works. The page | |
499 | * fault code immediately after we return from | |
500 | * this function is going to release the | |
501 | * mmap_sem and it's not depending on it | |
502 | * (unlike gup would if we were not to return | |
503 | * VM_FAULT_RETRY). | |
504 | * | |
505 | * If a fatal signal is pending we still take | |
506 | * the streamlined VM_FAULT_RETRY failure path | |
507 | * and there's no need to retake the mmap_sem | |
508 | * in such case. | |
509 | */ | |
510 | down_read(&mm->mmap_sem); | |
6bbc4a41 | 511 | ret = VM_FAULT_NOPAGE; |
dfa37dc3 AA |
512 | } |
513 | } | |
514 | ||
15b726ef AA |
515 | /* |
516 | * Here we race with the list_del; list_add in | |
517 | * userfaultfd_ctx_read(), however because we don't ever run | |
518 | * list_del_init() to refile across the two lists, the prev | |
519 | * and next pointers will never point to self. list_add also | |
520 | * would never let any of the two pointers to point to | |
521 | * self. So list_empty_careful won't risk to see both pointers | |
522 | * pointing to self at any time during the list refile. The | |
523 | * only case where list_del_init() is called is the full | |
524 | * removal in the wake function and there we don't re-list_add | |
525 | * and it's fine not to block on the spinlock. The uwq on this | |
526 | * kernel stack can be released after the list_del_init. | |
527 | */ | |
2055da97 | 528 | if (!list_empty_careful(&uwq.wq.entry)) { |
15b726ef AA |
529 | spin_lock(&ctx->fault_pending_wqh.lock); |
530 | /* | |
531 | * No need of list_del_init(), the uwq on the stack | |
532 | * will be freed shortly anyway. | |
533 | */ | |
2055da97 | 534 | list_del(&uwq.wq.entry); |
15b726ef | 535 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 536 | } |
86039bd3 AA |
537 | |
538 | /* | |
539 | * ctx may go away after this if the userfault pseudo fd is | |
540 | * already released. | |
541 | */ | |
542 | userfaultfd_ctx_put(ctx); | |
543 | ||
ba85c702 AA |
544 | out: |
545 | return ret; | |
86039bd3 AA |
546 | } |
547 | ||
8c9e7bb7 AA |
548 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
549 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 550 | { |
9a69a829 AA |
551 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
552 | goto out; | |
9cd75c3c PE |
553 | |
554 | ewq->ctx = ctx; | |
555 | init_waitqueue_entry(&ewq->wq, current); | |
556 | ||
557 | spin_lock(&ctx->event_wqh.lock); | |
558 | /* | |
559 | * After the __add_wait_queue the uwq is visible to userland | |
560 | * through poll/read(). | |
561 | */ | |
562 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
563 | for (;;) { | |
564 | set_current_state(TASK_KILLABLE); | |
565 | if (ewq->msg.event == 0) | |
566 | break; | |
567 | if (ACCESS_ONCE(ctx->released) || | |
568 | fatal_signal_pending(current)) { | |
9cd75c3c | 569 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
570 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
571 | struct userfaultfd_ctx *new; | |
572 | ||
573 | new = (struct userfaultfd_ctx *) | |
574 | (unsigned long) | |
575 | ewq->msg.arg.reserved.reserved1; | |
576 | ||
577 | userfaultfd_ctx_put(new); | |
578 | } | |
9cd75c3c PE |
579 | break; |
580 | } | |
581 | ||
582 | spin_unlock(&ctx->event_wqh.lock); | |
583 | ||
584 | wake_up_poll(&ctx->fd_wqh, POLLIN); | |
585 | schedule(); | |
586 | ||
587 | spin_lock(&ctx->event_wqh.lock); | |
588 | } | |
589 | __set_current_state(TASK_RUNNING); | |
590 | spin_unlock(&ctx->event_wqh.lock); | |
591 | ||
592 | /* | |
593 | * ctx may go away after this if the userfault pseudo fd is | |
594 | * already released. | |
595 | */ | |
9a69a829 | 596 | out: |
9cd75c3c | 597 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
598 | } |
599 | ||
600 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
601 | struct userfaultfd_wait_queue *ewq) | |
602 | { | |
603 | ewq->msg.event = 0; | |
604 | wake_up_locked(&ctx->event_wqh); | |
605 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
606 | } | |
607 | ||
893e26e6 PE |
608 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
609 | { | |
610 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
611 | struct userfaultfd_fork_ctx *fctx; | |
612 | ||
613 | octx = vma->vm_userfaultfd_ctx.ctx; | |
614 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
615 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
616 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
617 | return 0; | |
618 | } | |
619 | ||
620 | list_for_each_entry(fctx, fcs, list) | |
621 | if (fctx->orig == octx) { | |
622 | ctx = fctx->new; | |
623 | break; | |
624 | } | |
625 | ||
626 | if (!ctx) { | |
627 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
628 | if (!fctx) | |
629 | return -ENOMEM; | |
630 | ||
631 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
632 | if (!ctx) { | |
633 | kfree(fctx); | |
634 | return -ENOMEM; | |
635 | } | |
636 | ||
637 | atomic_set(&ctx->refcount, 1); | |
638 | ctx->flags = octx->flags; | |
639 | ctx->state = UFFD_STATE_RUNNING; | |
640 | ctx->features = octx->features; | |
641 | ctx->released = false; | |
642 | ctx->mm = vma->vm_mm; | |
d3aadc8e | 643 | atomic_inc(&ctx->mm->mm_count); |
893e26e6 PE |
644 | |
645 | userfaultfd_ctx_get(octx); | |
646 | fctx->orig = octx; | |
647 | fctx->new = ctx; | |
648 | list_add_tail(&fctx->list, fcs); | |
649 | } | |
650 | ||
651 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
652 | return 0; | |
653 | } | |
654 | ||
8c9e7bb7 | 655 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
656 | { |
657 | struct userfaultfd_ctx *ctx = fctx->orig; | |
658 | struct userfaultfd_wait_queue ewq; | |
659 | ||
660 | msg_init(&ewq.msg); | |
661 | ||
662 | ewq.msg.event = UFFD_EVENT_FORK; | |
663 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
664 | ||
8c9e7bb7 | 665 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
666 | } |
667 | ||
668 | void dup_userfaultfd_complete(struct list_head *fcs) | |
669 | { | |
893e26e6 PE |
670 | struct userfaultfd_fork_ctx *fctx, *n; |
671 | ||
672 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 673 | dup_fctx(fctx); |
893e26e6 PE |
674 | list_del(&fctx->list); |
675 | kfree(fctx); | |
676 | } | |
677 | } | |
678 | ||
72f87654 PE |
679 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
680 | struct vm_userfaultfd_ctx *vm_ctx) | |
681 | { | |
682 | struct userfaultfd_ctx *ctx; | |
683 | ||
684 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
685 | if (ctx && (ctx->features & UFFD_FEATURE_EVENT_REMAP)) { | |
686 | vm_ctx->ctx = ctx; | |
687 | userfaultfd_ctx_get(ctx); | |
688 | } | |
689 | } | |
690 | ||
90794bf1 | 691 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
692 | unsigned long from, unsigned long to, |
693 | unsigned long len) | |
694 | { | |
90794bf1 | 695 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
696 | struct userfaultfd_wait_queue ewq; |
697 | ||
698 | if (!ctx) | |
699 | return; | |
700 | ||
701 | if (to & ~PAGE_MASK) { | |
702 | userfaultfd_ctx_put(ctx); | |
703 | return; | |
704 | } | |
705 | ||
706 | msg_init(&ewq.msg); | |
707 | ||
708 | ewq.msg.event = UFFD_EVENT_REMAP; | |
709 | ewq.msg.arg.remap.from = from; | |
710 | ewq.msg.arg.remap.to = to; | |
711 | ewq.msg.arg.remap.len = len; | |
712 | ||
713 | userfaultfd_event_wait_completion(ctx, &ewq); | |
714 | } | |
715 | ||
70ccb92f | 716 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 717 | unsigned long start, unsigned long end) |
05ce7724 PE |
718 | { |
719 | struct mm_struct *mm = vma->vm_mm; | |
720 | struct userfaultfd_ctx *ctx; | |
721 | struct userfaultfd_wait_queue ewq; | |
722 | ||
723 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 724 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 725 | return true; |
05ce7724 PE |
726 | |
727 | userfaultfd_ctx_get(ctx); | |
728 | up_read(&mm->mmap_sem); | |
729 | ||
05ce7724 PE |
730 | msg_init(&ewq.msg); |
731 | ||
d811914d MR |
732 | ewq.msg.event = UFFD_EVENT_REMOVE; |
733 | ewq.msg.arg.remove.start = start; | |
734 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
735 | |
736 | userfaultfd_event_wait_completion(ctx, &ewq); | |
737 | ||
70ccb92f | 738 | return false; |
05ce7724 PE |
739 | } |
740 | ||
897ab3e0 MR |
741 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
742 | unsigned long start, unsigned long end) | |
743 | { | |
744 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
745 | ||
746 | list_for_each_entry(unmap_ctx, unmaps, list) | |
747 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
748 | unmap_ctx->end == end) | |
749 | return true; | |
750 | ||
751 | return false; | |
752 | } | |
753 | ||
754 | int userfaultfd_unmap_prep(struct vm_area_struct *vma, | |
755 | unsigned long start, unsigned long end, | |
756 | struct list_head *unmaps) | |
757 | { | |
758 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) { | |
759 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
760 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
761 | ||
762 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
763 | has_unmap_ctx(ctx, unmaps, start, end)) | |
764 | continue; | |
765 | ||
766 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
767 | if (!unmap_ctx) | |
768 | return -ENOMEM; | |
769 | ||
770 | userfaultfd_ctx_get(ctx); | |
771 | unmap_ctx->ctx = ctx; | |
772 | unmap_ctx->start = start; | |
773 | unmap_ctx->end = end; | |
774 | list_add_tail(&unmap_ctx->list, unmaps); | |
775 | } | |
776 | ||
777 | return 0; | |
778 | } | |
779 | ||
780 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
781 | { | |
782 | struct userfaultfd_unmap_ctx *ctx, *n; | |
783 | struct userfaultfd_wait_queue ewq; | |
784 | ||
785 | list_for_each_entry_safe(ctx, n, uf, list) { | |
786 | msg_init(&ewq.msg); | |
787 | ||
788 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
789 | ewq.msg.arg.remove.start = ctx->start; | |
790 | ewq.msg.arg.remove.end = ctx->end; | |
791 | ||
792 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
793 | ||
794 | list_del(&ctx->list); | |
795 | kfree(ctx); | |
796 | } | |
797 | } | |
798 | ||
86039bd3 AA |
799 | static int userfaultfd_release(struct inode *inode, struct file *file) |
800 | { | |
801 | struct userfaultfd_ctx *ctx = file->private_data; | |
802 | struct mm_struct *mm = ctx->mm; | |
803 | struct vm_area_struct *vma, *prev; | |
804 | /* len == 0 means wake all */ | |
805 | struct userfaultfd_wake_range range = { .len = 0, }; | |
806 | unsigned long new_flags; | |
807 | ||
808 | ACCESS_ONCE(ctx->released) = true; | |
809 | ||
d2005e3f ON |
810 | if (!mmget_not_zero(mm)) |
811 | goto wakeup; | |
812 | ||
86039bd3 AA |
813 | /* |
814 | * Flush page faults out of all CPUs. NOTE: all page faults | |
815 | * must be retried without returning VM_FAULT_SIGBUS if | |
816 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
817 | * changes while handle_userfault released the mmap_sem. So | |
818 | * it's critical that released is set to true (above), before | |
819 | * taking the mmap_sem for writing. | |
820 | */ | |
821 | down_write(&mm->mmap_sem); | |
822 | prev = NULL; | |
823 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
824 | cond_resched(); | |
825 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
826 | !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
827 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { | |
828 | prev = vma; | |
829 | continue; | |
830 | } | |
831 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
832 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, | |
833 | new_flags, vma->anon_vma, | |
834 | vma->vm_file, vma->vm_pgoff, | |
835 | vma_policy(vma), | |
836 | NULL_VM_UFFD_CTX); | |
837 | if (prev) | |
838 | vma = prev; | |
839 | else | |
840 | prev = vma; | |
841 | vma->vm_flags = new_flags; | |
842 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
843 | } | |
844 | up_write(&mm->mmap_sem); | |
d2005e3f ON |
845 | mmput(mm); |
846 | wakeup: | |
86039bd3 | 847 | /* |
15b726ef | 848 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 849 | * the last page faults that may have been already waiting on |
15b726ef | 850 | * the fault_*wqh. |
86039bd3 | 851 | */ |
15b726ef | 852 | spin_lock(&ctx->fault_pending_wqh.lock); |
ac5be6b4 AA |
853 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
854 | __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, &range); | |
15b726ef | 855 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 856 | |
5a18b64e MR |
857 | /* Flush pending events that may still wait on event_wqh */ |
858 | wake_up_all(&ctx->event_wqh); | |
859 | ||
86039bd3 AA |
860 | wake_up_poll(&ctx->fd_wqh, POLLHUP); |
861 | userfaultfd_ctx_put(ctx); | |
862 | return 0; | |
863 | } | |
864 | ||
15b726ef | 865 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
866 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
867 | wait_queue_head_t *wqh) | |
86039bd3 | 868 | { |
ac6424b9 | 869 | wait_queue_entry_t *wq; |
15b726ef | 870 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 871 | |
6dcc27fd | 872 | VM_BUG_ON(!spin_is_locked(&wqh->lock)); |
86039bd3 | 873 | |
15b726ef | 874 | uwq = NULL; |
6dcc27fd | 875 | if (!waitqueue_active(wqh)) |
15b726ef AA |
876 | goto out; |
877 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 878 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
879 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
880 | out: | |
881 | return uwq; | |
86039bd3 | 882 | } |
6dcc27fd PE |
883 | |
884 | static inline struct userfaultfd_wait_queue *find_userfault( | |
885 | struct userfaultfd_ctx *ctx) | |
886 | { | |
887 | return find_userfault_in(&ctx->fault_pending_wqh); | |
888 | } | |
86039bd3 | 889 | |
9cd75c3c PE |
890 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
891 | struct userfaultfd_ctx *ctx) | |
892 | { | |
893 | return find_userfault_in(&ctx->event_wqh); | |
894 | } | |
895 | ||
86039bd3 AA |
896 | static unsigned int userfaultfd_poll(struct file *file, poll_table *wait) |
897 | { | |
898 | struct userfaultfd_ctx *ctx = file->private_data; | |
899 | unsigned int ret; | |
900 | ||
901 | poll_wait(file, &ctx->fd_wqh, wait); | |
902 | ||
903 | switch (ctx->state) { | |
904 | case UFFD_STATE_WAIT_API: | |
905 | return POLLERR; | |
906 | case UFFD_STATE_RUNNING: | |
ba85c702 AA |
907 | /* |
908 | * poll() never guarantees that read won't block. | |
909 | * userfaults can be waken before they're read(). | |
910 | */ | |
911 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
912 | return POLLERR; | |
15b726ef AA |
913 | /* |
914 | * lockless access to see if there are pending faults | |
915 | * __pollwait last action is the add_wait_queue but | |
916 | * the spin_unlock would allow the waitqueue_active to | |
917 | * pass above the actual list_add inside | |
918 | * add_wait_queue critical section. So use a full | |
919 | * memory barrier to serialize the list_add write of | |
920 | * add_wait_queue() with the waitqueue_active read | |
921 | * below. | |
922 | */ | |
923 | ret = 0; | |
924 | smp_mb(); | |
925 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
926 | ret = POLLIN; | |
9cd75c3c PE |
927 | else if (waitqueue_active(&ctx->event_wqh)) |
928 | ret = POLLIN; | |
929 | ||
86039bd3 AA |
930 | return ret; |
931 | default: | |
8474901a AA |
932 | WARN_ON_ONCE(1); |
933 | return POLLERR; | |
86039bd3 AA |
934 | } |
935 | } | |
936 | ||
893e26e6 PE |
937 | static const struct file_operations userfaultfd_fops; |
938 | ||
939 | static int resolve_userfault_fork(struct userfaultfd_ctx *ctx, | |
940 | struct userfaultfd_ctx *new, | |
941 | struct uffd_msg *msg) | |
942 | { | |
943 | int fd; | |
944 | struct file *file; | |
945 | unsigned int flags = new->flags & UFFD_SHARED_FCNTL_FLAGS; | |
946 | ||
947 | fd = get_unused_fd_flags(flags); | |
948 | if (fd < 0) | |
949 | return fd; | |
950 | ||
951 | file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, new, | |
952 | O_RDWR | flags); | |
953 | if (IS_ERR(file)) { | |
954 | put_unused_fd(fd); | |
955 | return PTR_ERR(file); | |
956 | } | |
957 | ||
958 | fd_install(fd, file); | |
959 | msg->arg.reserved.reserved1 = 0; | |
960 | msg->arg.fork.ufd = fd; | |
961 | ||
962 | return 0; | |
963 | } | |
964 | ||
86039bd3 | 965 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
a9b85f94 | 966 | struct uffd_msg *msg) |
86039bd3 AA |
967 | { |
968 | ssize_t ret; | |
969 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 970 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
971 | /* |
972 | * Handling fork event requires sleeping operations, so | |
973 | * we drop the event_wqh lock, then do these ops, then | |
974 | * lock it back and wake up the waiter. While the lock is | |
975 | * dropped the ewq may go away so we keep track of it | |
976 | * carefully. | |
977 | */ | |
978 | LIST_HEAD(fork_event); | |
979 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 980 | |
15b726ef | 981 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
86039bd3 AA |
982 | spin_lock(&ctx->fd_wqh.lock); |
983 | __add_wait_queue(&ctx->fd_wqh, &wait); | |
984 | for (;;) { | |
985 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
986 | spin_lock(&ctx->fault_pending_wqh.lock); |
987 | uwq = find_userfault(ctx); | |
988 | if (uwq) { | |
2c5b7e1b AA |
989 | /* |
990 | * Use a seqcount to repeat the lockless check | |
991 | * in wake_userfault() to avoid missing | |
992 | * wakeups because during the refile both | |
993 | * waitqueue could become empty if this is the | |
994 | * only userfault. | |
995 | */ | |
996 | write_seqcount_begin(&ctx->refile_seq); | |
997 | ||
86039bd3 | 998 | /* |
15b726ef AA |
999 | * The fault_pending_wqh.lock prevents the uwq |
1000 | * to disappear from under us. | |
1001 | * | |
1002 | * Refile this userfault from | |
1003 | * fault_pending_wqh to fault_wqh, it's not | |
1004 | * pending anymore after we read it. | |
1005 | * | |
1006 | * Use list_del() by hand (as | |
1007 | * userfaultfd_wake_function also uses | |
1008 | * list_del_init() by hand) to be sure nobody | |
1009 | * changes __remove_wait_queue() to use | |
1010 | * list_del_init() in turn breaking the | |
1011 | * !list_empty_careful() check in | |
2055da97 | 1012 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1013 | * must never be empty at any time during the |
1014 | * refile, or the waitqueue could disappear | |
1015 | * from under us. The "wait_queue_head_t" | |
1016 | * parameter of __remove_wait_queue() is unused | |
1017 | * anyway. | |
86039bd3 | 1018 | */ |
2055da97 | 1019 | list_del(&uwq->wq.entry); |
15b726ef AA |
1020 | __add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
1021 | ||
2c5b7e1b AA |
1022 | write_seqcount_end(&ctx->refile_seq); |
1023 | ||
a9b85f94 AA |
1024 | /* careful to always initialize msg if ret == 0 */ |
1025 | *msg = uwq->msg; | |
15b726ef | 1026 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1027 | ret = 0; |
1028 | break; | |
1029 | } | |
15b726ef | 1030 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1031 | |
1032 | spin_lock(&ctx->event_wqh.lock); | |
1033 | uwq = find_userfault_evt(ctx); | |
1034 | if (uwq) { | |
1035 | *msg = uwq->msg; | |
1036 | ||
893e26e6 PE |
1037 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1038 | fork_nctx = (struct userfaultfd_ctx *) | |
1039 | (unsigned long) | |
1040 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1041 | list_move(&uwq->wq.entry, &fork_event); |
893e26e6 PE |
1042 | spin_unlock(&ctx->event_wqh.lock); |
1043 | ret = 0; | |
1044 | break; | |
1045 | } | |
1046 | ||
9cd75c3c PE |
1047 | userfaultfd_event_complete(ctx, uwq); |
1048 | spin_unlock(&ctx->event_wqh.lock); | |
1049 | ret = 0; | |
1050 | break; | |
1051 | } | |
1052 | spin_unlock(&ctx->event_wqh.lock); | |
1053 | ||
86039bd3 AA |
1054 | if (signal_pending(current)) { |
1055 | ret = -ERESTARTSYS; | |
1056 | break; | |
1057 | } | |
1058 | if (no_wait) { | |
1059 | ret = -EAGAIN; | |
1060 | break; | |
1061 | } | |
1062 | spin_unlock(&ctx->fd_wqh.lock); | |
1063 | schedule(); | |
1064 | spin_lock(&ctx->fd_wqh.lock); | |
1065 | } | |
1066 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1067 | __set_current_state(TASK_RUNNING); | |
1068 | spin_unlock(&ctx->fd_wqh.lock); | |
1069 | ||
893e26e6 PE |
1070 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
1071 | ret = resolve_userfault_fork(ctx, fork_nctx, msg); | |
1072 | ||
1073 | if (!ret) { | |
1074 | spin_lock(&ctx->event_wqh.lock); | |
1075 | if (!list_empty(&fork_event)) { | |
1076 | uwq = list_first_entry(&fork_event, | |
1077 | typeof(*uwq), | |
2055da97 IM |
1078 | wq.entry); |
1079 | list_del(&uwq->wq.entry); | |
893e26e6 PE |
1080 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); |
1081 | userfaultfd_event_complete(ctx, uwq); | |
1082 | } | |
1083 | spin_unlock(&ctx->event_wqh.lock); | |
1084 | } | |
1085 | } | |
1086 | ||
86039bd3 AA |
1087 | return ret; |
1088 | } | |
1089 | ||
1090 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1091 | size_t count, loff_t *ppos) | |
1092 | { | |
1093 | struct userfaultfd_ctx *ctx = file->private_data; | |
1094 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1095 | struct uffd_msg msg; |
86039bd3 AA |
1096 | int no_wait = file->f_flags & O_NONBLOCK; |
1097 | ||
1098 | if (ctx->state == UFFD_STATE_WAIT_API) | |
1099 | return -EINVAL; | |
86039bd3 AA |
1100 | |
1101 | for (;;) { | |
a9b85f94 | 1102 | if (count < sizeof(msg)) |
86039bd3 | 1103 | return ret ? ret : -EINVAL; |
a9b85f94 | 1104 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg); |
86039bd3 AA |
1105 | if (_ret < 0) |
1106 | return ret ? ret : _ret; | |
a9b85f94 | 1107 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1108 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1109 | ret += sizeof(msg); |
1110 | buf += sizeof(msg); | |
1111 | count -= sizeof(msg); | |
86039bd3 AA |
1112 | /* |
1113 | * Allow to read more than one fault at time but only | |
1114 | * block if waiting for the very first one. | |
1115 | */ | |
1116 | no_wait = O_NONBLOCK; | |
1117 | } | |
1118 | } | |
1119 | ||
1120 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1121 | struct userfaultfd_wake_range *range) | |
1122 | { | |
15b726ef | 1123 | spin_lock(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1124 | /* wake all in the range and autoremove */ |
15b726ef | 1125 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1126 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1127 | range); |
1128 | if (waitqueue_active(&ctx->fault_wqh)) | |
ac5be6b4 | 1129 | __wake_up_locked_key(&ctx->fault_wqh, TASK_NORMAL, range); |
15b726ef | 1130 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1131 | } |
1132 | ||
1133 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1134 | struct userfaultfd_wake_range *range) | |
1135 | { | |
2c5b7e1b AA |
1136 | unsigned seq; |
1137 | bool need_wakeup; | |
1138 | ||
86039bd3 AA |
1139 | /* |
1140 | * To be sure waitqueue_active() is not reordered by the CPU | |
1141 | * before the pagetable update, use an explicit SMP memory | |
1142 | * barrier here. PT lock release or up_read(mmap_sem) still | |
1143 | * have release semantics that can allow the | |
1144 | * waitqueue_active() to be reordered before the pte update. | |
1145 | */ | |
1146 | smp_mb(); | |
1147 | ||
1148 | /* | |
1149 | * Use waitqueue_active because it's very frequent to | |
1150 | * change the address space atomically even if there are no | |
1151 | * userfaults yet. So we take the spinlock only when we're | |
1152 | * sure we've userfaults to wake. | |
1153 | */ | |
2c5b7e1b AA |
1154 | do { |
1155 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1156 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1157 | waitqueue_active(&ctx->fault_wqh); | |
1158 | cond_resched(); | |
1159 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1160 | if (need_wakeup) | |
86039bd3 AA |
1161 | __wake_userfault(ctx, range); |
1162 | } | |
1163 | ||
1164 | static __always_inline int validate_range(struct mm_struct *mm, | |
1165 | __u64 start, __u64 len) | |
1166 | { | |
1167 | __u64 task_size = mm->task_size; | |
1168 | ||
1169 | if (start & ~PAGE_MASK) | |
1170 | return -EINVAL; | |
1171 | if (len & ~PAGE_MASK) | |
1172 | return -EINVAL; | |
1173 | if (!len) | |
1174 | return -EINVAL; | |
1175 | if (start < mmap_min_addr) | |
1176 | return -EINVAL; | |
1177 | if (start >= task_size) | |
1178 | return -EINVAL; | |
1179 | if (len > task_size - start) | |
1180 | return -EINVAL; | |
1181 | return 0; | |
1182 | } | |
1183 | ||
ba6907db MR |
1184 | static inline bool vma_can_userfault(struct vm_area_struct *vma) |
1185 | { | |
cac67329 MR |
1186 | return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) || |
1187 | vma_is_shmem(vma); | |
ba6907db MR |
1188 | } |
1189 | ||
86039bd3 AA |
1190 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, |
1191 | unsigned long arg) | |
1192 | { | |
1193 | struct mm_struct *mm = ctx->mm; | |
1194 | struct vm_area_struct *vma, *prev, *cur; | |
1195 | int ret; | |
1196 | struct uffdio_register uffdio_register; | |
1197 | struct uffdio_register __user *user_uffdio_register; | |
1198 | unsigned long vm_flags, new_flags; | |
1199 | bool found; | |
ce53e8e6 | 1200 | bool basic_ioctls; |
86039bd3 AA |
1201 | unsigned long start, end, vma_end; |
1202 | ||
1203 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1204 | ||
1205 | ret = -EFAULT; | |
1206 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1207 | sizeof(uffdio_register)-sizeof(__u64))) | |
1208 | goto out; | |
1209 | ||
1210 | ret = -EINVAL; | |
1211 | if (!uffdio_register.mode) | |
1212 | goto out; | |
1213 | if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING| | |
1214 | UFFDIO_REGISTER_MODE_WP)) | |
1215 | goto out; | |
1216 | vm_flags = 0; | |
1217 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1218 | vm_flags |= VM_UFFD_MISSING; | |
1219 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { | |
1220 | vm_flags |= VM_UFFD_WP; | |
1221 | /* | |
1222 | * FIXME: remove the below error constraint by | |
1223 | * implementing the wprotect tracking mode. | |
1224 | */ | |
1225 | ret = -EINVAL; | |
1226 | goto out; | |
1227 | } | |
1228 | ||
1229 | ret = validate_range(mm, uffdio_register.range.start, | |
1230 | uffdio_register.range.len); | |
1231 | if (ret) | |
1232 | goto out; | |
1233 | ||
1234 | start = uffdio_register.range.start; | |
1235 | end = start + uffdio_register.range.len; | |
1236 | ||
d2005e3f ON |
1237 | ret = -ENOMEM; |
1238 | if (!mmget_not_zero(mm)) | |
1239 | goto out; | |
1240 | ||
86039bd3 AA |
1241 | down_write(&mm->mmap_sem); |
1242 | vma = find_vma_prev(mm, start, &prev); | |
86039bd3 AA |
1243 | if (!vma) |
1244 | goto out_unlock; | |
1245 | ||
1246 | /* check that there's at least one vma in the range */ | |
1247 | ret = -EINVAL; | |
1248 | if (vma->vm_start >= end) | |
1249 | goto out_unlock; | |
1250 | ||
cab350af MK |
1251 | /* |
1252 | * If the first vma contains huge pages, make sure start address | |
1253 | * is aligned to huge page size. | |
1254 | */ | |
1255 | if (is_vm_hugetlb_page(vma)) { | |
1256 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1257 | ||
1258 | if (start & (vma_hpagesize - 1)) | |
1259 | goto out_unlock; | |
1260 | } | |
1261 | ||
86039bd3 AA |
1262 | /* |
1263 | * Search for not compatible vmas. | |
86039bd3 AA |
1264 | */ |
1265 | found = false; | |
ce53e8e6 | 1266 | basic_ioctls = false; |
86039bd3 AA |
1267 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { |
1268 | cond_resched(); | |
1269 | ||
1270 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1271 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1272 | ||
1273 | /* check not compatible vmas */ | |
1274 | ret = -EINVAL; | |
ba6907db | 1275 | if (!vma_can_userfault(cur)) |
86039bd3 | 1276 | goto out_unlock; |
cab350af MK |
1277 | /* |
1278 | * If this vma contains ending address, and huge pages | |
1279 | * check alignment. | |
1280 | */ | |
1281 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1282 | end > cur->vm_start) { | |
1283 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1284 | ||
1285 | ret = -EINVAL; | |
1286 | ||
1287 | if (end & (vma_hpagesize - 1)) | |
1288 | goto out_unlock; | |
1289 | } | |
86039bd3 AA |
1290 | |
1291 | /* | |
1292 | * Check that this vma isn't already owned by a | |
1293 | * different userfaultfd. We can't allow more than one | |
1294 | * userfaultfd to own a single vma simultaneously or we | |
1295 | * wouldn't know which one to deliver the userfaults to. | |
1296 | */ | |
1297 | ret = -EBUSY; | |
1298 | if (cur->vm_userfaultfd_ctx.ctx && | |
1299 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1300 | goto out_unlock; | |
1301 | ||
cab350af MK |
1302 | /* |
1303 | * Note vmas containing huge pages | |
1304 | */ | |
ce53e8e6 MR |
1305 | if (is_vm_hugetlb_page(cur)) |
1306 | basic_ioctls = true; | |
cab350af | 1307 | |
86039bd3 AA |
1308 | found = true; |
1309 | } | |
1310 | BUG_ON(!found); | |
1311 | ||
1312 | if (vma->vm_start < start) | |
1313 | prev = vma; | |
1314 | ||
1315 | ret = 0; | |
1316 | do { | |
1317 | cond_resched(); | |
1318 | ||
ba6907db | 1319 | BUG_ON(!vma_can_userfault(vma)); |
86039bd3 AA |
1320 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1321 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
1322 | ||
1323 | /* | |
1324 | * Nothing to do: this vma is already registered into this | |
1325 | * userfaultfd and with the right tracking mode too. | |
1326 | */ | |
1327 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1328 | (vma->vm_flags & vm_flags) == vm_flags) | |
1329 | goto skip; | |
1330 | ||
1331 | if (vma->vm_start > start) | |
1332 | start = vma->vm_start; | |
1333 | vma_end = min(end, vma->vm_end); | |
1334 | ||
1335 | new_flags = (vma->vm_flags & ~vm_flags) | vm_flags; | |
1336 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
1337 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1338 | vma_policy(vma), | |
1339 | ((struct vm_userfaultfd_ctx){ ctx })); | |
1340 | if (prev) { | |
1341 | vma = prev; | |
1342 | goto next; | |
1343 | } | |
1344 | if (vma->vm_start < start) { | |
1345 | ret = split_vma(mm, vma, start, 1); | |
1346 | if (ret) | |
1347 | break; | |
1348 | } | |
1349 | if (vma->vm_end > end) { | |
1350 | ret = split_vma(mm, vma, end, 0); | |
1351 | if (ret) | |
1352 | break; | |
1353 | } | |
1354 | next: | |
1355 | /* | |
1356 | * In the vma_merge() successful mprotect-like case 8: | |
1357 | * the next vma was merged into the current one and | |
1358 | * the current one has not been updated yet. | |
1359 | */ | |
1360 | vma->vm_flags = new_flags; | |
1361 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
1362 | ||
1363 | skip: | |
1364 | prev = vma; | |
1365 | start = vma->vm_end; | |
1366 | vma = vma->vm_next; | |
1367 | } while (vma && vma->vm_start < end); | |
1368 | out_unlock: | |
1369 | up_write(&mm->mmap_sem); | |
d2005e3f | 1370 | mmput(mm); |
86039bd3 AA |
1371 | if (!ret) { |
1372 | /* | |
1373 | * Now that we scanned all vmas we can already tell | |
1374 | * userland which ioctls methods are guaranteed to | |
1375 | * succeed on this range. | |
1376 | */ | |
ce53e8e6 | 1377 | if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : |
cab350af | 1378 | UFFD_API_RANGE_IOCTLS, |
86039bd3 AA |
1379 | &user_uffdio_register->ioctls)) |
1380 | ret = -EFAULT; | |
1381 | } | |
1382 | out: | |
1383 | return ret; | |
1384 | } | |
1385 | ||
1386 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1387 | unsigned long arg) | |
1388 | { | |
1389 | struct mm_struct *mm = ctx->mm; | |
1390 | struct vm_area_struct *vma, *prev, *cur; | |
1391 | int ret; | |
1392 | struct uffdio_range uffdio_unregister; | |
1393 | unsigned long new_flags; | |
1394 | bool found; | |
1395 | unsigned long start, end, vma_end; | |
1396 | const void __user *buf = (void __user *)arg; | |
1397 | ||
1398 | ret = -EFAULT; | |
1399 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1400 | goto out; | |
1401 | ||
1402 | ret = validate_range(mm, uffdio_unregister.start, | |
1403 | uffdio_unregister.len); | |
1404 | if (ret) | |
1405 | goto out; | |
1406 | ||
1407 | start = uffdio_unregister.start; | |
1408 | end = start + uffdio_unregister.len; | |
1409 | ||
d2005e3f ON |
1410 | ret = -ENOMEM; |
1411 | if (!mmget_not_zero(mm)) | |
1412 | goto out; | |
1413 | ||
86039bd3 AA |
1414 | down_write(&mm->mmap_sem); |
1415 | vma = find_vma_prev(mm, start, &prev); | |
86039bd3 AA |
1416 | if (!vma) |
1417 | goto out_unlock; | |
1418 | ||
1419 | /* check that there's at least one vma in the range */ | |
1420 | ret = -EINVAL; | |
1421 | if (vma->vm_start >= end) | |
1422 | goto out_unlock; | |
1423 | ||
cab350af MK |
1424 | /* |
1425 | * If the first vma contains huge pages, make sure start address | |
1426 | * is aligned to huge page size. | |
1427 | */ | |
1428 | if (is_vm_hugetlb_page(vma)) { | |
1429 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1430 | ||
1431 | if (start & (vma_hpagesize - 1)) | |
1432 | goto out_unlock; | |
1433 | } | |
1434 | ||
86039bd3 AA |
1435 | /* |
1436 | * Search for not compatible vmas. | |
86039bd3 AA |
1437 | */ |
1438 | found = false; | |
1439 | ret = -EINVAL; | |
1440 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
1441 | cond_resched(); | |
1442 | ||
1443 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1444 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1445 | ||
1446 | /* | |
1447 | * Check not compatible vmas, not strictly required | |
1448 | * here as not compatible vmas cannot have an | |
1449 | * userfaultfd_ctx registered on them, but this | |
1450 | * provides for more strict behavior to notice | |
1451 | * unregistration errors. | |
1452 | */ | |
ba6907db | 1453 | if (!vma_can_userfault(cur)) |
86039bd3 AA |
1454 | goto out_unlock; |
1455 | ||
1456 | found = true; | |
1457 | } | |
1458 | BUG_ON(!found); | |
1459 | ||
1460 | if (vma->vm_start < start) | |
1461 | prev = vma; | |
1462 | ||
1463 | ret = 0; | |
1464 | do { | |
1465 | cond_resched(); | |
1466 | ||
ba6907db | 1467 | BUG_ON(!vma_can_userfault(vma)); |
86039bd3 AA |
1468 | |
1469 | /* | |
1470 | * Nothing to do: this vma is already registered into this | |
1471 | * userfaultfd and with the right tracking mode too. | |
1472 | */ | |
1473 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1474 | goto skip; | |
1475 | ||
1476 | if (vma->vm_start > start) | |
1477 | start = vma->vm_start; | |
1478 | vma_end = min(end, vma->vm_end); | |
1479 | ||
09fa5296 AA |
1480 | if (userfaultfd_missing(vma)) { |
1481 | /* | |
1482 | * Wake any concurrent pending userfault while | |
1483 | * we unregister, so they will not hang | |
1484 | * permanently and it avoids userland to call | |
1485 | * UFFDIO_WAKE explicitly. | |
1486 | */ | |
1487 | struct userfaultfd_wake_range range; | |
1488 | range.start = start; | |
1489 | range.len = vma_end - start; | |
1490 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1491 | } | |
1492 | ||
86039bd3 AA |
1493 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); |
1494 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
1495 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1496 | vma_policy(vma), | |
1497 | NULL_VM_UFFD_CTX); | |
1498 | if (prev) { | |
1499 | vma = prev; | |
1500 | goto next; | |
1501 | } | |
1502 | if (vma->vm_start < start) { | |
1503 | ret = split_vma(mm, vma, start, 1); | |
1504 | if (ret) | |
1505 | break; | |
1506 | } | |
1507 | if (vma->vm_end > end) { | |
1508 | ret = split_vma(mm, vma, end, 0); | |
1509 | if (ret) | |
1510 | break; | |
1511 | } | |
1512 | next: | |
1513 | /* | |
1514 | * In the vma_merge() successful mprotect-like case 8: | |
1515 | * the next vma was merged into the current one and | |
1516 | * the current one has not been updated yet. | |
1517 | */ | |
1518 | vma->vm_flags = new_flags; | |
1519 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
1520 | ||
1521 | skip: | |
1522 | prev = vma; | |
1523 | start = vma->vm_end; | |
1524 | vma = vma->vm_next; | |
1525 | } while (vma && vma->vm_start < end); | |
1526 | out_unlock: | |
1527 | up_write(&mm->mmap_sem); | |
d2005e3f | 1528 | mmput(mm); |
86039bd3 AA |
1529 | out: |
1530 | return ret; | |
1531 | } | |
1532 | ||
1533 | /* | |
ba85c702 AA |
1534 | * userfaultfd_wake may be used in combination with the |
1535 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1536 | */ |
1537 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1538 | unsigned long arg) | |
1539 | { | |
1540 | int ret; | |
1541 | struct uffdio_range uffdio_wake; | |
1542 | struct userfaultfd_wake_range range; | |
1543 | const void __user *buf = (void __user *)arg; | |
1544 | ||
1545 | ret = -EFAULT; | |
1546 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1547 | goto out; | |
1548 | ||
1549 | ret = validate_range(ctx->mm, uffdio_wake.start, uffdio_wake.len); | |
1550 | if (ret) | |
1551 | goto out; | |
1552 | ||
1553 | range.start = uffdio_wake.start; | |
1554 | range.len = uffdio_wake.len; | |
1555 | ||
1556 | /* | |
1557 | * len == 0 means wake all and we don't want to wake all here, | |
1558 | * so check it again to be sure. | |
1559 | */ | |
1560 | VM_BUG_ON(!range.len); | |
1561 | ||
1562 | wake_userfault(ctx, &range); | |
1563 | ret = 0; | |
1564 | ||
1565 | out: | |
1566 | return ret; | |
1567 | } | |
1568 | ||
ad465cae AA |
1569 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1570 | unsigned long arg) | |
1571 | { | |
1572 | __s64 ret; | |
1573 | struct uffdio_copy uffdio_copy; | |
1574 | struct uffdio_copy __user *user_uffdio_copy; | |
1575 | struct userfaultfd_wake_range range; | |
1576 | ||
1577 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1578 | ||
1579 | ret = -EFAULT; | |
1580 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1581 | /* don't copy "copy" last field */ | |
1582 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1583 | goto out; | |
1584 | ||
1585 | ret = validate_range(ctx->mm, uffdio_copy.dst, uffdio_copy.len); | |
1586 | if (ret) | |
1587 | goto out; | |
1588 | /* | |
1589 | * double check for wraparound just in case. copy_from_user() | |
1590 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1591 | * in the userland range. | |
1592 | */ | |
1593 | ret = -EINVAL; | |
1594 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1595 | goto out; | |
1596 | if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE) | |
1597 | goto out; | |
d2005e3f ON |
1598 | if (mmget_not_zero(ctx->mm)) { |
1599 | ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src, | |
1600 | uffdio_copy.len); | |
1601 | mmput(ctx->mm); | |
96333187 | 1602 | } else { |
e86b298b | 1603 | return -ESRCH; |
d2005e3f | 1604 | } |
ad465cae AA |
1605 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1606 | return -EFAULT; | |
1607 | if (ret < 0) | |
1608 | goto out; | |
1609 | BUG_ON(!ret); | |
1610 | /* len == 0 would wake all */ | |
1611 | range.len = ret; | |
1612 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1613 | range.start = uffdio_copy.dst; | |
1614 | wake_userfault(ctx, &range); | |
1615 | } | |
1616 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1617 | out: | |
1618 | return ret; | |
1619 | } | |
1620 | ||
1621 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1622 | unsigned long arg) | |
1623 | { | |
1624 | __s64 ret; | |
1625 | struct uffdio_zeropage uffdio_zeropage; | |
1626 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1627 | struct userfaultfd_wake_range range; | |
1628 | ||
1629 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1630 | ||
1631 | ret = -EFAULT; | |
1632 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1633 | /* don't copy "zeropage" last field */ | |
1634 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1635 | goto out; | |
1636 | ||
1637 | ret = validate_range(ctx->mm, uffdio_zeropage.range.start, | |
1638 | uffdio_zeropage.range.len); | |
1639 | if (ret) | |
1640 | goto out; | |
1641 | ret = -EINVAL; | |
1642 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1643 | goto out; | |
1644 | ||
d2005e3f ON |
1645 | if (mmget_not_zero(ctx->mm)) { |
1646 | ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start, | |
1647 | uffdio_zeropage.range.len); | |
1648 | mmput(ctx->mm); | |
9d95aa4b | 1649 | } else { |
e86b298b | 1650 | return -ESRCH; |
d2005e3f | 1651 | } |
ad465cae AA |
1652 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1653 | return -EFAULT; | |
1654 | if (ret < 0) | |
1655 | goto out; | |
1656 | /* len == 0 would wake all */ | |
1657 | BUG_ON(!ret); | |
1658 | range.len = ret; | |
1659 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1660 | range.start = uffdio_zeropage.range.start; | |
1661 | wake_userfault(ctx, &range); | |
1662 | } | |
1663 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1664 | out: | |
1665 | return ret; | |
1666 | } | |
1667 | ||
9cd75c3c PE |
1668 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1669 | { | |
1670 | /* | |
1671 | * For the current set of features the bits just coincide | |
1672 | */ | |
1673 | return (unsigned int)user_features; | |
1674 | } | |
1675 | ||
86039bd3 AA |
1676 | /* |
1677 | * userland asks for a certain API version and we return which bits | |
1678 | * and ioctl commands are implemented in this kernel for such API | |
1679 | * version or -EINVAL if unknown. | |
1680 | */ | |
1681 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1682 | unsigned long arg) | |
1683 | { | |
1684 | struct uffdio_api uffdio_api; | |
1685 | void __user *buf = (void __user *)arg; | |
1686 | int ret; | |
65603144 | 1687 | __u64 features; |
86039bd3 AA |
1688 | |
1689 | ret = -EINVAL; | |
1690 | if (ctx->state != UFFD_STATE_WAIT_API) | |
1691 | goto out; | |
1692 | ret = -EFAULT; | |
a9b85f94 | 1693 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1694 | goto out; |
65603144 AA |
1695 | features = uffdio_api.features; |
1696 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) { | |
86039bd3 AA |
1697 | memset(&uffdio_api, 0, sizeof(uffdio_api)); |
1698 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1699 | goto out; | |
1700 | ret = -EINVAL; | |
1701 | goto out; | |
1702 | } | |
65603144 AA |
1703 | /* report all available features and ioctls to userland */ |
1704 | uffdio_api.features = UFFD_API_FEATURES; | |
86039bd3 AA |
1705 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1706 | ret = -EFAULT; | |
1707 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1708 | goto out; | |
1709 | ctx->state = UFFD_STATE_RUNNING; | |
65603144 AA |
1710 | /* only enable the requested features for this uffd context */ |
1711 | ctx->features = uffd_ctx_features(features); | |
86039bd3 AA |
1712 | ret = 0; |
1713 | out: | |
1714 | return ret; | |
1715 | } | |
1716 | ||
1717 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
1718 | unsigned long arg) | |
1719 | { | |
1720 | int ret = -EINVAL; | |
1721 | struct userfaultfd_ctx *ctx = file->private_data; | |
1722 | ||
e6485a47 AA |
1723 | if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API) |
1724 | return -EINVAL; | |
1725 | ||
86039bd3 AA |
1726 | switch(cmd) { |
1727 | case UFFDIO_API: | |
1728 | ret = userfaultfd_api(ctx, arg); | |
1729 | break; | |
1730 | case UFFDIO_REGISTER: | |
1731 | ret = userfaultfd_register(ctx, arg); | |
1732 | break; | |
1733 | case UFFDIO_UNREGISTER: | |
1734 | ret = userfaultfd_unregister(ctx, arg); | |
1735 | break; | |
1736 | case UFFDIO_WAKE: | |
1737 | ret = userfaultfd_wake(ctx, arg); | |
1738 | break; | |
ad465cae AA |
1739 | case UFFDIO_COPY: |
1740 | ret = userfaultfd_copy(ctx, arg); | |
1741 | break; | |
1742 | case UFFDIO_ZEROPAGE: | |
1743 | ret = userfaultfd_zeropage(ctx, arg); | |
1744 | break; | |
86039bd3 AA |
1745 | } |
1746 | return ret; | |
1747 | } | |
1748 | ||
1749 | #ifdef CONFIG_PROC_FS | |
1750 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
1751 | { | |
1752 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 1753 | wait_queue_entry_t *wq; |
86039bd3 AA |
1754 | struct userfaultfd_wait_queue *uwq; |
1755 | unsigned long pending = 0, total = 0; | |
1756 | ||
15b726ef | 1757 | spin_lock(&ctx->fault_pending_wqh.lock); |
2055da97 | 1758 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
1759 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
1760 | pending++; | |
1761 | total++; | |
1762 | } | |
2055da97 | 1763 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 | 1764 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
86039bd3 AA |
1765 | total++; |
1766 | } | |
15b726ef | 1767 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1768 | |
1769 | /* | |
1770 | * If more protocols will be added, there will be all shown | |
1771 | * separated by a space. Like this: | |
1772 | * protocols: aa:... bb:... | |
1773 | */ | |
1774 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 1775 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
1776 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
1777 | } | |
1778 | #endif | |
1779 | ||
1780 | static const struct file_operations userfaultfd_fops = { | |
1781 | #ifdef CONFIG_PROC_FS | |
1782 | .show_fdinfo = userfaultfd_show_fdinfo, | |
1783 | #endif | |
1784 | .release = userfaultfd_release, | |
1785 | .poll = userfaultfd_poll, | |
1786 | .read = userfaultfd_read, | |
1787 | .unlocked_ioctl = userfaultfd_ioctl, | |
1788 | .compat_ioctl = userfaultfd_ioctl, | |
1789 | .llseek = noop_llseek, | |
1790 | }; | |
1791 | ||
3004ec9c AA |
1792 | static void init_once_userfaultfd_ctx(void *mem) |
1793 | { | |
1794 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
1795 | ||
1796 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
1797 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 1798 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 1799 | init_waitqueue_head(&ctx->fd_wqh); |
2c5b7e1b | 1800 | seqcount_init(&ctx->refile_seq); |
3004ec9c AA |
1801 | } |
1802 | ||
86039bd3 | 1803 | /** |
9332ef9d | 1804 | * userfaultfd_file_create - Creates a userfaultfd file pointer. |
86039bd3 AA |
1805 | * @flags: Flags for the userfaultfd file. |
1806 | * | |
9332ef9d | 1807 | * This function creates a userfaultfd file pointer, w/out installing |
86039bd3 AA |
1808 | * it into the fd table. This is useful when the userfaultfd file is |
1809 | * used during the initialization of data structures that require | |
1810 | * extra setup after the userfaultfd creation. So the userfaultfd | |
1811 | * creation is split into the file pointer creation phase, and the | |
1812 | * file descriptor installation phase. In this way races with | |
1813 | * userspace closing the newly installed file descriptor can be | |
9332ef9d | 1814 | * avoided. Returns a userfaultfd file pointer, or a proper error |
86039bd3 AA |
1815 | * pointer. |
1816 | */ | |
1817 | static struct file *userfaultfd_file_create(int flags) | |
1818 | { | |
1819 | struct file *file; | |
1820 | struct userfaultfd_ctx *ctx; | |
1821 | ||
1822 | BUG_ON(!current->mm); | |
1823 | ||
1824 | /* Check the UFFD_* constants for consistency. */ | |
1825 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); | |
1826 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
1827 | ||
1828 | file = ERR_PTR(-EINVAL); | |
1829 | if (flags & ~UFFD_SHARED_FCNTL_FLAGS) | |
1830 | goto out; | |
1831 | ||
1832 | file = ERR_PTR(-ENOMEM); | |
3004ec9c | 1833 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 AA |
1834 | if (!ctx) |
1835 | goto out; | |
1836 | ||
1837 | atomic_set(&ctx->refcount, 1); | |
86039bd3 | 1838 | ctx->flags = flags; |
9cd75c3c | 1839 | ctx->features = 0; |
86039bd3 AA |
1840 | ctx->state = UFFD_STATE_WAIT_API; |
1841 | ctx->released = false; | |
1842 | ctx->mm = current->mm; | |
1843 | /* prevent the mm struct to be freed */ | |
f1f10076 | 1844 | mmgrab(ctx->mm); |
86039bd3 AA |
1845 | |
1846 | file = anon_inode_getfile("[userfaultfd]", &userfaultfd_fops, ctx, | |
1847 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS)); | |
c03e946f | 1848 | if (IS_ERR(file)) { |
d2005e3f | 1849 | mmdrop(ctx->mm); |
3004ec9c | 1850 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 1851 | } |
86039bd3 AA |
1852 | out: |
1853 | return file; | |
1854 | } | |
1855 | ||
1856 | SYSCALL_DEFINE1(userfaultfd, int, flags) | |
1857 | { | |
1858 | int fd, error; | |
1859 | struct file *file; | |
1860 | ||
1861 | error = get_unused_fd_flags(flags & UFFD_SHARED_FCNTL_FLAGS); | |
1862 | if (error < 0) | |
1863 | return error; | |
1864 | fd = error; | |
1865 | ||
1866 | file = userfaultfd_file_create(flags); | |
1867 | if (IS_ERR(file)) { | |
1868 | error = PTR_ERR(file); | |
1869 | goto err_put_unused_fd; | |
1870 | } | |
1871 | fd_install(fd, file); | |
1872 | ||
1873 | return fd; | |
1874 | ||
1875 | err_put_unused_fd: | |
1876 | put_unused_fd(fd); | |
1877 | ||
1878 | return error; | |
1879 | } | |
3004ec9c AA |
1880 | |
1881 | static int __init userfaultfd_init(void) | |
1882 | { | |
1883 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", | |
1884 | sizeof(struct userfaultfd_ctx), | |
1885 | 0, | |
1886 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
1887 | init_once_userfaultfd_ctx); | |
1888 | return 0; | |
1889 | } | |
1890 | __initcall(userfaultfd_init); |