Commit | Line | Data |
---|---|---|
20c8ccb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
86039bd3 AA |
2 | /* |
3 | * fs/userfaultfd.c | |
4 | * | |
5 | * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org> | |
6 | * Copyright (C) 2008-2009 Red Hat, Inc. | |
7 | * Copyright (C) 2015 Red Hat, Inc. | |
8 | * | |
86039bd3 AA |
9 | * Some part derived from fs/eventfd.c (anon inode setup) and |
10 | * mm/ksm.c (mm hashing). | |
11 | */ | |
12 | ||
9cd75c3c | 13 | #include <linux/list.h> |
86039bd3 | 14 | #include <linux/hashtable.h> |
174cd4b1 | 15 | #include <linux/sched/signal.h> |
6e84f315 | 16 | #include <linux/sched/mm.h> |
86039bd3 AA |
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> | |
cab350af | 29 | #include <linux/hugetlb.h> |
86039bd3 | 30 | |
cefdca0a PX |
31 | int sysctl_unprivileged_userfaultfd __read_mostly = 1; |
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. | |
cbcfa130 EB |
43 | * |
44 | * Locking order: | |
45 | * fd_wqh.lock | |
46 | * fault_pending_wqh.lock | |
47 | * fault_wqh.lock | |
48 | * event_wqh.lock | |
49 | * | |
50 | * To avoid deadlocks, IRQs must be disabled when taking any of the above locks, | |
51 | * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's | |
52 | * also taken in IRQ context. | |
3004ec9c | 53 | */ |
86039bd3 | 54 | struct userfaultfd_ctx { |
15b726ef AA |
55 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
56 | wait_queue_head_t fault_pending_wqh; | |
57 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
58 | wait_queue_head_t fault_wqh; |
59 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
60 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
61 | /* waitqueue head for events */ |
62 | wait_queue_head_t event_wqh; | |
2c5b7e1b | 63 | /* a refile sequence protected by fault_pending_wqh lock */ |
2ca97ac8 | 64 | seqcount_spinlock_t refile_seq; |
3004ec9c | 65 | /* pseudo fd refcounting */ |
ca880420 | 66 | refcount_t refcount; |
86039bd3 AA |
67 | /* userfaultfd syscall flags */ |
68 | unsigned int flags; | |
9cd75c3c PE |
69 | /* features requested from the userspace */ |
70 | unsigned int features; | |
86039bd3 AA |
71 | /* state machine */ |
72 | enum userfaultfd_state state; | |
73 | /* released */ | |
74 | bool released; | |
df2cc96e MR |
75 | /* memory mappings are changing because of non-cooperative event */ |
76 | bool mmap_changing; | |
86039bd3 AA |
77 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ |
78 | struct mm_struct *mm; | |
79 | }; | |
80 | ||
893e26e6 PE |
81 | struct userfaultfd_fork_ctx { |
82 | struct userfaultfd_ctx *orig; | |
83 | struct userfaultfd_ctx *new; | |
84 | struct list_head list; | |
85 | }; | |
86 | ||
897ab3e0 MR |
87 | struct userfaultfd_unmap_ctx { |
88 | struct userfaultfd_ctx *ctx; | |
89 | unsigned long start; | |
90 | unsigned long end; | |
91 | struct list_head list; | |
92 | }; | |
93 | ||
86039bd3 | 94 | struct userfaultfd_wait_queue { |
a9b85f94 | 95 | struct uffd_msg msg; |
ac6424b9 | 96 | wait_queue_entry_t wq; |
86039bd3 | 97 | struct userfaultfd_ctx *ctx; |
15a77c6f | 98 | bool waken; |
86039bd3 AA |
99 | }; |
100 | ||
101 | struct userfaultfd_wake_range { | |
102 | unsigned long start; | |
103 | unsigned long len; | |
104 | }; | |
105 | ||
ac6424b9 | 106 | static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode, |
86039bd3 AA |
107 | int wake_flags, void *key) |
108 | { | |
109 | struct userfaultfd_wake_range *range = key; | |
110 | int ret; | |
111 | struct userfaultfd_wait_queue *uwq; | |
112 | unsigned long start, len; | |
113 | ||
114 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
115 | ret = 0; | |
86039bd3 AA |
116 | /* len == 0 means wake all */ |
117 | start = range->start; | |
118 | len = range->len; | |
a9b85f94 AA |
119 | if (len && (start > uwq->msg.arg.pagefault.address || |
120 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 121 | goto out; |
15a77c6f AA |
122 | WRITE_ONCE(uwq->waken, true); |
123 | /* | |
a9668cd6 PZ |
124 | * The Program-Order guarantees provided by the scheduler |
125 | * ensure uwq->waken is visible before the task is woken. | |
15a77c6f | 126 | */ |
86039bd3 | 127 | ret = wake_up_state(wq->private, mode); |
a9668cd6 | 128 | if (ret) { |
86039bd3 AA |
129 | /* |
130 | * Wake only once, autoremove behavior. | |
131 | * | |
a9668cd6 PZ |
132 | * After the effect of list_del_init is visible to the other |
133 | * CPUs, the waitqueue may disappear from under us, see the | |
134 | * !list_empty_careful() in handle_userfault(). | |
135 | * | |
136 | * try_to_wake_up() has an implicit smp_mb(), and the | |
137 | * wq->private is read before calling the extern function | |
138 | * "wake_up_state" (which in turns calls try_to_wake_up). | |
86039bd3 | 139 | */ |
2055da97 | 140 | list_del_init(&wq->entry); |
a9668cd6 | 141 | } |
86039bd3 AA |
142 | out: |
143 | return ret; | |
144 | } | |
145 | ||
146 | /** | |
147 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
148 | * context. | |
149 | * @ctx: [in] Pointer to the userfaultfd context. | |
86039bd3 AA |
150 | */ |
151 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
152 | { | |
ca880420 | 153 | refcount_inc(&ctx->refcount); |
86039bd3 AA |
154 | } |
155 | ||
156 | /** | |
157 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
158 | * context. | |
159 | * @ctx: [in] Pointer to userfaultfd context. | |
160 | * | |
161 | * The userfaultfd context reference must have been previously acquired either | |
162 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
163 | */ | |
164 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
165 | { | |
ca880420 | 166 | if (refcount_dec_and_test(&ctx->refcount)) { |
86039bd3 AA |
167 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); |
168 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
169 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
170 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
171 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
172 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
173 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
174 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 175 | mmdrop(ctx->mm); |
3004ec9c | 176 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
177 | } |
178 | } | |
179 | ||
a9b85f94 | 180 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 181 | { |
a9b85f94 AA |
182 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
183 | /* | |
184 | * Must use memset to zero out the paddings or kernel data is | |
185 | * leaked to userland. | |
186 | */ | |
187 | memset(msg, 0, sizeof(struct uffd_msg)); | |
188 | } | |
189 | ||
190 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
191 | unsigned int flags, | |
9d4ac934 AP |
192 | unsigned long reason, |
193 | unsigned int features) | |
a9b85f94 AA |
194 | { |
195 | struct uffd_msg msg; | |
196 | msg_init(&msg); | |
197 | msg.event = UFFD_EVENT_PAGEFAULT; | |
198 | msg.arg.pagefault.address = address; | |
86039bd3 AA |
199 | if (flags & FAULT_FLAG_WRITE) |
200 | /* | |
a4605a61 | 201 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
a9b85f94 AA |
202 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE |
203 | * was not set in a UFFD_EVENT_PAGEFAULT, it means it | |
204 | * was a read fault, otherwise if set it means it's | |
205 | * a write fault. | |
86039bd3 | 206 | */ |
a9b85f94 | 207 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 AA |
208 | if (reason & VM_UFFD_WP) |
209 | /* | |
a9b85f94 AA |
210 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
211 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was | |
212 | * not set in a UFFD_EVENT_PAGEFAULT, it means it was | |
213 | * a missing fault, otherwise if set it means it's a | |
214 | * write protect fault. | |
86039bd3 | 215 | */ |
a9b85f94 | 216 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
9d4ac934 | 217 | if (features & UFFD_FEATURE_THREAD_ID) |
a36985d3 | 218 | msg.arg.pagefault.feat.ptid = task_pid_vnr(current); |
a9b85f94 | 219 | return msg; |
86039bd3 AA |
220 | } |
221 | ||
369cd212 MK |
222 | #ifdef CONFIG_HUGETLB_PAGE |
223 | /* | |
224 | * Same functionality as userfaultfd_must_wait below with modifications for | |
225 | * hugepmd ranges. | |
226 | */ | |
227 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 228 | struct vm_area_struct *vma, |
369cd212 MK |
229 | unsigned long address, |
230 | unsigned long flags, | |
231 | unsigned long reason) | |
232 | { | |
233 | struct mm_struct *mm = ctx->mm; | |
1e2c0436 | 234 | pte_t *ptep, pte; |
369cd212 MK |
235 | bool ret = true; |
236 | ||
42fc5414 | 237 | mmap_assert_locked(mm); |
369cd212 | 238 | |
1e2c0436 JF |
239 | ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma)); |
240 | ||
241 | if (!ptep) | |
369cd212 MK |
242 | goto out; |
243 | ||
244 | ret = false; | |
1e2c0436 | 245 | pte = huge_ptep_get(ptep); |
369cd212 MK |
246 | |
247 | /* | |
248 | * Lockless access: we're in a wait_event so it's ok if it | |
249 | * changes under us. | |
250 | */ | |
1e2c0436 | 251 | if (huge_pte_none(pte)) |
369cd212 | 252 | ret = true; |
1e2c0436 | 253 | if (!huge_pte_write(pte) && (reason & VM_UFFD_WP)) |
369cd212 MK |
254 | ret = true; |
255 | out: | |
256 | return ret; | |
257 | } | |
258 | #else | |
259 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 260 | struct vm_area_struct *vma, |
369cd212 MK |
261 | unsigned long address, |
262 | unsigned long flags, | |
263 | unsigned long reason) | |
264 | { | |
265 | return false; /* should never get here */ | |
266 | } | |
267 | #endif /* CONFIG_HUGETLB_PAGE */ | |
268 | ||
8d2afd96 AA |
269 | /* |
270 | * Verify the pagetables are still not ok after having reigstered into | |
271 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
272 | * userfault that has already been resolved, if userfaultfd_read and | |
273 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
274 | * threads. | |
275 | */ | |
276 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
277 | unsigned long address, | |
278 | unsigned long flags, | |
279 | unsigned long reason) | |
280 | { | |
281 | struct mm_struct *mm = ctx->mm; | |
282 | pgd_t *pgd; | |
c2febafc | 283 | p4d_t *p4d; |
8d2afd96 AA |
284 | pud_t *pud; |
285 | pmd_t *pmd, _pmd; | |
286 | pte_t *pte; | |
287 | bool ret = true; | |
288 | ||
42fc5414 | 289 | mmap_assert_locked(mm); |
8d2afd96 AA |
290 | |
291 | pgd = pgd_offset(mm, address); | |
292 | if (!pgd_present(*pgd)) | |
293 | goto out; | |
c2febafc KS |
294 | p4d = p4d_offset(pgd, address); |
295 | if (!p4d_present(*p4d)) | |
296 | goto out; | |
297 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
298 | if (!pud_present(*pud)) |
299 | goto out; | |
300 | pmd = pmd_offset(pud, address); | |
301 | /* | |
302 | * READ_ONCE must function as a barrier with narrower scope | |
303 | * and it must be equivalent to: | |
304 | * _pmd = *pmd; barrier(); | |
305 | * | |
306 | * This is to deal with the instability (as in | |
307 | * pmd_trans_unstable) of the pmd. | |
308 | */ | |
309 | _pmd = READ_ONCE(*pmd); | |
a365ac09 | 310 | if (pmd_none(_pmd)) |
8d2afd96 AA |
311 | goto out; |
312 | ||
313 | ret = false; | |
a365ac09 HY |
314 | if (!pmd_present(_pmd)) |
315 | goto out; | |
316 | ||
63b2d417 AA |
317 | if (pmd_trans_huge(_pmd)) { |
318 | if (!pmd_write(_pmd) && (reason & VM_UFFD_WP)) | |
319 | ret = true; | |
8d2afd96 | 320 | goto out; |
63b2d417 | 321 | } |
8d2afd96 AA |
322 | |
323 | /* | |
324 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
325 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
326 | */ | |
327 | pte = pte_offset_map(pmd, address); | |
328 | /* | |
329 | * Lockless access: we're in a wait_event so it's ok if it | |
330 | * changes under us. | |
331 | */ | |
332 | if (pte_none(*pte)) | |
333 | ret = true; | |
63b2d417 AA |
334 | if (!pte_write(*pte) && (reason & VM_UFFD_WP)) |
335 | ret = true; | |
8d2afd96 AA |
336 | pte_unmap(pte); |
337 | ||
338 | out: | |
339 | return ret; | |
340 | } | |
341 | ||
3e69ad08 PX |
342 | static inline long userfaultfd_get_blocking_state(unsigned int flags) |
343 | { | |
344 | if (flags & FAULT_FLAG_INTERRUPTIBLE) | |
345 | return TASK_INTERRUPTIBLE; | |
346 | ||
347 | if (flags & FAULT_FLAG_KILLABLE) | |
348 | return TASK_KILLABLE; | |
349 | ||
350 | return TASK_UNINTERRUPTIBLE; | |
351 | } | |
352 | ||
86039bd3 AA |
353 | /* |
354 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
355 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
356 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
357 | * recommendation in __lock_page_or_retry is not an understatement. | |
358 | * | |
c1e8d7c6 | 359 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released |
86039bd3 AA |
360 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is |
361 | * not set. | |
362 | * | |
363 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
364 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
c1e8d7c6 | 365 | * fatal_signal_pending()s, and the mmap_lock must be released before |
86039bd3 AA |
366 | * returning it. |
367 | */ | |
2b740303 | 368 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 369 | { |
82b0f8c3 | 370 | struct mm_struct *mm = vmf->vma->vm_mm; |
86039bd3 AA |
371 | struct userfaultfd_ctx *ctx; |
372 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 373 | vm_fault_t ret = VM_FAULT_SIGBUS; |
3e69ad08 | 374 | bool must_wait; |
15a77c6f | 375 | long blocking_state; |
86039bd3 | 376 | |
64c2b203 AA |
377 | /* |
378 | * We don't do userfault handling for the final child pid update. | |
379 | * | |
380 | * We also don't do userfault handling during | |
381 | * coredumping. hugetlbfs has the special | |
382 | * follow_hugetlb_page() to skip missing pages in the | |
383 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
384 | * the no_page_table() helper in follow_page_mask(), but the | |
385 | * shmem_vm_ops->fault method is invoked even during | |
c1e8d7c6 | 386 | * coredumping without mmap_lock and it ends up here. |
64c2b203 AA |
387 | */ |
388 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
389 | goto out; | |
390 | ||
391 | /* | |
c1e8d7c6 ML |
392 | * Coredumping runs without mmap_lock so we can only check that |
393 | * the mmap_lock is held, if PF_DUMPCORE was not set. | |
64c2b203 | 394 | */ |
42fc5414 | 395 | mmap_assert_locked(mm); |
64c2b203 | 396 | |
82b0f8c3 | 397 | ctx = vmf->vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 398 | if (!ctx) |
ba85c702 | 399 | goto out; |
86039bd3 AA |
400 | |
401 | BUG_ON(ctx->mm != mm); | |
402 | ||
403 | VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP)); | |
404 | VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP)); | |
405 | ||
2d6d6f5a PS |
406 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
407 | goto out; | |
408 | ||
86039bd3 AA |
409 | /* |
410 | * If it's already released don't get it. This avoids to loop | |
411 | * in __get_user_pages if userfaultfd_release waits on the | |
c1e8d7c6 | 412 | * caller of handle_userfault to release the mmap_lock. |
86039bd3 | 413 | */ |
6aa7de05 | 414 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
415 | /* |
416 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
417 | * cooperative manager can close the uffd after the | |
418 | * last UFFDIO_COPY, without risking to trigger an | |
419 | * involuntary SIGBUS if the process was starting the | |
420 | * userfaultfd while the userfaultfd was still armed | |
421 | * (but after the last UFFDIO_COPY). If the uffd | |
422 | * wasn't already closed when the userfault reached | |
423 | * this point, that would normally be solved by | |
424 | * userfaultfd_must_wait returning 'false'. | |
425 | * | |
426 | * If we were to return VM_FAULT_SIGBUS here, the non | |
427 | * cooperative manager would be instead forced to | |
428 | * always call UFFDIO_UNREGISTER before it can safely | |
429 | * close the uffd. | |
430 | */ | |
431 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 432 | goto out; |
656710a6 | 433 | } |
86039bd3 AA |
434 | |
435 | /* | |
436 | * Check that we can return VM_FAULT_RETRY. | |
437 | * | |
438 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
439 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
440 | * -EBUSY failures, if the userfaultfd is to be extended for | |
441 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
442 | * without first stopping userland access to the memory. For | |
443 | * VM_UFFD_MISSING userfaults this is enough for now. | |
444 | */ | |
82b0f8c3 | 445 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
446 | /* |
447 | * Validate the invariant that nowait must allow retry | |
448 | * to be sure not to return SIGBUS erroneously on | |
449 | * nowait invocations. | |
450 | */ | |
82b0f8c3 | 451 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
452 | #ifdef CONFIG_DEBUG_VM |
453 | if (printk_ratelimit()) { | |
454 | printk(KERN_WARNING | |
82b0f8c3 JK |
455 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
456 | vmf->flags); | |
86039bd3 AA |
457 | dump_stack(); |
458 | } | |
459 | #endif | |
ba85c702 | 460 | goto out; |
86039bd3 AA |
461 | } |
462 | ||
463 | /* | |
464 | * Handle nowait, not much to do other than tell it to retry | |
465 | * and wait. | |
466 | */ | |
ba85c702 | 467 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 468 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 469 | goto out; |
86039bd3 | 470 | |
c1e8d7c6 | 471 | /* take the reference before dropping the mmap_lock */ |
86039bd3 AA |
472 | userfaultfd_ctx_get(ctx); |
473 | ||
86039bd3 AA |
474 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
475 | uwq.wq.private = current; | |
9d4ac934 AP |
476 | uwq.msg = userfault_msg(vmf->address, vmf->flags, reason, |
477 | ctx->features); | |
86039bd3 | 478 | uwq.ctx = ctx; |
15a77c6f | 479 | uwq.waken = false; |
86039bd3 | 480 | |
3e69ad08 | 481 | blocking_state = userfaultfd_get_blocking_state(vmf->flags); |
dfa37dc3 | 482 | |
cbcfa130 | 483 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
484 | /* |
485 | * After the __add_wait_queue the uwq is visible to userland | |
486 | * through poll/read(). | |
487 | */ | |
15b726ef AA |
488 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
489 | /* | |
490 | * The smp_mb() after __set_current_state prevents the reads | |
491 | * following the spin_unlock to happen before the list_add in | |
492 | * __add_wait_queue. | |
493 | */ | |
15a77c6f | 494 | set_current_state(blocking_state); |
cbcfa130 | 495 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 496 | |
369cd212 MK |
497 | if (!is_vm_hugetlb_page(vmf->vma)) |
498 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, | |
499 | reason); | |
500 | else | |
7868a208 PA |
501 | must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma, |
502 | vmf->address, | |
369cd212 | 503 | vmf->flags, reason); |
d8ed45c5 | 504 | mmap_read_unlock(mm); |
8d2afd96 | 505 | |
f9bf3522 | 506 | if (likely(must_wait && !READ_ONCE(ctx->released))) { |
a9a08845 | 507 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 508 | schedule(); |
ba85c702 | 509 | } |
86039bd3 | 510 | |
ba85c702 | 511 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
512 | |
513 | /* | |
514 | * Here we race with the list_del; list_add in | |
515 | * userfaultfd_ctx_read(), however because we don't ever run | |
516 | * list_del_init() to refile across the two lists, the prev | |
517 | * and next pointers will never point to self. list_add also | |
518 | * would never let any of the two pointers to point to | |
519 | * self. So list_empty_careful won't risk to see both pointers | |
520 | * pointing to self at any time during the list refile. The | |
521 | * only case where list_del_init() is called is the full | |
522 | * removal in the wake function and there we don't re-list_add | |
523 | * and it's fine not to block on the spinlock. The uwq on this | |
524 | * kernel stack can be released after the list_del_init. | |
525 | */ | |
2055da97 | 526 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 527 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
528 | /* |
529 | * No need of list_del_init(), the uwq on the stack | |
530 | * will be freed shortly anyway. | |
531 | */ | |
2055da97 | 532 | list_del(&uwq.wq.entry); |
cbcfa130 | 533 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 534 | } |
86039bd3 AA |
535 | |
536 | /* | |
537 | * ctx may go away after this if the userfault pseudo fd is | |
538 | * already released. | |
539 | */ | |
540 | userfaultfd_ctx_put(ctx); | |
541 | ||
ba85c702 AA |
542 | out: |
543 | return ret; | |
86039bd3 AA |
544 | } |
545 | ||
8c9e7bb7 AA |
546 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
547 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 548 | { |
0cbb4b4f AA |
549 | struct userfaultfd_ctx *release_new_ctx; |
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); | |
0cbb4b4f | 556 | release_new_ctx = NULL; |
9cd75c3c | 557 | |
cbcfa130 | 558 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
559 | /* |
560 | * After the __add_wait_queue the uwq is visible to userland | |
561 | * through poll/read(). | |
562 | */ | |
563 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
564 | for (;;) { | |
565 | set_current_state(TASK_KILLABLE); | |
566 | if (ewq->msg.event == 0) | |
567 | break; | |
6aa7de05 | 568 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 569 | fatal_signal_pending(current)) { |
384632e6 AA |
570 | /* |
571 | * &ewq->wq may be queued in fork_event, but | |
572 | * __remove_wait_queue ignores the head | |
573 | * parameter. It would be a problem if it | |
574 | * didn't. | |
575 | */ | |
9cd75c3c | 576 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
577 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
578 | struct userfaultfd_ctx *new; | |
579 | ||
580 | new = (struct userfaultfd_ctx *) | |
581 | (unsigned long) | |
582 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 583 | release_new_ctx = new; |
7eb76d45 | 584 | } |
9cd75c3c PE |
585 | break; |
586 | } | |
587 | ||
cbcfa130 | 588 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 589 | |
a9a08845 | 590 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
591 | schedule(); |
592 | ||
cbcfa130 | 593 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
594 | } |
595 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 596 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 597 | |
0cbb4b4f AA |
598 | if (release_new_ctx) { |
599 | struct vm_area_struct *vma; | |
600 | struct mm_struct *mm = release_new_ctx->mm; | |
601 | ||
602 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
d8ed45c5 | 603 | mmap_write_lock(mm); |
0cbb4b4f | 604 | for (vma = mm->mmap; vma; vma = vma->vm_next) |
31e810aa | 605 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 606 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
31e810aa MR |
607 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); |
608 | } | |
d8ed45c5 | 609 | mmap_write_unlock(mm); |
0cbb4b4f AA |
610 | |
611 | userfaultfd_ctx_put(release_new_ctx); | |
612 | } | |
613 | ||
9cd75c3c PE |
614 | /* |
615 | * ctx may go away after this if the userfault pseudo fd is | |
616 | * already released. | |
617 | */ | |
9a69a829 | 618 | out: |
df2cc96e | 619 | WRITE_ONCE(ctx->mmap_changing, false); |
9cd75c3c | 620 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
621 | } |
622 | ||
623 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
624 | struct userfaultfd_wait_queue *ewq) | |
625 | { | |
626 | ewq->msg.event = 0; | |
627 | wake_up_locked(&ctx->event_wqh); | |
628 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
629 | } | |
630 | ||
893e26e6 PE |
631 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
632 | { | |
633 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
634 | struct userfaultfd_fork_ctx *fctx; | |
635 | ||
636 | octx = vma->vm_userfaultfd_ctx.ctx; | |
637 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
638 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
639 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
640 | return 0; | |
641 | } | |
642 | ||
643 | list_for_each_entry(fctx, fcs, list) | |
644 | if (fctx->orig == octx) { | |
645 | ctx = fctx->new; | |
646 | break; | |
647 | } | |
648 | ||
649 | if (!ctx) { | |
650 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
651 | if (!fctx) | |
652 | return -ENOMEM; | |
653 | ||
654 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
655 | if (!ctx) { | |
656 | kfree(fctx); | |
657 | return -ENOMEM; | |
658 | } | |
659 | ||
ca880420 | 660 | refcount_set(&ctx->refcount, 1); |
893e26e6 PE |
661 | ctx->flags = octx->flags; |
662 | ctx->state = UFFD_STATE_RUNNING; | |
663 | ctx->features = octx->features; | |
664 | ctx->released = false; | |
df2cc96e | 665 | ctx->mmap_changing = false; |
893e26e6 | 666 | ctx->mm = vma->vm_mm; |
00bb31fa | 667 | mmgrab(ctx->mm); |
893e26e6 PE |
668 | |
669 | userfaultfd_ctx_get(octx); | |
df2cc96e | 670 | WRITE_ONCE(octx->mmap_changing, true); |
893e26e6 PE |
671 | fctx->orig = octx; |
672 | fctx->new = ctx; | |
673 | list_add_tail(&fctx->list, fcs); | |
674 | } | |
675 | ||
676 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
677 | return 0; | |
678 | } | |
679 | ||
8c9e7bb7 | 680 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
681 | { |
682 | struct userfaultfd_ctx *ctx = fctx->orig; | |
683 | struct userfaultfd_wait_queue ewq; | |
684 | ||
685 | msg_init(&ewq.msg); | |
686 | ||
687 | ewq.msg.event = UFFD_EVENT_FORK; | |
688 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
689 | ||
8c9e7bb7 | 690 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
691 | } |
692 | ||
693 | void dup_userfaultfd_complete(struct list_head *fcs) | |
694 | { | |
893e26e6 PE |
695 | struct userfaultfd_fork_ctx *fctx, *n; |
696 | ||
697 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 698 | dup_fctx(fctx); |
893e26e6 PE |
699 | list_del(&fctx->list); |
700 | kfree(fctx); | |
701 | } | |
702 | } | |
703 | ||
72f87654 PE |
704 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
705 | struct vm_userfaultfd_ctx *vm_ctx) | |
706 | { | |
707 | struct userfaultfd_ctx *ctx; | |
708 | ||
709 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
710 | |
711 | if (!ctx) | |
712 | return; | |
713 | ||
714 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
715 | vm_ctx->ctx = ctx; |
716 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 717 | WRITE_ONCE(ctx->mmap_changing, true); |
3cfd22be PX |
718 | } else { |
719 | /* Drop uffd context if remap feature not enabled */ | |
720 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
721 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
72f87654 PE |
722 | } |
723 | } | |
724 | ||
90794bf1 | 725 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
726 | unsigned long from, unsigned long to, |
727 | unsigned long len) | |
728 | { | |
90794bf1 | 729 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
730 | struct userfaultfd_wait_queue ewq; |
731 | ||
732 | if (!ctx) | |
733 | return; | |
734 | ||
735 | if (to & ~PAGE_MASK) { | |
736 | userfaultfd_ctx_put(ctx); | |
737 | return; | |
738 | } | |
739 | ||
740 | msg_init(&ewq.msg); | |
741 | ||
742 | ewq.msg.event = UFFD_EVENT_REMAP; | |
743 | ewq.msg.arg.remap.from = from; | |
744 | ewq.msg.arg.remap.to = to; | |
745 | ewq.msg.arg.remap.len = len; | |
746 | ||
747 | userfaultfd_event_wait_completion(ctx, &ewq); | |
748 | } | |
749 | ||
70ccb92f | 750 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 751 | unsigned long start, unsigned long end) |
05ce7724 PE |
752 | { |
753 | struct mm_struct *mm = vma->vm_mm; | |
754 | struct userfaultfd_ctx *ctx; | |
755 | struct userfaultfd_wait_queue ewq; | |
756 | ||
757 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 758 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 759 | return true; |
05ce7724 PE |
760 | |
761 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 762 | WRITE_ONCE(ctx->mmap_changing, true); |
d8ed45c5 | 763 | mmap_read_unlock(mm); |
05ce7724 | 764 | |
05ce7724 PE |
765 | msg_init(&ewq.msg); |
766 | ||
d811914d MR |
767 | ewq.msg.event = UFFD_EVENT_REMOVE; |
768 | ewq.msg.arg.remove.start = start; | |
769 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
770 | |
771 | userfaultfd_event_wait_completion(ctx, &ewq); | |
772 | ||
70ccb92f | 773 | return false; |
05ce7724 PE |
774 | } |
775 | ||
897ab3e0 MR |
776 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
777 | unsigned long start, unsigned long end) | |
778 | { | |
779 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
780 | ||
781 | list_for_each_entry(unmap_ctx, unmaps, list) | |
782 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
783 | unmap_ctx->end == end) | |
784 | return true; | |
785 | ||
786 | return false; | |
787 | } | |
788 | ||
789 | int userfaultfd_unmap_prep(struct vm_area_struct *vma, | |
790 | unsigned long start, unsigned long end, | |
791 | struct list_head *unmaps) | |
792 | { | |
793 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) { | |
794 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
795 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
796 | ||
797 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
798 | has_unmap_ctx(ctx, unmaps, start, end)) | |
799 | continue; | |
800 | ||
801 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
802 | if (!unmap_ctx) | |
803 | return -ENOMEM; | |
804 | ||
805 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 806 | WRITE_ONCE(ctx->mmap_changing, true); |
897ab3e0 MR |
807 | unmap_ctx->ctx = ctx; |
808 | unmap_ctx->start = start; | |
809 | unmap_ctx->end = end; | |
810 | list_add_tail(&unmap_ctx->list, unmaps); | |
811 | } | |
812 | ||
813 | return 0; | |
814 | } | |
815 | ||
816 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
817 | { | |
818 | struct userfaultfd_unmap_ctx *ctx, *n; | |
819 | struct userfaultfd_wait_queue ewq; | |
820 | ||
821 | list_for_each_entry_safe(ctx, n, uf, list) { | |
822 | msg_init(&ewq.msg); | |
823 | ||
824 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
825 | ewq.msg.arg.remove.start = ctx->start; | |
826 | ewq.msg.arg.remove.end = ctx->end; | |
827 | ||
828 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
829 | ||
830 | list_del(&ctx->list); | |
831 | kfree(ctx); | |
832 | } | |
833 | } | |
834 | ||
86039bd3 AA |
835 | static int userfaultfd_release(struct inode *inode, struct file *file) |
836 | { | |
837 | struct userfaultfd_ctx *ctx = file->private_data; | |
838 | struct mm_struct *mm = ctx->mm; | |
839 | struct vm_area_struct *vma, *prev; | |
840 | /* len == 0 means wake all */ | |
841 | struct userfaultfd_wake_range range = { .len = 0, }; | |
842 | unsigned long new_flags; | |
843 | ||
6aa7de05 | 844 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 845 | |
d2005e3f ON |
846 | if (!mmget_not_zero(mm)) |
847 | goto wakeup; | |
848 | ||
86039bd3 AA |
849 | /* |
850 | * Flush page faults out of all CPUs. NOTE: all page faults | |
851 | * must be retried without returning VM_FAULT_SIGBUS if | |
852 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
c1e8d7c6 | 853 | * changes while handle_userfault released the mmap_lock. So |
86039bd3 | 854 | * it's critical that released is set to true (above), before |
c1e8d7c6 | 855 | * taking the mmap_lock for writing. |
86039bd3 | 856 | */ |
d8ed45c5 | 857 | mmap_write_lock(mm); |
86039bd3 AA |
858 | prev = NULL; |
859 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
860 | cond_resched(); | |
861 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
862 | !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
863 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { | |
864 | prev = vma; | |
865 | continue; | |
866 | } | |
867 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
4d45e75a JH |
868 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, |
869 | new_flags, vma->anon_vma, | |
870 | vma->vm_file, vma->vm_pgoff, | |
871 | vma_policy(vma), | |
872 | NULL_VM_UFFD_CTX); | |
873 | if (prev) | |
874 | vma = prev; | |
875 | else | |
876 | prev = vma; | |
86039bd3 AA |
877 | vma->vm_flags = new_flags; |
878 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
879 | } | |
d8ed45c5 | 880 | mmap_write_unlock(mm); |
d2005e3f ON |
881 | mmput(mm); |
882 | wakeup: | |
86039bd3 | 883 | /* |
15b726ef | 884 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 885 | * the last page faults that may have been already waiting on |
15b726ef | 886 | * the fault_*wqh. |
86039bd3 | 887 | */ |
cbcfa130 | 888 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 889 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 890 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 891 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 892 | |
5a18b64e MR |
893 | /* Flush pending events that may still wait on event_wqh */ |
894 | wake_up_all(&ctx->event_wqh); | |
895 | ||
a9a08845 | 896 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
897 | userfaultfd_ctx_put(ctx); |
898 | return 0; | |
899 | } | |
900 | ||
15b726ef | 901 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
902 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
903 | wait_queue_head_t *wqh) | |
86039bd3 | 904 | { |
ac6424b9 | 905 | wait_queue_entry_t *wq; |
15b726ef | 906 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 907 | |
456a7378 | 908 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 909 | |
15b726ef | 910 | uwq = NULL; |
6dcc27fd | 911 | if (!waitqueue_active(wqh)) |
15b726ef AA |
912 | goto out; |
913 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 914 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
915 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
916 | out: | |
917 | return uwq; | |
86039bd3 | 918 | } |
6dcc27fd PE |
919 | |
920 | static inline struct userfaultfd_wait_queue *find_userfault( | |
921 | struct userfaultfd_ctx *ctx) | |
922 | { | |
923 | return find_userfault_in(&ctx->fault_pending_wqh); | |
924 | } | |
86039bd3 | 925 | |
9cd75c3c PE |
926 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
927 | struct userfaultfd_ctx *ctx) | |
928 | { | |
929 | return find_userfault_in(&ctx->event_wqh); | |
930 | } | |
931 | ||
076ccb76 | 932 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
933 | { |
934 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 935 | __poll_t ret; |
86039bd3 AA |
936 | |
937 | poll_wait(file, &ctx->fd_wqh, wait); | |
938 | ||
939 | switch (ctx->state) { | |
940 | case UFFD_STATE_WAIT_API: | |
a9a08845 | 941 | return EPOLLERR; |
86039bd3 | 942 | case UFFD_STATE_RUNNING: |
ba85c702 AA |
943 | /* |
944 | * poll() never guarantees that read won't block. | |
945 | * userfaults can be waken before they're read(). | |
946 | */ | |
947 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 948 | return EPOLLERR; |
15b726ef AA |
949 | /* |
950 | * lockless access to see if there are pending faults | |
951 | * __pollwait last action is the add_wait_queue but | |
952 | * the spin_unlock would allow the waitqueue_active to | |
953 | * pass above the actual list_add inside | |
954 | * add_wait_queue critical section. So use a full | |
955 | * memory barrier to serialize the list_add write of | |
956 | * add_wait_queue() with the waitqueue_active read | |
957 | * below. | |
958 | */ | |
959 | ret = 0; | |
960 | smp_mb(); | |
961 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
a9a08845 | 962 | ret = EPOLLIN; |
9cd75c3c | 963 | else if (waitqueue_active(&ctx->event_wqh)) |
a9a08845 | 964 | ret = EPOLLIN; |
9cd75c3c | 965 | |
86039bd3 AA |
966 | return ret; |
967 | default: | |
8474901a | 968 | WARN_ON_ONCE(1); |
a9a08845 | 969 | return EPOLLERR; |
86039bd3 AA |
970 | } |
971 | } | |
972 | ||
893e26e6 PE |
973 | static const struct file_operations userfaultfd_fops; |
974 | ||
975 | static int resolve_userfault_fork(struct userfaultfd_ctx *ctx, | |
976 | struct userfaultfd_ctx *new, | |
977 | struct uffd_msg *msg) | |
978 | { | |
979 | int fd; | |
893e26e6 | 980 | |
284cd241 EB |
981 | fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new, |
982 | O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS)); | |
893e26e6 PE |
983 | if (fd < 0) |
984 | return fd; | |
985 | ||
893e26e6 PE |
986 | msg->arg.reserved.reserved1 = 0; |
987 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
988 | return 0; |
989 | } | |
990 | ||
86039bd3 | 991 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
a9b85f94 | 992 | struct uffd_msg *msg) |
86039bd3 AA |
993 | { |
994 | ssize_t ret; | |
995 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 996 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
997 | /* |
998 | * Handling fork event requires sleeping operations, so | |
999 | * we drop the event_wqh lock, then do these ops, then | |
1000 | * lock it back and wake up the waiter. While the lock is | |
1001 | * dropped the ewq may go away so we keep track of it | |
1002 | * carefully. | |
1003 | */ | |
1004 | LIST_HEAD(fork_event); | |
1005 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1006 | |
15b726ef | 1007 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1008 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1009 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1010 | for (;;) { | |
1011 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1012 | spin_lock(&ctx->fault_pending_wqh.lock); |
1013 | uwq = find_userfault(ctx); | |
1014 | if (uwq) { | |
2c5b7e1b AA |
1015 | /* |
1016 | * Use a seqcount to repeat the lockless check | |
1017 | * in wake_userfault() to avoid missing | |
1018 | * wakeups because during the refile both | |
1019 | * waitqueue could become empty if this is the | |
1020 | * only userfault. | |
1021 | */ | |
1022 | write_seqcount_begin(&ctx->refile_seq); | |
1023 | ||
86039bd3 | 1024 | /* |
15b726ef AA |
1025 | * The fault_pending_wqh.lock prevents the uwq |
1026 | * to disappear from under us. | |
1027 | * | |
1028 | * Refile this userfault from | |
1029 | * fault_pending_wqh to fault_wqh, it's not | |
1030 | * pending anymore after we read it. | |
1031 | * | |
1032 | * Use list_del() by hand (as | |
1033 | * userfaultfd_wake_function also uses | |
1034 | * list_del_init() by hand) to be sure nobody | |
1035 | * changes __remove_wait_queue() to use | |
1036 | * list_del_init() in turn breaking the | |
1037 | * !list_empty_careful() check in | |
2055da97 | 1038 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1039 | * must never be empty at any time during the |
1040 | * refile, or the waitqueue could disappear | |
1041 | * from under us. The "wait_queue_head_t" | |
1042 | * parameter of __remove_wait_queue() is unused | |
1043 | * anyway. | |
86039bd3 | 1044 | */ |
2055da97 | 1045 | list_del(&uwq->wq.entry); |
c430d1e8 | 1046 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1047 | |
2c5b7e1b AA |
1048 | write_seqcount_end(&ctx->refile_seq); |
1049 | ||
a9b85f94 AA |
1050 | /* careful to always initialize msg if ret == 0 */ |
1051 | *msg = uwq->msg; | |
15b726ef | 1052 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1053 | ret = 0; |
1054 | break; | |
1055 | } | |
15b726ef | 1056 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1057 | |
1058 | spin_lock(&ctx->event_wqh.lock); | |
1059 | uwq = find_userfault_evt(ctx); | |
1060 | if (uwq) { | |
1061 | *msg = uwq->msg; | |
1062 | ||
893e26e6 PE |
1063 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1064 | fork_nctx = (struct userfaultfd_ctx *) | |
1065 | (unsigned long) | |
1066 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1067 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1068 | /* |
1069 | * fork_nctx can be freed as soon as | |
1070 | * we drop the lock, unless we take a | |
1071 | * reference on it. | |
1072 | */ | |
1073 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1074 | spin_unlock(&ctx->event_wqh.lock); |
1075 | ret = 0; | |
1076 | break; | |
1077 | } | |
1078 | ||
9cd75c3c PE |
1079 | userfaultfd_event_complete(ctx, uwq); |
1080 | spin_unlock(&ctx->event_wqh.lock); | |
1081 | ret = 0; | |
1082 | break; | |
1083 | } | |
1084 | spin_unlock(&ctx->event_wqh.lock); | |
1085 | ||
86039bd3 AA |
1086 | if (signal_pending(current)) { |
1087 | ret = -ERESTARTSYS; | |
1088 | break; | |
1089 | } | |
1090 | if (no_wait) { | |
1091 | ret = -EAGAIN; | |
1092 | break; | |
1093 | } | |
ae62c16e | 1094 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1095 | schedule(); |
ae62c16e | 1096 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1097 | } |
1098 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1099 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1100 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1101 | |
893e26e6 PE |
1102 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
1103 | ret = resolve_userfault_fork(ctx, fork_nctx, msg); | |
cbcfa130 | 1104 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1105 | if (!list_empty(&fork_event)) { |
1106 | /* | |
1107 | * The fork thread didn't abort, so we can | |
1108 | * drop the temporary refcount. | |
1109 | */ | |
1110 | userfaultfd_ctx_put(fork_nctx); | |
1111 | ||
1112 | uwq = list_first_entry(&fork_event, | |
1113 | typeof(*uwq), | |
1114 | wq.entry); | |
1115 | /* | |
1116 | * If fork_event list wasn't empty and in turn | |
1117 | * the event wasn't already released by fork | |
1118 | * (the event is allocated on fork kernel | |
1119 | * stack), put the event back to its place in | |
1120 | * the event_wq. fork_event head will be freed | |
1121 | * as soon as we return so the event cannot | |
1122 | * stay queued there no matter the current | |
1123 | * "ret" value. | |
1124 | */ | |
1125 | list_del(&uwq->wq.entry); | |
1126 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1127 | |
384632e6 AA |
1128 | /* |
1129 | * Leave the event in the waitqueue and report | |
1130 | * error to userland if we failed to resolve | |
1131 | * the userfault fork. | |
1132 | */ | |
1133 | if (likely(!ret)) | |
893e26e6 | 1134 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1135 | } else { |
1136 | /* | |
1137 | * Here the fork thread aborted and the | |
1138 | * refcount from the fork thread on fork_nctx | |
1139 | * has already been released. We still hold | |
1140 | * the reference we took before releasing the | |
1141 | * lock above. If resolve_userfault_fork | |
1142 | * failed we've to drop it because the | |
1143 | * fork_nctx has to be freed in such case. If | |
1144 | * it succeeded we'll hold it because the new | |
1145 | * uffd references it. | |
1146 | */ | |
1147 | if (ret) | |
1148 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1149 | } |
cbcfa130 | 1150 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1151 | } |
1152 | ||
86039bd3 AA |
1153 | return ret; |
1154 | } | |
1155 | ||
1156 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1157 | size_t count, loff_t *ppos) | |
1158 | { | |
1159 | struct userfaultfd_ctx *ctx = file->private_data; | |
1160 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1161 | struct uffd_msg msg; |
86039bd3 AA |
1162 | int no_wait = file->f_flags & O_NONBLOCK; |
1163 | ||
1164 | if (ctx->state == UFFD_STATE_WAIT_API) | |
1165 | return -EINVAL; | |
86039bd3 AA |
1166 | |
1167 | for (;;) { | |
a9b85f94 | 1168 | if (count < sizeof(msg)) |
86039bd3 | 1169 | return ret ? ret : -EINVAL; |
a9b85f94 | 1170 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg); |
86039bd3 AA |
1171 | if (_ret < 0) |
1172 | return ret ? ret : _ret; | |
a9b85f94 | 1173 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1174 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1175 | ret += sizeof(msg); |
1176 | buf += sizeof(msg); | |
1177 | count -= sizeof(msg); | |
86039bd3 AA |
1178 | /* |
1179 | * Allow to read more than one fault at time but only | |
1180 | * block if waiting for the very first one. | |
1181 | */ | |
1182 | no_wait = O_NONBLOCK; | |
1183 | } | |
1184 | } | |
1185 | ||
1186 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1187 | struct userfaultfd_wake_range *range) | |
1188 | { | |
cbcfa130 | 1189 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1190 | /* wake all in the range and autoremove */ |
15b726ef | 1191 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1192 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1193 | range); |
1194 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1195 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1196 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1197 | } |
1198 | ||
1199 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1200 | struct userfaultfd_wake_range *range) | |
1201 | { | |
2c5b7e1b AA |
1202 | unsigned seq; |
1203 | bool need_wakeup; | |
1204 | ||
86039bd3 AA |
1205 | /* |
1206 | * To be sure waitqueue_active() is not reordered by the CPU | |
1207 | * before the pagetable update, use an explicit SMP memory | |
3e4e28c5 | 1208 | * barrier here. PT lock release or mmap_read_unlock(mm) still |
86039bd3 AA |
1209 | * have release semantics that can allow the |
1210 | * waitqueue_active() to be reordered before the pte update. | |
1211 | */ | |
1212 | smp_mb(); | |
1213 | ||
1214 | /* | |
1215 | * Use waitqueue_active because it's very frequent to | |
1216 | * change the address space atomically even if there are no | |
1217 | * userfaults yet. So we take the spinlock only when we're | |
1218 | * sure we've userfaults to wake. | |
1219 | */ | |
2c5b7e1b AA |
1220 | do { |
1221 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1222 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1223 | waitqueue_active(&ctx->fault_wqh); | |
1224 | cond_resched(); | |
1225 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1226 | if (need_wakeup) | |
86039bd3 AA |
1227 | __wake_userfault(ctx, range); |
1228 | } | |
1229 | ||
1230 | static __always_inline int validate_range(struct mm_struct *mm, | |
7d032574 | 1231 | __u64 *start, __u64 len) |
86039bd3 AA |
1232 | { |
1233 | __u64 task_size = mm->task_size; | |
1234 | ||
7d032574 AK |
1235 | *start = untagged_addr(*start); |
1236 | ||
1237 | if (*start & ~PAGE_MASK) | |
86039bd3 AA |
1238 | return -EINVAL; |
1239 | if (len & ~PAGE_MASK) | |
1240 | return -EINVAL; | |
1241 | if (!len) | |
1242 | return -EINVAL; | |
7d032574 | 1243 | if (*start < mmap_min_addr) |
86039bd3 | 1244 | return -EINVAL; |
7d032574 | 1245 | if (*start >= task_size) |
86039bd3 | 1246 | return -EINVAL; |
7d032574 | 1247 | if (len > task_size - *start) |
86039bd3 AA |
1248 | return -EINVAL; |
1249 | return 0; | |
1250 | } | |
1251 | ||
63b2d417 AA |
1252 | static inline bool vma_can_userfault(struct vm_area_struct *vma, |
1253 | unsigned long vm_flags) | |
ba6907db | 1254 | { |
63b2d417 AA |
1255 | /* FIXME: add WP support to hugetlbfs and shmem */ |
1256 | return vma_is_anonymous(vma) || | |
1257 | ((is_vm_hugetlb_page(vma) || vma_is_shmem(vma)) && | |
1258 | !(vm_flags & VM_UFFD_WP)); | |
ba6907db MR |
1259 | } |
1260 | ||
86039bd3 AA |
1261 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, |
1262 | unsigned long arg) | |
1263 | { | |
1264 | struct mm_struct *mm = ctx->mm; | |
1265 | struct vm_area_struct *vma, *prev, *cur; | |
1266 | int ret; | |
1267 | struct uffdio_register uffdio_register; | |
1268 | struct uffdio_register __user *user_uffdio_register; | |
1269 | unsigned long vm_flags, new_flags; | |
1270 | bool found; | |
ce53e8e6 | 1271 | bool basic_ioctls; |
86039bd3 AA |
1272 | unsigned long start, end, vma_end; |
1273 | ||
1274 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1275 | ||
1276 | ret = -EFAULT; | |
1277 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1278 | sizeof(uffdio_register)-sizeof(__u64))) | |
1279 | goto out; | |
1280 | ||
1281 | ret = -EINVAL; | |
1282 | if (!uffdio_register.mode) | |
1283 | goto out; | |
1284 | if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING| | |
1285 | UFFDIO_REGISTER_MODE_WP)) | |
1286 | goto out; | |
1287 | vm_flags = 0; | |
1288 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1289 | vm_flags |= VM_UFFD_MISSING; | |
63b2d417 | 1290 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) |
86039bd3 | 1291 | vm_flags |= VM_UFFD_WP; |
86039bd3 | 1292 | |
7d032574 | 1293 | ret = validate_range(mm, &uffdio_register.range.start, |
86039bd3 AA |
1294 | uffdio_register.range.len); |
1295 | if (ret) | |
1296 | goto out; | |
1297 | ||
1298 | start = uffdio_register.range.start; | |
1299 | end = start + uffdio_register.range.len; | |
1300 | ||
d2005e3f ON |
1301 | ret = -ENOMEM; |
1302 | if (!mmget_not_zero(mm)) | |
1303 | goto out; | |
1304 | ||
d8ed45c5 | 1305 | mmap_write_lock(mm); |
86039bd3 | 1306 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1307 | if (!vma) |
1308 | goto out_unlock; | |
1309 | ||
1310 | /* check that there's at least one vma in the range */ | |
1311 | ret = -EINVAL; | |
1312 | if (vma->vm_start >= end) | |
1313 | goto out_unlock; | |
1314 | ||
cab350af MK |
1315 | /* |
1316 | * If the first vma contains huge pages, make sure start address | |
1317 | * is aligned to huge page size. | |
1318 | */ | |
1319 | if (is_vm_hugetlb_page(vma)) { | |
1320 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1321 | ||
1322 | if (start & (vma_hpagesize - 1)) | |
1323 | goto out_unlock; | |
1324 | } | |
1325 | ||
86039bd3 AA |
1326 | /* |
1327 | * Search for not compatible vmas. | |
86039bd3 AA |
1328 | */ |
1329 | found = false; | |
ce53e8e6 | 1330 | basic_ioctls = false; |
86039bd3 AA |
1331 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { |
1332 | cond_resched(); | |
1333 | ||
1334 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1335 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1336 | ||
1337 | /* check not compatible vmas */ | |
1338 | ret = -EINVAL; | |
63b2d417 | 1339 | if (!vma_can_userfault(cur, vm_flags)) |
86039bd3 | 1340 | goto out_unlock; |
29ec9066 AA |
1341 | |
1342 | /* | |
1343 | * UFFDIO_COPY will fill file holes even without | |
1344 | * PROT_WRITE. This check enforces that if this is a | |
1345 | * MAP_SHARED, the process has write permission to the backing | |
1346 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1347 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1348 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1349 | */ | |
1350 | ret = -EPERM; | |
1351 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1352 | goto out_unlock; | |
1353 | ||
cab350af MK |
1354 | /* |
1355 | * If this vma contains ending address, and huge pages | |
1356 | * check alignment. | |
1357 | */ | |
1358 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1359 | end > cur->vm_start) { | |
1360 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1361 | ||
1362 | ret = -EINVAL; | |
1363 | ||
1364 | if (end & (vma_hpagesize - 1)) | |
1365 | goto out_unlock; | |
1366 | } | |
63b2d417 AA |
1367 | if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE)) |
1368 | goto out_unlock; | |
86039bd3 AA |
1369 | |
1370 | /* | |
1371 | * Check that this vma isn't already owned by a | |
1372 | * different userfaultfd. We can't allow more than one | |
1373 | * userfaultfd to own a single vma simultaneously or we | |
1374 | * wouldn't know which one to deliver the userfaults to. | |
1375 | */ | |
1376 | ret = -EBUSY; | |
1377 | if (cur->vm_userfaultfd_ctx.ctx && | |
1378 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1379 | goto out_unlock; | |
1380 | ||
cab350af MK |
1381 | /* |
1382 | * Note vmas containing huge pages | |
1383 | */ | |
ce53e8e6 MR |
1384 | if (is_vm_hugetlb_page(cur)) |
1385 | basic_ioctls = true; | |
cab350af | 1386 | |
86039bd3 AA |
1387 | found = true; |
1388 | } | |
1389 | BUG_ON(!found); | |
1390 | ||
1391 | if (vma->vm_start < start) | |
1392 | prev = vma; | |
1393 | ||
1394 | ret = 0; | |
1395 | do { | |
1396 | cond_resched(); | |
1397 | ||
63b2d417 | 1398 | BUG_ON(!vma_can_userfault(vma, vm_flags)); |
86039bd3 AA |
1399 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1400 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1401 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1402 | |
1403 | /* | |
1404 | * Nothing to do: this vma is already registered into this | |
1405 | * userfaultfd and with the right tracking mode too. | |
1406 | */ | |
1407 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1408 | (vma->vm_flags & vm_flags) == vm_flags) | |
1409 | goto skip; | |
1410 | ||
1411 | if (vma->vm_start > start) | |
1412 | start = vma->vm_start; | |
1413 | vma_end = min(end, vma->vm_end); | |
1414 | ||
9d4678eb AA |
1415 | new_flags = (vma->vm_flags & |
1416 | ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags; | |
86039bd3 AA |
1417 | prev = vma_merge(mm, prev, start, vma_end, new_flags, |
1418 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1419 | vma_policy(vma), | |
1420 | ((struct vm_userfaultfd_ctx){ ctx })); | |
1421 | if (prev) { | |
1422 | vma = prev; | |
1423 | goto next; | |
1424 | } | |
1425 | if (vma->vm_start < start) { | |
1426 | ret = split_vma(mm, vma, start, 1); | |
1427 | if (ret) | |
1428 | break; | |
1429 | } | |
1430 | if (vma->vm_end > end) { | |
1431 | ret = split_vma(mm, vma, end, 0); | |
1432 | if (ret) | |
1433 | break; | |
1434 | } | |
1435 | next: | |
1436 | /* | |
1437 | * In the vma_merge() successful mprotect-like case 8: | |
1438 | * the next vma was merged into the current one and | |
1439 | * the current one has not been updated yet. | |
1440 | */ | |
1441 | vma->vm_flags = new_flags; | |
1442 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
1443 | ||
1444 | skip: | |
1445 | prev = vma; | |
1446 | start = vma->vm_end; | |
1447 | vma = vma->vm_next; | |
1448 | } while (vma && vma->vm_start < end); | |
1449 | out_unlock: | |
d8ed45c5 | 1450 | mmap_write_unlock(mm); |
d2005e3f | 1451 | mmput(mm); |
86039bd3 | 1452 | if (!ret) { |
14819305 PX |
1453 | __u64 ioctls_out; |
1454 | ||
1455 | ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : | |
1456 | UFFD_API_RANGE_IOCTLS; | |
1457 | ||
1458 | /* | |
1459 | * Declare the WP ioctl only if the WP mode is | |
1460 | * specified and all checks passed with the range | |
1461 | */ | |
1462 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)) | |
1463 | ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT); | |
1464 | ||
86039bd3 AA |
1465 | /* |
1466 | * Now that we scanned all vmas we can already tell | |
1467 | * userland which ioctls methods are guaranteed to | |
1468 | * succeed on this range. | |
1469 | */ | |
14819305 | 1470 | if (put_user(ioctls_out, &user_uffdio_register->ioctls)) |
86039bd3 AA |
1471 | ret = -EFAULT; |
1472 | } | |
1473 | out: | |
1474 | return ret; | |
1475 | } | |
1476 | ||
1477 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1478 | unsigned long arg) | |
1479 | { | |
1480 | struct mm_struct *mm = ctx->mm; | |
1481 | struct vm_area_struct *vma, *prev, *cur; | |
1482 | int ret; | |
1483 | struct uffdio_range uffdio_unregister; | |
1484 | unsigned long new_flags; | |
1485 | bool found; | |
1486 | unsigned long start, end, vma_end; | |
1487 | const void __user *buf = (void __user *)arg; | |
1488 | ||
1489 | ret = -EFAULT; | |
1490 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1491 | goto out; | |
1492 | ||
7d032574 | 1493 | ret = validate_range(mm, &uffdio_unregister.start, |
86039bd3 AA |
1494 | uffdio_unregister.len); |
1495 | if (ret) | |
1496 | goto out; | |
1497 | ||
1498 | start = uffdio_unregister.start; | |
1499 | end = start + uffdio_unregister.len; | |
1500 | ||
d2005e3f ON |
1501 | ret = -ENOMEM; |
1502 | if (!mmget_not_zero(mm)) | |
1503 | goto out; | |
1504 | ||
d8ed45c5 | 1505 | mmap_write_lock(mm); |
86039bd3 | 1506 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1507 | if (!vma) |
1508 | goto out_unlock; | |
1509 | ||
1510 | /* check that there's at least one vma in the range */ | |
1511 | ret = -EINVAL; | |
1512 | if (vma->vm_start >= end) | |
1513 | goto out_unlock; | |
1514 | ||
cab350af MK |
1515 | /* |
1516 | * If the first vma contains huge pages, make sure start address | |
1517 | * is aligned to huge page size. | |
1518 | */ | |
1519 | if (is_vm_hugetlb_page(vma)) { | |
1520 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1521 | ||
1522 | if (start & (vma_hpagesize - 1)) | |
1523 | goto out_unlock; | |
1524 | } | |
1525 | ||
86039bd3 AA |
1526 | /* |
1527 | * Search for not compatible vmas. | |
86039bd3 AA |
1528 | */ |
1529 | found = false; | |
1530 | ret = -EINVAL; | |
1531 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
1532 | cond_resched(); | |
1533 | ||
1534 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1535 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1536 | ||
1537 | /* | |
1538 | * Check not compatible vmas, not strictly required | |
1539 | * here as not compatible vmas cannot have an | |
1540 | * userfaultfd_ctx registered on them, but this | |
1541 | * provides for more strict behavior to notice | |
1542 | * unregistration errors. | |
1543 | */ | |
63b2d417 | 1544 | if (!vma_can_userfault(cur, cur->vm_flags)) |
86039bd3 AA |
1545 | goto out_unlock; |
1546 | ||
1547 | found = true; | |
1548 | } | |
1549 | BUG_ON(!found); | |
1550 | ||
1551 | if (vma->vm_start < start) | |
1552 | prev = vma; | |
1553 | ||
1554 | ret = 0; | |
1555 | do { | |
1556 | cond_resched(); | |
1557 | ||
63b2d417 | 1558 | BUG_ON(!vma_can_userfault(vma, vma->vm_flags)); |
86039bd3 AA |
1559 | |
1560 | /* | |
1561 | * Nothing to do: this vma is already registered into this | |
1562 | * userfaultfd and with the right tracking mode too. | |
1563 | */ | |
1564 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1565 | goto skip; | |
1566 | ||
01e881f5 AA |
1567 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1568 | ||
86039bd3 AA |
1569 | if (vma->vm_start > start) |
1570 | start = vma->vm_start; | |
1571 | vma_end = min(end, vma->vm_end); | |
1572 | ||
09fa5296 AA |
1573 | if (userfaultfd_missing(vma)) { |
1574 | /* | |
1575 | * Wake any concurrent pending userfault while | |
1576 | * we unregister, so they will not hang | |
1577 | * permanently and it avoids userland to call | |
1578 | * UFFDIO_WAKE explicitly. | |
1579 | */ | |
1580 | struct userfaultfd_wake_range range; | |
1581 | range.start = start; | |
1582 | range.len = vma_end - start; | |
1583 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1584 | } | |
1585 | ||
86039bd3 AA |
1586 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); |
1587 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
1588 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1589 | vma_policy(vma), | |
1590 | NULL_VM_UFFD_CTX); | |
1591 | if (prev) { | |
1592 | vma = prev; | |
1593 | goto next; | |
1594 | } | |
1595 | if (vma->vm_start < start) { | |
1596 | ret = split_vma(mm, vma, start, 1); | |
1597 | if (ret) | |
1598 | break; | |
1599 | } | |
1600 | if (vma->vm_end > end) { | |
1601 | ret = split_vma(mm, vma, end, 0); | |
1602 | if (ret) | |
1603 | break; | |
1604 | } | |
1605 | next: | |
1606 | /* | |
1607 | * In the vma_merge() successful mprotect-like case 8: | |
1608 | * the next vma was merged into the current one and | |
1609 | * the current one has not been updated yet. | |
1610 | */ | |
1611 | vma->vm_flags = new_flags; | |
1612 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
1613 | ||
1614 | skip: | |
1615 | prev = vma; | |
1616 | start = vma->vm_end; | |
1617 | vma = vma->vm_next; | |
1618 | } while (vma && vma->vm_start < end); | |
1619 | out_unlock: | |
d8ed45c5 | 1620 | mmap_write_unlock(mm); |
d2005e3f | 1621 | mmput(mm); |
86039bd3 AA |
1622 | out: |
1623 | return ret; | |
1624 | } | |
1625 | ||
1626 | /* | |
ba85c702 AA |
1627 | * userfaultfd_wake may be used in combination with the |
1628 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1629 | */ |
1630 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1631 | unsigned long arg) | |
1632 | { | |
1633 | int ret; | |
1634 | struct uffdio_range uffdio_wake; | |
1635 | struct userfaultfd_wake_range range; | |
1636 | const void __user *buf = (void __user *)arg; | |
1637 | ||
1638 | ret = -EFAULT; | |
1639 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1640 | goto out; | |
1641 | ||
7d032574 | 1642 | ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1643 | if (ret) |
1644 | goto out; | |
1645 | ||
1646 | range.start = uffdio_wake.start; | |
1647 | range.len = uffdio_wake.len; | |
1648 | ||
1649 | /* | |
1650 | * len == 0 means wake all and we don't want to wake all here, | |
1651 | * so check it again to be sure. | |
1652 | */ | |
1653 | VM_BUG_ON(!range.len); | |
1654 | ||
1655 | wake_userfault(ctx, &range); | |
1656 | ret = 0; | |
1657 | ||
1658 | out: | |
1659 | return ret; | |
1660 | } | |
1661 | ||
ad465cae AA |
1662 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1663 | unsigned long arg) | |
1664 | { | |
1665 | __s64 ret; | |
1666 | struct uffdio_copy uffdio_copy; | |
1667 | struct uffdio_copy __user *user_uffdio_copy; | |
1668 | struct userfaultfd_wake_range range; | |
1669 | ||
1670 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1671 | ||
df2cc96e MR |
1672 | ret = -EAGAIN; |
1673 | if (READ_ONCE(ctx->mmap_changing)) | |
1674 | goto out; | |
1675 | ||
ad465cae AA |
1676 | ret = -EFAULT; |
1677 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1678 | /* don't copy "copy" last field */ | |
1679 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1680 | goto out; | |
1681 | ||
7d032574 | 1682 | ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1683 | if (ret) |
1684 | goto out; | |
1685 | /* | |
1686 | * double check for wraparound just in case. copy_from_user() | |
1687 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1688 | * in the userland range. | |
1689 | */ | |
1690 | ret = -EINVAL; | |
1691 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1692 | goto out; | |
72981e0e | 1693 | if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP)) |
ad465cae | 1694 | goto out; |
d2005e3f ON |
1695 | if (mmget_not_zero(ctx->mm)) { |
1696 | ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src, | |
72981e0e AA |
1697 | uffdio_copy.len, &ctx->mmap_changing, |
1698 | uffdio_copy.mode); | |
d2005e3f | 1699 | mmput(ctx->mm); |
96333187 | 1700 | } else { |
e86b298b | 1701 | return -ESRCH; |
d2005e3f | 1702 | } |
ad465cae AA |
1703 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1704 | return -EFAULT; | |
1705 | if (ret < 0) | |
1706 | goto out; | |
1707 | BUG_ON(!ret); | |
1708 | /* len == 0 would wake all */ | |
1709 | range.len = ret; | |
1710 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1711 | range.start = uffdio_copy.dst; | |
1712 | wake_userfault(ctx, &range); | |
1713 | } | |
1714 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1715 | out: | |
1716 | return ret; | |
1717 | } | |
1718 | ||
1719 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1720 | unsigned long arg) | |
1721 | { | |
1722 | __s64 ret; | |
1723 | struct uffdio_zeropage uffdio_zeropage; | |
1724 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1725 | struct userfaultfd_wake_range range; | |
1726 | ||
1727 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1728 | ||
df2cc96e MR |
1729 | ret = -EAGAIN; |
1730 | if (READ_ONCE(ctx->mmap_changing)) | |
1731 | goto out; | |
1732 | ||
ad465cae AA |
1733 | ret = -EFAULT; |
1734 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1735 | /* don't copy "zeropage" last field */ | |
1736 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1737 | goto out; | |
1738 | ||
7d032574 | 1739 | ret = validate_range(ctx->mm, &uffdio_zeropage.range.start, |
ad465cae AA |
1740 | uffdio_zeropage.range.len); |
1741 | if (ret) | |
1742 | goto out; | |
1743 | ret = -EINVAL; | |
1744 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1745 | goto out; | |
1746 | ||
d2005e3f ON |
1747 | if (mmget_not_zero(ctx->mm)) { |
1748 | ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start, | |
df2cc96e MR |
1749 | uffdio_zeropage.range.len, |
1750 | &ctx->mmap_changing); | |
d2005e3f | 1751 | mmput(ctx->mm); |
9d95aa4b | 1752 | } else { |
e86b298b | 1753 | return -ESRCH; |
d2005e3f | 1754 | } |
ad465cae AA |
1755 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1756 | return -EFAULT; | |
1757 | if (ret < 0) | |
1758 | goto out; | |
1759 | /* len == 0 would wake all */ | |
1760 | BUG_ON(!ret); | |
1761 | range.len = ret; | |
1762 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1763 | range.start = uffdio_zeropage.range.start; | |
1764 | wake_userfault(ctx, &range); | |
1765 | } | |
1766 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1767 | out: | |
1768 | return ret; | |
1769 | } | |
1770 | ||
63b2d417 AA |
1771 | static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx, |
1772 | unsigned long arg) | |
1773 | { | |
1774 | int ret; | |
1775 | struct uffdio_writeprotect uffdio_wp; | |
1776 | struct uffdio_writeprotect __user *user_uffdio_wp; | |
1777 | struct userfaultfd_wake_range range; | |
23080e27 | 1778 | bool mode_wp, mode_dontwake; |
63b2d417 AA |
1779 | |
1780 | if (READ_ONCE(ctx->mmap_changing)) | |
1781 | return -EAGAIN; | |
1782 | ||
1783 | user_uffdio_wp = (struct uffdio_writeprotect __user *) arg; | |
1784 | ||
1785 | if (copy_from_user(&uffdio_wp, user_uffdio_wp, | |
1786 | sizeof(struct uffdio_writeprotect))) | |
1787 | return -EFAULT; | |
1788 | ||
1789 | ret = validate_range(ctx->mm, &uffdio_wp.range.start, | |
1790 | uffdio_wp.range.len); | |
1791 | if (ret) | |
1792 | return ret; | |
1793 | ||
1794 | if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE | | |
1795 | UFFDIO_WRITEPROTECT_MODE_WP)) | |
1796 | return -EINVAL; | |
23080e27 PX |
1797 | |
1798 | mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP; | |
1799 | mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE; | |
1800 | ||
1801 | if (mode_wp && mode_dontwake) | |
63b2d417 AA |
1802 | return -EINVAL; |
1803 | ||
1804 | ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start, | |
23080e27 | 1805 | uffdio_wp.range.len, mode_wp, |
63b2d417 AA |
1806 | &ctx->mmap_changing); |
1807 | if (ret) | |
1808 | return ret; | |
1809 | ||
23080e27 | 1810 | if (!mode_wp && !mode_dontwake) { |
63b2d417 AA |
1811 | range.start = uffdio_wp.range.start; |
1812 | range.len = uffdio_wp.range.len; | |
1813 | wake_userfault(ctx, &range); | |
1814 | } | |
1815 | return ret; | |
1816 | } | |
1817 | ||
9cd75c3c PE |
1818 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1819 | { | |
1820 | /* | |
1821 | * For the current set of features the bits just coincide | |
1822 | */ | |
1823 | return (unsigned int)user_features; | |
1824 | } | |
1825 | ||
86039bd3 AA |
1826 | /* |
1827 | * userland asks for a certain API version and we return which bits | |
1828 | * and ioctl commands are implemented in this kernel for such API | |
1829 | * version or -EINVAL if unknown. | |
1830 | */ | |
1831 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1832 | unsigned long arg) | |
1833 | { | |
1834 | struct uffdio_api uffdio_api; | |
1835 | void __user *buf = (void __user *)arg; | |
1836 | int ret; | |
65603144 | 1837 | __u64 features; |
86039bd3 AA |
1838 | |
1839 | ret = -EINVAL; | |
1840 | if (ctx->state != UFFD_STATE_WAIT_API) | |
1841 | goto out; | |
1842 | ret = -EFAULT; | |
a9b85f94 | 1843 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1844 | goto out; |
65603144 | 1845 | features = uffdio_api.features; |
3c1c24d9 MR |
1846 | ret = -EINVAL; |
1847 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) | |
1848 | goto err_out; | |
1849 | ret = -EPERM; | |
1850 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
1851 | goto err_out; | |
65603144 AA |
1852 | /* report all available features and ioctls to userland */ |
1853 | uffdio_api.features = UFFD_API_FEATURES; | |
86039bd3 AA |
1854 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1855 | ret = -EFAULT; | |
1856 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1857 | goto out; | |
1858 | ctx->state = UFFD_STATE_RUNNING; | |
65603144 AA |
1859 | /* only enable the requested features for this uffd context */ |
1860 | ctx->features = uffd_ctx_features(features); | |
86039bd3 AA |
1861 | ret = 0; |
1862 | out: | |
1863 | return ret; | |
3c1c24d9 MR |
1864 | err_out: |
1865 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
1866 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1867 | ret = -EFAULT; | |
1868 | goto out; | |
86039bd3 AA |
1869 | } |
1870 | ||
1871 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
1872 | unsigned long arg) | |
1873 | { | |
1874 | int ret = -EINVAL; | |
1875 | struct userfaultfd_ctx *ctx = file->private_data; | |
1876 | ||
e6485a47 AA |
1877 | if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API) |
1878 | return -EINVAL; | |
1879 | ||
86039bd3 AA |
1880 | switch(cmd) { |
1881 | case UFFDIO_API: | |
1882 | ret = userfaultfd_api(ctx, arg); | |
1883 | break; | |
1884 | case UFFDIO_REGISTER: | |
1885 | ret = userfaultfd_register(ctx, arg); | |
1886 | break; | |
1887 | case UFFDIO_UNREGISTER: | |
1888 | ret = userfaultfd_unregister(ctx, arg); | |
1889 | break; | |
1890 | case UFFDIO_WAKE: | |
1891 | ret = userfaultfd_wake(ctx, arg); | |
1892 | break; | |
ad465cae AA |
1893 | case UFFDIO_COPY: |
1894 | ret = userfaultfd_copy(ctx, arg); | |
1895 | break; | |
1896 | case UFFDIO_ZEROPAGE: | |
1897 | ret = userfaultfd_zeropage(ctx, arg); | |
1898 | break; | |
63b2d417 AA |
1899 | case UFFDIO_WRITEPROTECT: |
1900 | ret = userfaultfd_writeprotect(ctx, arg); | |
1901 | break; | |
86039bd3 AA |
1902 | } |
1903 | return ret; | |
1904 | } | |
1905 | ||
1906 | #ifdef CONFIG_PROC_FS | |
1907 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
1908 | { | |
1909 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 1910 | wait_queue_entry_t *wq; |
86039bd3 AA |
1911 | unsigned long pending = 0, total = 0; |
1912 | ||
cbcfa130 | 1913 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 1914 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
1915 | pending++; |
1916 | total++; | |
1917 | } | |
2055da97 | 1918 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
1919 | total++; |
1920 | } | |
cbcfa130 | 1921 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1922 | |
1923 | /* | |
1924 | * If more protocols will be added, there will be all shown | |
1925 | * separated by a space. Like this: | |
1926 | * protocols: aa:... bb:... | |
1927 | */ | |
1928 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 1929 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
1930 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
1931 | } | |
1932 | #endif | |
1933 | ||
1934 | static const struct file_operations userfaultfd_fops = { | |
1935 | #ifdef CONFIG_PROC_FS | |
1936 | .show_fdinfo = userfaultfd_show_fdinfo, | |
1937 | #endif | |
1938 | .release = userfaultfd_release, | |
1939 | .poll = userfaultfd_poll, | |
1940 | .read = userfaultfd_read, | |
1941 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 1942 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
1943 | .llseek = noop_llseek, |
1944 | }; | |
1945 | ||
3004ec9c AA |
1946 | static void init_once_userfaultfd_ctx(void *mem) |
1947 | { | |
1948 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
1949 | ||
1950 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
1951 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 1952 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 1953 | init_waitqueue_head(&ctx->fd_wqh); |
2ca97ac8 | 1954 | seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock); |
3004ec9c AA |
1955 | } |
1956 | ||
284cd241 | 1957 | SYSCALL_DEFINE1(userfaultfd, int, flags) |
86039bd3 | 1958 | { |
86039bd3 | 1959 | struct userfaultfd_ctx *ctx; |
284cd241 | 1960 | int fd; |
86039bd3 | 1961 | |
cefdca0a PX |
1962 | if (!sysctl_unprivileged_userfaultfd && !capable(CAP_SYS_PTRACE)) |
1963 | return -EPERM; | |
1964 | ||
86039bd3 AA |
1965 | BUG_ON(!current->mm); |
1966 | ||
1967 | /* Check the UFFD_* constants for consistency. */ | |
1968 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); | |
1969 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
1970 | ||
86039bd3 | 1971 | if (flags & ~UFFD_SHARED_FCNTL_FLAGS) |
284cd241 | 1972 | return -EINVAL; |
86039bd3 | 1973 | |
3004ec9c | 1974 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 1975 | if (!ctx) |
284cd241 | 1976 | return -ENOMEM; |
86039bd3 | 1977 | |
ca880420 | 1978 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 1979 | ctx->flags = flags; |
9cd75c3c | 1980 | ctx->features = 0; |
86039bd3 AA |
1981 | ctx->state = UFFD_STATE_WAIT_API; |
1982 | ctx->released = false; | |
df2cc96e | 1983 | ctx->mmap_changing = false; |
86039bd3 AA |
1984 | ctx->mm = current->mm; |
1985 | /* prevent the mm struct to be freed */ | |
f1f10076 | 1986 | mmgrab(ctx->mm); |
86039bd3 | 1987 | |
284cd241 EB |
1988 | fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx, |
1989 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS)); | |
1990 | if (fd < 0) { | |
d2005e3f | 1991 | mmdrop(ctx->mm); |
3004ec9c | 1992 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 1993 | } |
86039bd3 | 1994 | return fd; |
86039bd3 | 1995 | } |
3004ec9c AA |
1996 | |
1997 | static int __init userfaultfd_init(void) | |
1998 | { | |
1999 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", | |
2000 | sizeof(struct userfaultfd_ctx), | |
2001 | 0, | |
2002 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
2003 | init_once_userfaultfd_ctx); | |
2004 | return 0; | |
2005 | } | |
2006 | __initcall(userfaultfd_init); |