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 | 17 | #include <linux/mm.h> |
6dfeaff9 | 18 | #include <linux/mmu_notifier.h> |
86039bd3 AA |
19 | #include <linux/poll.h> |
20 | #include <linux/slab.h> | |
21 | #include <linux/seq_file.h> | |
22 | #include <linux/file.h> | |
23 | #include <linux/bug.h> | |
24 | #include <linux/anon_inodes.h> | |
25 | #include <linux/syscalls.h> | |
26 | #include <linux/userfaultfd_k.h> | |
27 | #include <linux/mempolicy.h> | |
28 | #include <linux/ioctl.h> | |
29 | #include <linux/security.h> | |
cab350af | 30 | #include <linux/hugetlb.h> |
86039bd3 | 31 | |
d0d4730a | 32 | int sysctl_unprivileged_userfaultfd __read_mostly; |
cefdca0a | 33 | |
3004ec9c AA |
34 | static struct kmem_cache *userfaultfd_ctx_cachep __read_mostly; |
35 | ||
86039bd3 AA |
36 | enum userfaultfd_state { |
37 | UFFD_STATE_WAIT_API, | |
38 | UFFD_STATE_RUNNING, | |
39 | }; | |
40 | ||
3004ec9c AA |
41 | /* |
42 | * Start with fault_pending_wqh and fault_wqh so they're more likely | |
43 | * to be in the same cacheline. | |
cbcfa130 EB |
44 | * |
45 | * Locking order: | |
46 | * fd_wqh.lock | |
47 | * fault_pending_wqh.lock | |
48 | * fault_wqh.lock | |
49 | * event_wqh.lock | |
50 | * | |
51 | * To avoid deadlocks, IRQs must be disabled when taking any of the above locks, | |
52 | * since fd_wqh.lock is taken by aio_poll() while it's holding a lock that's | |
53 | * also taken in IRQ context. | |
3004ec9c | 54 | */ |
86039bd3 | 55 | struct userfaultfd_ctx { |
15b726ef AA |
56 | /* waitqueue head for the pending (i.e. not read) userfaults */ |
57 | wait_queue_head_t fault_pending_wqh; | |
58 | /* waitqueue head for the userfaults */ | |
86039bd3 AA |
59 | wait_queue_head_t fault_wqh; |
60 | /* waitqueue head for the pseudo fd to wakeup poll/read */ | |
61 | wait_queue_head_t fd_wqh; | |
9cd75c3c PE |
62 | /* waitqueue head for events */ |
63 | wait_queue_head_t event_wqh; | |
2c5b7e1b | 64 | /* a refile sequence protected by fault_pending_wqh lock */ |
2ca97ac8 | 65 | seqcount_spinlock_t refile_seq; |
3004ec9c | 66 | /* pseudo fd refcounting */ |
ca880420 | 67 | refcount_t refcount; |
86039bd3 AA |
68 | /* userfaultfd syscall flags */ |
69 | unsigned int flags; | |
9cd75c3c PE |
70 | /* features requested from the userspace */ |
71 | unsigned int features; | |
86039bd3 AA |
72 | /* state machine */ |
73 | enum userfaultfd_state state; | |
74 | /* released */ | |
75 | bool released; | |
df2cc96e MR |
76 | /* memory mappings are changing because of non-cooperative event */ |
77 | bool mmap_changing; | |
86039bd3 AA |
78 | /* mm with one ore more vmas attached to this userfaultfd_ctx */ |
79 | struct mm_struct *mm; | |
80 | }; | |
81 | ||
893e26e6 PE |
82 | struct userfaultfd_fork_ctx { |
83 | struct userfaultfd_ctx *orig; | |
84 | struct userfaultfd_ctx *new; | |
85 | struct list_head list; | |
86 | }; | |
87 | ||
897ab3e0 MR |
88 | struct userfaultfd_unmap_ctx { |
89 | struct userfaultfd_ctx *ctx; | |
90 | unsigned long start; | |
91 | unsigned long end; | |
92 | struct list_head list; | |
93 | }; | |
94 | ||
86039bd3 | 95 | struct userfaultfd_wait_queue { |
a9b85f94 | 96 | struct uffd_msg msg; |
ac6424b9 | 97 | wait_queue_entry_t wq; |
86039bd3 | 98 | struct userfaultfd_ctx *ctx; |
15a77c6f | 99 | bool waken; |
86039bd3 AA |
100 | }; |
101 | ||
102 | struct userfaultfd_wake_range { | |
103 | unsigned long start; | |
104 | unsigned long len; | |
105 | }; | |
106 | ||
ac6424b9 | 107 | static int userfaultfd_wake_function(wait_queue_entry_t *wq, unsigned mode, |
86039bd3 AA |
108 | int wake_flags, void *key) |
109 | { | |
110 | struct userfaultfd_wake_range *range = key; | |
111 | int ret; | |
112 | struct userfaultfd_wait_queue *uwq; | |
113 | unsigned long start, len; | |
114 | ||
115 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); | |
116 | ret = 0; | |
86039bd3 AA |
117 | /* len == 0 means wake all */ |
118 | start = range->start; | |
119 | len = range->len; | |
a9b85f94 AA |
120 | if (len && (start > uwq->msg.arg.pagefault.address || |
121 | start + len <= uwq->msg.arg.pagefault.address)) | |
86039bd3 | 122 | goto out; |
15a77c6f AA |
123 | WRITE_ONCE(uwq->waken, true); |
124 | /* | |
a9668cd6 PZ |
125 | * The Program-Order guarantees provided by the scheduler |
126 | * ensure uwq->waken is visible before the task is woken. | |
15a77c6f | 127 | */ |
86039bd3 | 128 | ret = wake_up_state(wq->private, mode); |
a9668cd6 | 129 | if (ret) { |
86039bd3 AA |
130 | /* |
131 | * Wake only once, autoremove behavior. | |
132 | * | |
a9668cd6 PZ |
133 | * After the effect of list_del_init is visible to the other |
134 | * CPUs, the waitqueue may disappear from under us, see the | |
135 | * !list_empty_careful() in handle_userfault(). | |
136 | * | |
137 | * try_to_wake_up() has an implicit smp_mb(), and the | |
138 | * wq->private is read before calling the extern function | |
139 | * "wake_up_state" (which in turns calls try_to_wake_up). | |
86039bd3 | 140 | */ |
2055da97 | 141 | list_del_init(&wq->entry); |
a9668cd6 | 142 | } |
86039bd3 AA |
143 | out: |
144 | return ret; | |
145 | } | |
146 | ||
147 | /** | |
148 | * userfaultfd_ctx_get - Acquires a reference to the internal userfaultfd | |
149 | * context. | |
150 | * @ctx: [in] Pointer to the userfaultfd context. | |
86039bd3 AA |
151 | */ |
152 | static void userfaultfd_ctx_get(struct userfaultfd_ctx *ctx) | |
153 | { | |
ca880420 | 154 | refcount_inc(&ctx->refcount); |
86039bd3 AA |
155 | } |
156 | ||
157 | /** | |
158 | * userfaultfd_ctx_put - Releases a reference to the internal userfaultfd | |
159 | * context. | |
160 | * @ctx: [in] Pointer to userfaultfd context. | |
161 | * | |
162 | * The userfaultfd context reference must have been previously acquired either | |
163 | * with userfaultfd_ctx_get() or userfaultfd_ctx_fdget(). | |
164 | */ | |
165 | static void userfaultfd_ctx_put(struct userfaultfd_ctx *ctx) | |
166 | { | |
ca880420 | 167 | if (refcount_dec_and_test(&ctx->refcount)) { |
86039bd3 AA |
168 | VM_BUG_ON(spin_is_locked(&ctx->fault_pending_wqh.lock)); |
169 | VM_BUG_ON(waitqueue_active(&ctx->fault_pending_wqh)); | |
170 | VM_BUG_ON(spin_is_locked(&ctx->fault_wqh.lock)); | |
171 | VM_BUG_ON(waitqueue_active(&ctx->fault_wqh)); | |
9cd75c3c PE |
172 | VM_BUG_ON(spin_is_locked(&ctx->event_wqh.lock)); |
173 | VM_BUG_ON(waitqueue_active(&ctx->event_wqh)); | |
86039bd3 AA |
174 | VM_BUG_ON(spin_is_locked(&ctx->fd_wqh.lock)); |
175 | VM_BUG_ON(waitqueue_active(&ctx->fd_wqh)); | |
d2005e3f | 176 | mmdrop(ctx->mm); |
3004ec9c | 177 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
86039bd3 AA |
178 | } |
179 | } | |
180 | ||
a9b85f94 | 181 | static inline void msg_init(struct uffd_msg *msg) |
86039bd3 | 182 | { |
a9b85f94 AA |
183 | BUILD_BUG_ON(sizeof(struct uffd_msg) != 32); |
184 | /* | |
185 | * Must use memset to zero out the paddings or kernel data is | |
186 | * leaked to userland. | |
187 | */ | |
188 | memset(msg, 0, sizeof(struct uffd_msg)); | |
189 | } | |
190 | ||
191 | static inline struct uffd_msg userfault_msg(unsigned long address, | |
192 | unsigned int flags, | |
9d4ac934 AP |
193 | unsigned long reason, |
194 | unsigned int features) | |
a9b85f94 AA |
195 | { |
196 | struct uffd_msg msg; | |
197 | msg_init(&msg); | |
198 | msg.event = UFFD_EVENT_PAGEFAULT; | |
199 | msg.arg.pagefault.address = address; | |
86039bd3 AA |
200 | if (flags & FAULT_FLAG_WRITE) |
201 | /* | |
a4605a61 | 202 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
a9b85f94 AA |
203 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WRITE |
204 | * was not set in a UFFD_EVENT_PAGEFAULT, it means it | |
205 | * was a read fault, otherwise if set it means it's | |
206 | * a write fault. | |
86039bd3 | 207 | */ |
a9b85f94 | 208 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WRITE; |
86039bd3 AA |
209 | if (reason & VM_UFFD_WP) |
210 | /* | |
a9b85f94 AA |
211 | * If UFFD_FEATURE_PAGEFAULT_FLAG_WP was set in the |
212 | * uffdio_api.features and UFFD_PAGEFAULT_FLAG_WP was | |
213 | * not set in a UFFD_EVENT_PAGEFAULT, it means it was | |
214 | * a missing fault, otherwise if set it means it's a | |
215 | * write protect fault. | |
86039bd3 | 216 | */ |
a9b85f94 | 217 | msg.arg.pagefault.flags |= UFFD_PAGEFAULT_FLAG_WP; |
9d4ac934 | 218 | if (features & UFFD_FEATURE_THREAD_ID) |
a36985d3 | 219 | msg.arg.pagefault.feat.ptid = task_pid_vnr(current); |
a9b85f94 | 220 | return msg; |
86039bd3 AA |
221 | } |
222 | ||
369cd212 MK |
223 | #ifdef CONFIG_HUGETLB_PAGE |
224 | /* | |
225 | * Same functionality as userfaultfd_must_wait below with modifications for | |
226 | * hugepmd ranges. | |
227 | */ | |
228 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 229 | struct vm_area_struct *vma, |
369cd212 MK |
230 | unsigned long address, |
231 | unsigned long flags, | |
232 | unsigned long reason) | |
233 | { | |
234 | struct mm_struct *mm = ctx->mm; | |
1e2c0436 | 235 | pte_t *ptep, pte; |
369cd212 MK |
236 | bool ret = true; |
237 | ||
42fc5414 | 238 | mmap_assert_locked(mm); |
369cd212 | 239 | |
1e2c0436 JF |
240 | ptep = huge_pte_offset(mm, address, vma_mmu_pagesize(vma)); |
241 | ||
242 | if (!ptep) | |
369cd212 MK |
243 | goto out; |
244 | ||
245 | ret = false; | |
1e2c0436 | 246 | pte = huge_ptep_get(ptep); |
369cd212 MK |
247 | |
248 | /* | |
249 | * Lockless access: we're in a wait_event so it's ok if it | |
250 | * changes under us. | |
251 | */ | |
1e2c0436 | 252 | if (huge_pte_none(pte)) |
369cd212 | 253 | ret = true; |
1e2c0436 | 254 | if (!huge_pte_write(pte) && (reason & VM_UFFD_WP)) |
369cd212 MK |
255 | ret = true; |
256 | out: | |
257 | return ret; | |
258 | } | |
259 | #else | |
260 | static inline bool userfaultfd_huge_must_wait(struct userfaultfd_ctx *ctx, | |
7868a208 | 261 | struct vm_area_struct *vma, |
369cd212 MK |
262 | unsigned long address, |
263 | unsigned long flags, | |
264 | unsigned long reason) | |
265 | { | |
266 | return false; /* should never get here */ | |
267 | } | |
268 | #endif /* CONFIG_HUGETLB_PAGE */ | |
269 | ||
8d2afd96 AA |
270 | /* |
271 | * Verify the pagetables are still not ok after having reigstered into | |
272 | * the fault_pending_wqh to avoid userland having to UFFDIO_WAKE any | |
273 | * userfault that has already been resolved, if userfaultfd_read and | |
274 | * UFFDIO_COPY|ZEROPAGE are being run simultaneously on two different | |
275 | * threads. | |
276 | */ | |
277 | static inline bool userfaultfd_must_wait(struct userfaultfd_ctx *ctx, | |
278 | unsigned long address, | |
279 | unsigned long flags, | |
280 | unsigned long reason) | |
281 | { | |
282 | struct mm_struct *mm = ctx->mm; | |
283 | pgd_t *pgd; | |
c2febafc | 284 | p4d_t *p4d; |
8d2afd96 AA |
285 | pud_t *pud; |
286 | pmd_t *pmd, _pmd; | |
287 | pte_t *pte; | |
288 | bool ret = true; | |
289 | ||
42fc5414 | 290 | mmap_assert_locked(mm); |
8d2afd96 AA |
291 | |
292 | pgd = pgd_offset(mm, address); | |
293 | if (!pgd_present(*pgd)) | |
294 | goto out; | |
c2febafc KS |
295 | p4d = p4d_offset(pgd, address); |
296 | if (!p4d_present(*p4d)) | |
297 | goto out; | |
298 | pud = pud_offset(p4d, address); | |
8d2afd96 AA |
299 | if (!pud_present(*pud)) |
300 | goto out; | |
301 | pmd = pmd_offset(pud, address); | |
302 | /* | |
303 | * READ_ONCE must function as a barrier with narrower scope | |
304 | * and it must be equivalent to: | |
305 | * _pmd = *pmd; barrier(); | |
306 | * | |
307 | * This is to deal with the instability (as in | |
308 | * pmd_trans_unstable) of the pmd. | |
309 | */ | |
310 | _pmd = READ_ONCE(*pmd); | |
a365ac09 | 311 | if (pmd_none(_pmd)) |
8d2afd96 AA |
312 | goto out; |
313 | ||
314 | ret = false; | |
a365ac09 HY |
315 | if (!pmd_present(_pmd)) |
316 | goto out; | |
317 | ||
63b2d417 AA |
318 | if (pmd_trans_huge(_pmd)) { |
319 | if (!pmd_write(_pmd) && (reason & VM_UFFD_WP)) | |
320 | ret = true; | |
8d2afd96 | 321 | goto out; |
63b2d417 | 322 | } |
8d2afd96 AA |
323 | |
324 | /* | |
325 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
326 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
327 | */ | |
328 | pte = pte_offset_map(pmd, address); | |
329 | /* | |
330 | * Lockless access: we're in a wait_event so it's ok if it | |
331 | * changes under us. | |
332 | */ | |
333 | if (pte_none(*pte)) | |
334 | ret = true; | |
63b2d417 AA |
335 | if (!pte_write(*pte) && (reason & VM_UFFD_WP)) |
336 | ret = true; | |
8d2afd96 AA |
337 | pte_unmap(pte); |
338 | ||
339 | out: | |
340 | return ret; | |
341 | } | |
342 | ||
3e69ad08 PX |
343 | static inline long userfaultfd_get_blocking_state(unsigned int flags) |
344 | { | |
345 | if (flags & FAULT_FLAG_INTERRUPTIBLE) | |
346 | return TASK_INTERRUPTIBLE; | |
347 | ||
348 | if (flags & FAULT_FLAG_KILLABLE) | |
349 | return TASK_KILLABLE; | |
350 | ||
351 | return TASK_UNINTERRUPTIBLE; | |
352 | } | |
353 | ||
86039bd3 AA |
354 | /* |
355 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
356 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
357 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
358 | * recommendation in __lock_page_or_retry is not an understatement. | |
359 | * | |
c1e8d7c6 | 360 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_lock must be released |
86039bd3 AA |
361 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is |
362 | * not set. | |
363 | * | |
364 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
365 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
c1e8d7c6 | 366 | * fatal_signal_pending()s, and the mmap_lock must be released before |
86039bd3 AA |
367 | * returning it. |
368 | */ | |
2b740303 | 369 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 370 | { |
82b0f8c3 | 371 | struct mm_struct *mm = vmf->vma->vm_mm; |
86039bd3 AA |
372 | struct userfaultfd_ctx *ctx; |
373 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 374 | vm_fault_t ret = VM_FAULT_SIGBUS; |
3e69ad08 | 375 | bool must_wait; |
15a77c6f | 376 | long blocking_state; |
86039bd3 | 377 | |
64c2b203 AA |
378 | /* |
379 | * We don't do userfault handling for the final child pid update. | |
380 | * | |
381 | * We also don't do userfault handling during | |
382 | * coredumping. hugetlbfs has the special | |
383 | * follow_hugetlb_page() to skip missing pages in the | |
384 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
385 | * the no_page_table() helper in follow_page_mask(), but the | |
386 | * shmem_vm_ops->fault method is invoked even during | |
c1e8d7c6 | 387 | * coredumping without mmap_lock and it ends up here. |
64c2b203 AA |
388 | */ |
389 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
390 | goto out; | |
391 | ||
392 | /* | |
c1e8d7c6 ML |
393 | * Coredumping runs without mmap_lock so we can only check that |
394 | * the mmap_lock is held, if PF_DUMPCORE was not set. | |
64c2b203 | 395 | */ |
42fc5414 | 396 | mmap_assert_locked(mm); |
64c2b203 | 397 | |
82b0f8c3 | 398 | ctx = vmf->vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 399 | if (!ctx) |
ba85c702 | 400 | goto out; |
86039bd3 AA |
401 | |
402 | BUG_ON(ctx->mm != mm); | |
403 | ||
404 | VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP)); | |
405 | VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP)); | |
406 | ||
2d6d6f5a PS |
407 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
408 | goto out; | |
37cd0575 LG |
409 | if ((vmf->flags & FAULT_FLAG_USER) == 0 && |
410 | ctx->flags & UFFD_USER_MODE_ONLY) { | |
411 | printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd " | |
412 | "sysctl knob to 1 if kernel faults must be handled " | |
413 | "without obtaining CAP_SYS_PTRACE capability\n"); | |
414 | goto out; | |
415 | } | |
2d6d6f5a | 416 | |
86039bd3 AA |
417 | /* |
418 | * If it's already released don't get it. This avoids to loop | |
419 | * in __get_user_pages if userfaultfd_release waits on the | |
c1e8d7c6 | 420 | * caller of handle_userfault to release the mmap_lock. |
86039bd3 | 421 | */ |
6aa7de05 | 422 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
423 | /* |
424 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
425 | * cooperative manager can close the uffd after the | |
426 | * last UFFDIO_COPY, without risking to trigger an | |
427 | * involuntary SIGBUS if the process was starting the | |
428 | * userfaultfd while the userfaultfd was still armed | |
429 | * (but after the last UFFDIO_COPY). If the uffd | |
430 | * wasn't already closed when the userfault reached | |
431 | * this point, that would normally be solved by | |
432 | * userfaultfd_must_wait returning 'false'. | |
433 | * | |
434 | * If we were to return VM_FAULT_SIGBUS here, the non | |
435 | * cooperative manager would be instead forced to | |
436 | * always call UFFDIO_UNREGISTER before it can safely | |
437 | * close the uffd. | |
438 | */ | |
439 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 440 | goto out; |
656710a6 | 441 | } |
86039bd3 AA |
442 | |
443 | /* | |
444 | * Check that we can return VM_FAULT_RETRY. | |
445 | * | |
446 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
447 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
448 | * -EBUSY failures, if the userfaultfd is to be extended for | |
449 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
450 | * without first stopping userland access to the memory. For | |
451 | * VM_UFFD_MISSING userfaults this is enough for now. | |
452 | */ | |
82b0f8c3 | 453 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
454 | /* |
455 | * Validate the invariant that nowait must allow retry | |
456 | * to be sure not to return SIGBUS erroneously on | |
457 | * nowait invocations. | |
458 | */ | |
82b0f8c3 | 459 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
460 | #ifdef CONFIG_DEBUG_VM |
461 | if (printk_ratelimit()) { | |
462 | printk(KERN_WARNING | |
82b0f8c3 JK |
463 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
464 | vmf->flags); | |
86039bd3 AA |
465 | dump_stack(); |
466 | } | |
467 | #endif | |
ba85c702 | 468 | goto out; |
86039bd3 AA |
469 | } |
470 | ||
471 | /* | |
472 | * Handle nowait, not much to do other than tell it to retry | |
473 | * and wait. | |
474 | */ | |
ba85c702 | 475 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 476 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 477 | goto out; |
86039bd3 | 478 | |
c1e8d7c6 | 479 | /* take the reference before dropping the mmap_lock */ |
86039bd3 AA |
480 | userfaultfd_ctx_get(ctx); |
481 | ||
86039bd3 AA |
482 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
483 | uwq.wq.private = current; | |
9d4ac934 AP |
484 | uwq.msg = userfault_msg(vmf->address, vmf->flags, reason, |
485 | ctx->features); | |
86039bd3 | 486 | uwq.ctx = ctx; |
15a77c6f | 487 | uwq.waken = false; |
86039bd3 | 488 | |
3e69ad08 | 489 | blocking_state = userfaultfd_get_blocking_state(vmf->flags); |
dfa37dc3 | 490 | |
cbcfa130 | 491 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
492 | /* |
493 | * After the __add_wait_queue the uwq is visible to userland | |
494 | * through poll/read(). | |
495 | */ | |
15b726ef AA |
496 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
497 | /* | |
498 | * The smp_mb() after __set_current_state prevents the reads | |
499 | * following the spin_unlock to happen before the list_add in | |
500 | * __add_wait_queue. | |
501 | */ | |
15a77c6f | 502 | set_current_state(blocking_state); |
cbcfa130 | 503 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 504 | |
369cd212 MK |
505 | if (!is_vm_hugetlb_page(vmf->vma)) |
506 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, | |
507 | reason); | |
508 | else | |
7868a208 PA |
509 | must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma, |
510 | vmf->address, | |
369cd212 | 511 | vmf->flags, reason); |
d8ed45c5 | 512 | mmap_read_unlock(mm); |
8d2afd96 | 513 | |
f9bf3522 | 514 | if (likely(must_wait && !READ_ONCE(ctx->released))) { |
a9a08845 | 515 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 516 | schedule(); |
ba85c702 | 517 | } |
86039bd3 | 518 | |
ba85c702 | 519 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
520 | |
521 | /* | |
522 | * Here we race with the list_del; list_add in | |
523 | * userfaultfd_ctx_read(), however because we don't ever run | |
524 | * list_del_init() to refile across the two lists, the prev | |
525 | * and next pointers will never point to self. list_add also | |
526 | * would never let any of the two pointers to point to | |
527 | * self. So list_empty_careful won't risk to see both pointers | |
528 | * pointing to self at any time during the list refile. The | |
529 | * only case where list_del_init() is called is the full | |
530 | * removal in the wake function and there we don't re-list_add | |
531 | * and it's fine not to block on the spinlock. The uwq on this | |
532 | * kernel stack can be released after the list_del_init. | |
533 | */ | |
2055da97 | 534 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 535 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
536 | /* |
537 | * No need of list_del_init(), the uwq on the stack | |
538 | * will be freed shortly anyway. | |
539 | */ | |
2055da97 | 540 | list_del(&uwq.wq.entry); |
cbcfa130 | 541 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 542 | } |
86039bd3 AA |
543 | |
544 | /* | |
545 | * ctx may go away after this if the userfault pseudo fd is | |
546 | * already released. | |
547 | */ | |
548 | userfaultfd_ctx_put(ctx); | |
549 | ||
ba85c702 AA |
550 | out: |
551 | return ret; | |
86039bd3 AA |
552 | } |
553 | ||
8c9e7bb7 AA |
554 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
555 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 556 | { |
0cbb4b4f AA |
557 | struct userfaultfd_ctx *release_new_ctx; |
558 | ||
9a69a829 AA |
559 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
560 | goto out; | |
9cd75c3c PE |
561 | |
562 | ewq->ctx = ctx; | |
563 | init_waitqueue_entry(&ewq->wq, current); | |
0cbb4b4f | 564 | release_new_ctx = NULL; |
9cd75c3c | 565 | |
cbcfa130 | 566 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
567 | /* |
568 | * After the __add_wait_queue the uwq is visible to userland | |
569 | * through poll/read(). | |
570 | */ | |
571 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
572 | for (;;) { | |
573 | set_current_state(TASK_KILLABLE); | |
574 | if (ewq->msg.event == 0) | |
575 | break; | |
6aa7de05 | 576 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 577 | fatal_signal_pending(current)) { |
384632e6 AA |
578 | /* |
579 | * &ewq->wq may be queued in fork_event, but | |
580 | * __remove_wait_queue ignores the head | |
581 | * parameter. It would be a problem if it | |
582 | * didn't. | |
583 | */ | |
9cd75c3c | 584 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
585 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
586 | struct userfaultfd_ctx *new; | |
587 | ||
588 | new = (struct userfaultfd_ctx *) | |
589 | (unsigned long) | |
590 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 591 | release_new_ctx = new; |
7eb76d45 | 592 | } |
9cd75c3c PE |
593 | break; |
594 | } | |
595 | ||
cbcfa130 | 596 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 597 | |
a9a08845 | 598 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
599 | schedule(); |
600 | ||
cbcfa130 | 601 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
602 | } |
603 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 604 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 605 | |
0cbb4b4f AA |
606 | if (release_new_ctx) { |
607 | struct vm_area_struct *vma; | |
608 | struct mm_struct *mm = release_new_ctx->mm; | |
609 | ||
610 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
d8ed45c5 | 611 | mmap_write_lock(mm); |
0cbb4b4f | 612 | for (vma = mm->mmap; vma; vma = vma->vm_next) |
31e810aa | 613 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 614 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
31e810aa MR |
615 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); |
616 | } | |
d8ed45c5 | 617 | mmap_write_unlock(mm); |
0cbb4b4f AA |
618 | |
619 | userfaultfd_ctx_put(release_new_ctx); | |
620 | } | |
621 | ||
9cd75c3c PE |
622 | /* |
623 | * ctx may go away after this if the userfault pseudo fd is | |
624 | * already released. | |
625 | */ | |
9a69a829 | 626 | out: |
df2cc96e | 627 | WRITE_ONCE(ctx->mmap_changing, false); |
9cd75c3c | 628 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
629 | } |
630 | ||
631 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
632 | struct userfaultfd_wait_queue *ewq) | |
633 | { | |
634 | ewq->msg.event = 0; | |
635 | wake_up_locked(&ctx->event_wqh); | |
636 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
637 | } | |
638 | ||
893e26e6 PE |
639 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
640 | { | |
641 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
642 | struct userfaultfd_fork_ctx *fctx; | |
643 | ||
644 | octx = vma->vm_userfaultfd_ctx.ctx; | |
645 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
646 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
647 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
648 | return 0; | |
649 | } | |
650 | ||
651 | list_for_each_entry(fctx, fcs, list) | |
652 | if (fctx->orig == octx) { | |
653 | ctx = fctx->new; | |
654 | break; | |
655 | } | |
656 | ||
657 | if (!ctx) { | |
658 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
659 | if (!fctx) | |
660 | return -ENOMEM; | |
661 | ||
662 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
663 | if (!ctx) { | |
664 | kfree(fctx); | |
665 | return -ENOMEM; | |
666 | } | |
667 | ||
ca880420 | 668 | refcount_set(&ctx->refcount, 1); |
893e26e6 PE |
669 | ctx->flags = octx->flags; |
670 | ctx->state = UFFD_STATE_RUNNING; | |
671 | ctx->features = octx->features; | |
672 | ctx->released = false; | |
df2cc96e | 673 | ctx->mmap_changing = false; |
893e26e6 | 674 | ctx->mm = vma->vm_mm; |
00bb31fa | 675 | mmgrab(ctx->mm); |
893e26e6 PE |
676 | |
677 | userfaultfd_ctx_get(octx); | |
df2cc96e | 678 | WRITE_ONCE(octx->mmap_changing, true); |
893e26e6 PE |
679 | fctx->orig = octx; |
680 | fctx->new = ctx; | |
681 | list_add_tail(&fctx->list, fcs); | |
682 | } | |
683 | ||
684 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
685 | return 0; | |
686 | } | |
687 | ||
8c9e7bb7 | 688 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
689 | { |
690 | struct userfaultfd_ctx *ctx = fctx->orig; | |
691 | struct userfaultfd_wait_queue ewq; | |
692 | ||
693 | msg_init(&ewq.msg); | |
694 | ||
695 | ewq.msg.event = UFFD_EVENT_FORK; | |
696 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
697 | ||
8c9e7bb7 | 698 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
699 | } |
700 | ||
701 | void dup_userfaultfd_complete(struct list_head *fcs) | |
702 | { | |
893e26e6 PE |
703 | struct userfaultfd_fork_ctx *fctx, *n; |
704 | ||
705 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 706 | dup_fctx(fctx); |
893e26e6 PE |
707 | list_del(&fctx->list); |
708 | kfree(fctx); | |
709 | } | |
710 | } | |
711 | ||
72f87654 PE |
712 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
713 | struct vm_userfaultfd_ctx *vm_ctx) | |
714 | { | |
715 | struct userfaultfd_ctx *ctx; | |
716 | ||
717 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
718 | |
719 | if (!ctx) | |
720 | return; | |
721 | ||
722 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
723 | vm_ctx->ctx = ctx; |
724 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 725 | WRITE_ONCE(ctx->mmap_changing, true); |
3cfd22be PX |
726 | } else { |
727 | /* Drop uffd context if remap feature not enabled */ | |
728 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
729 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
72f87654 PE |
730 | } |
731 | } | |
732 | ||
90794bf1 | 733 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
734 | unsigned long from, unsigned long to, |
735 | unsigned long len) | |
736 | { | |
90794bf1 | 737 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
738 | struct userfaultfd_wait_queue ewq; |
739 | ||
740 | if (!ctx) | |
741 | return; | |
742 | ||
743 | if (to & ~PAGE_MASK) { | |
744 | userfaultfd_ctx_put(ctx); | |
745 | return; | |
746 | } | |
747 | ||
748 | msg_init(&ewq.msg); | |
749 | ||
750 | ewq.msg.event = UFFD_EVENT_REMAP; | |
751 | ewq.msg.arg.remap.from = from; | |
752 | ewq.msg.arg.remap.to = to; | |
753 | ewq.msg.arg.remap.len = len; | |
754 | ||
755 | userfaultfd_event_wait_completion(ctx, &ewq); | |
756 | } | |
757 | ||
70ccb92f | 758 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 759 | unsigned long start, unsigned long end) |
05ce7724 PE |
760 | { |
761 | struct mm_struct *mm = vma->vm_mm; | |
762 | struct userfaultfd_ctx *ctx; | |
763 | struct userfaultfd_wait_queue ewq; | |
764 | ||
765 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 766 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 767 | return true; |
05ce7724 PE |
768 | |
769 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 770 | WRITE_ONCE(ctx->mmap_changing, true); |
d8ed45c5 | 771 | mmap_read_unlock(mm); |
05ce7724 | 772 | |
05ce7724 PE |
773 | msg_init(&ewq.msg); |
774 | ||
d811914d MR |
775 | ewq.msg.event = UFFD_EVENT_REMOVE; |
776 | ewq.msg.arg.remove.start = start; | |
777 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
778 | |
779 | userfaultfd_event_wait_completion(ctx, &ewq); | |
780 | ||
70ccb92f | 781 | return false; |
05ce7724 PE |
782 | } |
783 | ||
897ab3e0 MR |
784 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
785 | unsigned long start, unsigned long end) | |
786 | { | |
787 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
788 | ||
789 | list_for_each_entry(unmap_ctx, unmaps, list) | |
790 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
791 | unmap_ctx->end == end) | |
792 | return true; | |
793 | ||
794 | return false; | |
795 | } | |
796 | ||
797 | int userfaultfd_unmap_prep(struct vm_area_struct *vma, | |
798 | unsigned long start, unsigned long end, | |
799 | struct list_head *unmaps) | |
800 | { | |
801 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) { | |
802 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
803 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
804 | ||
805 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
806 | has_unmap_ctx(ctx, unmaps, start, end)) | |
807 | continue; | |
808 | ||
809 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
810 | if (!unmap_ctx) | |
811 | return -ENOMEM; | |
812 | ||
813 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 814 | WRITE_ONCE(ctx->mmap_changing, true); |
897ab3e0 MR |
815 | unmap_ctx->ctx = ctx; |
816 | unmap_ctx->start = start; | |
817 | unmap_ctx->end = end; | |
818 | list_add_tail(&unmap_ctx->list, unmaps); | |
819 | } | |
820 | ||
821 | return 0; | |
822 | } | |
823 | ||
824 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
825 | { | |
826 | struct userfaultfd_unmap_ctx *ctx, *n; | |
827 | struct userfaultfd_wait_queue ewq; | |
828 | ||
829 | list_for_each_entry_safe(ctx, n, uf, list) { | |
830 | msg_init(&ewq.msg); | |
831 | ||
832 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
833 | ewq.msg.arg.remove.start = ctx->start; | |
834 | ewq.msg.arg.remove.end = ctx->end; | |
835 | ||
836 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
837 | ||
838 | list_del(&ctx->list); | |
839 | kfree(ctx); | |
840 | } | |
841 | } | |
842 | ||
86039bd3 AA |
843 | static int userfaultfd_release(struct inode *inode, struct file *file) |
844 | { | |
845 | struct userfaultfd_ctx *ctx = file->private_data; | |
846 | struct mm_struct *mm = ctx->mm; | |
847 | struct vm_area_struct *vma, *prev; | |
848 | /* len == 0 means wake all */ | |
849 | struct userfaultfd_wake_range range = { .len = 0, }; | |
850 | unsigned long new_flags; | |
851 | ||
6aa7de05 | 852 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 853 | |
d2005e3f ON |
854 | if (!mmget_not_zero(mm)) |
855 | goto wakeup; | |
856 | ||
86039bd3 AA |
857 | /* |
858 | * Flush page faults out of all CPUs. NOTE: all page faults | |
859 | * must be retried without returning VM_FAULT_SIGBUS if | |
860 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
c1e8d7c6 | 861 | * changes while handle_userfault released the mmap_lock. So |
86039bd3 | 862 | * it's critical that released is set to true (above), before |
c1e8d7c6 | 863 | * taking the mmap_lock for writing. |
86039bd3 | 864 | */ |
d8ed45c5 | 865 | mmap_write_lock(mm); |
86039bd3 AA |
866 | prev = NULL; |
867 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
868 | cond_resched(); | |
869 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
870 | !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
871 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { | |
872 | prev = vma; | |
873 | continue; | |
874 | } | |
875 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
4d45e75a JH |
876 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, |
877 | new_flags, vma->anon_vma, | |
878 | vma->vm_file, vma->vm_pgoff, | |
879 | vma_policy(vma), | |
880 | NULL_VM_UFFD_CTX); | |
881 | if (prev) | |
882 | vma = prev; | |
883 | else | |
884 | prev = vma; | |
86039bd3 AA |
885 | vma->vm_flags = new_flags; |
886 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
887 | } | |
d8ed45c5 | 888 | mmap_write_unlock(mm); |
d2005e3f ON |
889 | mmput(mm); |
890 | wakeup: | |
86039bd3 | 891 | /* |
15b726ef | 892 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 893 | * the last page faults that may have been already waiting on |
15b726ef | 894 | * the fault_*wqh. |
86039bd3 | 895 | */ |
cbcfa130 | 896 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 897 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 898 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 899 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 900 | |
5a18b64e MR |
901 | /* Flush pending events that may still wait on event_wqh */ |
902 | wake_up_all(&ctx->event_wqh); | |
903 | ||
a9a08845 | 904 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
905 | userfaultfd_ctx_put(ctx); |
906 | return 0; | |
907 | } | |
908 | ||
15b726ef | 909 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
910 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
911 | wait_queue_head_t *wqh) | |
86039bd3 | 912 | { |
ac6424b9 | 913 | wait_queue_entry_t *wq; |
15b726ef | 914 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 915 | |
456a7378 | 916 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 917 | |
15b726ef | 918 | uwq = NULL; |
6dcc27fd | 919 | if (!waitqueue_active(wqh)) |
15b726ef AA |
920 | goto out; |
921 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 922 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
923 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
924 | out: | |
925 | return uwq; | |
86039bd3 | 926 | } |
6dcc27fd PE |
927 | |
928 | static inline struct userfaultfd_wait_queue *find_userfault( | |
929 | struct userfaultfd_ctx *ctx) | |
930 | { | |
931 | return find_userfault_in(&ctx->fault_pending_wqh); | |
932 | } | |
86039bd3 | 933 | |
9cd75c3c PE |
934 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
935 | struct userfaultfd_ctx *ctx) | |
936 | { | |
937 | return find_userfault_in(&ctx->event_wqh); | |
938 | } | |
939 | ||
076ccb76 | 940 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
941 | { |
942 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 943 | __poll_t ret; |
86039bd3 AA |
944 | |
945 | poll_wait(file, &ctx->fd_wqh, wait); | |
946 | ||
947 | switch (ctx->state) { | |
948 | case UFFD_STATE_WAIT_API: | |
a9a08845 | 949 | return EPOLLERR; |
86039bd3 | 950 | case UFFD_STATE_RUNNING: |
ba85c702 AA |
951 | /* |
952 | * poll() never guarantees that read won't block. | |
953 | * userfaults can be waken before they're read(). | |
954 | */ | |
955 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 956 | return EPOLLERR; |
15b726ef AA |
957 | /* |
958 | * lockless access to see if there are pending faults | |
959 | * __pollwait last action is the add_wait_queue but | |
960 | * the spin_unlock would allow the waitqueue_active to | |
961 | * pass above the actual list_add inside | |
962 | * add_wait_queue critical section. So use a full | |
963 | * memory barrier to serialize the list_add write of | |
964 | * add_wait_queue() with the waitqueue_active read | |
965 | * below. | |
966 | */ | |
967 | ret = 0; | |
968 | smp_mb(); | |
969 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
a9a08845 | 970 | ret = EPOLLIN; |
9cd75c3c | 971 | else if (waitqueue_active(&ctx->event_wqh)) |
a9a08845 | 972 | ret = EPOLLIN; |
9cd75c3c | 973 | |
86039bd3 AA |
974 | return ret; |
975 | default: | |
8474901a | 976 | WARN_ON_ONCE(1); |
a9a08845 | 977 | return EPOLLERR; |
86039bd3 AA |
978 | } |
979 | } | |
980 | ||
893e26e6 PE |
981 | static const struct file_operations userfaultfd_fops; |
982 | ||
b537900f DC |
983 | static int resolve_userfault_fork(struct userfaultfd_ctx *new, |
984 | struct inode *inode, | |
893e26e6 PE |
985 | struct uffd_msg *msg) |
986 | { | |
987 | int fd; | |
893e26e6 | 988 | |
b537900f DC |
989 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, new, |
990 | O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS), inode); | |
893e26e6 PE |
991 | if (fd < 0) |
992 | return fd; | |
993 | ||
893e26e6 PE |
994 | msg->arg.reserved.reserved1 = 0; |
995 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
996 | return 0; |
997 | } | |
998 | ||
86039bd3 | 999 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
b537900f | 1000 | struct uffd_msg *msg, struct inode *inode) |
86039bd3 AA |
1001 | { |
1002 | ssize_t ret; | |
1003 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 1004 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
1005 | /* |
1006 | * Handling fork event requires sleeping operations, so | |
1007 | * we drop the event_wqh lock, then do these ops, then | |
1008 | * lock it back and wake up the waiter. While the lock is | |
1009 | * dropped the ewq may go away so we keep track of it | |
1010 | * carefully. | |
1011 | */ | |
1012 | LIST_HEAD(fork_event); | |
1013 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1014 | |
15b726ef | 1015 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1016 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1017 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1018 | for (;;) { | |
1019 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1020 | spin_lock(&ctx->fault_pending_wqh.lock); |
1021 | uwq = find_userfault(ctx); | |
1022 | if (uwq) { | |
2c5b7e1b AA |
1023 | /* |
1024 | * Use a seqcount to repeat the lockless check | |
1025 | * in wake_userfault() to avoid missing | |
1026 | * wakeups because during the refile both | |
1027 | * waitqueue could become empty if this is the | |
1028 | * only userfault. | |
1029 | */ | |
1030 | write_seqcount_begin(&ctx->refile_seq); | |
1031 | ||
86039bd3 | 1032 | /* |
15b726ef AA |
1033 | * The fault_pending_wqh.lock prevents the uwq |
1034 | * to disappear from under us. | |
1035 | * | |
1036 | * Refile this userfault from | |
1037 | * fault_pending_wqh to fault_wqh, it's not | |
1038 | * pending anymore after we read it. | |
1039 | * | |
1040 | * Use list_del() by hand (as | |
1041 | * userfaultfd_wake_function also uses | |
1042 | * list_del_init() by hand) to be sure nobody | |
1043 | * changes __remove_wait_queue() to use | |
1044 | * list_del_init() in turn breaking the | |
1045 | * !list_empty_careful() check in | |
2055da97 | 1046 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1047 | * must never be empty at any time during the |
1048 | * refile, or the waitqueue could disappear | |
1049 | * from under us. The "wait_queue_head_t" | |
1050 | * parameter of __remove_wait_queue() is unused | |
1051 | * anyway. | |
86039bd3 | 1052 | */ |
2055da97 | 1053 | list_del(&uwq->wq.entry); |
c430d1e8 | 1054 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1055 | |
2c5b7e1b AA |
1056 | write_seqcount_end(&ctx->refile_seq); |
1057 | ||
a9b85f94 AA |
1058 | /* careful to always initialize msg if ret == 0 */ |
1059 | *msg = uwq->msg; | |
15b726ef | 1060 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1061 | ret = 0; |
1062 | break; | |
1063 | } | |
15b726ef | 1064 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1065 | |
1066 | spin_lock(&ctx->event_wqh.lock); | |
1067 | uwq = find_userfault_evt(ctx); | |
1068 | if (uwq) { | |
1069 | *msg = uwq->msg; | |
1070 | ||
893e26e6 PE |
1071 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1072 | fork_nctx = (struct userfaultfd_ctx *) | |
1073 | (unsigned long) | |
1074 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1075 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1076 | /* |
1077 | * fork_nctx can be freed as soon as | |
1078 | * we drop the lock, unless we take a | |
1079 | * reference on it. | |
1080 | */ | |
1081 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1082 | spin_unlock(&ctx->event_wqh.lock); |
1083 | ret = 0; | |
1084 | break; | |
1085 | } | |
1086 | ||
9cd75c3c PE |
1087 | userfaultfd_event_complete(ctx, uwq); |
1088 | spin_unlock(&ctx->event_wqh.lock); | |
1089 | ret = 0; | |
1090 | break; | |
1091 | } | |
1092 | spin_unlock(&ctx->event_wqh.lock); | |
1093 | ||
86039bd3 AA |
1094 | if (signal_pending(current)) { |
1095 | ret = -ERESTARTSYS; | |
1096 | break; | |
1097 | } | |
1098 | if (no_wait) { | |
1099 | ret = -EAGAIN; | |
1100 | break; | |
1101 | } | |
ae62c16e | 1102 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1103 | schedule(); |
ae62c16e | 1104 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1105 | } |
1106 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1107 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1108 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1109 | |
893e26e6 | 1110 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
b537900f | 1111 | ret = resolve_userfault_fork(fork_nctx, inode, msg); |
cbcfa130 | 1112 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1113 | if (!list_empty(&fork_event)) { |
1114 | /* | |
1115 | * The fork thread didn't abort, so we can | |
1116 | * drop the temporary refcount. | |
1117 | */ | |
1118 | userfaultfd_ctx_put(fork_nctx); | |
1119 | ||
1120 | uwq = list_first_entry(&fork_event, | |
1121 | typeof(*uwq), | |
1122 | wq.entry); | |
1123 | /* | |
1124 | * If fork_event list wasn't empty and in turn | |
1125 | * the event wasn't already released by fork | |
1126 | * (the event is allocated on fork kernel | |
1127 | * stack), put the event back to its place in | |
1128 | * the event_wq. fork_event head will be freed | |
1129 | * as soon as we return so the event cannot | |
1130 | * stay queued there no matter the current | |
1131 | * "ret" value. | |
1132 | */ | |
1133 | list_del(&uwq->wq.entry); | |
1134 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1135 | |
384632e6 AA |
1136 | /* |
1137 | * Leave the event in the waitqueue and report | |
1138 | * error to userland if we failed to resolve | |
1139 | * the userfault fork. | |
1140 | */ | |
1141 | if (likely(!ret)) | |
893e26e6 | 1142 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1143 | } else { |
1144 | /* | |
1145 | * Here the fork thread aborted and the | |
1146 | * refcount from the fork thread on fork_nctx | |
1147 | * has already been released. We still hold | |
1148 | * the reference we took before releasing the | |
1149 | * lock above. If resolve_userfault_fork | |
1150 | * failed we've to drop it because the | |
1151 | * fork_nctx has to be freed in such case. If | |
1152 | * it succeeded we'll hold it because the new | |
1153 | * uffd references it. | |
1154 | */ | |
1155 | if (ret) | |
1156 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1157 | } |
cbcfa130 | 1158 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1159 | } |
1160 | ||
86039bd3 AA |
1161 | return ret; |
1162 | } | |
1163 | ||
1164 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1165 | size_t count, loff_t *ppos) | |
1166 | { | |
1167 | struct userfaultfd_ctx *ctx = file->private_data; | |
1168 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1169 | struct uffd_msg msg; |
86039bd3 | 1170 | int no_wait = file->f_flags & O_NONBLOCK; |
b537900f | 1171 | struct inode *inode = file_inode(file); |
86039bd3 AA |
1172 | |
1173 | if (ctx->state == UFFD_STATE_WAIT_API) | |
1174 | return -EINVAL; | |
86039bd3 AA |
1175 | |
1176 | for (;;) { | |
a9b85f94 | 1177 | if (count < sizeof(msg)) |
86039bd3 | 1178 | return ret ? ret : -EINVAL; |
b537900f | 1179 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg, inode); |
86039bd3 AA |
1180 | if (_ret < 0) |
1181 | return ret ? ret : _ret; | |
a9b85f94 | 1182 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1183 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1184 | ret += sizeof(msg); |
1185 | buf += sizeof(msg); | |
1186 | count -= sizeof(msg); | |
86039bd3 AA |
1187 | /* |
1188 | * Allow to read more than one fault at time but only | |
1189 | * block if waiting for the very first one. | |
1190 | */ | |
1191 | no_wait = O_NONBLOCK; | |
1192 | } | |
1193 | } | |
1194 | ||
1195 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1196 | struct userfaultfd_wake_range *range) | |
1197 | { | |
cbcfa130 | 1198 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1199 | /* wake all in the range and autoremove */ |
15b726ef | 1200 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1201 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1202 | range); |
1203 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1204 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1205 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1206 | } |
1207 | ||
1208 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1209 | struct userfaultfd_wake_range *range) | |
1210 | { | |
2c5b7e1b AA |
1211 | unsigned seq; |
1212 | bool need_wakeup; | |
1213 | ||
86039bd3 AA |
1214 | /* |
1215 | * To be sure waitqueue_active() is not reordered by the CPU | |
1216 | * before the pagetable update, use an explicit SMP memory | |
3e4e28c5 | 1217 | * barrier here. PT lock release or mmap_read_unlock(mm) still |
86039bd3 AA |
1218 | * have release semantics that can allow the |
1219 | * waitqueue_active() to be reordered before the pte update. | |
1220 | */ | |
1221 | smp_mb(); | |
1222 | ||
1223 | /* | |
1224 | * Use waitqueue_active because it's very frequent to | |
1225 | * change the address space atomically even if there are no | |
1226 | * userfaults yet. So we take the spinlock only when we're | |
1227 | * sure we've userfaults to wake. | |
1228 | */ | |
2c5b7e1b AA |
1229 | do { |
1230 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1231 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1232 | waitqueue_active(&ctx->fault_wqh); | |
1233 | cond_resched(); | |
1234 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1235 | if (need_wakeup) | |
86039bd3 AA |
1236 | __wake_userfault(ctx, range); |
1237 | } | |
1238 | ||
1239 | static __always_inline int validate_range(struct mm_struct *mm, | |
7d032574 | 1240 | __u64 *start, __u64 len) |
86039bd3 AA |
1241 | { |
1242 | __u64 task_size = mm->task_size; | |
1243 | ||
7d032574 AK |
1244 | *start = untagged_addr(*start); |
1245 | ||
1246 | if (*start & ~PAGE_MASK) | |
86039bd3 AA |
1247 | return -EINVAL; |
1248 | if (len & ~PAGE_MASK) | |
1249 | return -EINVAL; | |
1250 | if (!len) | |
1251 | return -EINVAL; | |
7d032574 | 1252 | if (*start < mmap_min_addr) |
86039bd3 | 1253 | return -EINVAL; |
7d032574 | 1254 | if (*start >= task_size) |
86039bd3 | 1255 | return -EINVAL; |
7d032574 | 1256 | if (len > task_size - *start) |
86039bd3 AA |
1257 | return -EINVAL; |
1258 | return 0; | |
1259 | } | |
1260 | ||
63b2d417 AA |
1261 | static inline bool vma_can_userfault(struct vm_area_struct *vma, |
1262 | unsigned long vm_flags) | |
ba6907db | 1263 | { |
63b2d417 AA |
1264 | /* FIXME: add WP support to hugetlbfs and shmem */ |
1265 | return vma_is_anonymous(vma) || | |
1266 | ((is_vm_hugetlb_page(vma) || vma_is_shmem(vma)) && | |
1267 | !(vm_flags & VM_UFFD_WP)); | |
ba6907db MR |
1268 | } |
1269 | ||
86039bd3 AA |
1270 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, |
1271 | unsigned long arg) | |
1272 | { | |
1273 | struct mm_struct *mm = ctx->mm; | |
1274 | struct vm_area_struct *vma, *prev, *cur; | |
1275 | int ret; | |
1276 | struct uffdio_register uffdio_register; | |
1277 | struct uffdio_register __user *user_uffdio_register; | |
1278 | unsigned long vm_flags, new_flags; | |
1279 | bool found; | |
ce53e8e6 | 1280 | bool basic_ioctls; |
86039bd3 AA |
1281 | unsigned long start, end, vma_end; |
1282 | ||
1283 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1284 | ||
1285 | ret = -EFAULT; | |
1286 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1287 | sizeof(uffdio_register)-sizeof(__u64))) | |
1288 | goto out; | |
1289 | ||
1290 | ret = -EINVAL; | |
1291 | if (!uffdio_register.mode) | |
1292 | goto out; | |
1293 | if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING| | |
1294 | UFFDIO_REGISTER_MODE_WP)) | |
1295 | goto out; | |
1296 | vm_flags = 0; | |
1297 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1298 | vm_flags |= VM_UFFD_MISSING; | |
63b2d417 | 1299 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) |
86039bd3 | 1300 | vm_flags |= VM_UFFD_WP; |
86039bd3 | 1301 | |
7d032574 | 1302 | ret = validate_range(mm, &uffdio_register.range.start, |
86039bd3 AA |
1303 | uffdio_register.range.len); |
1304 | if (ret) | |
1305 | goto out; | |
1306 | ||
1307 | start = uffdio_register.range.start; | |
1308 | end = start + uffdio_register.range.len; | |
1309 | ||
d2005e3f ON |
1310 | ret = -ENOMEM; |
1311 | if (!mmget_not_zero(mm)) | |
1312 | goto out; | |
1313 | ||
d8ed45c5 | 1314 | mmap_write_lock(mm); |
86039bd3 | 1315 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1316 | if (!vma) |
1317 | goto out_unlock; | |
1318 | ||
1319 | /* check that there's at least one vma in the range */ | |
1320 | ret = -EINVAL; | |
1321 | if (vma->vm_start >= end) | |
1322 | goto out_unlock; | |
1323 | ||
cab350af MK |
1324 | /* |
1325 | * If the first vma contains huge pages, make sure start address | |
1326 | * is aligned to huge page size. | |
1327 | */ | |
1328 | if (is_vm_hugetlb_page(vma)) { | |
1329 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1330 | ||
1331 | if (start & (vma_hpagesize - 1)) | |
1332 | goto out_unlock; | |
1333 | } | |
1334 | ||
86039bd3 AA |
1335 | /* |
1336 | * Search for not compatible vmas. | |
86039bd3 AA |
1337 | */ |
1338 | found = false; | |
ce53e8e6 | 1339 | basic_ioctls = false; |
86039bd3 AA |
1340 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { |
1341 | cond_resched(); | |
1342 | ||
1343 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1344 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1345 | ||
1346 | /* check not compatible vmas */ | |
1347 | ret = -EINVAL; | |
63b2d417 | 1348 | if (!vma_can_userfault(cur, vm_flags)) |
86039bd3 | 1349 | goto out_unlock; |
29ec9066 AA |
1350 | |
1351 | /* | |
1352 | * UFFDIO_COPY will fill file holes even without | |
1353 | * PROT_WRITE. This check enforces that if this is a | |
1354 | * MAP_SHARED, the process has write permission to the backing | |
1355 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1356 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1357 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1358 | */ | |
1359 | ret = -EPERM; | |
1360 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1361 | goto out_unlock; | |
1362 | ||
cab350af MK |
1363 | /* |
1364 | * If this vma contains ending address, and huge pages | |
1365 | * check alignment. | |
1366 | */ | |
1367 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1368 | end > cur->vm_start) { | |
1369 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1370 | ||
1371 | ret = -EINVAL; | |
1372 | ||
1373 | if (end & (vma_hpagesize - 1)) | |
1374 | goto out_unlock; | |
1375 | } | |
63b2d417 AA |
1376 | if ((vm_flags & VM_UFFD_WP) && !(cur->vm_flags & VM_MAYWRITE)) |
1377 | goto out_unlock; | |
86039bd3 AA |
1378 | |
1379 | /* | |
1380 | * Check that this vma isn't already owned by a | |
1381 | * different userfaultfd. We can't allow more than one | |
1382 | * userfaultfd to own a single vma simultaneously or we | |
1383 | * wouldn't know which one to deliver the userfaults to. | |
1384 | */ | |
1385 | ret = -EBUSY; | |
1386 | if (cur->vm_userfaultfd_ctx.ctx && | |
1387 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1388 | goto out_unlock; | |
1389 | ||
cab350af MK |
1390 | /* |
1391 | * Note vmas containing huge pages | |
1392 | */ | |
ce53e8e6 MR |
1393 | if (is_vm_hugetlb_page(cur)) |
1394 | basic_ioctls = true; | |
cab350af | 1395 | |
86039bd3 AA |
1396 | found = true; |
1397 | } | |
1398 | BUG_ON(!found); | |
1399 | ||
1400 | if (vma->vm_start < start) | |
1401 | prev = vma; | |
1402 | ||
1403 | ret = 0; | |
1404 | do { | |
1405 | cond_resched(); | |
1406 | ||
63b2d417 | 1407 | BUG_ON(!vma_can_userfault(vma, vm_flags)); |
86039bd3 AA |
1408 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1409 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1410 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1411 | |
1412 | /* | |
1413 | * Nothing to do: this vma is already registered into this | |
1414 | * userfaultfd and with the right tracking mode too. | |
1415 | */ | |
1416 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1417 | (vma->vm_flags & vm_flags) == vm_flags) | |
1418 | goto skip; | |
1419 | ||
1420 | if (vma->vm_start > start) | |
1421 | start = vma->vm_start; | |
1422 | vma_end = min(end, vma->vm_end); | |
1423 | ||
9d4678eb AA |
1424 | new_flags = (vma->vm_flags & |
1425 | ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags; | |
86039bd3 AA |
1426 | prev = vma_merge(mm, prev, start, vma_end, new_flags, |
1427 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1428 | vma_policy(vma), | |
1429 | ((struct vm_userfaultfd_ctx){ ctx })); | |
1430 | if (prev) { | |
1431 | vma = prev; | |
1432 | goto next; | |
1433 | } | |
1434 | if (vma->vm_start < start) { | |
1435 | ret = split_vma(mm, vma, start, 1); | |
1436 | if (ret) | |
1437 | break; | |
1438 | } | |
1439 | if (vma->vm_end > end) { | |
1440 | ret = split_vma(mm, vma, end, 0); | |
1441 | if (ret) | |
1442 | break; | |
1443 | } | |
1444 | next: | |
1445 | /* | |
1446 | * In the vma_merge() successful mprotect-like case 8: | |
1447 | * the next vma was merged into the current one and | |
1448 | * the current one has not been updated yet. | |
1449 | */ | |
1450 | vma->vm_flags = new_flags; | |
1451 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
1452 | ||
6dfeaff9 PX |
1453 | if (is_vm_hugetlb_page(vma) && uffd_disable_huge_pmd_share(vma)) |
1454 | hugetlb_unshare_all_pmds(vma); | |
1455 | ||
86039bd3 AA |
1456 | skip: |
1457 | prev = vma; | |
1458 | start = vma->vm_end; | |
1459 | vma = vma->vm_next; | |
1460 | } while (vma && vma->vm_start < end); | |
1461 | out_unlock: | |
d8ed45c5 | 1462 | mmap_write_unlock(mm); |
d2005e3f | 1463 | mmput(mm); |
86039bd3 | 1464 | if (!ret) { |
14819305 PX |
1465 | __u64 ioctls_out; |
1466 | ||
1467 | ioctls_out = basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : | |
1468 | UFFD_API_RANGE_IOCTLS; | |
1469 | ||
1470 | /* | |
1471 | * Declare the WP ioctl only if the WP mode is | |
1472 | * specified and all checks passed with the range | |
1473 | */ | |
1474 | if (!(uffdio_register.mode & UFFDIO_REGISTER_MODE_WP)) | |
1475 | ioctls_out &= ~((__u64)1 << _UFFDIO_WRITEPROTECT); | |
1476 | ||
86039bd3 AA |
1477 | /* |
1478 | * Now that we scanned all vmas we can already tell | |
1479 | * userland which ioctls methods are guaranteed to | |
1480 | * succeed on this range. | |
1481 | */ | |
14819305 | 1482 | if (put_user(ioctls_out, &user_uffdio_register->ioctls)) |
86039bd3 AA |
1483 | ret = -EFAULT; |
1484 | } | |
1485 | out: | |
1486 | return ret; | |
1487 | } | |
1488 | ||
1489 | static int userfaultfd_unregister(struct userfaultfd_ctx *ctx, | |
1490 | unsigned long arg) | |
1491 | { | |
1492 | struct mm_struct *mm = ctx->mm; | |
1493 | struct vm_area_struct *vma, *prev, *cur; | |
1494 | int ret; | |
1495 | struct uffdio_range uffdio_unregister; | |
1496 | unsigned long new_flags; | |
1497 | bool found; | |
1498 | unsigned long start, end, vma_end; | |
1499 | const void __user *buf = (void __user *)arg; | |
1500 | ||
1501 | ret = -EFAULT; | |
1502 | if (copy_from_user(&uffdio_unregister, buf, sizeof(uffdio_unregister))) | |
1503 | goto out; | |
1504 | ||
7d032574 | 1505 | ret = validate_range(mm, &uffdio_unregister.start, |
86039bd3 AA |
1506 | uffdio_unregister.len); |
1507 | if (ret) | |
1508 | goto out; | |
1509 | ||
1510 | start = uffdio_unregister.start; | |
1511 | end = start + uffdio_unregister.len; | |
1512 | ||
d2005e3f ON |
1513 | ret = -ENOMEM; |
1514 | if (!mmget_not_zero(mm)) | |
1515 | goto out; | |
1516 | ||
d8ed45c5 | 1517 | mmap_write_lock(mm); |
86039bd3 | 1518 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1519 | if (!vma) |
1520 | goto out_unlock; | |
1521 | ||
1522 | /* check that there's at least one vma in the range */ | |
1523 | ret = -EINVAL; | |
1524 | if (vma->vm_start >= end) | |
1525 | goto out_unlock; | |
1526 | ||
cab350af MK |
1527 | /* |
1528 | * If the first vma contains huge pages, make sure start address | |
1529 | * is aligned to huge page size. | |
1530 | */ | |
1531 | if (is_vm_hugetlb_page(vma)) { | |
1532 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1533 | ||
1534 | if (start & (vma_hpagesize - 1)) | |
1535 | goto out_unlock; | |
1536 | } | |
1537 | ||
86039bd3 AA |
1538 | /* |
1539 | * Search for not compatible vmas. | |
86039bd3 AA |
1540 | */ |
1541 | found = false; | |
1542 | ret = -EINVAL; | |
1543 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
1544 | cond_resched(); | |
1545 | ||
1546 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1547 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1548 | ||
1549 | /* | |
1550 | * Check not compatible vmas, not strictly required | |
1551 | * here as not compatible vmas cannot have an | |
1552 | * userfaultfd_ctx registered on them, but this | |
1553 | * provides for more strict behavior to notice | |
1554 | * unregistration errors. | |
1555 | */ | |
63b2d417 | 1556 | if (!vma_can_userfault(cur, cur->vm_flags)) |
86039bd3 AA |
1557 | goto out_unlock; |
1558 | ||
1559 | found = true; | |
1560 | } | |
1561 | BUG_ON(!found); | |
1562 | ||
1563 | if (vma->vm_start < start) | |
1564 | prev = vma; | |
1565 | ||
1566 | ret = 0; | |
1567 | do { | |
1568 | cond_resched(); | |
1569 | ||
63b2d417 | 1570 | BUG_ON(!vma_can_userfault(vma, vma->vm_flags)); |
86039bd3 AA |
1571 | |
1572 | /* | |
1573 | * Nothing to do: this vma is already registered into this | |
1574 | * userfaultfd and with the right tracking mode too. | |
1575 | */ | |
1576 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1577 | goto skip; | |
1578 | ||
01e881f5 AA |
1579 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1580 | ||
86039bd3 AA |
1581 | if (vma->vm_start > start) |
1582 | start = vma->vm_start; | |
1583 | vma_end = min(end, vma->vm_end); | |
1584 | ||
09fa5296 AA |
1585 | if (userfaultfd_missing(vma)) { |
1586 | /* | |
1587 | * Wake any concurrent pending userfault while | |
1588 | * we unregister, so they will not hang | |
1589 | * permanently and it avoids userland to call | |
1590 | * UFFDIO_WAKE explicitly. | |
1591 | */ | |
1592 | struct userfaultfd_wake_range range; | |
1593 | range.start = start; | |
1594 | range.len = vma_end - start; | |
1595 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1596 | } | |
1597 | ||
86039bd3 AA |
1598 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); |
1599 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
1600 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1601 | vma_policy(vma), | |
1602 | NULL_VM_UFFD_CTX); | |
1603 | if (prev) { | |
1604 | vma = prev; | |
1605 | goto next; | |
1606 | } | |
1607 | if (vma->vm_start < start) { | |
1608 | ret = split_vma(mm, vma, start, 1); | |
1609 | if (ret) | |
1610 | break; | |
1611 | } | |
1612 | if (vma->vm_end > end) { | |
1613 | ret = split_vma(mm, vma, end, 0); | |
1614 | if (ret) | |
1615 | break; | |
1616 | } | |
1617 | next: | |
1618 | /* | |
1619 | * In the vma_merge() successful mprotect-like case 8: | |
1620 | * the next vma was merged into the current one and | |
1621 | * the current one has not been updated yet. | |
1622 | */ | |
1623 | vma->vm_flags = new_flags; | |
1624 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
1625 | ||
1626 | skip: | |
1627 | prev = vma; | |
1628 | start = vma->vm_end; | |
1629 | vma = vma->vm_next; | |
1630 | } while (vma && vma->vm_start < end); | |
1631 | out_unlock: | |
d8ed45c5 | 1632 | mmap_write_unlock(mm); |
d2005e3f | 1633 | mmput(mm); |
86039bd3 AA |
1634 | out: |
1635 | return ret; | |
1636 | } | |
1637 | ||
1638 | /* | |
ba85c702 AA |
1639 | * userfaultfd_wake may be used in combination with the |
1640 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1641 | */ |
1642 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1643 | unsigned long arg) | |
1644 | { | |
1645 | int ret; | |
1646 | struct uffdio_range uffdio_wake; | |
1647 | struct userfaultfd_wake_range range; | |
1648 | const void __user *buf = (void __user *)arg; | |
1649 | ||
1650 | ret = -EFAULT; | |
1651 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1652 | goto out; | |
1653 | ||
7d032574 | 1654 | ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1655 | if (ret) |
1656 | goto out; | |
1657 | ||
1658 | range.start = uffdio_wake.start; | |
1659 | range.len = uffdio_wake.len; | |
1660 | ||
1661 | /* | |
1662 | * len == 0 means wake all and we don't want to wake all here, | |
1663 | * so check it again to be sure. | |
1664 | */ | |
1665 | VM_BUG_ON(!range.len); | |
1666 | ||
1667 | wake_userfault(ctx, &range); | |
1668 | ret = 0; | |
1669 | ||
1670 | out: | |
1671 | return ret; | |
1672 | } | |
1673 | ||
ad465cae AA |
1674 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1675 | unsigned long arg) | |
1676 | { | |
1677 | __s64 ret; | |
1678 | struct uffdio_copy uffdio_copy; | |
1679 | struct uffdio_copy __user *user_uffdio_copy; | |
1680 | struct userfaultfd_wake_range range; | |
1681 | ||
1682 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1683 | ||
df2cc96e MR |
1684 | ret = -EAGAIN; |
1685 | if (READ_ONCE(ctx->mmap_changing)) | |
1686 | goto out; | |
1687 | ||
ad465cae AA |
1688 | ret = -EFAULT; |
1689 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1690 | /* don't copy "copy" last field */ | |
1691 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1692 | goto out; | |
1693 | ||
7d032574 | 1694 | ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1695 | if (ret) |
1696 | goto out; | |
1697 | /* | |
1698 | * double check for wraparound just in case. copy_from_user() | |
1699 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1700 | * in the userland range. | |
1701 | */ | |
1702 | ret = -EINVAL; | |
1703 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1704 | goto out; | |
72981e0e | 1705 | if (uffdio_copy.mode & ~(UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP)) |
ad465cae | 1706 | goto out; |
d2005e3f ON |
1707 | if (mmget_not_zero(ctx->mm)) { |
1708 | ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src, | |
72981e0e AA |
1709 | uffdio_copy.len, &ctx->mmap_changing, |
1710 | uffdio_copy.mode); | |
d2005e3f | 1711 | mmput(ctx->mm); |
96333187 | 1712 | } else { |
e86b298b | 1713 | return -ESRCH; |
d2005e3f | 1714 | } |
ad465cae AA |
1715 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1716 | return -EFAULT; | |
1717 | if (ret < 0) | |
1718 | goto out; | |
1719 | BUG_ON(!ret); | |
1720 | /* len == 0 would wake all */ | |
1721 | range.len = ret; | |
1722 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1723 | range.start = uffdio_copy.dst; | |
1724 | wake_userfault(ctx, &range); | |
1725 | } | |
1726 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1727 | out: | |
1728 | return ret; | |
1729 | } | |
1730 | ||
1731 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1732 | unsigned long arg) | |
1733 | { | |
1734 | __s64 ret; | |
1735 | struct uffdio_zeropage uffdio_zeropage; | |
1736 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1737 | struct userfaultfd_wake_range range; | |
1738 | ||
1739 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1740 | ||
df2cc96e MR |
1741 | ret = -EAGAIN; |
1742 | if (READ_ONCE(ctx->mmap_changing)) | |
1743 | goto out; | |
1744 | ||
ad465cae AA |
1745 | ret = -EFAULT; |
1746 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1747 | /* don't copy "zeropage" last field */ | |
1748 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1749 | goto out; | |
1750 | ||
7d032574 | 1751 | ret = validate_range(ctx->mm, &uffdio_zeropage.range.start, |
ad465cae AA |
1752 | uffdio_zeropage.range.len); |
1753 | if (ret) | |
1754 | goto out; | |
1755 | ret = -EINVAL; | |
1756 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1757 | goto out; | |
1758 | ||
d2005e3f ON |
1759 | if (mmget_not_zero(ctx->mm)) { |
1760 | ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start, | |
df2cc96e MR |
1761 | uffdio_zeropage.range.len, |
1762 | &ctx->mmap_changing); | |
d2005e3f | 1763 | mmput(ctx->mm); |
9d95aa4b | 1764 | } else { |
e86b298b | 1765 | return -ESRCH; |
d2005e3f | 1766 | } |
ad465cae AA |
1767 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1768 | return -EFAULT; | |
1769 | if (ret < 0) | |
1770 | goto out; | |
1771 | /* len == 0 would wake all */ | |
1772 | BUG_ON(!ret); | |
1773 | range.len = ret; | |
1774 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1775 | range.start = uffdio_zeropage.range.start; | |
1776 | wake_userfault(ctx, &range); | |
1777 | } | |
1778 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1779 | out: | |
1780 | return ret; | |
1781 | } | |
1782 | ||
63b2d417 AA |
1783 | static int userfaultfd_writeprotect(struct userfaultfd_ctx *ctx, |
1784 | unsigned long arg) | |
1785 | { | |
1786 | int ret; | |
1787 | struct uffdio_writeprotect uffdio_wp; | |
1788 | struct uffdio_writeprotect __user *user_uffdio_wp; | |
1789 | struct userfaultfd_wake_range range; | |
23080e27 | 1790 | bool mode_wp, mode_dontwake; |
63b2d417 AA |
1791 | |
1792 | if (READ_ONCE(ctx->mmap_changing)) | |
1793 | return -EAGAIN; | |
1794 | ||
1795 | user_uffdio_wp = (struct uffdio_writeprotect __user *) arg; | |
1796 | ||
1797 | if (copy_from_user(&uffdio_wp, user_uffdio_wp, | |
1798 | sizeof(struct uffdio_writeprotect))) | |
1799 | return -EFAULT; | |
1800 | ||
1801 | ret = validate_range(ctx->mm, &uffdio_wp.range.start, | |
1802 | uffdio_wp.range.len); | |
1803 | if (ret) | |
1804 | return ret; | |
1805 | ||
1806 | if (uffdio_wp.mode & ~(UFFDIO_WRITEPROTECT_MODE_DONTWAKE | | |
1807 | UFFDIO_WRITEPROTECT_MODE_WP)) | |
1808 | return -EINVAL; | |
23080e27 PX |
1809 | |
1810 | mode_wp = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_WP; | |
1811 | mode_dontwake = uffdio_wp.mode & UFFDIO_WRITEPROTECT_MODE_DONTWAKE; | |
1812 | ||
1813 | if (mode_wp && mode_dontwake) | |
63b2d417 AA |
1814 | return -EINVAL; |
1815 | ||
1816 | ret = mwriteprotect_range(ctx->mm, uffdio_wp.range.start, | |
23080e27 | 1817 | uffdio_wp.range.len, mode_wp, |
63b2d417 AA |
1818 | &ctx->mmap_changing); |
1819 | if (ret) | |
1820 | return ret; | |
1821 | ||
23080e27 | 1822 | if (!mode_wp && !mode_dontwake) { |
63b2d417 AA |
1823 | range.start = uffdio_wp.range.start; |
1824 | range.len = uffdio_wp.range.len; | |
1825 | wake_userfault(ctx, &range); | |
1826 | } | |
1827 | return ret; | |
1828 | } | |
1829 | ||
9cd75c3c PE |
1830 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1831 | { | |
1832 | /* | |
1833 | * For the current set of features the bits just coincide | |
1834 | */ | |
1835 | return (unsigned int)user_features; | |
1836 | } | |
1837 | ||
86039bd3 AA |
1838 | /* |
1839 | * userland asks for a certain API version and we return which bits | |
1840 | * and ioctl commands are implemented in this kernel for such API | |
1841 | * version or -EINVAL if unknown. | |
1842 | */ | |
1843 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1844 | unsigned long arg) | |
1845 | { | |
1846 | struct uffdio_api uffdio_api; | |
1847 | void __user *buf = (void __user *)arg; | |
1848 | int ret; | |
65603144 | 1849 | __u64 features; |
86039bd3 AA |
1850 | |
1851 | ret = -EINVAL; | |
1852 | if (ctx->state != UFFD_STATE_WAIT_API) | |
1853 | goto out; | |
1854 | ret = -EFAULT; | |
a9b85f94 | 1855 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1856 | goto out; |
65603144 | 1857 | features = uffdio_api.features; |
3c1c24d9 MR |
1858 | ret = -EINVAL; |
1859 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) | |
1860 | goto err_out; | |
1861 | ret = -EPERM; | |
1862 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
1863 | goto err_out; | |
65603144 AA |
1864 | /* report all available features and ioctls to userland */ |
1865 | uffdio_api.features = UFFD_API_FEATURES; | |
86039bd3 AA |
1866 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1867 | ret = -EFAULT; | |
1868 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1869 | goto out; | |
1870 | ctx->state = UFFD_STATE_RUNNING; | |
65603144 AA |
1871 | /* only enable the requested features for this uffd context */ |
1872 | ctx->features = uffd_ctx_features(features); | |
86039bd3 AA |
1873 | ret = 0; |
1874 | out: | |
1875 | return ret; | |
3c1c24d9 MR |
1876 | err_out: |
1877 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
1878 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1879 | ret = -EFAULT; | |
1880 | goto out; | |
86039bd3 AA |
1881 | } |
1882 | ||
1883 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
1884 | unsigned long arg) | |
1885 | { | |
1886 | int ret = -EINVAL; | |
1887 | struct userfaultfd_ctx *ctx = file->private_data; | |
1888 | ||
e6485a47 AA |
1889 | if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API) |
1890 | return -EINVAL; | |
1891 | ||
86039bd3 AA |
1892 | switch(cmd) { |
1893 | case UFFDIO_API: | |
1894 | ret = userfaultfd_api(ctx, arg); | |
1895 | break; | |
1896 | case UFFDIO_REGISTER: | |
1897 | ret = userfaultfd_register(ctx, arg); | |
1898 | break; | |
1899 | case UFFDIO_UNREGISTER: | |
1900 | ret = userfaultfd_unregister(ctx, arg); | |
1901 | break; | |
1902 | case UFFDIO_WAKE: | |
1903 | ret = userfaultfd_wake(ctx, arg); | |
1904 | break; | |
ad465cae AA |
1905 | case UFFDIO_COPY: |
1906 | ret = userfaultfd_copy(ctx, arg); | |
1907 | break; | |
1908 | case UFFDIO_ZEROPAGE: | |
1909 | ret = userfaultfd_zeropage(ctx, arg); | |
1910 | break; | |
63b2d417 AA |
1911 | case UFFDIO_WRITEPROTECT: |
1912 | ret = userfaultfd_writeprotect(ctx, arg); | |
1913 | break; | |
86039bd3 AA |
1914 | } |
1915 | return ret; | |
1916 | } | |
1917 | ||
1918 | #ifdef CONFIG_PROC_FS | |
1919 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
1920 | { | |
1921 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 1922 | wait_queue_entry_t *wq; |
86039bd3 AA |
1923 | unsigned long pending = 0, total = 0; |
1924 | ||
cbcfa130 | 1925 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 1926 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
1927 | pending++; |
1928 | total++; | |
1929 | } | |
2055da97 | 1930 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
1931 | total++; |
1932 | } | |
cbcfa130 | 1933 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1934 | |
1935 | /* | |
1936 | * If more protocols will be added, there will be all shown | |
1937 | * separated by a space. Like this: | |
1938 | * protocols: aa:... bb:... | |
1939 | */ | |
1940 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 1941 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
1942 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
1943 | } | |
1944 | #endif | |
1945 | ||
1946 | static const struct file_operations userfaultfd_fops = { | |
1947 | #ifdef CONFIG_PROC_FS | |
1948 | .show_fdinfo = userfaultfd_show_fdinfo, | |
1949 | #endif | |
1950 | .release = userfaultfd_release, | |
1951 | .poll = userfaultfd_poll, | |
1952 | .read = userfaultfd_read, | |
1953 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 1954 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
1955 | .llseek = noop_llseek, |
1956 | }; | |
1957 | ||
3004ec9c AA |
1958 | static void init_once_userfaultfd_ctx(void *mem) |
1959 | { | |
1960 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
1961 | ||
1962 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
1963 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 1964 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 1965 | init_waitqueue_head(&ctx->fd_wqh); |
2ca97ac8 | 1966 | seqcount_spinlock_init(&ctx->refile_seq, &ctx->fault_pending_wqh.lock); |
3004ec9c AA |
1967 | } |
1968 | ||
284cd241 | 1969 | SYSCALL_DEFINE1(userfaultfd, int, flags) |
86039bd3 | 1970 | { |
86039bd3 | 1971 | struct userfaultfd_ctx *ctx; |
284cd241 | 1972 | int fd; |
86039bd3 | 1973 | |
d0d4730a LG |
1974 | if (!sysctl_unprivileged_userfaultfd && |
1975 | (flags & UFFD_USER_MODE_ONLY) == 0 && | |
1976 | !capable(CAP_SYS_PTRACE)) { | |
1977 | printk_once(KERN_WARNING "uffd: Set unprivileged_userfaultfd " | |
1978 | "sysctl knob to 1 if kernel faults must be handled " | |
1979 | "without obtaining CAP_SYS_PTRACE capability\n"); | |
cefdca0a | 1980 | return -EPERM; |
d0d4730a | 1981 | } |
cefdca0a | 1982 | |
86039bd3 AA |
1983 | BUG_ON(!current->mm); |
1984 | ||
1985 | /* Check the UFFD_* constants for consistency. */ | |
37cd0575 | 1986 | BUILD_BUG_ON(UFFD_USER_MODE_ONLY & UFFD_SHARED_FCNTL_FLAGS); |
86039bd3 AA |
1987 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); |
1988 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
1989 | ||
37cd0575 | 1990 | if (flags & ~(UFFD_SHARED_FCNTL_FLAGS | UFFD_USER_MODE_ONLY)) |
284cd241 | 1991 | return -EINVAL; |
86039bd3 | 1992 | |
3004ec9c | 1993 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 1994 | if (!ctx) |
284cd241 | 1995 | return -ENOMEM; |
86039bd3 | 1996 | |
ca880420 | 1997 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 1998 | ctx->flags = flags; |
9cd75c3c | 1999 | ctx->features = 0; |
86039bd3 AA |
2000 | ctx->state = UFFD_STATE_WAIT_API; |
2001 | ctx->released = false; | |
df2cc96e | 2002 | ctx->mmap_changing = false; |
86039bd3 AA |
2003 | ctx->mm = current->mm; |
2004 | /* prevent the mm struct to be freed */ | |
f1f10076 | 2005 | mmgrab(ctx->mm); |
86039bd3 | 2006 | |
b537900f DC |
2007 | fd = anon_inode_getfd_secure("[userfaultfd]", &userfaultfd_fops, ctx, |
2008 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS), NULL); | |
284cd241 | 2009 | if (fd < 0) { |
d2005e3f | 2010 | mmdrop(ctx->mm); |
3004ec9c | 2011 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 2012 | } |
86039bd3 | 2013 | return fd; |
86039bd3 | 2014 | } |
3004ec9c AA |
2015 | |
2016 | static int __init userfaultfd_init(void) | |
2017 | { | |
2018 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", | |
2019 | sizeof(struct userfaultfd_ctx), | |
2020 | 0, | |
2021 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
2022 | init_once_userfaultfd_ctx); | |
2023 | return 0; | |
2024 | } | |
2025 | __initcall(userfaultfd_init); |