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 AA |
63 | /* a refile sequence protected by fault_pending_wqh lock */ |
64 | struct seqcount 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 | ||
237 | VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
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 | ||
289 | VM_BUG_ON(!rwsem_is_locked(&mm->mmap_sem)); | |
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 | ||
8d2afd96 AA |
317 | if (pmd_trans_huge(_pmd)) |
318 | goto out; | |
319 | ||
320 | /* | |
321 | * the pmd is stable (as in !pmd_trans_unstable) so we can re-read it | |
322 | * and use the standard pte_offset_map() instead of parsing _pmd. | |
323 | */ | |
324 | pte = pte_offset_map(pmd, address); | |
325 | /* | |
326 | * Lockless access: we're in a wait_event so it's ok if it | |
327 | * changes under us. | |
328 | */ | |
329 | if (pte_none(*pte)) | |
330 | ret = true; | |
331 | pte_unmap(pte); | |
332 | ||
333 | out: | |
334 | return ret; | |
335 | } | |
336 | ||
86039bd3 AA |
337 | /* |
338 | * The locking rules involved in returning VM_FAULT_RETRY depending on | |
339 | * FAULT_FLAG_ALLOW_RETRY, FAULT_FLAG_RETRY_NOWAIT and | |
340 | * FAULT_FLAG_KILLABLE are not straightforward. The "Caution" | |
341 | * recommendation in __lock_page_or_retry is not an understatement. | |
342 | * | |
343 | * If FAULT_FLAG_ALLOW_RETRY is set, the mmap_sem must be released | |
344 | * before returning VM_FAULT_RETRY only if FAULT_FLAG_RETRY_NOWAIT is | |
345 | * not set. | |
346 | * | |
347 | * If FAULT_FLAG_ALLOW_RETRY is set but FAULT_FLAG_KILLABLE is not | |
348 | * set, VM_FAULT_RETRY can still be returned if and only if there are | |
349 | * fatal_signal_pending()s, and the mmap_sem must be released before | |
350 | * returning it. | |
351 | */ | |
2b740303 | 352 | vm_fault_t handle_userfault(struct vm_fault *vmf, unsigned long reason) |
86039bd3 | 353 | { |
82b0f8c3 | 354 | struct mm_struct *mm = vmf->vma->vm_mm; |
86039bd3 AA |
355 | struct userfaultfd_ctx *ctx; |
356 | struct userfaultfd_wait_queue uwq; | |
2b740303 | 357 | vm_fault_t ret = VM_FAULT_SIGBUS; |
dfa37dc3 | 358 | bool must_wait, return_to_userland; |
15a77c6f | 359 | long blocking_state; |
86039bd3 | 360 | |
64c2b203 AA |
361 | /* |
362 | * We don't do userfault handling for the final child pid update. | |
363 | * | |
364 | * We also don't do userfault handling during | |
365 | * coredumping. hugetlbfs has the special | |
366 | * follow_hugetlb_page() to skip missing pages in the | |
367 | * FOLL_DUMP case, anon memory also checks for FOLL_DUMP with | |
368 | * the no_page_table() helper in follow_page_mask(), but the | |
369 | * shmem_vm_ops->fault method is invoked even during | |
370 | * coredumping without mmap_sem and it ends up here. | |
371 | */ | |
372 | if (current->flags & (PF_EXITING|PF_DUMPCORE)) | |
373 | goto out; | |
374 | ||
375 | /* | |
376 | * Coredumping runs without mmap_sem so we can only check that | |
377 | * the mmap_sem is held, if PF_DUMPCORE was not set. | |
378 | */ | |
379 | WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem)); | |
380 | ||
82b0f8c3 | 381 | ctx = vmf->vma->vm_userfaultfd_ctx.ctx; |
86039bd3 | 382 | if (!ctx) |
ba85c702 | 383 | goto out; |
86039bd3 AA |
384 | |
385 | BUG_ON(ctx->mm != mm); | |
386 | ||
387 | VM_BUG_ON(reason & ~(VM_UFFD_MISSING|VM_UFFD_WP)); | |
388 | VM_BUG_ON(!(reason & VM_UFFD_MISSING) ^ !!(reason & VM_UFFD_WP)); | |
389 | ||
2d6d6f5a PS |
390 | if (ctx->features & UFFD_FEATURE_SIGBUS) |
391 | goto out; | |
392 | ||
86039bd3 AA |
393 | /* |
394 | * If it's already released don't get it. This avoids to loop | |
395 | * in __get_user_pages if userfaultfd_release waits on the | |
396 | * caller of handle_userfault to release the mmap_sem. | |
397 | */ | |
6aa7de05 | 398 | if (unlikely(READ_ONCE(ctx->released))) { |
656710a6 AA |
399 | /* |
400 | * Don't return VM_FAULT_SIGBUS in this case, so a non | |
401 | * cooperative manager can close the uffd after the | |
402 | * last UFFDIO_COPY, without risking to trigger an | |
403 | * involuntary SIGBUS if the process was starting the | |
404 | * userfaultfd while the userfaultfd was still armed | |
405 | * (but after the last UFFDIO_COPY). If the uffd | |
406 | * wasn't already closed when the userfault reached | |
407 | * this point, that would normally be solved by | |
408 | * userfaultfd_must_wait returning 'false'. | |
409 | * | |
410 | * If we were to return VM_FAULT_SIGBUS here, the non | |
411 | * cooperative manager would be instead forced to | |
412 | * always call UFFDIO_UNREGISTER before it can safely | |
413 | * close the uffd. | |
414 | */ | |
415 | ret = VM_FAULT_NOPAGE; | |
ba85c702 | 416 | goto out; |
656710a6 | 417 | } |
86039bd3 AA |
418 | |
419 | /* | |
420 | * Check that we can return VM_FAULT_RETRY. | |
421 | * | |
422 | * NOTE: it should become possible to return VM_FAULT_RETRY | |
423 | * even if FAULT_FLAG_TRIED is set without leading to gup() | |
424 | * -EBUSY failures, if the userfaultfd is to be extended for | |
425 | * VM_UFFD_WP tracking and we intend to arm the userfault | |
426 | * without first stopping userland access to the memory. For | |
427 | * VM_UFFD_MISSING userfaults this is enough for now. | |
428 | */ | |
82b0f8c3 | 429 | if (unlikely(!(vmf->flags & FAULT_FLAG_ALLOW_RETRY))) { |
86039bd3 AA |
430 | /* |
431 | * Validate the invariant that nowait must allow retry | |
432 | * to be sure not to return SIGBUS erroneously on | |
433 | * nowait invocations. | |
434 | */ | |
82b0f8c3 | 435 | BUG_ON(vmf->flags & FAULT_FLAG_RETRY_NOWAIT); |
86039bd3 AA |
436 | #ifdef CONFIG_DEBUG_VM |
437 | if (printk_ratelimit()) { | |
438 | printk(KERN_WARNING | |
82b0f8c3 JK |
439 | "FAULT_FLAG_ALLOW_RETRY missing %x\n", |
440 | vmf->flags); | |
86039bd3 AA |
441 | dump_stack(); |
442 | } | |
443 | #endif | |
ba85c702 | 444 | goto out; |
86039bd3 AA |
445 | } |
446 | ||
447 | /* | |
448 | * Handle nowait, not much to do other than tell it to retry | |
449 | * and wait. | |
450 | */ | |
ba85c702 | 451 | ret = VM_FAULT_RETRY; |
82b0f8c3 | 452 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
ba85c702 | 453 | goto out; |
86039bd3 AA |
454 | |
455 | /* take the reference before dropping the mmap_sem */ | |
456 | userfaultfd_ctx_get(ctx); | |
457 | ||
86039bd3 AA |
458 | init_waitqueue_func_entry(&uwq.wq, userfaultfd_wake_function); |
459 | uwq.wq.private = current; | |
9d4ac934 AP |
460 | uwq.msg = userfault_msg(vmf->address, vmf->flags, reason, |
461 | ctx->features); | |
86039bd3 | 462 | uwq.ctx = ctx; |
15a77c6f | 463 | uwq.waken = false; |
86039bd3 | 464 | |
c270a7ee | 465 | return_to_userland = vmf->flags & FAULT_FLAG_INTERRUPTIBLE; |
15a77c6f AA |
466 | blocking_state = return_to_userland ? TASK_INTERRUPTIBLE : |
467 | TASK_KILLABLE; | |
dfa37dc3 | 468 | |
cbcfa130 | 469 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
470 | /* |
471 | * After the __add_wait_queue the uwq is visible to userland | |
472 | * through poll/read(). | |
473 | */ | |
15b726ef AA |
474 | __add_wait_queue(&ctx->fault_pending_wqh, &uwq.wq); |
475 | /* | |
476 | * The smp_mb() after __set_current_state prevents the reads | |
477 | * following the spin_unlock to happen before the list_add in | |
478 | * __add_wait_queue. | |
479 | */ | |
15a77c6f | 480 | set_current_state(blocking_state); |
cbcfa130 | 481 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 482 | |
369cd212 MK |
483 | if (!is_vm_hugetlb_page(vmf->vma)) |
484 | must_wait = userfaultfd_must_wait(ctx, vmf->address, vmf->flags, | |
485 | reason); | |
486 | else | |
7868a208 PA |
487 | must_wait = userfaultfd_huge_must_wait(ctx, vmf->vma, |
488 | vmf->address, | |
369cd212 | 489 | vmf->flags, reason); |
8d2afd96 AA |
490 | up_read(&mm->mmap_sem); |
491 | ||
6aa7de05 | 492 | if (likely(must_wait && !READ_ONCE(ctx->released) && |
dfa37dc3 AA |
493 | (return_to_userland ? !signal_pending(current) : |
494 | !fatal_signal_pending(current)))) { | |
a9a08845 | 495 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
86039bd3 | 496 | schedule(); |
ba85c702 | 497 | ret |= VM_FAULT_MAJOR; |
15a77c6f AA |
498 | |
499 | /* | |
500 | * False wakeups can orginate even from rwsem before | |
501 | * up_read() however userfaults will wait either for a | |
502 | * targeted wakeup on the specific uwq waitqueue from | |
503 | * wake_userfault() or for signals or for uffd | |
504 | * release. | |
505 | */ | |
506 | while (!READ_ONCE(uwq.waken)) { | |
507 | /* | |
508 | * This needs the full smp_store_mb() | |
509 | * guarantee as the state write must be | |
510 | * visible to other CPUs before reading | |
511 | * uwq.waken from other CPUs. | |
512 | */ | |
513 | set_current_state(blocking_state); | |
514 | if (READ_ONCE(uwq.waken) || | |
515 | READ_ONCE(ctx->released) || | |
516 | (return_to_userland ? signal_pending(current) : | |
517 | fatal_signal_pending(current))) | |
518 | break; | |
519 | schedule(); | |
520 | } | |
ba85c702 | 521 | } |
86039bd3 | 522 | |
ba85c702 | 523 | __set_current_state(TASK_RUNNING); |
15b726ef AA |
524 | |
525 | /* | |
526 | * Here we race with the list_del; list_add in | |
527 | * userfaultfd_ctx_read(), however because we don't ever run | |
528 | * list_del_init() to refile across the two lists, the prev | |
529 | * and next pointers will never point to self. list_add also | |
530 | * would never let any of the two pointers to point to | |
531 | * self. So list_empty_careful won't risk to see both pointers | |
532 | * pointing to self at any time during the list refile. The | |
533 | * only case where list_del_init() is called is the full | |
534 | * removal in the wake function and there we don't re-list_add | |
535 | * and it's fine not to block on the spinlock. The uwq on this | |
536 | * kernel stack can be released after the list_del_init. | |
537 | */ | |
2055da97 | 538 | if (!list_empty_careful(&uwq.wq.entry)) { |
cbcfa130 | 539 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
15b726ef AA |
540 | /* |
541 | * No need of list_del_init(), the uwq on the stack | |
542 | * will be freed shortly anyway. | |
543 | */ | |
2055da97 | 544 | list_del(&uwq.wq.entry); |
cbcfa130 | 545 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 546 | } |
86039bd3 AA |
547 | |
548 | /* | |
549 | * ctx may go away after this if the userfault pseudo fd is | |
550 | * already released. | |
551 | */ | |
552 | userfaultfd_ctx_put(ctx); | |
553 | ||
ba85c702 AA |
554 | out: |
555 | return ret; | |
86039bd3 AA |
556 | } |
557 | ||
8c9e7bb7 AA |
558 | static void userfaultfd_event_wait_completion(struct userfaultfd_ctx *ctx, |
559 | struct userfaultfd_wait_queue *ewq) | |
9cd75c3c | 560 | { |
0cbb4b4f AA |
561 | struct userfaultfd_ctx *release_new_ctx; |
562 | ||
9a69a829 AA |
563 | if (WARN_ON_ONCE(current->flags & PF_EXITING)) |
564 | goto out; | |
9cd75c3c PE |
565 | |
566 | ewq->ctx = ctx; | |
567 | init_waitqueue_entry(&ewq->wq, current); | |
0cbb4b4f | 568 | release_new_ctx = NULL; |
9cd75c3c | 569 | |
cbcfa130 | 570 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
571 | /* |
572 | * After the __add_wait_queue the uwq is visible to userland | |
573 | * through poll/read(). | |
574 | */ | |
575 | __add_wait_queue(&ctx->event_wqh, &ewq->wq); | |
576 | for (;;) { | |
577 | set_current_state(TASK_KILLABLE); | |
578 | if (ewq->msg.event == 0) | |
579 | break; | |
6aa7de05 | 580 | if (READ_ONCE(ctx->released) || |
9cd75c3c | 581 | fatal_signal_pending(current)) { |
384632e6 AA |
582 | /* |
583 | * &ewq->wq may be queued in fork_event, but | |
584 | * __remove_wait_queue ignores the head | |
585 | * parameter. It would be a problem if it | |
586 | * didn't. | |
587 | */ | |
9cd75c3c | 588 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); |
7eb76d45 MR |
589 | if (ewq->msg.event == UFFD_EVENT_FORK) { |
590 | struct userfaultfd_ctx *new; | |
591 | ||
592 | new = (struct userfaultfd_ctx *) | |
593 | (unsigned long) | |
594 | ewq->msg.arg.reserved.reserved1; | |
0cbb4b4f | 595 | release_new_ctx = new; |
7eb76d45 | 596 | } |
9cd75c3c PE |
597 | break; |
598 | } | |
599 | ||
cbcfa130 | 600 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 601 | |
a9a08845 | 602 | wake_up_poll(&ctx->fd_wqh, EPOLLIN); |
9cd75c3c PE |
603 | schedule(); |
604 | ||
cbcfa130 | 605 | spin_lock_irq(&ctx->event_wqh.lock); |
9cd75c3c PE |
606 | } |
607 | __set_current_state(TASK_RUNNING); | |
cbcfa130 | 608 | spin_unlock_irq(&ctx->event_wqh.lock); |
9cd75c3c | 609 | |
0cbb4b4f AA |
610 | if (release_new_ctx) { |
611 | struct vm_area_struct *vma; | |
612 | struct mm_struct *mm = release_new_ctx->mm; | |
613 | ||
614 | /* the various vma->vm_userfaultfd_ctx still points to it */ | |
615 | down_write(&mm->mmap_sem); | |
04f5866e AA |
616 | /* no task can run (and in turn coredump) yet */ |
617 | VM_WARN_ON(!mmget_still_valid(mm)); | |
0cbb4b4f | 618 | for (vma = mm->mmap; vma; vma = vma->vm_next) |
31e810aa | 619 | if (vma->vm_userfaultfd_ctx.ctx == release_new_ctx) { |
0cbb4b4f | 620 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; |
31e810aa MR |
621 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); |
622 | } | |
0cbb4b4f AA |
623 | up_write(&mm->mmap_sem); |
624 | ||
625 | userfaultfd_ctx_put(release_new_ctx); | |
626 | } | |
627 | ||
9cd75c3c PE |
628 | /* |
629 | * ctx may go away after this if the userfault pseudo fd is | |
630 | * already released. | |
631 | */ | |
9a69a829 | 632 | out: |
df2cc96e | 633 | WRITE_ONCE(ctx->mmap_changing, false); |
9cd75c3c | 634 | userfaultfd_ctx_put(ctx); |
9cd75c3c PE |
635 | } |
636 | ||
637 | static void userfaultfd_event_complete(struct userfaultfd_ctx *ctx, | |
638 | struct userfaultfd_wait_queue *ewq) | |
639 | { | |
640 | ewq->msg.event = 0; | |
641 | wake_up_locked(&ctx->event_wqh); | |
642 | __remove_wait_queue(&ctx->event_wqh, &ewq->wq); | |
643 | } | |
644 | ||
893e26e6 PE |
645 | int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs) |
646 | { | |
647 | struct userfaultfd_ctx *ctx = NULL, *octx; | |
648 | struct userfaultfd_fork_ctx *fctx; | |
649 | ||
650 | octx = vma->vm_userfaultfd_ctx.ctx; | |
651 | if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) { | |
652 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
653 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
654 | return 0; | |
655 | } | |
656 | ||
657 | list_for_each_entry(fctx, fcs, list) | |
658 | if (fctx->orig == octx) { | |
659 | ctx = fctx->new; | |
660 | break; | |
661 | } | |
662 | ||
663 | if (!ctx) { | |
664 | fctx = kmalloc(sizeof(*fctx), GFP_KERNEL); | |
665 | if (!fctx) | |
666 | return -ENOMEM; | |
667 | ||
668 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); | |
669 | if (!ctx) { | |
670 | kfree(fctx); | |
671 | return -ENOMEM; | |
672 | } | |
673 | ||
ca880420 | 674 | refcount_set(&ctx->refcount, 1); |
893e26e6 PE |
675 | ctx->flags = octx->flags; |
676 | ctx->state = UFFD_STATE_RUNNING; | |
677 | ctx->features = octx->features; | |
678 | ctx->released = false; | |
df2cc96e | 679 | ctx->mmap_changing = false; |
893e26e6 | 680 | ctx->mm = vma->vm_mm; |
00bb31fa | 681 | mmgrab(ctx->mm); |
893e26e6 PE |
682 | |
683 | userfaultfd_ctx_get(octx); | |
df2cc96e | 684 | WRITE_ONCE(octx->mmap_changing, true); |
893e26e6 PE |
685 | fctx->orig = octx; |
686 | fctx->new = ctx; | |
687 | list_add_tail(&fctx->list, fcs); | |
688 | } | |
689 | ||
690 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
691 | return 0; | |
692 | } | |
693 | ||
8c9e7bb7 | 694 | static void dup_fctx(struct userfaultfd_fork_ctx *fctx) |
893e26e6 PE |
695 | { |
696 | struct userfaultfd_ctx *ctx = fctx->orig; | |
697 | struct userfaultfd_wait_queue ewq; | |
698 | ||
699 | msg_init(&ewq.msg); | |
700 | ||
701 | ewq.msg.event = UFFD_EVENT_FORK; | |
702 | ewq.msg.arg.reserved.reserved1 = (unsigned long)fctx->new; | |
703 | ||
8c9e7bb7 | 704 | userfaultfd_event_wait_completion(ctx, &ewq); |
893e26e6 PE |
705 | } |
706 | ||
707 | void dup_userfaultfd_complete(struct list_head *fcs) | |
708 | { | |
893e26e6 PE |
709 | struct userfaultfd_fork_ctx *fctx, *n; |
710 | ||
711 | list_for_each_entry_safe(fctx, n, fcs, list) { | |
8c9e7bb7 | 712 | dup_fctx(fctx); |
893e26e6 PE |
713 | list_del(&fctx->list); |
714 | kfree(fctx); | |
715 | } | |
716 | } | |
717 | ||
72f87654 PE |
718 | void mremap_userfaultfd_prep(struct vm_area_struct *vma, |
719 | struct vm_userfaultfd_ctx *vm_ctx) | |
720 | { | |
721 | struct userfaultfd_ctx *ctx; | |
722 | ||
723 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
3cfd22be PX |
724 | |
725 | if (!ctx) | |
726 | return; | |
727 | ||
728 | if (ctx->features & UFFD_FEATURE_EVENT_REMAP) { | |
72f87654 PE |
729 | vm_ctx->ctx = ctx; |
730 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 731 | WRITE_ONCE(ctx->mmap_changing, true); |
3cfd22be PX |
732 | } else { |
733 | /* Drop uffd context if remap feature not enabled */ | |
734 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
735 | vma->vm_flags &= ~(VM_UFFD_WP | VM_UFFD_MISSING); | |
72f87654 PE |
736 | } |
737 | } | |
738 | ||
90794bf1 | 739 | void mremap_userfaultfd_complete(struct vm_userfaultfd_ctx *vm_ctx, |
72f87654 PE |
740 | unsigned long from, unsigned long to, |
741 | unsigned long len) | |
742 | { | |
90794bf1 | 743 | struct userfaultfd_ctx *ctx = vm_ctx->ctx; |
72f87654 PE |
744 | struct userfaultfd_wait_queue ewq; |
745 | ||
746 | if (!ctx) | |
747 | return; | |
748 | ||
749 | if (to & ~PAGE_MASK) { | |
750 | userfaultfd_ctx_put(ctx); | |
751 | return; | |
752 | } | |
753 | ||
754 | msg_init(&ewq.msg); | |
755 | ||
756 | ewq.msg.event = UFFD_EVENT_REMAP; | |
757 | ewq.msg.arg.remap.from = from; | |
758 | ewq.msg.arg.remap.to = to; | |
759 | ewq.msg.arg.remap.len = len; | |
760 | ||
761 | userfaultfd_event_wait_completion(ctx, &ewq); | |
762 | } | |
763 | ||
70ccb92f | 764 | bool userfaultfd_remove(struct vm_area_struct *vma, |
d811914d | 765 | unsigned long start, unsigned long end) |
05ce7724 PE |
766 | { |
767 | struct mm_struct *mm = vma->vm_mm; | |
768 | struct userfaultfd_ctx *ctx; | |
769 | struct userfaultfd_wait_queue ewq; | |
770 | ||
771 | ctx = vma->vm_userfaultfd_ctx.ctx; | |
d811914d | 772 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_REMOVE)) |
70ccb92f | 773 | return true; |
05ce7724 PE |
774 | |
775 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 776 | WRITE_ONCE(ctx->mmap_changing, true); |
05ce7724 PE |
777 | up_read(&mm->mmap_sem); |
778 | ||
05ce7724 PE |
779 | msg_init(&ewq.msg); |
780 | ||
d811914d MR |
781 | ewq.msg.event = UFFD_EVENT_REMOVE; |
782 | ewq.msg.arg.remove.start = start; | |
783 | ewq.msg.arg.remove.end = end; | |
05ce7724 PE |
784 | |
785 | userfaultfd_event_wait_completion(ctx, &ewq); | |
786 | ||
70ccb92f | 787 | return false; |
05ce7724 PE |
788 | } |
789 | ||
897ab3e0 MR |
790 | static bool has_unmap_ctx(struct userfaultfd_ctx *ctx, struct list_head *unmaps, |
791 | unsigned long start, unsigned long end) | |
792 | { | |
793 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
794 | ||
795 | list_for_each_entry(unmap_ctx, unmaps, list) | |
796 | if (unmap_ctx->ctx == ctx && unmap_ctx->start == start && | |
797 | unmap_ctx->end == end) | |
798 | return true; | |
799 | ||
800 | return false; | |
801 | } | |
802 | ||
803 | int userfaultfd_unmap_prep(struct vm_area_struct *vma, | |
804 | unsigned long start, unsigned long end, | |
805 | struct list_head *unmaps) | |
806 | { | |
807 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) { | |
808 | struct userfaultfd_unmap_ctx *unmap_ctx; | |
809 | struct userfaultfd_ctx *ctx = vma->vm_userfaultfd_ctx.ctx; | |
810 | ||
811 | if (!ctx || !(ctx->features & UFFD_FEATURE_EVENT_UNMAP) || | |
812 | has_unmap_ctx(ctx, unmaps, start, end)) | |
813 | continue; | |
814 | ||
815 | unmap_ctx = kzalloc(sizeof(*unmap_ctx), GFP_KERNEL); | |
816 | if (!unmap_ctx) | |
817 | return -ENOMEM; | |
818 | ||
819 | userfaultfd_ctx_get(ctx); | |
df2cc96e | 820 | WRITE_ONCE(ctx->mmap_changing, true); |
897ab3e0 MR |
821 | unmap_ctx->ctx = ctx; |
822 | unmap_ctx->start = start; | |
823 | unmap_ctx->end = end; | |
824 | list_add_tail(&unmap_ctx->list, unmaps); | |
825 | } | |
826 | ||
827 | return 0; | |
828 | } | |
829 | ||
830 | void userfaultfd_unmap_complete(struct mm_struct *mm, struct list_head *uf) | |
831 | { | |
832 | struct userfaultfd_unmap_ctx *ctx, *n; | |
833 | struct userfaultfd_wait_queue ewq; | |
834 | ||
835 | list_for_each_entry_safe(ctx, n, uf, list) { | |
836 | msg_init(&ewq.msg); | |
837 | ||
838 | ewq.msg.event = UFFD_EVENT_UNMAP; | |
839 | ewq.msg.arg.remove.start = ctx->start; | |
840 | ewq.msg.arg.remove.end = ctx->end; | |
841 | ||
842 | userfaultfd_event_wait_completion(ctx->ctx, &ewq); | |
843 | ||
844 | list_del(&ctx->list); | |
845 | kfree(ctx); | |
846 | } | |
847 | } | |
848 | ||
86039bd3 AA |
849 | static int userfaultfd_release(struct inode *inode, struct file *file) |
850 | { | |
851 | struct userfaultfd_ctx *ctx = file->private_data; | |
852 | struct mm_struct *mm = ctx->mm; | |
853 | struct vm_area_struct *vma, *prev; | |
854 | /* len == 0 means wake all */ | |
855 | struct userfaultfd_wake_range range = { .len = 0, }; | |
856 | unsigned long new_flags; | |
46d0b24c | 857 | bool still_valid; |
86039bd3 | 858 | |
6aa7de05 | 859 | WRITE_ONCE(ctx->released, true); |
86039bd3 | 860 | |
d2005e3f ON |
861 | if (!mmget_not_zero(mm)) |
862 | goto wakeup; | |
863 | ||
86039bd3 AA |
864 | /* |
865 | * Flush page faults out of all CPUs. NOTE: all page faults | |
866 | * must be retried without returning VM_FAULT_SIGBUS if | |
867 | * userfaultfd_ctx_get() succeeds but vma->vma_userfault_ctx | |
868 | * changes while handle_userfault released the mmap_sem. So | |
869 | * it's critical that released is set to true (above), before | |
870 | * taking the mmap_sem for writing. | |
871 | */ | |
872 | down_write(&mm->mmap_sem); | |
46d0b24c | 873 | still_valid = mmget_still_valid(mm); |
86039bd3 AA |
874 | prev = NULL; |
875 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
876 | cond_resched(); | |
877 | BUG_ON(!!vma->vm_userfaultfd_ctx.ctx ^ | |
878 | !!(vma->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
879 | if (vma->vm_userfaultfd_ctx.ctx != ctx) { | |
880 | prev = vma; | |
881 | continue; | |
882 | } | |
883 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); | |
46d0b24c ON |
884 | if (still_valid) { |
885 | prev = vma_merge(mm, prev, vma->vm_start, vma->vm_end, | |
886 | new_flags, vma->anon_vma, | |
887 | vma->vm_file, vma->vm_pgoff, | |
888 | vma_policy(vma), | |
889 | NULL_VM_UFFD_CTX); | |
890 | if (prev) | |
891 | vma = prev; | |
892 | else | |
893 | prev = vma; | |
894 | } | |
86039bd3 AA |
895 | vma->vm_flags = new_flags; |
896 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
897 | } | |
898 | up_write(&mm->mmap_sem); | |
d2005e3f ON |
899 | mmput(mm); |
900 | wakeup: | |
86039bd3 | 901 | /* |
15b726ef | 902 | * After no new page faults can wait on this fault_*wqh, flush |
86039bd3 | 903 | * the last page faults that may have been already waiting on |
15b726ef | 904 | * the fault_*wqh. |
86039bd3 | 905 | */ |
cbcfa130 | 906 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
ac5be6b4 | 907 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, &range); |
c430d1e8 | 908 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, &range); |
cbcfa130 | 909 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 910 | |
5a18b64e MR |
911 | /* Flush pending events that may still wait on event_wqh */ |
912 | wake_up_all(&ctx->event_wqh); | |
913 | ||
a9a08845 | 914 | wake_up_poll(&ctx->fd_wqh, EPOLLHUP); |
86039bd3 AA |
915 | userfaultfd_ctx_put(ctx); |
916 | return 0; | |
917 | } | |
918 | ||
15b726ef | 919 | /* fault_pending_wqh.lock must be hold by the caller */ |
6dcc27fd PE |
920 | static inline struct userfaultfd_wait_queue *find_userfault_in( |
921 | wait_queue_head_t *wqh) | |
86039bd3 | 922 | { |
ac6424b9 | 923 | wait_queue_entry_t *wq; |
15b726ef | 924 | struct userfaultfd_wait_queue *uwq; |
86039bd3 | 925 | |
456a7378 | 926 | lockdep_assert_held(&wqh->lock); |
86039bd3 | 927 | |
15b726ef | 928 | uwq = NULL; |
6dcc27fd | 929 | if (!waitqueue_active(wqh)) |
15b726ef AA |
930 | goto out; |
931 | /* walk in reverse to provide FIFO behavior to read userfaults */ | |
2055da97 | 932 | wq = list_last_entry(&wqh->head, typeof(*wq), entry); |
15b726ef AA |
933 | uwq = container_of(wq, struct userfaultfd_wait_queue, wq); |
934 | out: | |
935 | return uwq; | |
86039bd3 | 936 | } |
6dcc27fd PE |
937 | |
938 | static inline struct userfaultfd_wait_queue *find_userfault( | |
939 | struct userfaultfd_ctx *ctx) | |
940 | { | |
941 | return find_userfault_in(&ctx->fault_pending_wqh); | |
942 | } | |
86039bd3 | 943 | |
9cd75c3c PE |
944 | static inline struct userfaultfd_wait_queue *find_userfault_evt( |
945 | struct userfaultfd_ctx *ctx) | |
946 | { | |
947 | return find_userfault_in(&ctx->event_wqh); | |
948 | } | |
949 | ||
076ccb76 | 950 | static __poll_t userfaultfd_poll(struct file *file, poll_table *wait) |
86039bd3 AA |
951 | { |
952 | struct userfaultfd_ctx *ctx = file->private_data; | |
076ccb76 | 953 | __poll_t ret; |
86039bd3 AA |
954 | |
955 | poll_wait(file, &ctx->fd_wqh, wait); | |
956 | ||
957 | switch (ctx->state) { | |
958 | case UFFD_STATE_WAIT_API: | |
a9a08845 | 959 | return EPOLLERR; |
86039bd3 | 960 | case UFFD_STATE_RUNNING: |
ba85c702 AA |
961 | /* |
962 | * poll() never guarantees that read won't block. | |
963 | * userfaults can be waken before they're read(). | |
964 | */ | |
965 | if (unlikely(!(file->f_flags & O_NONBLOCK))) | |
a9a08845 | 966 | return EPOLLERR; |
15b726ef AA |
967 | /* |
968 | * lockless access to see if there are pending faults | |
969 | * __pollwait last action is the add_wait_queue but | |
970 | * the spin_unlock would allow the waitqueue_active to | |
971 | * pass above the actual list_add inside | |
972 | * add_wait_queue critical section. So use a full | |
973 | * memory barrier to serialize the list_add write of | |
974 | * add_wait_queue() with the waitqueue_active read | |
975 | * below. | |
976 | */ | |
977 | ret = 0; | |
978 | smp_mb(); | |
979 | if (waitqueue_active(&ctx->fault_pending_wqh)) | |
a9a08845 | 980 | ret = EPOLLIN; |
9cd75c3c | 981 | else if (waitqueue_active(&ctx->event_wqh)) |
a9a08845 | 982 | ret = EPOLLIN; |
9cd75c3c | 983 | |
86039bd3 AA |
984 | return ret; |
985 | default: | |
8474901a | 986 | WARN_ON_ONCE(1); |
a9a08845 | 987 | return EPOLLERR; |
86039bd3 AA |
988 | } |
989 | } | |
990 | ||
893e26e6 PE |
991 | static const struct file_operations userfaultfd_fops; |
992 | ||
993 | static int resolve_userfault_fork(struct userfaultfd_ctx *ctx, | |
994 | struct userfaultfd_ctx *new, | |
995 | struct uffd_msg *msg) | |
996 | { | |
997 | int fd; | |
893e26e6 | 998 | |
284cd241 EB |
999 | fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, new, |
1000 | O_RDWR | (new->flags & UFFD_SHARED_FCNTL_FLAGS)); | |
893e26e6 PE |
1001 | if (fd < 0) |
1002 | return fd; | |
1003 | ||
893e26e6 PE |
1004 | msg->arg.reserved.reserved1 = 0; |
1005 | msg->arg.fork.ufd = fd; | |
893e26e6 PE |
1006 | return 0; |
1007 | } | |
1008 | ||
86039bd3 | 1009 | static ssize_t userfaultfd_ctx_read(struct userfaultfd_ctx *ctx, int no_wait, |
a9b85f94 | 1010 | struct uffd_msg *msg) |
86039bd3 AA |
1011 | { |
1012 | ssize_t ret; | |
1013 | DECLARE_WAITQUEUE(wait, current); | |
15b726ef | 1014 | struct userfaultfd_wait_queue *uwq; |
893e26e6 PE |
1015 | /* |
1016 | * Handling fork event requires sleeping operations, so | |
1017 | * we drop the event_wqh lock, then do these ops, then | |
1018 | * lock it back and wake up the waiter. While the lock is | |
1019 | * dropped the ewq may go away so we keep track of it | |
1020 | * carefully. | |
1021 | */ | |
1022 | LIST_HEAD(fork_event); | |
1023 | struct userfaultfd_ctx *fork_nctx = NULL; | |
86039bd3 | 1024 | |
15b726ef | 1025 | /* always take the fd_wqh lock before the fault_pending_wqh lock */ |
ae62c16e | 1026 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1027 | __add_wait_queue(&ctx->fd_wqh, &wait); |
1028 | for (;;) { | |
1029 | set_current_state(TASK_INTERRUPTIBLE); | |
15b726ef AA |
1030 | spin_lock(&ctx->fault_pending_wqh.lock); |
1031 | uwq = find_userfault(ctx); | |
1032 | if (uwq) { | |
2c5b7e1b AA |
1033 | /* |
1034 | * Use a seqcount to repeat the lockless check | |
1035 | * in wake_userfault() to avoid missing | |
1036 | * wakeups because during the refile both | |
1037 | * waitqueue could become empty if this is the | |
1038 | * only userfault. | |
1039 | */ | |
1040 | write_seqcount_begin(&ctx->refile_seq); | |
1041 | ||
86039bd3 | 1042 | /* |
15b726ef AA |
1043 | * The fault_pending_wqh.lock prevents the uwq |
1044 | * to disappear from under us. | |
1045 | * | |
1046 | * Refile this userfault from | |
1047 | * fault_pending_wqh to fault_wqh, it's not | |
1048 | * pending anymore after we read it. | |
1049 | * | |
1050 | * Use list_del() by hand (as | |
1051 | * userfaultfd_wake_function also uses | |
1052 | * list_del_init() by hand) to be sure nobody | |
1053 | * changes __remove_wait_queue() to use | |
1054 | * list_del_init() in turn breaking the | |
1055 | * !list_empty_careful() check in | |
2055da97 | 1056 | * handle_userfault(). The uwq->wq.head list |
15b726ef AA |
1057 | * must never be empty at any time during the |
1058 | * refile, or the waitqueue could disappear | |
1059 | * from under us. The "wait_queue_head_t" | |
1060 | * parameter of __remove_wait_queue() is unused | |
1061 | * anyway. | |
86039bd3 | 1062 | */ |
2055da97 | 1063 | list_del(&uwq->wq.entry); |
c430d1e8 | 1064 | add_wait_queue(&ctx->fault_wqh, &uwq->wq); |
15b726ef | 1065 | |
2c5b7e1b AA |
1066 | write_seqcount_end(&ctx->refile_seq); |
1067 | ||
a9b85f94 AA |
1068 | /* careful to always initialize msg if ret == 0 */ |
1069 | *msg = uwq->msg; | |
15b726ef | 1070 | spin_unlock(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1071 | ret = 0; |
1072 | break; | |
1073 | } | |
15b726ef | 1074 | spin_unlock(&ctx->fault_pending_wqh.lock); |
9cd75c3c PE |
1075 | |
1076 | spin_lock(&ctx->event_wqh.lock); | |
1077 | uwq = find_userfault_evt(ctx); | |
1078 | if (uwq) { | |
1079 | *msg = uwq->msg; | |
1080 | ||
893e26e6 PE |
1081 | if (uwq->msg.event == UFFD_EVENT_FORK) { |
1082 | fork_nctx = (struct userfaultfd_ctx *) | |
1083 | (unsigned long) | |
1084 | uwq->msg.arg.reserved.reserved1; | |
2055da97 | 1085 | list_move(&uwq->wq.entry, &fork_event); |
384632e6 AA |
1086 | /* |
1087 | * fork_nctx can be freed as soon as | |
1088 | * we drop the lock, unless we take a | |
1089 | * reference on it. | |
1090 | */ | |
1091 | userfaultfd_ctx_get(fork_nctx); | |
893e26e6 PE |
1092 | spin_unlock(&ctx->event_wqh.lock); |
1093 | ret = 0; | |
1094 | break; | |
1095 | } | |
1096 | ||
9cd75c3c PE |
1097 | userfaultfd_event_complete(ctx, uwq); |
1098 | spin_unlock(&ctx->event_wqh.lock); | |
1099 | ret = 0; | |
1100 | break; | |
1101 | } | |
1102 | spin_unlock(&ctx->event_wqh.lock); | |
1103 | ||
86039bd3 AA |
1104 | if (signal_pending(current)) { |
1105 | ret = -ERESTARTSYS; | |
1106 | break; | |
1107 | } | |
1108 | if (no_wait) { | |
1109 | ret = -EAGAIN; | |
1110 | break; | |
1111 | } | |
ae62c16e | 1112 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1113 | schedule(); |
ae62c16e | 1114 | spin_lock_irq(&ctx->fd_wqh.lock); |
86039bd3 AA |
1115 | } |
1116 | __remove_wait_queue(&ctx->fd_wqh, &wait); | |
1117 | __set_current_state(TASK_RUNNING); | |
ae62c16e | 1118 | spin_unlock_irq(&ctx->fd_wqh.lock); |
86039bd3 | 1119 | |
893e26e6 PE |
1120 | if (!ret && msg->event == UFFD_EVENT_FORK) { |
1121 | ret = resolve_userfault_fork(ctx, fork_nctx, msg); | |
cbcfa130 | 1122 | spin_lock_irq(&ctx->event_wqh.lock); |
384632e6 AA |
1123 | if (!list_empty(&fork_event)) { |
1124 | /* | |
1125 | * The fork thread didn't abort, so we can | |
1126 | * drop the temporary refcount. | |
1127 | */ | |
1128 | userfaultfd_ctx_put(fork_nctx); | |
1129 | ||
1130 | uwq = list_first_entry(&fork_event, | |
1131 | typeof(*uwq), | |
1132 | wq.entry); | |
1133 | /* | |
1134 | * If fork_event list wasn't empty and in turn | |
1135 | * the event wasn't already released by fork | |
1136 | * (the event is allocated on fork kernel | |
1137 | * stack), put the event back to its place in | |
1138 | * the event_wq. fork_event head will be freed | |
1139 | * as soon as we return so the event cannot | |
1140 | * stay queued there no matter the current | |
1141 | * "ret" value. | |
1142 | */ | |
1143 | list_del(&uwq->wq.entry); | |
1144 | __add_wait_queue(&ctx->event_wqh, &uwq->wq); | |
893e26e6 | 1145 | |
384632e6 AA |
1146 | /* |
1147 | * Leave the event in the waitqueue and report | |
1148 | * error to userland if we failed to resolve | |
1149 | * the userfault fork. | |
1150 | */ | |
1151 | if (likely(!ret)) | |
893e26e6 | 1152 | userfaultfd_event_complete(ctx, uwq); |
384632e6 AA |
1153 | } else { |
1154 | /* | |
1155 | * Here the fork thread aborted and the | |
1156 | * refcount from the fork thread on fork_nctx | |
1157 | * has already been released. We still hold | |
1158 | * the reference we took before releasing the | |
1159 | * lock above. If resolve_userfault_fork | |
1160 | * failed we've to drop it because the | |
1161 | * fork_nctx has to be freed in such case. If | |
1162 | * it succeeded we'll hold it because the new | |
1163 | * uffd references it. | |
1164 | */ | |
1165 | if (ret) | |
1166 | userfaultfd_ctx_put(fork_nctx); | |
893e26e6 | 1167 | } |
cbcfa130 | 1168 | spin_unlock_irq(&ctx->event_wqh.lock); |
893e26e6 PE |
1169 | } |
1170 | ||
86039bd3 AA |
1171 | return ret; |
1172 | } | |
1173 | ||
1174 | static ssize_t userfaultfd_read(struct file *file, char __user *buf, | |
1175 | size_t count, loff_t *ppos) | |
1176 | { | |
1177 | struct userfaultfd_ctx *ctx = file->private_data; | |
1178 | ssize_t _ret, ret = 0; | |
a9b85f94 | 1179 | struct uffd_msg msg; |
86039bd3 AA |
1180 | int no_wait = file->f_flags & O_NONBLOCK; |
1181 | ||
1182 | if (ctx->state == UFFD_STATE_WAIT_API) | |
1183 | return -EINVAL; | |
86039bd3 AA |
1184 | |
1185 | for (;;) { | |
a9b85f94 | 1186 | if (count < sizeof(msg)) |
86039bd3 | 1187 | return ret ? ret : -EINVAL; |
a9b85f94 | 1188 | _ret = userfaultfd_ctx_read(ctx, no_wait, &msg); |
86039bd3 AA |
1189 | if (_ret < 0) |
1190 | return ret ? ret : _ret; | |
a9b85f94 | 1191 | if (copy_to_user((__u64 __user *) buf, &msg, sizeof(msg))) |
86039bd3 | 1192 | return ret ? ret : -EFAULT; |
a9b85f94 AA |
1193 | ret += sizeof(msg); |
1194 | buf += sizeof(msg); | |
1195 | count -= sizeof(msg); | |
86039bd3 AA |
1196 | /* |
1197 | * Allow to read more than one fault at time but only | |
1198 | * block if waiting for the very first one. | |
1199 | */ | |
1200 | no_wait = O_NONBLOCK; | |
1201 | } | |
1202 | } | |
1203 | ||
1204 | static void __wake_userfault(struct userfaultfd_ctx *ctx, | |
1205 | struct userfaultfd_wake_range *range) | |
1206 | { | |
cbcfa130 | 1207 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 | 1208 | /* wake all in the range and autoremove */ |
15b726ef | 1209 | if (waitqueue_active(&ctx->fault_pending_wqh)) |
ac5be6b4 | 1210 | __wake_up_locked_key(&ctx->fault_pending_wqh, TASK_NORMAL, |
15b726ef AA |
1211 | range); |
1212 | if (waitqueue_active(&ctx->fault_wqh)) | |
c430d1e8 | 1213 | __wake_up(&ctx->fault_wqh, TASK_NORMAL, 1, range); |
cbcfa130 | 1214 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1215 | } |
1216 | ||
1217 | static __always_inline void wake_userfault(struct userfaultfd_ctx *ctx, | |
1218 | struct userfaultfd_wake_range *range) | |
1219 | { | |
2c5b7e1b AA |
1220 | unsigned seq; |
1221 | bool need_wakeup; | |
1222 | ||
86039bd3 AA |
1223 | /* |
1224 | * To be sure waitqueue_active() is not reordered by the CPU | |
1225 | * before the pagetable update, use an explicit SMP memory | |
1226 | * barrier here. PT lock release or up_read(mmap_sem) still | |
1227 | * have release semantics that can allow the | |
1228 | * waitqueue_active() to be reordered before the pte update. | |
1229 | */ | |
1230 | smp_mb(); | |
1231 | ||
1232 | /* | |
1233 | * Use waitqueue_active because it's very frequent to | |
1234 | * change the address space atomically even if there are no | |
1235 | * userfaults yet. So we take the spinlock only when we're | |
1236 | * sure we've userfaults to wake. | |
1237 | */ | |
2c5b7e1b AA |
1238 | do { |
1239 | seq = read_seqcount_begin(&ctx->refile_seq); | |
1240 | need_wakeup = waitqueue_active(&ctx->fault_pending_wqh) || | |
1241 | waitqueue_active(&ctx->fault_wqh); | |
1242 | cond_resched(); | |
1243 | } while (read_seqcount_retry(&ctx->refile_seq, seq)); | |
1244 | if (need_wakeup) | |
86039bd3 AA |
1245 | __wake_userfault(ctx, range); |
1246 | } | |
1247 | ||
1248 | static __always_inline int validate_range(struct mm_struct *mm, | |
7d032574 | 1249 | __u64 *start, __u64 len) |
86039bd3 AA |
1250 | { |
1251 | __u64 task_size = mm->task_size; | |
1252 | ||
7d032574 AK |
1253 | *start = untagged_addr(*start); |
1254 | ||
1255 | if (*start & ~PAGE_MASK) | |
86039bd3 AA |
1256 | return -EINVAL; |
1257 | if (len & ~PAGE_MASK) | |
1258 | return -EINVAL; | |
1259 | if (!len) | |
1260 | return -EINVAL; | |
7d032574 | 1261 | if (*start < mmap_min_addr) |
86039bd3 | 1262 | return -EINVAL; |
7d032574 | 1263 | if (*start >= task_size) |
86039bd3 | 1264 | return -EINVAL; |
7d032574 | 1265 | if (len > task_size - *start) |
86039bd3 AA |
1266 | return -EINVAL; |
1267 | return 0; | |
1268 | } | |
1269 | ||
ba6907db MR |
1270 | static inline bool vma_can_userfault(struct vm_area_struct *vma) |
1271 | { | |
cac67329 MR |
1272 | return vma_is_anonymous(vma) || is_vm_hugetlb_page(vma) || |
1273 | vma_is_shmem(vma); | |
ba6907db MR |
1274 | } |
1275 | ||
86039bd3 AA |
1276 | static int userfaultfd_register(struct userfaultfd_ctx *ctx, |
1277 | unsigned long arg) | |
1278 | { | |
1279 | struct mm_struct *mm = ctx->mm; | |
1280 | struct vm_area_struct *vma, *prev, *cur; | |
1281 | int ret; | |
1282 | struct uffdio_register uffdio_register; | |
1283 | struct uffdio_register __user *user_uffdio_register; | |
1284 | unsigned long vm_flags, new_flags; | |
1285 | bool found; | |
ce53e8e6 | 1286 | bool basic_ioctls; |
86039bd3 AA |
1287 | unsigned long start, end, vma_end; |
1288 | ||
1289 | user_uffdio_register = (struct uffdio_register __user *) arg; | |
1290 | ||
1291 | ret = -EFAULT; | |
1292 | if (copy_from_user(&uffdio_register, user_uffdio_register, | |
1293 | sizeof(uffdio_register)-sizeof(__u64))) | |
1294 | goto out; | |
1295 | ||
1296 | ret = -EINVAL; | |
1297 | if (!uffdio_register.mode) | |
1298 | goto out; | |
1299 | if (uffdio_register.mode & ~(UFFDIO_REGISTER_MODE_MISSING| | |
1300 | UFFDIO_REGISTER_MODE_WP)) | |
1301 | goto out; | |
1302 | vm_flags = 0; | |
1303 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_MISSING) | |
1304 | vm_flags |= VM_UFFD_MISSING; | |
1305 | if (uffdio_register.mode & UFFDIO_REGISTER_MODE_WP) { | |
1306 | vm_flags |= VM_UFFD_WP; | |
1307 | /* | |
1308 | * FIXME: remove the below error constraint by | |
1309 | * implementing the wprotect tracking mode. | |
1310 | */ | |
1311 | ret = -EINVAL; | |
1312 | goto out; | |
1313 | } | |
1314 | ||
7d032574 | 1315 | ret = validate_range(mm, &uffdio_register.range.start, |
86039bd3 AA |
1316 | uffdio_register.range.len); |
1317 | if (ret) | |
1318 | goto out; | |
1319 | ||
1320 | start = uffdio_register.range.start; | |
1321 | end = start + uffdio_register.range.len; | |
1322 | ||
d2005e3f ON |
1323 | ret = -ENOMEM; |
1324 | if (!mmget_not_zero(mm)) | |
1325 | goto out; | |
1326 | ||
86039bd3 | 1327 | down_write(&mm->mmap_sem); |
04f5866e AA |
1328 | if (!mmget_still_valid(mm)) |
1329 | goto out_unlock; | |
86039bd3 | 1330 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1331 | if (!vma) |
1332 | goto out_unlock; | |
1333 | ||
1334 | /* check that there's at least one vma in the range */ | |
1335 | ret = -EINVAL; | |
1336 | if (vma->vm_start >= end) | |
1337 | goto out_unlock; | |
1338 | ||
cab350af MK |
1339 | /* |
1340 | * If the first vma contains huge pages, make sure start address | |
1341 | * is aligned to huge page size. | |
1342 | */ | |
1343 | if (is_vm_hugetlb_page(vma)) { | |
1344 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1345 | ||
1346 | if (start & (vma_hpagesize - 1)) | |
1347 | goto out_unlock; | |
1348 | } | |
1349 | ||
86039bd3 AA |
1350 | /* |
1351 | * Search for not compatible vmas. | |
86039bd3 AA |
1352 | */ |
1353 | found = false; | |
ce53e8e6 | 1354 | basic_ioctls = false; |
86039bd3 AA |
1355 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { |
1356 | cond_resched(); | |
1357 | ||
1358 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1359 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1360 | ||
1361 | /* check not compatible vmas */ | |
1362 | ret = -EINVAL; | |
ba6907db | 1363 | if (!vma_can_userfault(cur)) |
86039bd3 | 1364 | goto out_unlock; |
29ec9066 AA |
1365 | |
1366 | /* | |
1367 | * UFFDIO_COPY will fill file holes even without | |
1368 | * PROT_WRITE. This check enforces that if this is a | |
1369 | * MAP_SHARED, the process has write permission to the backing | |
1370 | * file. If VM_MAYWRITE is set it also enforces that on a | |
1371 | * MAP_SHARED vma: there is no F_WRITE_SEAL and no further | |
1372 | * F_WRITE_SEAL can be taken until the vma is destroyed. | |
1373 | */ | |
1374 | ret = -EPERM; | |
1375 | if (unlikely(!(cur->vm_flags & VM_MAYWRITE))) | |
1376 | goto out_unlock; | |
1377 | ||
cab350af MK |
1378 | /* |
1379 | * If this vma contains ending address, and huge pages | |
1380 | * check alignment. | |
1381 | */ | |
1382 | if (is_vm_hugetlb_page(cur) && end <= cur->vm_end && | |
1383 | end > cur->vm_start) { | |
1384 | unsigned long vma_hpagesize = vma_kernel_pagesize(cur); | |
1385 | ||
1386 | ret = -EINVAL; | |
1387 | ||
1388 | if (end & (vma_hpagesize - 1)) | |
1389 | goto out_unlock; | |
1390 | } | |
86039bd3 AA |
1391 | |
1392 | /* | |
1393 | * Check that this vma isn't already owned by a | |
1394 | * different userfaultfd. We can't allow more than one | |
1395 | * userfaultfd to own a single vma simultaneously or we | |
1396 | * wouldn't know which one to deliver the userfaults to. | |
1397 | */ | |
1398 | ret = -EBUSY; | |
1399 | if (cur->vm_userfaultfd_ctx.ctx && | |
1400 | cur->vm_userfaultfd_ctx.ctx != ctx) | |
1401 | goto out_unlock; | |
1402 | ||
cab350af MK |
1403 | /* |
1404 | * Note vmas containing huge pages | |
1405 | */ | |
ce53e8e6 MR |
1406 | if (is_vm_hugetlb_page(cur)) |
1407 | basic_ioctls = true; | |
cab350af | 1408 | |
86039bd3 AA |
1409 | found = true; |
1410 | } | |
1411 | BUG_ON(!found); | |
1412 | ||
1413 | if (vma->vm_start < start) | |
1414 | prev = vma; | |
1415 | ||
1416 | ret = 0; | |
1417 | do { | |
1418 | cond_resched(); | |
1419 | ||
ba6907db | 1420 | BUG_ON(!vma_can_userfault(vma)); |
86039bd3 AA |
1421 | BUG_ON(vma->vm_userfaultfd_ctx.ctx && |
1422 | vma->vm_userfaultfd_ctx.ctx != ctx); | |
29ec9066 | 1423 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
86039bd3 AA |
1424 | |
1425 | /* | |
1426 | * Nothing to do: this vma is already registered into this | |
1427 | * userfaultfd and with the right tracking mode too. | |
1428 | */ | |
1429 | if (vma->vm_userfaultfd_ctx.ctx == ctx && | |
1430 | (vma->vm_flags & vm_flags) == vm_flags) | |
1431 | goto skip; | |
1432 | ||
1433 | if (vma->vm_start > start) | |
1434 | start = vma->vm_start; | |
1435 | vma_end = min(end, vma->vm_end); | |
1436 | ||
9d4678eb AA |
1437 | new_flags = (vma->vm_flags & |
1438 | ~(VM_UFFD_MISSING|VM_UFFD_WP)) | vm_flags; | |
86039bd3 AA |
1439 | prev = vma_merge(mm, prev, start, vma_end, new_flags, |
1440 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1441 | vma_policy(vma), | |
1442 | ((struct vm_userfaultfd_ctx){ ctx })); | |
1443 | if (prev) { | |
1444 | vma = prev; | |
1445 | goto next; | |
1446 | } | |
1447 | if (vma->vm_start < start) { | |
1448 | ret = split_vma(mm, vma, start, 1); | |
1449 | if (ret) | |
1450 | break; | |
1451 | } | |
1452 | if (vma->vm_end > end) { | |
1453 | ret = split_vma(mm, vma, end, 0); | |
1454 | if (ret) | |
1455 | break; | |
1456 | } | |
1457 | next: | |
1458 | /* | |
1459 | * In the vma_merge() successful mprotect-like case 8: | |
1460 | * the next vma was merged into the current one and | |
1461 | * the current one has not been updated yet. | |
1462 | */ | |
1463 | vma->vm_flags = new_flags; | |
1464 | vma->vm_userfaultfd_ctx.ctx = ctx; | |
1465 | ||
1466 | skip: | |
1467 | prev = vma; | |
1468 | start = vma->vm_end; | |
1469 | vma = vma->vm_next; | |
1470 | } while (vma && vma->vm_start < end); | |
1471 | out_unlock: | |
1472 | up_write(&mm->mmap_sem); | |
d2005e3f | 1473 | mmput(mm); |
86039bd3 AA |
1474 | if (!ret) { |
1475 | /* | |
1476 | * Now that we scanned all vmas we can already tell | |
1477 | * userland which ioctls methods are guaranteed to | |
1478 | * succeed on this range. | |
1479 | */ | |
ce53e8e6 | 1480 | if (put_user(basic_ioctls ? UFFD_API_RANGE_IOCTLS_BASIC : |
cab350af | 1481 | UFFD_API_RANGE_IOCTLS, |
86039bd3 AA |
1482 | &user_uffdio_register->ioctls)) |
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 | ||
86039bd3 | 1517 | down_write(&mm->mmap_sem); |
04f5866e AA |
1518 | if (!mmget_still_valid(mm)) |
1519 | goto out_unlock; | |
86039bd3 | 1520 | vma = find_vma_prev(mm, start, &prev); |
86039bd3 AA |
1521 | if (!vma) |
1522 | goto out_unlock; | |
1523 | ||
1524 | /* check that there's at least one vma in the range */ | |
1525 | ret = -EINVAL; | |
1526 | if (vma->vm_start >= end) | |
1527 | goto out_unlock; | |
1528 | ||
cab350af MK |
1529 | /* |
1530 | * If the first vma contains huge pages, make sure start address | |
1531 | * is aligned to huge page size. | |
1532 | */ | |
1533 | if (is_vm_hugetlb_page(vma)) { | |
1534 | unsigned long vma_hpagesize = vma_kernel_pagesize(vma); | |
1535 | ||
1536 | if (start & (vma_hpagesize - 1)) | |
1537 | goto out_unlock; | |
1538 | } | |
1539 | ||
86039bd3 AA |
1540 | /* |
1541 | * Search for not compatible vmas. | |
86039bd3 AA |
1542 | */ |
1543 | found = false; | |
1544 | ret = -EINVAL; | |
1545 | for (cur = vma; cur && cur->vm_start < end; cur = cur->vm_next) { | |
1546 | cond_resched(); | |
1547 | ||
1548 | BUG_ON(!!cur->vm_userfaultfd_ctx.ctx ^ | |
1549 | !!(cur->vm_flags & (VM_UFFD_MISSING | VM_UFFD_WP))); | |
1550 | ||
1551 | /* | |
1552 | * Check not compatible vmas, not strictly required | |
1553 | * here as not compatible vmas cannot have an | |
1554 | * userfaultfd_ctx registered on them, but this | |
1555 | * provides for more strict behavior to notice | |
1556 | * unregistration errors. | |
1557 | */ | |
ba6907db | 1558 | if (!vma_can_userfault(cur)) |
86039bd3 AA |
1559 | goto out_unlock; |
1560 | ||
1561 | found = true; | |
1562 | } | |
1563 | BUG_ON(!found); | |
1564 | ||
1565 | if (vma->vm_start < start) | |
1566 | prev = vma; | |
1567 | ||
1568 | ret = 0; | |
1569 | do { | |
1570 | cond_resched(); | |
1571 | ||
ba6907db | 1572 | BUG_ON(!vma_can_userfault(vma)); |
86039bd3 AA |
1573 | |
1574 | /* | |
1575 | * Nothing to do: this vma is already registered into this | |
1576 | * userfaultfd and with the right tracking mode too. | |
1577 | */ | |
1578 | if (!vma->vm_userfaultfd_ctx.ctx) | |
1579 | goto skip; | |
1580 | ||
01e881f5 AA |
1581 | WARN_ON(!(vma->vm_flags & VM_MAYWRITE)); |
1582 | ||
86039bd3 AA |
1583 | if (vma->vm_start > start) |
1584 | start = vma->vm_start; | |
1585 | vma_end = min(end, vma->vm_end); | |
1586 | ||
09fa5296 AA |
1587 | if (userfaultfd_missing(vma)) { |
1588 | /* | |
1589 | * Wake any concurrent pending userfault while | |
1590 | * we unregister, so they will not hang | |
1591 | * permanently and it avoids userland to call | |
1592 | * UFFDIO_WAKE explicitly. | |
1593 | */ | |
1594 | struct userfaultfd_wake_range range; | |
1595 | range.start = start; | |
1596 | range.len = vma_end - start; | |
1597 | wake_userfault(vma->vm_userfaultfd_ctx.ctx, &range); | |
1598 | } | |
1599 | ||
86039bd3 AA |
1600 | new_flags = vma->vm_flags & ~(VM_UFFD_MISSING | VM_UFFD_WP); |
1601 | prev = vma_merge(mm, prev, start, vma_end, new_flags, | |
1602 | vma->anon_vma, vma->vm_file, vma->vm_pgoff, | |
1603 | vma_policy(vma), | |
1604 | NULL_VM_UFFD_CTX); | |
1605 | if (prev) { | |
1606 | vma = prev; | |
1607 | goto next; | |
1608 | } | |
1609 | if (vma->vm_start < start) { | |
1610 | ret = split_vma(mm, vma, start, 1); | |
1611 | if (ret) | |
1612 | break; | |
1613 | } | |
1614 | if (vma->vm_end > end) { | |
1615 | ret = split_vma(mm, vma, end, 0); | |
1616 | if (ret) | |
1617 | break; | |
1618 | } | |
1619 | next: | |
1620 | /* | |
1621 | * In the vma_merge() successful mprotect-like case 8: | |
1622 | * the next vma was merged into the current one and | |
1623 | * the current one has not been updated yet. | |
1624 | */ | |
1625 | vma->vm_flags = new_flags; | |
1626 | vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX; | |
1627 | ||
1628 | skip: | |
1629 | prev = vma; | |
1630 | start = vma->vm_end; | |
1631 | vma = vma->vm_next; | |
1632 | } while (vma && vma->vm_start < end); | |
1633 | out_unlock: | |
1634 | up_write(&mm->mmap_sem); | |
d2005e3f | 1635 | mmput(mm); |
86039bd3 AA |
1636 | out: |
1637 | return ret; | |
1638 | } | |
1639 | ||
1640 | /* | |
ba85c702 AA |
1641 | * userfaultfd_wake may be used in combination with the |
1642 | * UFFDIO_*_MODE_DONTWAKE to wakeup userfaults in batches. | |
86039bd3 AA |
1643 | */ |
1644 | static int userfaultfd_wake(struct userfaultfd_ctx *ctx, | |
1645 | unsigned long arg) | |
1646 | { | |
1647 | int ret; | |
1648 | struct uffdio_range uffdio_wake; | |
1649 | struct userfaultfd_wake_range range; | |
1650 | const void __user *buf = (void __user *)arg; | |
1651 | ||
1652 | ret = -EFAULT; | |
1653 | if (copy_from_user(&uffdio_wake, buf, sizeof(uffdio_wake))) | |
1654 | goto out; | |
1655 | ||
7d032574 | 1656 | ret = validate_range(ctx->mm, &uffdio_wake.start, uffdio_wake.len); |
86039bd3 AA |
1657 | if (ret) |
1658 | goto out; | |
1659 | ||
1660 | range.start = uffdio_wake.start; | |
1661 | range.len = uffdio_wake.len; | |
1662 | ||
1663 | /* | |
1664 | * len == 0 means wake all and we don't want to wake all here, | |
1665 | * so check it again to be sure. | |
1666 | */ | |
1667 | VM_BUG_ON(!range.len); | |
1668 | ||
1669 | wake_userfault(ctx, &range); | |
1670 | ret = 0; | |
1671 | ||
1672 | out: | |
1673 | return ret; | |
1674 | } | |
1675 | ||
ad465cae AA |
1676 | static int userfaultfd_copy(struct userfaultfd_ctx *ctx, |
1677 | unsigned long arg) | |
1678 | { | |
1679 | __s64 ret; | |
1680 | struct uffdio_copy uffdio_copy; | |
1681 | struct uffdio_copy __user *user_uffdio_copy; | |
1682 | struct userfaultfd_wake_range range; | |
1683 | ||
1684 | user_uffdio_copy = (struct uffdio_copy __user *) arg; | |
1685 | ||
df2cc96e MR |
1686 | ret = -EAGAIN; |
1687 | if (READ_ONCE(ctx->mmap_changing)) | |
1688 | goto out; | |
1689 | ||
ad465cae AA |
1690 | ret = -EFAULT; |
1691 | if (copy_from_user(&uffdio_copy, user_uffdio_copy, | |
1692 | /* don't copy "copy" last field */ | |
1693 | sizeof(uffdio_copy)-sizeof(__s64))) | |
1694 | goto out; | |
1695 | ||
7d032574 | 1696 | ret = validate_range(ctx->mm, &uffdio_copy.dst, uffdio_copy.len); |
ad465cae AA |
1697 | if (ret) |
1698 | goto out; | |
1699 | /* | |
1700 | * double check for wraparound just in case. copy_from_user() | |
1701 | * will later check uffdio_copy.src + uffdio_copy.len to fit | |
1702 | * in the userland range. | |
1703 | */ | |
1704 | ret = -EINVAL; | |
1705 | if (uffdio_copy.src + uffdio_copy.len <= uffdio_copy.src) | |
1706 | goto out; | |
1707 | if (uffdio_copy.mode & ~UFFDIO_COPY_MODE_DONTWAKE) | |
1708 | goto out; | |
d2005e3f ON |
1709 | if (mmget_not_zero(ctx->mm)) { |
1710 | ret = mcopy_atomic(ctx->mm, uffdio_copy.dst, uffdio_copy.src, | |
df2cc96e | 1711 | uffdio_copy.len, &ctx->mmap_changing); |
d2005e3f | 1712 | mmput(ctx->mm); |
96333187 | 1713 | } else { |
e86b298b | 1714 | return -ESRCH; |
d2005e3f | 1715 | } |
ad465cae AA |
1716 | if (unlikely(put_user(ret, &user_uffdio_copy->copy))) |
1717 | return -EFAULT; | |
1718 | if (ret < 0) | |
1719 | goto out; | |
1720 | BUG_ON(!ret); | |
1721 | /* len == 0 would wake all */ | |
1722 | range.len = ret; | |
1723 | if (!(uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE)) { | |
1724 | range.start = uffdio_copy.dst; | |
1725 | wake_userfault(ctx, &range); | |
1726 | } | |
1727 | ret = range.len == uffdio_copy.len ? 0 : -EAGAIN; | |
1728 | out: | |
1729 | return ret; | |
1730 | } | |
1731 | ||
1732 | static int userfaultfd_zeropage(struct userfaultfd_ctx *ctx, | |
1733 | unsigned long arg) | |
1734 | { | |
1735 | __s64 ret; | |
1736 | struct uffdio_zeropage uffdio_zeropage; | |
1737 | struct uffdio_zeropage __user *user_uffdio_zeropage; | |
1738 | struct userfaultfd_wake_range range; | |
1739 | ||
1740 | user_uffdio_zeropage = (struct uffdio_zeropage __user *) arg; | |
1741 | ||
df2cc96e MR |
1742 | ret = -EAGAIN; |
1743 | if (READ_ONCE(ctx->mmap_changing)) | |
1744 | goto out; | |
1745 | ||
ad465cae AA |
1746 | ret = -EFAULT; |
1747 | if (copy_from_user(&uffdio_zeropage, user_uffdio_zeropage, | |
1748 | /* don't copy "zeropage" last field */ | |
1749 | sizeof(uffdio_zeropage)-sizeof(__s64))) | |
1750 | goto out; | |
1751 | ||
7d032574 | 1752 | ret = validate_range(ctx->mm, &uffdio_zeropage.range.start, |
ad465cae AA |
1753 | uffdio_zeropage.range.len); |
1754 | if (ret) | |
1755 | goto out; | |
1756 | ret = -EINVAL; | |
1757 | if (uffdio_zeropage.mode & ~UFFDIO_ZEROPAGE_MODE_DONTWAKE) | |
1758 | goto out; | |
1759 | ||
d2005e3f ON |
1760 | if (mmget_not_zero(ctx->mm)) { |
1761 | ret = mfill_zeropage(ctx->mm, uffdio_zeropage.range.start, | |
df2cc96e MR |
1762 | uffdio_zeropage.range.len, |
1763 | &ctx->mmap_changing); | |
d2005e3f | 1764 | mmput(ctx->mm); |
9d95aa4b | 1765 | } else { |
e86b298b | 1766 | return -ESRCH; |
d2005e3f | 1767 | } |
ad465cae AA |
1768 | if (unlikely(put_user(ret, &user_uffdio_zeropage->zeropage))) |
1769 | return -EFAULT; | |
1770 | if (ret < 0) | |
1771 | goto out; | |
1772 | /* len == 0 would wake all */ | |
1773 | BUG_ON(!ret); | |
1774 | range.len = ret; | |
1775 | if (!(uffdio_zeropage.mode & UFFDIO_ZEROPAGE_MODE_DONTWAKE)) { | |
1776 | range.start = uffdio_zeropage.range.start; | |
1777 | wake_userfault(ctx, &range); | |
1778 | } | |
1779 | ret = range.len == uffdio_zeropage.range.len ? 0 : -EAGAIN; | |
1780 | out: | |
1781 | return ret; | |
1782 | } | |
1783 | ||
9cd75c3c PE |
1784 | static inline unsigned int uffd_ctx_features(__u64 user_features) |
1785 | { | |
1786 | /* | |
1787 | * For the current set of features the bits just coincide | |
1788 | */ | |
1789 | return (unsigned int)user_features; | |
1790 | } | |
1791 | ||
86039bd3 AA |
1792 | /* |
1793 | * userland asks for a certain API version and we return which bits | |
1794 | * and ioctl commands are implemented in this kernel for such API | |
1795 | * version or -EINVAL if unknown. | |
1796 | */ | |
1797 | static int userfaultfd_api(struct userfaultfd_ctx *ctx, | |
1798 | unsigned long arg) | |
1799 | { | |
1800 | struct uffdio_api uffdio_api; | |
1801 | void __user *buf = (void __user *)arg; | |
1802 | int ret; | |
65603144 | 1803 | __u64 features; |
86039bd3 AA |
1804 | |
1805 | ret = -EINVAL; | |
1806 | if (ctx->state != UFFD_STATE_WAIT_API) | |
1807 | goto out; | |
1808 | ret = -EFAULT; | |
a9b85f94 | 1809 | if (copy_from_user(&uffdio_api, buf, sizeof(uffdio_api))) |
86039bd3 | 1810 | goto out; |
65603144 | 1811 | features = uffdio_api.features; |
3c1c24d9 MR |
1812 | ret = -EINVAL; |
1813 | if (uffdio_api.api != UFFD_API || (features & ~UFFD_API_FEATURES)) | |
1814 | goto err_out; | |
1815 | ret = -EPERM; | |
1816 | if ((features & UFFD_FEATURE_EVENT_FORK) && !capable(CAP_SYS_PTRACE)) | |
1817 | goto err_out; | |
65603144 AA |
1818 | /* report all available features and ioctls to userland */ |
1819 | uffdio_api.features = UFFD_API_FEATURES; | |
86039bd3 AA |
1820 | uffdio_api.ioctls = UFFD_API_IOCTLS; |
1821 | ret = -EFAULT; | |
1822 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1823 | goto out; | |
1824 | ctx->state = UFFD_STATE_RUNNING; | |
65603144 AA |
1825 | /* only enable the requested features for this uffd context */ |
1826 | ctx->features = uffd_ctx_features(features); | |
86039bd3 AA |
1827 | ret = 0; |
1828 | out: | |
1829 | return ret; | |
3c1c24d9 MR |
1830 | err_out: |
1831 | memset(&uffdio_api, 0, sizeof(uffdio_api)); | |
1832 | if (copy_to_user(buf, &uffdio_api, sizeof(uffdio_api))) | |
1833 | ret = -EFAULT; | |
1834 | goto out; | |
86039bd3 AA |
1835 | } |
1836 | ||
1837 | static long userfaultfd_ioctl(struct file *file, unsigned cmd, | |
1838 | unsigned long arg) | |
1839 | { | |
1840 | int ret = -EINVAL; | |
1841 | struct userfaultfd_ctx *ctx = file->private_data; | |
1842 | ||
e6485a47 AA |
1843 | if (cmd != UFFDIO_API && ctx->state == UFFD_STATE_WAIT_API) |
1844 | return -EINVAL; | |
1845 | ||
86039bd3 AA |
1846 | switch(cmd) { |
1847 | case UFFDIO_API: | |
1848 | ret = userfaultfd_api(ctx, arg); | |
1849 | break; | |
1850 | case UFFDIO_REGISTER: | |
1851 | ret = userfaultfd_register(ctx, arg); | |
1852 | break; | |
1853 | case UFFDIO_UNREGISTER: | |
1854 | ret = userfaultfd_unregister(ctx, arg); | |
1855 | break; | |
1856 | case UFFDIO_WAKE: | |
1857 | ret = userfaultfd_wake(ctx, arg); | |
1858 | break; | |
ad465cae AA |
1859 | case UFFDIO_COPY: |
1860 | ret = userfaultfd_copy(ctx, arg); | |
1861 | break; | |
1862 | case UFFDIO_ZEROPAGE: | |
1863 | ret = userfaultfd_zeropage(ctx, arg); | |
1864 | break; | |
86039bd3 AA |
1865 | } |
1866 | return ret; | |
1867 | } | |
1868 | ||
1869 | #ifdef CONFIG_PROC_FS | |
1870 | static void userfaultfd_show_fdinfo(struct seq_file *m, struct file *f) | |
1871 | { | |
1872 | struct userfaultfd_ctx *ctx = f->private_data; | |
ac6424b9 | 1873 | wait_queue_entry_t *wq; |
86039bd3 AA |
1874 | unsigned long pending = 0, total = 0; |
1875 | ||
cbcfa130 | 1876 | spin_lock_irq(&ctx->fault_pending_wqh.lock); |
2055da97 | 1877 | list_for_each_entry(wq, &ctx->fault_pending_wqh.head, entry) { |
15b726ef AA |
1878 | pending++; |
1879 | total++; | |
1880 | } | |
2055da97 | 1881 | list_for_each_entry(wq, &ctx->fault_wqh.head, entry) { |
86039bd3 AA |
1882 | total++; |
1883 | } | |
cbcfa130 | 1884 | spin_unlock_irq(&ctx->fault_pending_wqh.lock); |
86039bd3 AA |
1885 | |
1886 | /* | |
1887 | * If more protocols will be added, there will be all shown | |
1888 | * separated by a space. Like this: | |
1889 | * protocols: aa:... bb:... | |
1890 | */ | |
1891 | seq_printf(m, "pending:\t%lu\ntotal:\t%lu\nAPI:\t%Lx:%x:%Lx\n", | |
045098e9 | 1892 | pending, total, UFFD_API, ctx->features, |
86039bd3 AA |
1893 | UFFD_API_IOCTLS|UFFD_API_RANGE_IOCTLS); |
1894 | } | |
1895 | #endif | |
1896 | ||
1897 | static const struct file_operations userfaultfd_fops = { | |
1898 | #ifdef CONFIG_PROC_FS | |
1899 | .show_fdinfo = userfaultfd_show_fdinfo, | |
1900 | #endif | |
1901 | .release = userfaultfd_release, | |
1902 | .poll = userfaultfd_poll, | |
1903 | .read = userfaultfd_read, | |
1904 | .unlocked_ioctl = userfaultfd_ioctl, | |
1832f2d8 | 1905 | .compat_ioctl = compat_ptr_ioctl, |
86039bd3 AA |
1906 | .llseek = noop_llseek, |
1907 | }; | |
1908 | ||
3004ec9c AA |
1909 | static void init_once_userfaultfd_ctx(void *mem) |
1910 | { | |
1911 | struct userfaultfd_ctx *ctx = (struct userfaultfd_ctx *) mem; | |
1912 | ||
1913 | init_waitqueue_head(&ctx->fault_pending_wqh); | |
1914 | init_waitqueue_head(&ctx->fault_wqh); | |
9cd75c3c | 1915 | init_waitqueue_head(&ctx->event_wqh); |
3004ec9c | 1916 | init_waitqueue_head(&ctx->fd_wqh); |
2c5b7e1b | 1917 | seqcount_init(&ctx->refile_seq); |
3004ec9c AA |
1918 | } |
1919 | ||
284cd241 | 1920 | SYSCALL_DEFINE1(userfaultfd, int, flags) |
86039bd3 | 1921 | { |
86039bd3 | 1922 | struct userfaultfd_ctx *ctx; |
284cd241 | 1923 | int fd; |
86039bd3 | 1924 | |
cefdca0a PX |
1925 | if (!sysctl_unprivileged_userfaultfd && !capable(CAP_SYS_PTRACE)) |
1926 | return -EPERM; | |
1927 | ||
86039bd3 AA |
1928 | BUG_ON(!current->mm); |
1929 | ||
1930 | /* Check the UFFD_* constants for consistency. */ | |
1931 | BUILD_BUG_ON(UFFD_CLOEXEC != O_CLOEXEC); | |
1932 | BUILD_BUG_ON(UFFD_NONBLOCK != O_NONBLOCK); | |
1933 | ||
86039bd3 | 1934 | if (flags & ~UFFD_SHARED_FCNTL_FLAGS) |
284cd241 | 1935 | return -EINVAL; |
86039bd3 | 1936 | |
3004ec9c | 1937 | ctx = kmem_cache_alloc(userfaultfd_ctx_cachep, GFP_KERNEL); |
86039bd3 | 1938 | if (!ctx) |
284cd241 | 1939 | return -ENOMEM; |
86039bd3 | 1940 | |
ca880420 | 1941 | refcount_set(&ctx->refcount, 1); |
86039bd3 | 1942 | ctx->flags = flags; |
9cd75c3c | 1943 | ctx->features = 0; |
86039bd3 AA |
1944 | ctx->state = UFFD_STATE_WAIT_API; |
1945 | ctx->released = false; | |
df2cc96e | 1946 | ctx->mmap_changing = false; |
86039bd3 AA |
1947 | ctx->mm = current->mm; |
1948 | /* prevent the mm struct to be freed */ | |
f1f10076 | 1949 | mmgrab(ctx->mm); |
86039bd3 | 1950 | |
284cd241 EB |
1951 | fd = anon_inode_getfd("[userfaultfd]", &userfaultfd_fops, ctx, |
1952 | O_RDWR | (flags & UFFD_SHARED_FCNTL_FLAGS)); | |
1953 | if (fd < 0) { | |
d2005e3f | 1954 | mmdrop(ctx->mm); |
3004ec9c | 1955 | kmem_cache_free(userfaultfd_ctx_cachep, ctx); |
c03e946f | 1956 | } |
86039bd3 | 1957 | return fd; |
86039bd3 | 1958 | } |
3004ec9c AA |
1959 | |
1960 | static int __init userfaultfd_init(void) | |
1961 | { | |
1962 | userfaultfd_ctx_cachep = kmem_cache_create("userfaultfd_ctx_cache", | |
1963 | sizeof(struct userfaultfd_ctx), | |
1964 | 0, | |
1965 | SLAB_HWCACHE_ALIGN|SLAB_PANIC, | |
1966 | init_once_userfaultfd_ctx); | |
1967 | return 0; | |
1968 | } | |
1969 | __initcall(userfaultfd_init); |